This episode is brought to you by Progressive Commercial Insurance.
As a business owner, you take on a lot of roles: Marketer, bookkeeper, CEO. But when it comes to small business insurance, Progressive has you covered. They offer:
Count on Progressive to handle your insurance while you do, well, everything else. Quote today in as little as seven minutes at progressivecommercial.com. Progressive Casualty Insurance Company coverage provided in service by affiliated and third-party insurers. Discounts and coverage selections not available in all states or situations.
Parlez tout français. Hablas español. Parlez italiano.
If you’ve used Babbel, you would. Babbel’s conversation-based techniques teach you useful words and phrases to get you speaking quickly about the things you actually talk about in the real world. With lessons handcrafted by over 200 language experts and voiced by real native speakers, Babbel is like having a private tutor in your pocket. Start speaking with Babbel today. Get up to 55% off your Babbel subscription right now at babbel.com/wandery. Spelled B-A-B-B-E-L.com/wandery. Rules and restrictions may apply.
Welcome to the Sinica Podcast, a weekly discussion of current affairs in China.
In this program, we’ll look at:
that can help us better understand what’s happening in China’s:
Join me each week for in-depth conversations that shed more light and bring less heat to how we think and talk about China. I’m Kaiser Kuo, coming to you this week from Chapel Hill, North Carolina.
Sinica is supported this year by the Center for East Asian Studies at the University of Wisconsin-Madison, a national resource center for the study of East Asia.
The Sinica Podcast will remain free, but if you work for an organization that believes in what I’m doing with the show, please consider lending your support.
You can reach me at SinicaPod@gmail.com. And listeners, please support my work by becoming a paying subscriber at SinicaPodcast.com. The semester is starting, two big tuition checks to write, so help a guy educate his kids.
You will enjoy, in addition to the podcast:
And of course, you will have the knowledge that you are helping me to do what I honestly believe is important work.
So do check out the page to see all that’s on offer, and please do consider helping me out.
Dan Wang has been on the Sinica Podcast a couple of times before, and I am delighted to have him back today.
He is one of the sharpest and most original observers of China’s technology sector and manufacturing landscape, having won a certain level of fame for his annual letters and other essays — writings that somehow managed to combine on-the-ground insights with big picture perspectives.
Dan has worked for Gavekal Dragonomics in Beijing since 2017. After a stint with the Paul Tsai China Law Center at Yale, he’s now at the Hoover Institute at Stanford.
If you’ve seen the PBS Nova documentary “Inside China’s Tech Boom,” which I had the pleasure of narrating — it’s a film by David Borenstein — you’ve already encountered Dan. He was a featured voice helping to explain the deeper drivers behind China’s technological rise and talked eloquently, I thought, about the importance of process knowledge, of what the Greeks called metis, which is an important idea that’s really stayed with me and has become quite foundational to my understanding of China and the importance of manufacturing.
Today, we’re going to be talking about his new book, which comes out just about the time you’ll be listening to this. It’s called:
“Breakneck: China’s Quest to Engineer the Future.”
It’s a book that posits — and here I’m greatly oversimplifying — that China is ruled by engineers and they do what engineers like to do: they build. America, on the other hand, is ruled by lawyers. It’s an engineering state on the one hand and a lawyerly society on the other.
Dan’s book is full of memorable witticisms and pithy, trenchant observations. Perhaps most importantly, it explores what each side might ideally learn from the other. They obviously each have their strengths and their weaknesses, so I’m really anxious to ask Dan about whether he thinks Americans are actually learning the right lessons or just burying their heads in the sand and inhaling big plumes of copium.
Before we jump in, I want to point out that this book was especially interesting for me as somebody whose abortive doctoral dissertation was specifically about the rise of this engineering state, about the… The emergence of technocrats in post-Mao China. So things might get a little in the weeds. I ask your forgiveness in advance and will do my best to keep it reasonably accessible.
Dan Wang, welcome back to Sinica and happy birthday, man.
Dan Wang: Thank you very much, Kaiser. And what better birthday present than to speak to old friends like this?
Dan Wang: Yeah, it’s great to have you.
We have to start with what, for me, was clearly the most important part of your entire book, which is that magical and totally improbable guitar-making hub in Guizhou that you stumbled upon as you and Christian Shepard from the Washington Post and another friend rode your bikes through that mountainous province toward Chongqing.
As a card-carrying guitar nerd, this totally blew my mind. I got to find this place. How does a little inland town end up just cranking out guitars for the whole world? I mean, is this just one of those serendipitous quirks of China’s industrial sprawl? Or is there something systematic in how the state, local governments, and entrepreneurial networks operate so that these clusters take root in the unlikeliest places?
And I guess more importantly, were there any of you guys who were guitar players? And if so, did you guys try out some of the local handiwork while you were there?
Kaiser, you’re much cooler than me. You are a guitar player. I am a clarinet player. And I think by coolness, that just really outranks me.
How indeed did kind of a third or fourth tier city in Guizhou become one of the great hubs of guitar making?
Well, in 2021, when I was stuck in China during the summer due to the success of the zero COVID strategy at the time, I asked two friends of mine,
“Hey, why don’t we go on a really long bike ride somewhere in the southwest, which I find the most beautiful part of China?”
Oh, for sure.
And so over five days, we cycled from Guiyang to Chongqing. It was four days in Guizhou, the province of Guizhou, and then until the fifth day when we reached Chongqing.
It was on our second or third day when we came across these giant guitar cymbals on the side of the road. So there were these guitars that were hanging off streetlights. There’s this giant guitar that was on a hill that was kind of this ornamental thing. And off in the distance, there was another big guitar that you could see on the town square.
And so we were very puzzled by this. We unfortunately didn’t stop to try out the handicraft. I’m pretty sure that neither Chris, Zheng, Tung, nor I are anything of real guitar players ourselves.
And afterwards, I went to find that Zhengan County in Guizhou is indeed the largest guitar-making hub in the world. I think it’s something like 30% of guitars in China is produced there. I have to get the exact figure right in my book.
And that happened due to a great accident in which a lot of folks in Guizhou were moving to Guangdong. In the 90s, Guangdong was making absolutely everything and anything. Some people were making guitars for export. And so a lot of people from Guizhou just happened to move to a particular guitar factory.
One of the things that we really found on our bike ride was when you’re going through China’s countryside, Tristan made this very astute observation that there are hardly any middle-aged or people in their 20s or 30s that you could find in Guizhou. It’s a lot of children being led with the grandparents. And that’s because anyone who is able to work has been moving over to the coastal areas where you could have a much better job producing guitars or whatever it is for export.
And something that the local government in Zhengan did was that it found that, well, there’s a lot of people making guitars here. Guitars are not really endemic to the local culture of people playing guitar. That’s not really a Guizhou thing. That’s not really necessarily a Chinese thing.
I’m working to change that, but yeah.
Well, you’re a big force, Kaiser. Maybe we can change that. But it just attracted a lot of people to try to say,
“Hey, why don’t you move back home to Guizhou? You can make a lot of guitars here.”
And somehow that strategy worked. And so a lot of people moved back to Guizhou from Guangdong, and now they’re producing guitars mostly on the lower end.
So this is not the sort of things that will be sold in, I think, the high-end guitar shops that you would probably frequent, Kaiser. But there is some innovation here, and I expect that they will get better and better.
Yeah. I mean, it’s amazing how good quality the Chinese guitars have. I mean, it’s astonishing. And all of the major brands are actually making a lot of their guitars in China now.
Yeah. No, that’d be great. You can pedal there. Right. Yeah. No, I’m not going to do that.
But yeah, was the enticements just the usual package of tax incentives, of steeply discounted infrastructure promises of raw materials? What do they do to entice people to a place like that? What do they typically do?
I think the typical enticement is:
A lot of folks in Guizhou, folks in the Southwest can’t necessarily love the Southeast and Guangdong where they were working. It’s too humid. They might say, “we don’t love the Cantonese food. Where’s all the spices? Where’s all the pickles? Where is the really pungent flavors that folks in Guizhou are used to?”
And so this coincided with sort of this rural revitalization program that Beijing has emphasized for quite a while now. And so I think it is just this big happy accident that I would say a pretty random place in Guizhou is just making so many guitars now.
Awesome. Dan, I know you’re going to end up on every major podcast talking about this book, so I want to avoid just asking you about the main themes or going through chapter by chapter. Instead, I was hoping that we could use the main themes of the book as kind of a jumping-off point to explore a lot of the questions that popped into my head as I read it, questions I’m sure you’ve thought about as well. Not necessarily things that made their way into the pages of the book itself, but let me start here.
I mean, we can all rattle off the obvious differences between an engineering state and a lawyerly society. You got speed versus procedure, certain social orderliness versus the chaos of pure market forces. But what are some of the more subtle trade-offs, the ones that most people don’t even know that they’re making that maybe shape daily life in each system? I’m thinking predictability, dignity, moral legitimacy. I mean, which of these things matters to people who live inside each system?
Yeah. Well, I want to push you a little bit on this, Kaiser. I wonder which is the system that delivers legitimacy. I could posit that the lawyerly society has some degree of legitimacy because there are some procedures in place that people expect that rules have to be followed, and maybe the lawyers are better at following the rules.
On the other hand, the Communist Party, I think, would say, well, we have much greater legitimacy. We have this, what is that term, whole process, substantive democracy, in which we are delivering much better things for the people. So I think legitimacy is a concept here that we can play around a little bit with.
What I’ll say is that the engineering state, I think I came onto this framework in part due to these excellent articles I found in 2001, I believe, that was written by an interesting analyst at the time called Kaiser Kuo, who pointed out that there were quite a lot of engineers that were being promoted into the Central Committee and the Politburo.
And I think there has been quite a lot of discussion since 2002, which is the really striking year when every member of the Standing Committee of the Politburo, notably Hu Jintao, as well as Wen Jiabao, had degrees in engineering.
Of course this was a really striking fact for a lot of people.
I think there has also been this kind of view and understanding that America is very lawyerly and that the government is of the lawyers, by the lawyers, and for the lawyers.
And so what I wanted to add onto this kind of general understanding that was in the air, so to speak, was that I felt like I really experienced the merits and the madness of the engineering state by living there from 2017 to 2023.
I was in China at a time when a lot of things were getting a lot better. The high-speed rail system had really come into fruition at that time. People were no longer shoving each other around to get in line. The system felt quite rational and well-organized.
Shanghai is a marvelously functional city where one is never really more than 15 or 20 minutes away from a subway stop. Shanghai was building all sorts of parks. It built about 500 parks by the year 2020. By the end of this year, the city targets that it will have a thousand parks. Shanghai is just this remarkably well-functional, livable place.
And so that was something that I really experienced by living there. But Shanghai is also infamously the city that suffered perhaps the worst lockdown ever. In the history of humanity, in which 25 million people were unable to leave their apartment compounds for about eight to ten weeks over the course of the spring in 2022. And so that was something that I felt very ethically myself.
When I moved to the Yale Law School’s Paul Tsai China Center, really being embedded in one of these most elite, elite-making institutions in the United States, really seeing that the US is run by lawyers, seeing how the Biden administration at that time had been really, really lawyerly. About 11 out of 15 cabinet members in Joe Biden’s administration had gone to law school. Many prominent folks went to Yale Law in particular.
That was sort of what I wanted to add, that this was something I lived and felt in both places.
Yeah, absolutely. We’ll talk a little bit about this idea of performance legitimacy down the road here. But so I want to dig into sort of maybe philosophical underpinnings of this contrast that you highlight.
In the West, we often reach for the trolley problem as a kind of shorthand for thinking about moral tradeoffs.
I mean, do you pull that lever to sacrifice one life in order to save five? I’ve often wondered how this dilemma looks different through the lens, you know, like the one that you’ve drawn, whether it looks different between an engineering state and a lawyerly society.
I would imagine an engineering-oriented society be more inclined to treat this as kind of a technical optimization problem. You just kind of minimize total loss, while a maybe more lawyerly society would insist on:
Kind of, you know, a utilitarian versus a deontological philosophical orientation.
Maybe that points to a deeper distinction. I mean, do these orientations that you’ve described, do they line up with the classic contrast between communitarian or group-oriented values on the one hand and on individualistic ones on the other?
Yeah. I think that’s actually a pretty fascinating question. I wonder if there is a systematically different way that Chinese tackle the trolley problem in a way that is pretty distinct from the way that Westerners think about the trolley problem.
I think the level that I was thinking a little bit more about was that I think part of the reason I wanted to come up with this framework of engineers and lawyers is that I think we’ve been reasoning about the US-China conflict in these 20th century terms like:
And all of these terms have some use, but I’m not really sure that they still really apply in very nice ways now.
You know, are we going to say that something like, is China fundamentally left-wing or right-wing? Well, I can make arguments on both sides. Is the US fundamentally more left-wing or right-wing? Again, this is something that we can debate and I’m not sure how far exactly we get up to these sort of frameworks.
And so the framework that I came up with of the engineering state and the lawyerly society, I would submit is just no worse than trying to figure out exactly to what extent China is Marxist today.
You know, I don’t think that Marxism is quite the right lens to try to understand the people’s republic. Maybe it is, but I think this is what we need to do is to have a plurality of frameworks here.
Maybe we should have something like the discussion of:
We just need to have more than one framework really to think about the great conflict of the moment.
Yeah, no, I completely agree. And that’s what I really like about this particular framework is it takes us beyond these sort of binaries of ideology, you know, China being just such an incredibly syncretic society that blends so many aspects.
But the one thing that I think it all circles around is this technocratic policy, and I think it feels like, to me, a very, very good explanatory lens. So I applaud that.
I’ve often used a concept I kind of borrow from economics when I think about what a society values. And that’s, you know, the concept of elasticity.
You know, I imagine that in every society, individuals have kind of an intuitive sense. I don’t think they have it mapped out really explicitly, but, you know, how much of one thing that they value, they’d be willing to give up to gain some amount of something else that they value.
You can, you know, kind of almost put numbers to it.
I'll trade you three points of administrative efficiency to get one point of procedural fairness, right?
Or I'll trade you two points of transparency for one point of speed.
I mean, it seems to me like for decades, Americans’ coefficient of elasticity has been really, really rigid. They’ve been very unwilling to trade down in civil and political rights for even for, you know, pretty markedly… Improved economic outcomes. But I mean, it’s my sense, Dan, I’m wondering if you agree that lately, because of China’s example in manufacturing strength, in infrastructure, in its energy build-out, in the energy transition, in education, in STEM education especially, I feel like there’s a shift happening and this is happening. And I think you note this, it’s both on the right and the left, in America, like within MAGA and among also, say the abundance bros, right? You know, Derek Thompson and Ezra Klein and those guys, more Americans seem willing to accept some erosion in rights or process in exchange for what they believe are better material outcomes.
Do you think that that coefficient is changing? And if it is, does it change the way that you think about the lawyerly versus engineering states, especially if we start seeing each side borrowing from the other’s value hierarchy?
Well, there’s certainly a lot of borrowing between the U.S. and China at the moment. But I’m not sure that they’re borrowing all the right things.
Yeah. That’s the big issue. Are they?
Well, I think what we are seeing with the Trump administration is a lot of authoritarianism without the good stuff—good stuff like functional subways, better transport infrastructure, and better infrastructure generally. I think you’re very right to point out that there is a sense of deep dissatisfaction in the U.S. I mean, that is always true everywhere at all times, but I think there is an especially big sense at the moment that the U.S. has not been very functional for quite a long while.
The U.S. has not been very functional because especially in the bigger cities, where things are just far too expensive. If we’re thinking about cities like New York, Boston, or San Francisco, housing prices are really unaffordable for too many people. These are cities that try to build new infrastructure—mass transit—and basically don’t do a very good job of it.
You know, I was really struck that it’s not just that New York is unable to build new subway stations and new subway lines with any sort of efficiency; it costs about $2 billion per mile to build a new mile of subway in New York City. They’re not even doing simpler stuff very well.
And so this is the sort of thing that looks kind of ridiculous. Why does it take several years to upgrade a bus station? I realize that’s kind of a complex structure. There are all sorts of intricacies with the tunnels, but still this is fundamentally a bus station that shouldn’t take more than five years to build out.
So, you know, we have broken mass transit. We have unaffordable housing. The pandemic revealed that the U.S. isn’t able to manufacture a lot of pretty basic goods. There were shortages of masks and cotton swabs. There were shortages of furniture, all sorts of simple consumer goods that weren’t easily exportable from China at the time.
And so there is a pretty big sense that nothing is working when we have to face this critical transition to decarbonize the economy and to build a lot more solar, wind, transmission lines, which all demand quite a lot of land.
And so I think I wonder if there is the case that the U.S. has even made a conscious decision to try to erode some of the elasticity of the proceduralism. Because I think one of my arguments in the book is that the proceduralism has encrusted itself throughout a long period of time without anyone’s real intention to create a lot of processes everywhere throughout the American government.
This is sort of a force that kind of took a life of its own. And this was something that a lot of homeowners and especially the NIMBY set exploited, I would say, to block new housing in Berkeley for students, to block a solar or a wind project as well as their transmission grids. This became something that richer people were able to access and exploit to block projects that they didn’t like.
And that isn’t even really a majoritarian demand for greater proceduralism. This was kind of an independent life force that grew upon itself and has a very vested interest of minoritarians that are really vested in trying to keep that system so they are able to block a new apartment building if it takes them with their light away, for example.
You know, you work to be very fair in the book and that’s something I really like about it. I mean, you don’t just heap praise though on the engineering state. You make a point of calling out the downsides. And they’re very real. Can we talk about some of those? What the problems are of the engineering state? What does it get wrong? You sort of channel the James Scott scene like a state thing and a lot of the excesses of that thinking.
There’s two chapters in particular of your book that really dwell on this. And they are about, of course, the one-child policy, which is a conspicuous failure of the engineering state mentality and also the zero COVID policy, which starts off as sort of a triumph, not right away, right? It, I guess, displays some of the pathologies of it, but by the spring of 2020 you see this V-shaped recovery. You see China really use its state capacity to wrangle the COVID epidemic.
But then of course, you talk quite a bit about the lockdown. So, talk a little bit about what some of the major downsides are. I think the engineering state has major upsides.
Um, so to be clear, I really want to articulate that the speed of construction of new housing in China, new roads, tall bridges, subway systems, nuclear, all sorts of construction in China, I would say is net positive. You could go to Guizhou as I did, look at these really tall bridges. It is pretty easy to say, well, this is a bridge to nowhere, but I think it is also true that a bridge to nowhere quickly turns nowhere into two somewheres at the ends of these bridges.
If you take a look at China’s major infrastructure, I would say that on net, it’s been extremely positive, that the benefits have way, way, way exceeded the costs.
Now, I would say that there have certainly been some costs:
But in spite of these costs — human, environmental, financial — I would still say that the benefits of infrastructure way exceeded the downsides of so much frenetic construction.
When I say that you talk about downsides, I don’t mean to suggest that you present a kind of moral equivalence between the systems. It’s pretty clear that you believe one side needs to learn more from the other right now.
It’s pretty clear where you think the osmotic gradient should flow.
The problem, I think, is that the Chinese leadership is not only physical engineers. They’re also fundamentally social engineers, and they cannot stop themselves from treating the population as just another building material to be remolded or torn down as the circumstances demand.
And so I think we can point to a lot of social engineering projects in China and we can point to the repression of ethno-religious minorities in Tibet as well as Xinjiang. Even with the Han majority, people have lived for a long while with the hukou system, which is not even fully abolished yet, in which it becomes really difficult for a migrant worker to move to Beijing or Shanghai and access educational facilities for her child.
What I really decided to focus on were these two big projects that you mentioned:
And I think you’re really right to point out that zero COVID follows an arc that isn’t very straightforward.
I think the first act of this big dramatic arc of zero COVID was the spring of 2020, or even earlier in the winter of 2020, when I was living in Beijing and we heard about this new pneumonia that was spreading through Wuhan.
And when we saw the Wuhan lockdown, which was in January, I believe January 23rd, you have these sort of dates that are emblazoned in your mind if you lived through the pandemic in China.
Wuhan lockdown, hearing the stories of the ophthalmologist, Dr. Li Wenliang, who raised valid concerns and was disciplined by the state for raising these sort of concerns, created a lot of anger among pretty much everyone I knew that there was yet another respiratory virus that was spreading from China.
This is the second one after 20 years with the first SARS crisis.
There had been some political suppression of bad news up until the state really tried to react and try to tamp it down. A big way. And so that was the great first act when a lot of commentators from the U.S. and parts of the West were sometimes even gleefully saying that this might be China’s Chernobyl moment in which a disaster triggers the political downfall of the entire regime. And so that was the first act.
And then the second act proved a lot of that wrong. So the second act of China’s COVID experience was the much longer time period when Beijing, Shanghai, central government, local governments proved that China was able to control the virus much more effectively than the U.S. can or much of the West could. And so the second act was people in China feeling relatively glad that they were living in China and able to be free of transmissions, able to carry on life relatively normally.
There were some costs. I wasn’t able to see my parents who were in Pennsylvania. My parents were telling me this very un-Chinese thing, which is to say,
“Stay there. Don’t come to visit us. Trump’s America in 2020 is a terrible mess. So, you should just stay in China where life is a lot better.”
They weren’t wrong. They weren’t wrong at the time.
But then there was the third act of China’s COVID experience. That third act was triggered by the much more transmissible Omicron variant of the virus, which overcame a lot of vaccines and was just extraordinarily transmissible. That was really the variant of the virus that forced Shanghai to go into lockdown for about eight weeks in the spring of 2022 when people could only go downstairs to their apartment compounds to have their noses and their throats swabbed. Otherwise, you couldn’t really go outside even for any sort of fresh air.
And so this was a time that drove a lot of people crazy. This was a time when a lot of families were suffering some degree of food insecurity because the Shanghai government had no logistical capacity to really try to deliver food to a lot of families. I knew a lot of families where the parents really tried to reduce their food intake so that they could save some food for their kids.
The food shortages resolved after, I believe, something like the second week of April. But, you know, this was something that was pretty extraordinary—that people were feeling food insecure in China’s largest city in the year 2022. That was really surprising.
And then the great denouement of the great dramatic act of China’s COVID experience was when in 2022’s December, Beijing decided to drop all COVID restrictions in the coldest month of the year, when people had very few fever reducers in stock to meet this great ending of the pandemic when zero COVID kind of became total COVID.
And so in Shanghai, I caught COVID around December 22nd, when I think everybody else was catching COVID at around the same time. So luckily, I had quite a fine experience with all of these things. But there were a lot of folks in Shanghai who didn’t have a very good time getting COVID at that point.
And so, you know, this is where the engineering state is pretty ambiguous, I think, in terms of its effect. So sometimes it looks pretty good that it was able to follow WHO recommendations and control the virus until it then collapsed under its own weight.
So the evidence here is pretty ambiguous, I would say.
Yeah, absolutely. But, you know, at the same time, I worry that there’s a certain type of American copium smoker who is taking these failures of the engineering state, assuming them to be inevitable consequences of adopting the sorts of things that you would like them to say. And, you know, they’re telling themselves these sort of self-soothing daily affirmations, like:
So, you know, actually, America is doing great. Thank you very much.
Yeah, I wonder.
I think I absolutely agree with you that the mood in the U.S. especially fluctuates way too wildly for what the situation actually is.
I remember at the end of 2022, there was just excessive triumphalism in the U.S. because China ended its zero COVID program in this horrible collapse in which a lot of people died and the state suppressed all of this data.
Russia then wasn’t doing very well in its fight against Ukraine. And so it looked like Ukraine was also winning against autocracy.
And the end of 2022 was also the years when it seemed like the U.S. had these great technological breakthroughs,
and the autocracies simply didn’t have these technologies in place.
And so the views have shifted quite a lot. And these views go up and down, I think, a little bit too wildly given the state of… The evidence. And one of the things that I’m always trying to say, you know, when I was at China, now when I’m at the Hoover Institution is always that this is going to be a really long struggle between the U.S. and China. This conflict, these tensions will go on for a very long time. I don’t think that it is anything like a static picture in which one country is winning and they will have any sort of a decisive advantage. I think that the struggle will take place over a very long time.
And there’s not going to be any scenario in which one country simply disappears off the face of the earth. That is a fantasy. And I think it is also a fantasy to imagine that either country will collapse and never get back on its feet. I think that both countries are going to be winning and losing. And when they’re winning, they’re going to be making a lot of mistakes. When they’re losing, they’re going to try to catch up. And that’s just going to be a dynamic process over the next few decades.
Do you agree?
I do agree. I think the language of existential threat and the framing of zero sum is foolish when you see it on either side. Let me get to the things that we ought to be, we as Americans ought to be learning from China. One of the things that you really emphasize is process knowledge. I mentioned that in the introduction. For you, is that primarily a cultural asset? That is the status of engineers, the kind of tolerance for iteration. Is it a firm level capability, having long patient capital, kind of shop floor autonomy? Or for you, is it kind of a policy environment with permitting and procurement and standards at the fore?
Where would you intervene first, in other words, to sort of rebuild process knowledge in the United States where it’s so sorely lacking?
I think it is all of the above, Kaiser, that it is cultural, it is policy driven, it is a matter of economics. So I think the most important thing to grasp about technology is not the actual physical instruments or tools that we can see, anything like a robotic arm. It’s also not a recipe or a blueprint or a patent, any sort of knowledge that’s really easy to write down.
I think the most important part of technology has to be the process knowledge, which is all of this meta and tacit knowledge that exists more on a population level. And so this is something that various hubs of knowledge production have been able to recreate in the past.
You know, at the start of the industrial revolution in the UK, there was just a lot of knowledge about how to build textiles in order and how to build engines.
Right.
When that moved from Britain to Germany, Germany had a lot of process knowledge about how to do interesting new fields like electrical engineering, as well as chemistry. And that has moved from country to country. The US has been a major industrial leader on something like automotives, on something like semiconductors in the past.
And right now, a lot of process knowledge with manufacturing is being built and activated and grown in China, where you could be a worker in Shenzhen, making iPhones in the first year, being poached to make Huawei phones the second year, then making a DJI drone the third year, and then making a CATL electric vehicle battery the fourth year.
And so there’s just so much knowledge that can’t be written down with technology that is necessary for the production of a lot of different goods.
So I think this is one of these things that the US didn’t sufficiently appreciate when a lot of corporates did offshore a lot of jobs to China. I want to be clear that a lot of the manufacturing job losses in the US have been triggered by automation and technological change, not so much by offshoring, but something like 10% of the manufacturing change is created by offshoring.
And one of these things that I wonder about is if Apple didn’t build all of its iPhones in Shenzhen, and rather built it in, let’s say:
What if all of that knowledge involved in building hardware was actually in the industrial Midwest in the US as well? Could it be that Wisconsin or Michigan or Ohio are actually major producers of:
that is present in Shenzhen as well?
And so this is one of these things that I think has been critically understated in the US that has been driven by an excessively financial profit-driven model that didn’t account for all of the most important things with process knowledge.
Right. I mean, this possibility, this hypothetical that you float of an Apple producing in Cleveland, that seems to place a little too much of the onus on Apple. It’s not as though that decision could have been undertaken in a vacuum. There were other factors that it had to consider rather than just simply the cost of labor. It was, as you say, you know, there was a policy… Environment. You know, there are other reasons they chose not to do that. And surely you would agree that it’s not just on Apple.
Absolutely. I think that the infrastructure wasn’t in place. The costs were much, much lower in the past. And so these are all real.
Yeah. Yeah. I mean, when you, when you talk about, when I asked you about process knowledge and, you know, whether it’s a cultural asset or a firm-level capability or policy environment thing, you said all of the above. I mean, that reminds you of something that you wrote recently. You just published in Foreign Affairs with your former boss, Arthur Kroeber, who is, by the way, one of the people in the China space who I just admire the most.
You guys wrote that, you know, China has taken in all of the above technology strategy. What would you include as the pieces of that strategy that perhaps people are less aware of?
I think that people know, you know, big pieces of it, but some of it, I think there is still a gap in our understanding of how China did this. What would you identify?
Arthur and I wrote that piece in Foreign Affairs called The Real China Model, in part to try to rebut the sense that China has succeeded technologically simply because it has stolen all the IP from the US. And so, you know, I, one of my favorite boogeymen is this tweet by Senator Tom Cotton, which he tweeted on World IP Day,
“China doesn’t innovate, it only steals.”
I think that is a flagrantly wrong presumption that I think we just need to discard because it is not helping us understand China any better.
There’s also this view out there that China succeeded simply by subsidizing its way into technological leadership. I think that’s not wrong, but I think it is woefully incomplete to say that the Chinese have been able to make central planning work and been able to select winners. I think they haven’t had a terrific track record on that.
What we point out in this piece is that China has actually built a lot of what we call deep infrastructure to be able to have its success.
Now, deep infrastructure goes beyond traditional infrastructure where China is superb — of trains and ports and highways to move goods around. What we point out are three big things:
Yeah, you noted, I mean, I think just to throw one stat, that China’s total electricity output is greater than the United States and the EU combined, and every year it adds another Britain’s worth of electrical production.
That’s right. Chinese people are on smartphones constantly, maybe even a little bit too much. And the Communist Party is very much in charge.
And so when you marry these three pieces of deep infrastructure —
- power
- connectivity
- process knowledge
— to the fierce dynamism among Chinese entrepreneurs who are really competitive in trying to build interesting new projects, build more cheaply than the other guy, not necessarily achieving a lot of profit, but creating new and worthwhile products, when you marry all of these things together, I think it is no surprise that China has become the technological superpower that it is today.
There are some elements of technology theft from the West. There is an obvious element of the state trying to pick winners, subsidizing all of these things.
What we can acknowledge is that China has both a strong state as well as strong entrepreneurs that have built a lot of these technological achievements.
Dan, I’ve often remarked on how China in the 21st century is a much less technophobic or techno-pessimistic society than America is today. You can see it in survey research on attitudes toward things like AI. But I mean, anyone who’s lived in China and the US, as both you and I have, we know this intuitively, right? Just in the posture that people have toward technology.
I mean, so years ago, I interviewed a philosopher named Anna Greenspan about a book that she wrote.
Called Shanghai Future, one I highly recommend.
Me too.
You’ve read this?
Yes, I’m a big admirer of Anna’s work.
Yeah, she’s great. So you remember, she talked about this big difference in attitudes toward futurity in the US and China. I’ve come to use kind of shorthand that I like. China is still in its Star Trek phase and the US is in its Black Mirror phase, right?
So the question I have for you is, what is the causal direction, if indeed you see any causality at work here, between China’s technocratic engineer-dominated polity and its technophilic society? Does the technocracy create the technophilia or does the technophilia create the technocracy?
I think that the technocracy creates the technophilia. I’m willing to change my mind on this, but I think it is definitely the case that China’s leadership uses mega projects, big prestige projects, really to try to rally the population into doing something better. And I think there are some ways in which this could be a little bit insidious.
One theory that I’ve come across is that one of the reasons that Li Peng, the premier throughout the 1990s, was so heavily invested in the Three Gorges Dam was in part to try to distract from his own image as what the Western media labeled as “the butcher of Beijing” for having ordered the Tiananmen crackdowns.
And so the Chinese government decided that it is going to try to build its way out of this political crisis of 1989 and to really invest in a lot of technology here. There should be a forthcoming book about this. And so once that book is out, maybe we can point to it.
I think it is definitely the case that the Chinese government loves pointing at pictures of great infrastructure. You can’t open an issue of Tioshe, which I was fervently reading when I was living in China, without coming across some amazing new bridge that the government has built, some great new port, which always looks very telegenic, or some speeding high-speed rail going through the countryside.
And so they definitely love to create these sort of images. There is a sense, I think, in which the Chinese government really likes to promote these big novels like Wandering Earth, which has been adapted into a film, and Three-Body Problem, in which there is kind of this emphasis on a world government that is entirely run by engineers working together to overcome a great threat to humanity.
That is, I think, a common theme to Liu Cixin. I think he is one of these progenitors of the engineering state’s mindset.
Right. Of the so-called Industrial Party, the Gongyedang.
That’s right. It’s sort of the Ur text of the Gongyedang.
And I spent a lot of time talking about the Gongyedang in my chapter on tech power.
Right.
And I think the contrast is with the United States, which has had a pretty major tech clash. I think we saw a lot of skepticism of social media, especially after 2017. There right now is still a lot of worries about what smartphones are doing to young people, what social media is doing to young people, what AI might be doing to all of us.
That is all real. And that strain is less present in China, I think, in part because the state loves to create new engineering projects, and in part because I think the Chinese have naturally been more optimistic over the last 40 years than Americans have because they’ve seen their lives improve in such obvious ways.
In lockstep with the improvement of technology. So yeah, it’s reinforcing, right?
And I wonder to what extent the Chinese government might actually be actively censoring some of these views. There has been extensive censorship of opposition to the Three Gorges Dam. And there may even now be some censorship to the big new dam that is being built in Tibet as well.
And so I think there is, on the one hand, the leadership itself is technophilic and trying to engineer their way out of every problem. On the other hand, they may also be censoring some of the perhaps merited, humanistic, critical backlash against what technologies are doing to us.
I want to get into how maybe the technophilia has enabled the technocracy in just a little bit, but because I do think there’s a little bit of bidirectional causality here.
But I want to first ask you whether you think that things like the fact that so many of the leaders are themselves engineers, it sets up a ladder of success, right? I mean, where high status and access to resources and power are kind of enabled by technical, technological prowess, right? So it sets up an incentive system.
So if you are a parent, you’re raising children, you’re going to want to push your children into STEM education. And that itself kind of reinforces that technophilia in society, you know, to your point.
I feel like that’s a big piece of it. Have you given much thought to that as well, to the sort of social forces that work in reinforcing technocratic politics? I think there is definitely a sense that Chinese parents prefer that their kids study STEM degrees. And that is definitely much more obvious that many more Chinese kids are studying math relative to American kids, which I think is a shame. Many more Americans need to be much, much better than the pathetic math capabilities that they presently possess through a lackluster education focused on STEM. I think that should definitely be the case.
So Vivek Ramaswamy was right. Maybe Vivek was right. The issue, I think, is that the slight wrinkle that I would present to you, Kaiser, is I wonder if it is the case that though parents encourage kids to study STEM, they’re not necessarily encouraging the kids to become engineers.
I think the allure of working in tech and consumer internet, especially for one of these big, prestigious firms like
is still much more alluring than working as an engineer. Maybe it is so much more alluring to work in the financial sector rather than in some sort of a technical engineering field, in part because they pay so much better. And so I think the kids are still facing the same tug of incentives that smart kids in the U.S. also feel in being drawn to Silicon Valley as well as Wall Street.
And something else I wonder about, I’m really curious for your take on this, Kaiser. You’ve been spending a little bit more time in China than I have over the last few years. But I was in China in December of 2024. And one of these things that I become really cranky and annoyed by is just how much people are on their phones all the time.
So people are texting other folks over the middle of a dinner. You know, you can see over a hot pot and a hot pot restaurant, many people are just on their phones instead of speaking to their dinner companions. Every trendy cafe shop is better photographed rather than a place to sit and have coffee with other people.
Maybe I’m just getting too old and cranky here, Kaiser. Maybe you can talk me into, you know, being a little bit more sympathetic.
No, you’re only going to hear the same crankiness from me. I mean, it’s something I freaking hate. And I’m also probably guilty of it. I mean, I find myself just having that tug. I mean, I can’t even conceive of taking a subway ride without having my headphones. I’ll walk three blocks back to my apartment if I’ve forgotten my headphones. Yeah, I’m terrible about it. But yeah, this is like the plight of modern homo sapiens. It’s not just a China or America thing. I see it in the States almost just as bad.
But yeah, I mean, I’ve remarked on this before. I used to, you know, you’re standing on the sidelines of a soccer game and you turned another parent of one of your kids’ classmates and you say,
“What are you doing about juniors’ screen time?”
And they’re too busy on their own damn phone to even hear your question. And yeah, it’s a problem.
It’s a problem. I wonder if it might be slightly worse in China because everything has to be turned into a Wang Hong spot and everything has to be photographed as well.
Oh, Christ. Yeah. I mean, I was in Shaxi in Yunnan and it’s becoming that way, you know, because Li Weifei shot a television show called
“Chiou Fung Le Di Phang”
and everyone has to, you know, like have their picture taken where she was and where that scene was shot. Christ.
Maybe let’s check our crankiness and get back to some techno-optimism.
Yeah. You know, actually, I want to dig into history here. I mean, you don’t explore this so much in the book, but I’m sure you’ve given us a lot of thought, which is, you know, the question of what gave rise to the engineering state in China?
I mean, when do we start to see it emerge? Was it a deliberate policy choice or something that just sort of happened? I mean, because this is something I explored quite a bit in my own work as a graduate student I mentioned in the intro and that you so kindly name-checked. I was really inspired to write on this question because by the early nineties, when I was doing this work, it was already, you know, China was already so thoroughly technocratic.
It was already so dominated by engineers. It hadn’t even peaked yet, but already you could see it. I mean, there were already books about this, but like Lee Chung, who’s now at HKU, Lynn White of Princeton, they did a lot of work on, on technocracy. But what struck me was that it had become so technocratic, but somehow it had gone unremarked upon in China itself.
There were foreigners who were looking at this fact and marveling at it, but it was in China itself. It was like, “yeah, of course.”
I wonder if there’s something deeper in China’s history, maybe the imperial civil service examination system, or this, you know, oriental despotism idea of Karl Wittfogel in his hydraulic theory of civilization. You know, he posits that… The technical demands of water management in China created both the opportunity and the necessity for centralized political control. So you have engineers sort of running the state. These were the things that I was exploring and I was wondering what you think about this. What are the historical and maybe cultural roots of the engineering state?
Yeah, I think there are definitely deeper roots in both the engineering state as well as the lawyerly society. That was my next question.
The part of America being very lawyerly, you can read the Declaration of Independence as almost a legal document. So many of the founding fathers were lawyers: first 16 U.S. presidents from Washington to Lincoln—13 of them have been lawyers at some point. And so in the U.S. there is definitely this very obvious legal tradition.
And I think that you can say the same about China as well. I don’t want to take this too literally. I think the work of Karl Wittfogel on oriental hydraulic despotism was a product of the time. He was this strange cold warrior that was trying to discredit the Soviet Union. I don’t refer to Wittfogel at all in my book. But I am definitely a big fan of the work on the clergy system. In particular, Professor Huang Yashun’s book, The Rise and Fall of the East. What is it? Examination.
The Rise and Fall of the East. Examination, autocracy, science and technology. That might be right. We have to fact-check that one. But I think the examination system is very real.
And so I do want to trace a lineage of the engineering state to imperial times. Without being too literal about this, but one might be able to say that imperial China was a proto-engineering state in part because the emperors ordered so many people to build Great Walls or Grand Canals.
So many people died trying to build this canal. The historical records here may be exaggerating some things, but so many people were supposed to have fallen in the course of building this Grand Canal. One might be able to say that the emperors rarely hesitated to almost completely reorder a peasant’s relationship to her land. So there was some social engineering here as well.
Again, I don’t want to be too literal to say that the emperors were straightforwardly engineers, but I think one can trace the sort of lineage because of the state’s management of the imperial exam or the Keji system.
And I think one of these differences I want to trace between the West and China is that I think the Chinese were practicing a source of a sense of absolutism starting from the first Qin dynasty with Qin Shi Huang, in which the state really tried to control quite a lot of things.
This is someone that we label today in China as a despot who buried the scholars and standardized the weights. And so there’s this sense of autocracy stretching back for about 2,000 years now. The Chinese had been practicing absolutism way before the European monarchs ever whiffed this idea in the 17th and 18th centuries.
And so one of my ideas here is that one of the reasons, perhaps, that China did not develop a liberal tradition was that the court administered the exams, which was how one became an intellectual in the first place. And so it becomes really difficult for an intellectual to become a court intellectual by advocating for constraints on the power of the emperor. So mostly all of the mandarins were encouraged to just say,
“How do we govern better? How do we increase the discretion of the sovereign?”
You don’t really get very far by saying,
“Well, what we need is some sort of property rights. What we need is to protect the business people.”
You never really quite had that. And so you didn’t have as vibrant a sense of a liberal intellectual tradition emerging out of China. Rather, that was much more of an absolute sense of trying to increase the power of the sovereign.
Yeah, absolutely. I think that you’ve put your finger on it right there. This cooptation of the entire literati class just by making their advancement contingent on their support for a state orthodoxy.
Right. And I think we see parallels to that today in the Communist Party. I actually think that, I mean, I could spend a lot of time talking about this, but that there’s always been this sort of privileging of knowledge elites. And that assumes, of course, that there’s some objective knowledge in the universe against which you can be tested.
So I mean, at all points, there is this sort of a paradigm of what is true. And there is some canonical set of texts. They could be the Confucian classics or they could be, you know, engineering texts. And if you have demonstrable knowledge of that, somehow that qualifies you for office. I mean, that seems to be sort of the common… Thread. Yeah. So something that I, so, you know, the U.S. obsession with process, in its best form, protects the weak, which is really good. But, as we’ve discussed, it can impede the provision of public goods, the building of infrastructure that can really hurt the weak.
So China’s obsession with outcomes often lifts the many, but can screw the few or occasionally, as in the case of the one-child policy and zero COVID, which you talked about, it can screw the many as well.
So I guess the big question is, how do you build or design institutions that kind of somehow bind outcomes to rights? That is, build fast without trampling people. And what are the kinds of small practical reforms that can move either system in that direction?
Maybe we can start with China. What are some ways where these institutions can be bound up more in rights? And then we can move to the U.S. because you’re very hard on U.S. proceduralism. You’re very generous about its civic function, but maybe we could talk a little bit about the reforms that lawyers could champion that would improve build speed without betraying that kind of ethical core.
Yeah. Well, here is where I would give a plug to my friend, Nick Bagley’s work. He is a law professor at the University of Michigan. He has a book that will be coming out that I think is a perfect encapsulation of the problems of the lawyerly society. He doesn’t quite call it that. And he proposes these tangible legal reforms such that:
So that is one of these books that will be coming out sometime next year.
You know, I think there is actually a kind of a simple answer to a lot of construction. It’s not that the U.S. and China are the only countries that are unable to hit the right balance. I think actually a lot of countries have hit the sweet spot in terms of constructing mass transit while protecting the public interest. And so this is most of Europe. This is Japan. And we can just take a look at what these other countries do.
You know, I was just back to the U.S. after spending much of the summer in Europe. My wife and I spent a month in Denmark. Denmark is really highly functional in terms of public transit. You can go down to the subway systems that are completely spotless. They’re cleaner than anything in Shanghai. They’re fully automated and they just work really well. And you don’t even have to buy any tickets or go in any turnstiles. It’s such a high trust society that people know that you will have bought your tickets beforehand.
And so, countries like Denmark, countries like Japan, which has built a lot of high-speed rail, these are not shining exemplars of human rights abuses.
I would say that, you know, we can just take a look at:
They are able to build trains and subways and all sorts of infrastructure at really reasonable costs without having violated a lot of rights. And so it is mostly the Chinese and the Americans that have gotten the balance wrong.
Yeah, that’s a good point. And do you see efforts now on either side to try to learn from these better examples?
I wonder to what extent China is learning better examples of public interest. I think there have been some ways in which China is learning good lessons. I think it is not the case that environmental reviews for high-speed rail, for example, are entirely perfunctory. I think that the builders are actually trying to do their best to mitigate a lot of environmental issues.
What’s just not available in China is endless lawsuits that can delay absolutely everything on purely procedural bases.
And I think the Chinese have also had some examples of protests that achieved the delay or the cancellation of projects. Remember, I think it was in 2020, when folks in some bigger city—may have even been in Shanghai—went onto the streets to protest the construction of a new trash processing site near their home.
Now, maybe that’s nimbyism. Maybe that is misbegotten. But, you know, we do see that there have been some protests of people trying to maintain their neighborhoods and tell what they like. Maybe that’s positive. Maybe that’s negative.
And I think there is definitely this big sense in the U.S., as we mentioned before, that the U.S. has been dysfunctional for the many, and we need to get much better at building housing, mass transit, all sorts of infrastructure to get the country moving again.
Now, for the most part, I would say that the U.S. government now isn’t learning the right lessons from China. Rather, it’s learning most of the bad lessons from China.
Yeah, as you said. So on the topic of learning lessons, you know, the COVID lockdowns… showed the extreme downsides of the engineering state. I mean, a good engineer, a good scientist, presumably learns from mistakes. I think it’s widely accepted that there were a lot of mistakes made during that time.
What lessons do you think China’s leaders themselves drew from the experience?
That’s a great question. And I haven’t given that too much thought. And I wonder whether there is a lot of studies here. Now, how did enforcing these lockdowns really change the leadership’s mind? Now, I wonder whether they have also learned some of the wrong lessons with COVID.
I mean, one of the things that really struck me was that the Shanghai lockdown, locking down 25 million people in 2022 for eight weeks was accomplished through just the normal police systems. You know, you just had the regular police actually enforce COVID lockdowns.
As best as I can tell, no officers of the People’s Armed Police, which is the paramilitary force that wear what looked like army uniforms, were really deployed to try to enforce a lockdown of that magnitude. And they certainly didn’t have to bring out the People’s Liberation Army to try to suppress the desire to be free.
And so I wonder whether the leadership has learned a lesson that actually the coercive internal security apparatus doesn’t have to be so large in order for the people to be pretty obedient about what are really extraordinary controls that no one had expected at that time. That could be a potential lesson there.
Perhaps other lessons have been that the Chinese surveillance state grew very extensively, that people were tracked on their phones all the time for contact tracing purposes. And there were some issues about privacy concerns. But for the most part, people went along with all sorts of these projects.
And I wonder if the Chinese state has just learned that autocracy is actually much more possible. It’s even more possible than they thought. And I’m hopeful that they learn some good lessons out of this as well. Off the top of my head, I’m not sure I can name any, but I’m wondering, what do you think?
Yeah, no, I mean, I think that you touched on something that I wanted to ask you about, because you know, a lot of people who believe that, you know, the COVID era biosecurity state that was coming into being — you know, the controls that the health code apps and the checkpoints created — that this was just never going to be set aside once the pandemic passed, that this was going to be a regular feature of life.
They thought that the leadership was going to get so addicted to this level of control that they just never let go of it. But it seems like they have. I mean, the app is gone. The checkpoints, the screenings, they’re all, you know, a thing of the past.
Indeed.
I think that’s a pretty good example of maybe a lesson, if not a lesson learned, at least that they exercise a little bit of restraint. Touch wood.
Right. The question is whether they have very long memories and built up this muscle such that if they ever need to exercise these muscles again, they’re going to be able to roll these things out.
It doesn’t surprise me that a lot of the checkpoints, a lot of these apps, and a lot of these COVID testing facilities have been torn down because they became these hated symbols of enforcement. So it could be the case that they took away these highly visible symbols of enforcement, but they have the memory and the muscles to try to bring them back really quickly if necessary.
Yeah, that’s a very good point. I think they certainly have that muscle memory now.
Dan, you write about fortress capabilities, the kind of redundancy, the overcapacity that Western analysts often dismiss or disparage as wasteful. But you actually make the argument in your book, I thought that was a really interesting one, how inefficiency can actually be kind of a source of resilience in China’s system.
How should we be thinking about the trade-off between resilience and efficiency when comparing China’s fortress model with America’s maybe leaner, but possibly more fragile system?
I think one of the things that the pandemic revealed was exactly how fragile a lot of America’s supply chains really were. They were poised for perfection, and it didn’t take much for everything to be ruined.
And there has been this…
Depending on just-in-time delivery and…
Exactly. Just-in-time delivery is something that creates a lot of profitability because you’re reducing your flow of inventory. I think this is also really attributed to Tim Cook of Apple that created these hyper-optimized supply chains. Things were moving around all the time, and so they built very little inventory in order to prepare for shocks.
And I think one of the benefits, mostly a benefit of the engineering state, is that they do build a lot of redundancy. It creates a lot of inefficiency. You can find… There is extreme inefficiency in the Chinese state with the state-owned enterprise sector. There’s just so many redundant jobs. You would just have too many people doing the same things. You have dragging down profitability in all sorts of ways. But that turns out to be really useful in a crisis, that you have the capacity to retool your manufacturing lines in order to build not electronics, but cotton masks, as was the case with JD, Jindong, as well as Foxconn as well.
And so China has a lot of redundancy. China is trying to build up its own oil and gas sector, even though it’s much more costly to tap Chinese gas and oil relative to Russian or American gas or oil because Xi Jinping really treasures energy sovereignty. They’re building a lot more farmland in less than optimal places. I think it is also very striking that as soon as you take the high-speed train out of Beijing or Shanghai, you run into farmland really quickly. And that is because they want to set aside a lot of land for provinces and major municipalities to be food self-sufficient.
And so all of the redundancy involved with manufacturing, all of this overcapacity, there’s also a way to maintain process knowledge that they are constantly training their workers to make sure that their skills don’t go rusty. And I think, again, this is where I think for the most part, China’s engineering state in a lot of economics, you can point to a lot of flaws with debt, with environmental destruction, with all sorts of profitability costs. But there are also some benefits, and these are really revealed during a crisis which you can never predict could emerge.
So Dan, I mean, shame on me. I have not yet finished reading Derek Thompson and Ezra Klein’s Abundance, but I think I get the gist of their argument. It’s interesting to me how little they actually talk about China. But where do your ideas sit in relation to their ideas?
I would want to be a card-carrying member of the Abundance movement. I am slated to speak at the Abundance Conference in Washington, D.C. in the first week of September. So I think I am proximate enough to that.
Now, I think my challenge to Ezra and Derek are to speak a little bit more about China. I think the first parts of the Abundance book, there’s a lot of discussion of how the U.S. isn’t building enough mass transit and infrastructure.
And then the second part of Abundance is talking a little bit more about the scientific failings of the U.S., in which they’re not really taking advantage of being able to scale up and commercialize a lot of American scientific innovations. So China is a good operating model for Abundance. It’s not the best. It is not the most amazing, shining example for the U.S. to follow.
I would love people to ask them whether that was a tactical choice on their part to avoid making the China comparison just to, you know, I mean, because the optics of it aren’t necessarily good. It no longer looks like rah-rah, go America. It erodes some of the, I think, the patriotic oomph that the book otherwise has.
I suspect that what is the case is that, I mean, it’s not only, I mean, it’s not the case that China is avoided entirely. Both Abundance as well as Breakneck talk about California high-speed rail and its awful failings relative to the Beijing-Shanghai line. I suspect what is the case is that Ezra and Derek believe, as I do, that America doesn’t need to become like China in order to build infrastructure. It would be good enough to be like France, Denmark, or Japan.
And so I think we really don’t need to reach the China model. There’s just much better models for the U.S. to reach. And so this is why I say that China is a good operating model of abundance, not the best.
It is good because China has demonstrated that there are virtues to overcapacity, that it is really good to have a hyper-competitive solar sector that is driving prices down, not making a lot of money for investors, but, you know, creating a lot of consumer surplus and building a lot of mass transit for a country that desperately needed it.
There were a lot of costs, but, you know, again, we don’t have to fully copy the Chinese model wholesale in order to get to a better mode of abundance.
You know, you close your book down by emphasizing lived experience, what ordinary citizens feel day to day in terms of dignity, of fairness, of security. I mean, I’ve argued for a long time that Chinese people, like all people, most people at least, anchor their feelings about a given government and its legitimacy, not just in performance, however important that is, but also in whether the state feels to them intuitively morally upright or whether it feels just.
States that emphasize procedural legitimacy obviously tend to foreground this. You know, in China, when you have local corruption, You have arbitrary crackdowns, you have unequal treatment. It can definitely undermine legitimacy, legitimacy on the ground. And when people see the state standing up to bullies or ensuring national dignity, it can bolster this type of legitimacy, which I would love together. You know, it’s this sort of sense of moral uprightness or justice, and that can be domestic or foreign.
How much do you think legitimacy in China actually rests on what I would call the moral dimension, the state, you know, being just or upright and defending dignity? In other words, when you have corruption or arbitrary crackdowns and this stuff eats away at moral standing versus when the state asserts itself against bullies or delivers on fairness, how decisive is that in shaping how people experience the party’s legitimacy day to day?
I ask this because so often there’s this idea that China’s only about performance legitimacy and that somehow an economic downturn or slowdown could deliver a death blow to performance legitimacy. I feel like that’s only a part of the story.
I certainly agree that it has been a persistent fantasy in the U.S. and some parts of the West that China’s political legitimacy depends entirely on economic growth. You’ve seen this narrative come again and again:
I think that is just a silly argument that we see even in 2025. I think that China’s legitimacy is more broadly based than that. I wonder to what extent moral legitimacy, the sort of Confucian virtue, is very much present in China.
I think certainly there is a view that the leadership would try to act as if they are very good Confucians in China. And I wonder to what extent that is actually very effective. Because I think one of the issues I have with China, and it was Professor Huang Yasheng who laid this out very well, is that the state tries to increase a lot of legitimacy in the virtue of the rulers, but they’re not thinking in terms of incentives and constraints and systems that really try to police behavior and induce better governance.
When they reduce things into a matter of morality and virtue, it becomes more about the person rather than about the system. And I think Huang Yasheng has been really good at pointing out how virtue has been a distraction to better governance. What do you think?
Yeah, no, I think that he’s not wrong, that that is a problem. It’s not systematized; that it’s still subject to a lot of kind of patrimonialization. And I think that he’s absolutely right, that if you look at patterns of protest in Chinese history, the way that it is voiced often is in terms of moral failings of leaders rather than particular policies. That is not always a helpful framing when it comes from above or from below. So I tend to agree with him there.
I want to move on though and talk about legitimacy itself. I think there’s this inability among many Americans, and I think you just hinted at it just now, to see beyond procedural legitimacy as the only possible foundation for proper political authority.
I have long believed that this fundamental refusal—it’s not always articulated, but it’s often really present in the American habitus, just in the language that we use—is a big part of the problem when it comes to forming a good understanding of China. It produces a very unhelpful moral framing, and it makes us interpret everything that Beijing does in the most negative possible light.
I think it fuels escalation. It’s not like Beijing is unaware also that there is this kind of assumption of illegitimacy on the American part. I mean, it’s pretty obvious from China’s point of view, and it makes them very defensive. It makes them very anxious. It makes them also assume the worst: that they assume America’s real goal is to destabilize China, which, yeah, they’re not necessarily wrong.
Maybe you’re not.
So my question is, does this appear to you to be changing? Do you think that there is now an appeal to the American public of this idea of performance legitimacy, especially since procedural legitimacy no longer appears in America to deliver the goods when it seems to be so badly eroded? Is there kind of an uptick in appreciation for performance legitimacy?
Because I mean, just to put my cards on the table, I mean, I’ve noticed since January of this year a vibe shift, especially among younger people, in their attitudes toward China. And often it seems to be on the grounds that, hey, look, they deliver the goods.
I think there absolutely is a sense even within the American elite to say, well, we design all of these… Procedures in place in order to ensure some sort of fairness and making sure that the public interest is consulted. And I think there has been a sense even within the Democratic Party that, you know, we take a look at these blue states and blue cities, big cities, which are almost unanimously governed by Democrats. And they don’t seem to be working all that well.
You know, there’s tremendous public disorder in a lot of cities. Mass transit isn’t functioning very well. A lot of politicians are much more interested to govern on social issues rather than delivering economic issues that many families, working-class families care the most about. And I think there is a sense that we can’t just rely on processes in order to deliver the sort of legitimacy that we’re talking about.
I think that that is a very vibrant debate within the left now, that we can’t simply be the lawyerly society anymore. How do we actually deliver the goods? And so this is where, to put my own cards on the table, I am in favor of abundance. I am in favor of Ezra and Derek’s program to create much better cities, show that California and New York are not deeply broken things.
That when voters point to the track record of Democratic mayors as well as governors, there is something real here to be able to say that they’re actually meeting the needs of the people rather than just making sort of statements and performative gestures that don’t actually deliver the goods for anyone.
So in the end, and here, I mean, we’ll kind of wrap up with this, but you know, the engineering mindset can be way too literal, right? And the lawyerly mindset can be way too formal. I guess what I want is some kind of conceptual pluralism. I want like this set of institutional practices that somehow are able to switch frames, you know, to use the right frame in the right moment.
I guess what I’d like to see is somehow that we build the muscle inside China, its one-party state, to build that muscle inside polarized democracies like the one we live in right now, to be able to do that, to be able to be, you know, conceptually plural in that way. And I feel like that’s what your book gets at.
Is that a fair characterization? And what are the ways we can build toward that kind of, you know, conceptual pluralism?
You’re absolutely right, Kaiser. And I’m glad that you picked up on this point, that one of the things I really craved after spending six years in China was some degree of pluralism, that, you know, it wasn’t just one official register speaking above all the rest. That was really eagerly censoring all of these different viewpoints.
And I think I’ve said so many cancelable remarks on this podcast, Kaiser, but let me offer a yet more cancelable remark. I think there is a better profession rather than engineers and lawyers to govern the population, and that is dentists. No, I joke.
I think that the right profession to govern the population, if we had to choose but one, would be something like economists. I think that economists have a sense of procedure, they have a sense of getting things done, and they have a sense of social science, not to engage in really stupid things.
Unfortunately, I think economists are the most reviled academic profession on the planet. They certainly have gotten into a sticky wicket for themselves. But I think one thing that I will always be glad for for economists is that they were the people most actively pushing back against things like policies like the one-child policy.
That was the case in China, in which it was the economist who was the head of Peking University that really pushed back against the one-child policy in earlier formulations in the 1950s. And it was mostly the economic profession in the West that pushed back against the population bomb by Ehrlich.
And so I think that economists are the happy go-between. But I think that economists certainly need to be supplemented by degrees of pluralism on themselves. There should be lawyers in government. Absolutely. There should also be engineers in government rather than the U.S. Senate, which has 47 people who went to law school and one person trained in engineering.
I think there should be some sort of a balance with all of these things. I certainly don’t want to be entirely ruled by humanists. Mao Zedong was many things. He was, I think, primarily a poet. And if you take a look at earlier iterations of the Soviet Union, you had all these fantastic writers around Joseph Stalin. They were such good writers. They were such good literary critics.
And look at what a mess they made. So I don’t want to be governed by poets and literary critics. That sounds like an absolutely terrible paradigm. I think what we need are people who understand social science. And so my nomination is to be ruled by economists.
I’m going to put my vote in for historians. I think they have that sort of… Perspicacity and then that broader frame. And they’re not as paralyzed as economists are. And if we have to go with economists, I’m going to go with the Arthur Krobers over the Michael Pettises to rule us. That’s a better economist, perhaps. I think I, as someone who belongs to an institution called the Hoover History Lab, think that historians would not be so bad either.
Yeah, not so bad at all. Well, Dan, what a fantastically fun and wide-ranging conversation I’ve had. I cannot recommend the book more highly. Make sure that you get out and buy it right away. It comes out on the 26th, on August 26th. I encourage you all to pick up a copy. Above all, it’s a really fun read. It’s full, like I said, of really great turns of phrase. I had a long list of memorable quotes from it that I put together as I was reading it.
Let’s move on now, though, down to the segment I call “Paying It Forward,” where I ask you to name-check a younger colleague, maybe somebody at Hoover. I mean, Hoover was full of villains as far as I can tell, but there’s got to be one person worth name-checking there before we move on to recommendations. So who do you offer “Paying It Forward”?
I will offer two names:
So those are my two names, Afra Wong as well as He Liu. He Liu and I have crossed swords a little bit on Substack. He’s extremely committed to the liberal project when it comes to China, and nothing wrong with that. But like I said, we’ve crossed swords a bit. But great recommendations both. Afra, I’ve seen some of her work as well, and it’s excellent.
What about recommendations, Dan? Do you have a book you’ve read recently that you would like to recommend or anything, film, music, anything at all?
Well, I think over the course of book writing, I really got myself back into the classics, the things that I have really enjoyed. And so I guess I will recommend two sets of things.
The first set are the Mozart’s Italian operas written with Lorenzo da Ponte. These are:
I found myself, over the course of book writing, listening to these highly pleasurable, fun, and inventive operas that I think will stay with me for the rest of my life. So these are the Italian operas by Mozart.
Yeah.
And I think what I will do is also recommend my quartet of favorite novels. I have four novels that I’ve been rereading recently. And so the first one is The Red and the Black by Stendhal, which has these incredible depictions of the mistakes and stupidities that one commits in the act of love. This is a French novel that was published in 1830.
I will also throw in another French novel, the Proust. And so these are really wonderful, intoxicating tales of love that Marcel Proust has created for us. The entire series of In Search of Lost Time.
That’s right. And the Penguin translations are all quite good in English.
A third novel is that everybody is reading Moby Dick this summer.
Yeah. Why is that? Why is everyone reading Moby Dick? I mean, I know that Joe Weisenthal from the Bloomberg Odd Lots podcast seems to be leading the charge on this. But I reread Moby Dick about, well, maybe six or seven years ago. Yeah, fantastic novel. But what do you think explains it being such a zeitgeist thing this summer?
It is just like a strange and bizarre marvelous white whale. You never know at which corners of the four seas that Moby Dick will shoot his spout up. And so I think that’s a little bit of a mystery to me. But I am I am a dickhead. And I love the depictions of mesmerizing whale lore.
And my favorite final novel is Bleak House by Charles Dickens. It is just this very fun, inventive, clever book that is a miracle of construction. So I commend it to your listeners.
Operas by Mozart, as well as this quartet of novels. Fantastic. Great, great, great, great.
I have a couple of recommendations. One is by Yun Sun from the Stimson Center. She heads their China practice. It’s in Foreign Affairs. It’s called China is Enjoying Trump 2.0, which I thought did a really good job of sort of channeling Beijing’s perspective on what’s happened in the time since Trump took office now. It’s like seven months now.
It’s really good. She’s always solid. And this is a particularly, I think, Excellent view into the Chinese mind on this. I also want to plug a book I’m reading right now. It’s called Revolutionary Spring by Christopher Clarke, who is one of my favorite historians. It’s just an amazing work of history.
Hopefully, you’ve read his earlier book, The Sleepwalkers, which is about the run-up to the First World War, which I also highly recommend. I’ve actually recommended it before on Seneca.
I’m hard-pressed to think of a working historian who has all the things that Clarke brings to the table, which is just an obvious facility in so many languages and this ability to just zoom in. Because the revolutions of 1848, which is what Revolutionary Spring is about, these happen all over Europe and at the same time.
So if you’ve got to write a book on this, you need to be able to:
And the other thing, of course, is that Clarke is just a brilliant, brilliant writer. His prose is just delicious.
I think it’s such a good book. It’s a really hardcore history. I mean, it’s not for the faint of heart. There’s more detail than I think a lot of people are used to, and it’s just great. So that and Sleepwalkers—my recommendations.
Dan, once again, thank you so much for taking so much time to talk to me. And happy birthday.
“Thank you.”
What is it? Happy birthday.
“It is my 33rd birthday.”
What is three? Is that an auspicious number?
“I’m not sure.”
Well, it’s half of 66, which is an auspicious number.
“Okay, that’s good.”
Yeah. Thank you so much for taking the time. And congrats on the book, which is, again, just so terrific. It’s been a total delight.
“Thank you again, Kaiser.”
Looking forward to seeing you again.
You’ve been listening to The Seneca Podcast. The show is produced, recorded, engineered, edited, and mastered by me, Kaiser Kuo. Support the show through Substack at www.sinecapodcast.com, where there is a terrific offering of original China-related writing and audio.
Email me at sinecapod@gmail.com if you’ve got ideas on how you can help out with the show. Don’t forget to leave a review on Apple Podcasts.
Enormous gratitude to the University of Wisconsin-Madison Center for East Asian Studies for supporting the show. Huge thanks to my guest, Dan Wong.
Thanks for listening, and we’ll see you again next week. Take care.
Bye.
]]>Okay, we’re here at OpenAI with some exciting news from the AI for Science team. With us is Kevin Weil, from, I guess, your VP of AI for Science.
VP of OpenAI for Science, yeah. OpenAI for Science, and Victor Powell, who is the product lead on the new product that we’re talking about today. And with me is our new AI for Science host, RJ. Welcome.
“Thanks for having us.”
“Thanks for having us. Yeah, it’s very good to be here.”
“Thanks for hosting us as well. It’s always nice to come over to the office.”
What are we announcing today?
So we’re launching Prism, which is a free AI-native LaTeX editor.
What does all that mean? Because probably a lot of people on the pod haven’t worked with LaTeX in the past. LaTeX is a language, effectively, for typesetting mathematics, physics, and science in general.
So if you’re a scientist writing a paper, you’re probably not using Google Docs because you need to — you have diagrams, you have equations, et cetera. But it’s — and it’s been the standard for decades. But the tools that people use to actually write LaTeX, write their papers, haven’t changed in a long time.
And in particular, AI can help with a lot of the tasks, right? Because you spend your time doing the science, you need to write it up. That’s an important part of communicating your work. But you want that to be fast, and you want that to be accelerated, and AI can help in a ton of ways. And we’ll talk about some of those.
But if you step back, right, it is OpenAI for Science. Our goal is to accelerate science. And the surface area of science is very large. So we’re trying to build tools and products that help every scientist move faster with AI.
Some of that is obviously the work that we can do with the model, making the model able to solve really hard scientific frontier kind of problems, allowing it to think for a long time. But it’s not only that, right?
If there was a lesson from what happened over the last year with software engineering, it’s that part of the acceleration in software engineering came from better models. But part of it also came from the fact that you now have AI embedded into the workflows, into the products that you use as a software engineer, right?
And so that’s what we’re doing here. So OpenAI for Science, it’s both building great models for scientists and also speeding them up by bringing AI into the workflow. That’s what we’re doing with Prism.
I often say like every million copy and paste done in ChatGPT, there’s probably some product to be built.
“Right, exactly.”
That’s a good analogy.
“Yeah.”
That’s a good way to look at it. Especially with LaTeX, having written a lot of LaTeX papers.
“Yes.”
“Yeah, me too.”
The number of hours as a grad student I spent trying to get some diagram to line up exactly.
“Exactly.”
“Oh, man.”
Yeah. Cool. And Victor, this is your sort of baby.
“Yeah, I guess it started off as just a project. I left Meta about three years ago trying to look for various different projects to start. And this was one that like when I sort of presented it to people, they’re like, oh, I get it. That’s, I see what you’re doing.”
And so I’ve just been focused on that, building it for about a year and a half. And, you know, it has now become part of OpenAI and that’s been very exciting.
“Congrats.”
“Thank you.”
Yeah. So it’s kind of a fun story, right? I mean, we, as we were thinking, we had this thesis around, it’s not just models. It’s also building models into the workflow and accelerating scientists in that way.
And this is, there are obviously a lot of different ways that you can do that, but the scientific collaboration and publishing thing is definitely one of them. And I was looking around like, what is there in this space? And there hadn’t been a lot of innovation for a long time.
Like it wasn’t that different from when I was writing up my assignments and papers in tech and grad school. And then I found on this Reddit forum, maybe it was /r/LaTeX. I don’t remember, but somewhere on this Reddit forum, I found this thing about a company called Cricket.
And I was looking around, I couldn’t find who the founder was. It took me a little while. And then I think I found you on Twitter and DM’d you out of the blue and just said,
“Hey, I don’t know if you want to talk about this, but I would love to talk about this if you’re open to it,”
and gave you my number. And we talked on the phone and then jumped on a Zoom and eventually met in San Francisco and made it happen.
“That’s right.”
It’s awesome to have you guys here, but it’s just, yeah, I have a ton of respect. For what you, what you started to build. I actually never heard that full story from you until now. You gotta find that Reddit user and thank them because, you know, it might have been me.
I thought you were totally in stealth because it was the hardest thing to actually figure out who the founder of this thing was. And then I was like, “Oh, for sure. He’s not going to respond to my random DM.”
I mean, I guess that’s a part of, part of our focus has always just been entirely on product, and to the point where it’s almost embarrassing how little we focus on anything else.
Yeah. It worked out for you.
Also full circle for a moment for you using Twitter to do your business development.
Yeah, that’s right. So that’s kind of interesting.
Like I actually, yeah, probably one of the most important social network innovations, I guess, is those, that stuff. And I’m sure you know a lot about that.
Shall we go right into a demo or talk about it?
Yeah, always fun to show it.
I’m a fan of show, don’t tell. Push people to the video.
All right. I’ll try and arrange this so you guys can see a little bit.
Yes.
So what you have here, so this is, this is Prism. And what you can see is on the left here, this is actual LaTeX. You can see why you might want AI to help you write it because it’s a little bit, it’s a language. It’s a little bit messy.
And then on the right, this is my colleague’s paper. Alex Lipsoske is a physicist. This is a paper that he wrote on black holes. And so you see it over here, all the, all the, you know, you can imagine trying to write this in Google Docs or something — it’d be impossible.
This is why LaTeX is super powerful.
And then, you’ve got kind of your files here that make up the project:
You can go through and change it and then you compile that into the PDF itself. But here I can say, at the bottom, you can use the AI using GPT 5.2. And I could say, you know, this introduction, maybe I want a little help writing the introduction.
So,
“Help me proofread the introduction section paragraph by paragraph, suggest places where I can simplify.”
There’s a lot of demo and we’re working on it pretty heavily. So just, you can’t be nervous.
Spoken like a true founder.
One of the nice things is you could do this in ChatGPT, but you’d have to go upload your files into a chat, right? And you’re going back and forth here because the AI is built into the product. It has all of the files that are part of your project. It automatically puts them in context. It works the way you think it would work.
So here it’s looking at the files.
And it’s given us kind of a diff here. So it’s suggesting changes. You’ve got:
You can see the different places where it is suggesting that we change things.
So, okay, we can, we’ll just keep all of them, right? YOLO.
Here’s the thing — we’re changing Alex’s paper. What’s the big deal?
So here’s another thing we were talking about: diagrams in LaTeX.
So, I’ve got a, say, I wanted to input a commutative diagram, right? It’s really easy to draw a commutative diagram like this. Yeah, it is an absolute nightmare to put these things into LaTeX.
So I will upload this photo and I’ll say here, whoops.
Is there a tech bench for this kind of stuff? Like a set of evals?
So here’s a commutative diagram that I drew on the whiteboard:
“Can you make it into a LaTeX diagram and put it right after the, I don’t know, right after, right before, right at the top of the introduction section? Make sure you get the details right.”
So, I didn’t want to interrupt you while you were typing, but why don’t you use voice?
Oh, actually I should. I totally could.
Yeah.
No, but isn’t it interesting that we all have these voice buttons and we don’t use it?
Yeah, it’s not second nature yet. Like it’s interesting.
And that one I totally should have. I was going to also show something. So here I am in the LaTeX and it’s working.
You also can create new parallel chats. So you can have whole sessions with ChatGPT that can be going in parallel.
So here I’ll ask it, there are all these equations. We’re talking about symmetries of this black hole wave equation. And in particular, there’s this complex symmetry here. I like how it. Notice how, yeah. Yeah. Notice how it sinks when I highlight it, but I’ll say like, why don’t you, I’ll go to my chat so I can start doing this in parallel.
I’ll say, please make sure, or please verify that the H plus operator in the new symmetries section is indeed a symmetry of the stationary axisymmetric black hole. Do you understand those questions? You lost me. Are there a whole wave of equations? I have, but after that. I don’t know if Brandon is actually a natural physics person.
Yeah. I’ll say, don’t do it in the paper. Show it here. I don’t want it to actually edit the paper. I just wanted to prove it here. Right. Yeah. Okay. So I’ll get that going.
Now, while we’re waiting for the diagram to finish, we can also get another thing going in parallel. So I’ll say, I need to write up a set of lecture notes on general relativity. You know, say I’m a professor, right? I’ve got, I’m teaching a class or something, put together a 30-minute set of lecture notes on a Riemannian curvature.
Wow. That’s a very different task. Put it into the file. I made this gr_lecture.tex. Okay. And so I’ve got this going.
All right. Well, it came back on my earlier one — H plus symmetry. Is it really here? You got ChatGPT doing a whole bunch of work to verify that this is indeed a symmetry of the equation. Okay. It does. It confirms it.
Right. So you’ve got the full power of a reasoning model that can think deeply about frontier science. And now we can go back while it works on the other thing.
Okay. So this was where I was making the diagram, right? It put it right below the introduction. I’ll compile it again. So this is an auto compile. You can turn that on.
Yeah. Okay. And it nailed it. So it looks like it got it pretty much exactly. Just a small check. Check the details.
Oh yeah. Good enough for me. Yeah. It’s pretty good, but all right, we can see if it’ll get it right. Let’s say, the C vertex should be directly…
To your point about voice though, I do think maybe over time the code might recede into the background more as you’re really interacting with the paper.
When you started this product, how were you envisioning it would be used? Or were there other design choices you were considering that you didn’t take?
By the way, before you answer it, we have our general relativity lecture notes here, but that was quick. So 30 minutes, this is a — yeah — so 30 pages there, 30-minute section.
Okay. So we got curvature, covariant derivatives. This looks like a reasonable set of notes if you were going to go teach a class, right? It just did it for you.
Or you can think like, you know, generate the problem set for this week.
So it’s got some examples here. We could tell it to work out solutions to the examples. That’s sort of a hidden feature of LightTeX too, that it actually makes it pretty easy to generate problem sets with answer sheets and things like this.
There’s so many cool features of LightTeX that I think are underutilized.
So anyways, you could see we:
And that’s just basically all in parallel.
And you can imagine lots of other things you can do.
For example, if you have a proof and maybe just have the bullet points on a proof, you can say, “Here are the bullet points. Now flesh it out for me.”
You can imagine checking all of your references before you publish, making sure all of them are real and up to date. You can imagine having it generate your references based on the topic.
There are just so many areas where AI can help. That’s a big problem when you’re trying to put together a paper: get all the references right.
This is time that used to go to typing a paper, not science. And now it can go back to science.
And that’s just one of the ways that we look at accelerating scientists all over the world.
I would say definitely be careful about including references you haven’t read.
Like that’s the point: you can put a hundred references, but if you didn’t read them, you might as well not have them.
But yeah, I think that web connection is very important.
And like, is this stock GPT five or is this like a fine tune?
It’s GPT 5.2.
Yeah.
Yeah.
But, and by the way, when you’re looking at references, you can also ask ChatGPT to help you understand the reference, you know, read this paper, tell me the relevance. So all of the things that you might want to do to accelerate your work, you can just do from within this interface.
You still have to do your work, but it should make it faster, especially like even linking to the references. So you can go and verify like, okay, this is this one. So this might also make it easier to write the paper as you do the work, right? Rather than, rather than, oh, okay. Now I got to spend two days in LaTeX land, like trying to get my paper.
Right. Like a tool for thought rather than just a publishing tool.
Yeah.
What about collaboration?
It’s great.
Yeah.
So it’s built for, I mean, you can speak to this. Well, it’s built for collaboration. So you can bring on as many collaborators as you want, which is nice. I think most other tools in the space have hard limits and charge you money and other things. In Prism, it’s as many collaborators as you want for free.
Commenting.
Yeah. So you’ve got commenting, you’ve got all the kind of collaboration tools that you would want.
Yeah.
Good.
And then any of the like engineering choices, like, you know, what might engineers not appreciate when just looking at a tool like this?
Often it would be like multi-line diff generation that you need to do because you’re editing a pretty complex document. It does get pretty complicated. I mean, we’re using, let me know if I’m getting too technical into the weeds, but, you know, we’re relying heavily on the Monaco JavaScript framework.
So that I’m very familiar with the lack of documentation of Monaco. That’s actually interesting you say that because it’s very true that it’s an extremely powerful library that is almost entirely undocumented.
Yeah. It’s just types. But you can use codecs now to generate the documentation for you.
Yeah. You think Microsoft should get on that.
But yeah, yeah.
You know, like just stuff like that. Like I like to hear about the behind the scenes of like building something like this.
I think initially one of the, one interesting challenges was that we really pushed on it being WebAssembly and fully just running in the browser at first, the whole entire LaTeX compilation. That did help us in the sense that we were able to flesh out the design and the AI capabilities early on without having to invest heavily in the backend infrastructure.
But eventually we did hit a wall with that approach. Once we switched it to a backend PDF rendering, that’s when we really started to hit an inflection point with usage.
Now fast.
Yeah.
Yeah.
I think we also, the AI in here benefits a lot from everything that we’ve learned building codecs. And as we go forward, I think we’ll likely just integrate the full codecs harness into the application here.
So you get all the benefits of the tools and the skills and all the things that codecs can do today, and you just sort of automatically can bring that into your environment here.
Yeah.
Are they just the same app?
Maybe. I think potentially it depends on…
I mean, here’s the reason I’m hesitating: I think the interesting thing with this and with codecs is we’re still mostly in a world today where:
But the more that AI improves, people trust it and they’re just YOLOing it, right? You’re generating code and you’re looking at the code sort of secondary to instructing the AI and driving from that.
The UI probably changes for all of these things, right? You don’t need your document front and center because you’re actually not looking at your document as much. That’s sort of your backup and your interaction with your AI is primary.
And as that happens, I think you might see these UIs kind of converge over time. So we’ll see.
But I definitely would love to see a world where people needed to spend less time thinking about the actual syntax and much more about what they’re trying to create.
Yeah.
I feel like this plus a notebook would be amazing.
Yeah.
Because, because, and something that AI can run quite, run a analysis, generate plots. So stick that in the paper here. Like, “Oh, read, you know, like this paper, like this part of the paper, like take that equation and like, you know, do something with it.” That would be a really amazing integration.
Yeah. Like think through the different corollaries of this thing from this paper and produce some alternatives. And then like, yeah, I completely agree. Yeah. Yeah.
I do think that’s sort of the progression where it’s like doing, doing maybe work for a few seconds versus maybe we’re already at a point where it’s doing work for a few minutes, eventually doing work for hours, days, coming back with very complicated analysis.
Yeah. I mean, that, that’s actually maybe a good segue into some of the other questions that I had about your initiative.
I mean, so stepping back to AI for science in general, can you talk a little bit? I have a million questions, but maybe start with what I… okay. I feel that validation of AI for science is critical to its success, right? You have to have some sort of real world validation of the results that you produce with your AI, right?
So what are the, I know that there’s been some publicity in the past. What are the latest and greatest hits of the things that big labs or any lab is doing with open AIs?
I mean, when you step back and look at the trend, I think that’s the biggest thing. Because we can debate exactly – like you’ve probably seen in the last few weeks, even – there’ve been a bunch of different examples of like GPT 5.2 contributing to open research problems and things like that.
And then you get into this debate of:
And you know, that’s a legitimate discussion. But when you step back two years ago, we were like, you know, this thing can pass the SAT. That’s amazing. And then you progress to like, it can do a little bit of contest math and it can start to solve harder problems. Wow.
And then you keep going and it’s starting to solve graduate level problems. And then you have a model that gets a gold medal at the IMO. And now we’re sitting here talking about, you know, it solving open problems at the frontier of math and physics and biology and other fields.
So it’s just, I mean, the progression is incredible. And if you think about where we are today, then you fast forward six months, 12 months. I am very optimistic about what the models are going to be able to do to accelerate science.
Yeah. It’s like, it’s already happening. If there’s one thing that I’ve learned from my like two-ish years at OpenAI, it’s:
“You go very quickly from this thing is just impossible for AI to do. Like it’s too hard. I can’t do it” to “Hey, I can just barely do it. And it kind of doesn’t work. Only early adopters are doing it because it’s not particularly reliable yet, but it sort of works,” to “Oh my God, AI does this thing really well. And I could never imagine not using AI for this in the future.”
It’s like, once you start to get to, you know, five, 10% on some particular eval, you very quickly go to like 60, 70, 80. And we’re just at the phase where AI can help in some — not all, but in some elements of frontier science, math, biology, chemistry, et cetera. And it just means we’re like right at the cusp and it’s super exciting.
So, I mean, fast forward a year or, you know, the end of the year, and we have AIs that can do a lot of this discovery process. Then the bottleneck becomes the wet lab or the lab, right?
Yeah. So what are you seeing in that domain?
Yeah. By the way, I totally — we were talking a little bit about software engineering before and the analogies. I think 2026 for AI and science is going to look a lot like what 2025 looked like for AI and software engineering.
Where if you go back to the beginning of 2025, if you were using AI heavily to write your code, you were sort of an early adopter and it kind of worked, but it certainly wasn’t like everybody was doing it. And then you fast forward 12 months and at the end of 2025, if you were not using AI to write a lot of your code, you’re probably falling behind. I think we’re going to see that same kind of progression in AI and science where, today it’s early adopters, but you’re really starting to see some proof points and solving open problems, developing new kinds of proteins and things like that.
But you’re right, as it really starts to work. I think this is the year that it’s really going to start to work. It shifts the bottleneck.
And I think we’re going to be starting to talk a lot more about robotic labs and other things. Like, do you need to have a grad student pipetting things? No, probably not. Right now you do, but why shouldn’t we have robotic labs where you have AI models doing what they do best—reasoning over a huge amount of different information.
They have read substantially every paper in every field and can bring a lot of information to bear to help prune the search tree on, for example, a new material that you’re trying to create. Then you have a robotic lab that can roll out a bunch of experiments in parallel, do them while we sleep, and feed the results back into the AI, let it learn from them, design the next set of experiments, and go.
So, it’s hard to imagine that doesn’t even have to be yellow science, right? To your point, you’re verifying it as you go because you have an actual lab building it in real life. But you can just do so much more in parallel. You can think harder upfront with AI to design the experiments, prune the search tree, search over a smaller number of higher-value targets, then automate the experimentation and turn it around faster.
And again, like this is acceleration: if we’re successful, you end up doing maybe the next 25 years of science in five years instead. So in 2030, we could be doing 2050 level science, and that would be an awesome outcome. The world is a better place if that happens.
Absolutely. I guess we spoke recently with Heather Kulik at MIT, and one of the things she pointed out was that there’s an element of serendipity to working in a lab that you lose. She was of the opinion that there’s
So again, you’re at a bottleneck, but humans need something to do.
Well, what she said sounds totally reasonable to me. There are probably places where humans are adding no value because they’re literally just trying to pipette a certain amount of a thing and do another thing, or do some repeated motion in a bunch of different ways.
And then there are places where it’s less well understood. You want the full flexibility of a really smart human thinking about the work they’re doing.
By the way, the same is true in the more theoretical fields as well, where it’s not about automating all humans out of their jobs. This is about accelerating scientists. It’s scientists plus AI together being better than scientists alone or AI alone.
I think the same is true whether you’re talking about:
- something happening in silico proving a theoretical problem
- something happening in the real world with a lab
Find the parts that you don’t need a human to do and try to automate them as much as possible so the humans can spend their time on the most valuable things.
I’m very pro the in silico acceleration, because you have more control over that and you can parallelize, repeat, and do all those things.
I think there will be huge value because a lot of fields are heavily simulatable. For example, nuclear fusion runs a lot of simulations before experiments because experiments are very time-consuming and expensive.
But I’m excited to see what you can do when you have a loop between a very intelligent reasoning model that understands fusion and a simulation: the model thinks about what parameters to set, runs a bunch of simulations in parallel, feeds that back, and you have the same sort of lab loop—except it’s all in silico, running on a giant GPU cluster.
Then, when you’ve really gotten to the end of that calculation, you go run it IRL.
This is bringing it back to prism. This is sort of a nice aspect that you’re getting a more sophisticated view of your result, right? Instead of just, you know, like a chat output in it, I would hope as it develops, it’s a way for a scientist to be able to interact with the information before you kick off your nuclear fusion experiment for, you know, $10 million or whatever.
And the human can learn from more things, right? You just get more data that you can look at and evaluate. So, yeah.
So this, by the way, this fusion discussion makes me think like, you know, if one day opening after science, you know, it gets serious enough and starts to self-accelerate, you should solve cold fusion and, you know, be your own power source.
Well, I mean, this is why we’re so excited about this, right? I mean, imagine our mission is to bring AGI to the world in a way that’s beneficial to all of humanity.
It’s right there at the lobby. Yeah. It’s amazing. You see it every day you walk in, you see it. Yeah, absolutely.
And imagine, I mean, if we had GPT-9 inside of ChatGPT today, it would be awesome. You could do lots of things. But if you had GPT-9, which I’m using as a stand-in for AGI, and it could:
Like, that’s the real benefit of AGI. That’s, I think, maybe the most tangible way that we’re all going to feel AGI as it starts to be real.
Yeah. And that’s why this work is so mission-driven for us.
So, that brings up two questions in my mind:
Because this is how—though you laugh, it’s a little bit serious—all the AI for drug discovery companies ended up being drug companies because they couldn’t sell the drug, so far, with some exceptions now like Noetic, for example.
But they end up being drug companies because they can’t sell the drug. In any event, there’s a lot of precedence for using AI to basically build your own portfolio.
So, are you thinking about that angle or this is right now just about enabling scientists outside of OpenAI?
Yeah, I mean, my personal belief as we drive towards AGI is not that we’re going to create AGI and then all sit back and enjoy our universal basic income and write poetry. The future will involve, especially advanced science, experts helping to drive these models.
I don’t believe any one company is going to do everything. That’s why we’re focusing, first and foremost, on accelerating scientists outside of these walls. Our goal is not to win a Nobel Prize ourselves, it is for a hundred scientists to win Nobel Prizes using our technology.
Yeah.
At the same time, I think there are places where sometimes, when you’re building for other people, you learn best if you actually go end to end on something.
Yeah.
Because then you’re your own customer and you understand it in a tighter loop than you would if you were purely building for people outside the walls.
So, I think it makes sense for us to take a handful of bets like that, but by and large, we’re going to partner because the surface area of science is massive.
Yeah.
And we want to accelerate all of science.
Yeah.
We’re covering all sorts of disciplines from chemistry to structural biology to material science. It’s all over the place. There’s a lot to do.
One thing I did want to bring across also was that AI for Science sits within the broader research org at OpenAI. One of the more interesting things is like self-acceleration, let’s call it.
Where Jakub has very publicly declared that we’ll have an automated researcher by September 2026.
Yeah. The beginnings of one, I think you said, right? Like the intern version this year?
Right.
First product.
Yep.
And I’m sure you have more cooking internally, but why so soon? That’s eight months away. What’s the goal there? Anything you can share?
Yeah, I mean, eight months feels like forever in this industry. AGI by then? Basically infinite time.
I mean, no, it’s exactly what you said, right? It’s if we can create a a model, an AI researcher that can actually do novel AI research, then we can move way faster, right? We will self-accelerate. We can discover more things quickly. We can apply GPUs and compute to moving our own research faster. And that just means that we can improve our models at a faster rate.
And every bit that we improve our models means that we are a step closer to bringing AGI and all the things that we were talking about with personalized medicine and new materials. And, like, we can bring these amazing things into the world faster. So it is about self-acceleration.
Yeah.
I think one thing I’m most trying to figure out is how closely is machine learning research, which is a science, or high-performance compute, which is also something that you guys are doing a lot of, close to the traditional hard sciences, let’s call it, like physics and chemistry.
I think in a lot of ways it’s sort of a parallel effort to this. Like, it is the work that we’re trying to do with AI, OpenAI for Science, and accelerating other scientists. The parallel internally is they’re trying to build products and models for AI researchers to accelerate them.
So there’s a lot of sort of parallelism to these two work streams. They’re similar in goal, just for a different set of users.
Yeah.
Okay.
Any parting thoughts, questions, anything we should have asked?
Well, I hope everybody tries Prism. It’s available today at prism.openai.com. It’s totally free. You log in with your ChatGPT account, and you can go build anything you would like. We’re really excited to see what people use it for, and if you run into issues or have any feedback, let us know.
I have a paper I’m going to write really soon on that.
Amazing.
We’ll just show notes on this thing. I don’t know. Let’s see what it does in LaTeX.
Yeah. Totally.
Yeah.
Congrats on your first OpenAI launch.
There you go. Congratulations.
Congrats.
Thanks for having us.
Yeah.
Thank you.
]]>This episode is brought to you by Progressive Commercial Insurance. As a business owner, you take a lot of roles:
But when it comes to small business insurance, Progressive has you covered. They offer discounts on commercial auto insurance, customizable coverages that can grow with your business, and reliable protection for whatever comes your way.
Count on Progressive to handle your insurance, while you do, well, everything else. Quote today in as little as eight minutes at ProgressiveCommercial.com.
Progressive Casualty Insurance Company. Coverage provided and serviced by affiliated and third-party insurers. Discounts and coverage selections not available in all states or situations.
The American Airlines Advantage Business Program is changing the way companies book travel and get rewarded. Designed for fast-growing businesses, the program makes it easy to earn rewards. And it’s free to join.
Your company earns one Advantage Mile for every dollar spent on business travel booked anywhere with American. Use these miles to help offset future travel expenses, transfer to employees, and more.
You’ll also gain access to a suite of tools to streamline travel management, including:
And it’s a win-win. Travelers can earn additional loyalty points on top of what they already earned through the Advantage Program, helping them reach status faster.
Earn more on business travel you’re already taking with the American Airlines Advantage Business Program. Register today at aa.com/AdvantageBusiness.
Welcome to the Seneca Podcast, a weekly discussion of current affairs in China. In this program, we look at books, ideas, new research, intellectual currents, and cultural trends that can help us better understand what’s happening in China’s politics, foreign relations, economics, and society.
Join me each week for in-depth conversations that shed more light and bring less heat to how we think and talk about China.
I’m Kaiser Guo, coming to you this week from my home in Chapel Hill, North Carolina.
Seneca is supported this year by the Center for East Asian Studies at the University of Wisconsin-Madison, a national resource center for the study of East Asia. The Seneca Podcast will remain free, but if you work for an organization that believes in what I’m doing with the show and with the newsletter, please do consider lending your support.
You can reach me at sinecapod@gmail.com. And listeners, please support my work by becoming a paying subscriber at sinecapodcast.com. You’ll enjoy, in addition to the pod:
And of course, you can bask in the knowledge that you’re helping me do what I honestly believe is important work. So, do check out the page, see all it is on offer, and consider helping me out.
Today, my guest is Afra Wong. I suspect many of you will already have come across her work through her podcast, through appearances on other China-focused shows, or through the many provocative, beautifully written, and fascinating essays she’s published.
Afra is a writer working between London and the Bay Area, currently a fellow with Gov.ai, and previously with the Roots of Progress Institute. Before going full-time as an independent writer last year, she spent six years in Silicon Valley covering AI and crypto, running newsrooms, building developer communities, and absorbing the Valley’s growth logic from the inside.
She writes about China and about Silicon Valley — the latter sometimes metaphorically — but about neither of these places ever as mere abstractions.
She writes about them as overlapping systems, how China’s technological interiority shows up in Western debates about AI, industrial policy, and even progress itself.
She’s also the host of the Chinese language podcast Pipei Jiao Wah, Cyber Pink, and part of the Baihua podcasting community.
We’re talking today about her recent Wired piece on what might be China’s most influential science fiction project that you’ve never heard of: the Morning Star of Ling Gao, or Ling Gao Qi Ming, and the worldview behind it, something known as the Industrial Party or the Gung Yedang.
If you haven’t read that yet, click the link, read the piece. It’s one of actually several China-focused pieces in this issue of the magazine — some really good stuff. Come back when you’ve finished. We will still be here.
This isn’t just going to be a conversation about time travel sci-fi — though that would be a lot of fun — but actually about:
About how a country explains to itself why it fell behind. and what it thinks salvation looks like.
Afra Wang, a very, very warm welcome to Seneca.
Oh, wow. Thank you so much, Kaiser.
When you were describing my work experiences, it’s almost like I’m reliving my past life, especially my time doing a lot of growth stuff for tech companies and crypto. And actually, I discovered the Morning Star of Ling Gao, or Ling Gao Qi Ming, as a collective science fiction novel writing project from my crypto phase.
Really?
Yeah. I was told by a lot of nerdy technologists, people who are Chinese cypherpunks, saying there is the greatest DAO experiment ever, which is a sci-fi story collectively written by many people, like hundreds of thousands of people. I was like, “wow, what do you mean?”
Because DAO in crypto represents decentralized, autonomous organization. Referring to this science fiction writing as a DAO experimentation is really fascinating. It also sort of reflects on the demographic — the people who are reading this story, right? Who are reading the Morning Star of Lingao? Who are reading Lingao Qiming? And it turns out to be:
Yeah, not surprising at all. A lot of overlap with sci-fi.
But before we get into sci-fi and about that essay, this is your first time on the show, so I’d like to give listeners a chance to get to know you a bit better.
You describe yourself as a kind of cultural in-betweener, and that really resonates obviously with me. For people who move between China and the West, especially when writing about technology and about power, translation isn’t just a linguistic exercise. It’s actually epistemic, but it’s also moral and maybe even aesthetic. I mean, it covers pretty much all of philosophy.
One thing that struck me reading your essay is how effortlessly you seem to do this, just to kind of code switch, not just in language, but also in your moral and emotional register, especially when you’re writing about something as charged as the industrial party. Is that something you experience as deliberate, or does it feel almost second nature to you at this point?
I think probably I am a somewhat open-minded and perceiving person, so I don’t know, people have been telling me that I tend to kind of like be able to make friends with all kinds of people. I think that’s, in a sense, like a good trade for me to be a more discerning writer because I think I’m really sensitive to vibes.
Also, I like to use the vibe because this is how I feel. I’m really sensitive to the aesthetic, the sensations when I encounter something, for example, the Silicon Valley mental model versus the Hangzhou-Shenzhen-Beijing mental model, right? I was really fascinated by the sort of the cognitive infrastructure, like the intellectual backbone of the Chinese version.
So I, you know, last year I wrote something called the China Tech Canon, which is a response.
Yeah, that was great. Thank you so much.
Yeah, I think it’s like, it’s all come to the sense that I want to like deeply, contextually translate certain, you can say:
I want to translate something to the Western discourse, but in a much more like humanistic and personal way because I think I am somehow constantly digesting cultures from both sides. I am native in Chinese, but I feel really native in English as well, in the Silicon Valley discourse as well. So I think that I’m just kind of like naturally juggled in between.
Do you go the other direction as well? Do you translate the Silicon Valley kind of tech canon into Chinese as well? Or do you find yourself doing more sort of the explanation in the direction of explaining China to the West?
Yeah, so not about technology, but I’ve been doing this Chinese language podcast for many years with my amazing co-host. I think all of us are cultural in-betweeners and we actually translate the Western popular culture and then talk about those Western popular culture in Chinese language. You know, for example, the popular movie Hamnet is a golden global hit. And we recorded a podcast about Hamnet in Chinese language, but the whole context, the theme, and the reaction, the catharsis we experienced — we were basically discussing this movie in Chinese language, although it’s a quintessential English movie.
Yeah, I read the novel. I have not seen the movie yet. Is it good?
Oh, it’s absolutely good. It’s so moving. It’s very touching, and you do experience this Greek tragedy style catharsis at the very end because it’s like a movie to Force you to confront a lot of eternal questions like death, like loss. Like, yeah, it’s such a layered movie, I can’t really explain it. It’s beautiful. It absolutely changes some part, like, deepest part of you.
So do you ever find yourself judging things differently depending on which context you’re inhabiting? I mean, because, I mean, not because you think one side is right, but because, you know, different histories seem to demand different weights, different priorities. You know, I mean, this is something I’m constantly wrestling with.
How conscious is that process for you when you’re writing? So, you know, you might have one view of the industrial party, say, as a Chinese person living in China and another entirely looking at them from the outside and talking about that to Americans. So, do you find yourself sort of having different standards?
I think I do. I think I’ve been having double consciousness since I grew up as a kid in China. I have double consciousness in a sense that a lot of stuff can coexist although they look like contradict to each other but they could both be true.
Like, you know, in a sense I went through the whole Chinese education, right? I finished high school in China and then I only went to U.S. for college. And I think, I guess, like, accepting a lot of contradictory views and philosophies, as you said, abstemious knowledge systems is part of reality to me, I would say.
But I still think the Chinese Chinese me and English me or the sensible me and anxious immigrant me, when they’re coexisting, I think there is a converging aesthetic standards or sensibility that I uphold. For example,
I think there is certain sensibilities and aesthetics that’s always true and always, I could always try to stay true to that.
Wow, that sounds so healthy and grounded. That’s fantastic. It seems like you experience this kind of ability to code switch and to experience sort of two whole different moral and epistemic systems as more of a freedom than a burden, then.
I would say so. Like, for example, this piece for Wired, it’s about Industrial Party, it’s about this poorly written, crowdsourced science fiction writing. I do not like reading this piece. I do not like reading this story at all because it’s so poorly written.
But at the same time, it gives this energy and spirit of what people are actually craving for in the rapidly developing, urbanizing China and why people feel so strongly about this developmentalism. And in a sense, maybe U.S. needs more poorly written collective science fiction like Lingao because U.S. right now kind of needs some industrial party people.
I mean, I hate the story. I hate the, you know, like the greatest Chinese science fiction as the title of this Wired piece is actually an irony, right? It’s not actually greatest because it’s like honestly really bad, but it speaks to so many things that I, yeah.
We’re going to get really deep into Ling Gao in just a second here, but there are a couple other things I still want to ask you about because there’s another divide that I see you moving across really fluently and that’s the one between STEM and the humanities, between, you know, the engineering ways of seeing the world and the more humanistic or cultural ways of seeing the world.
So reading your work, I get the sense that you’re genuinely at home in both of these registers. You’re able to translate between them without, you know, romanticizing the one or condescending to the other. Is that, again, something that you’re conscious of when you’re writing or does it feel like a natural part of, you know, how you make sense of tech and society?
I’m not sure if I’m really fluent in the STEM language. First, I am not a technologist. I don’t code except, you know, like right now, live coding makes everything easier. Everyone’s doing that.
Yeah, everyone’s doing it. Not me.
Yeah, I honestly don’t think I speak the KPI coded language, like optimizing everything, improve everything, because I do have a lot of friends that are like that, but I do think working in tech company gives me a sense that an entire corporation, like hundreds of people could just like grind really hard, iterate the product really hard just to like improve
- 2% of user retention or
- 1% of daily active user
because I’ve been there and… I was one of the people who were trying really hard to retain users, study the users, or try to improve the recommendation algorithm, so our app has more revenue that day. You can see this is all correlated, right? I was a content manager, a growth manager during my first job. When you put out a certain content or adjust the algorithm a little bit, there’s an instant bump in your revenue that day. It’s almost like it’s extremely correlated.
If you spend more money on Facebook’s advertisement, you will just get more new users. It is so direct in the tech world, and I do think I understand that eagerness or straightforwardness in the tech landscape.
This divide, though, between the STEM view and the humanities view, do you feel like that divide is even more acutely felt in Chinese life than in the Western context? I mean, the gap between the engineering dude and the artsy fartsy literati type—do you think that’s an outdated caricature by this point, or is that still something very much a dividing line in Chinese life?
I think China’s society logic was dictated by the STEM optimization logic, or like industrialization logic for a long time until the young people are so tired, people are so tired, and then this sort of optimizing bubble bursted.
So back then, maybe 10 years ago, optimizing everything — trying hard. There was an internet slang for people trying too hard, trying to get promoted, make a lot of money during the economic boom — during the Chinese economic boom and internet attack boom. This was admirable.
But right now, this bubble bursted, so people proactively do not want to participate, such as Nuli lore, Nuli fairy tales. Instead, you see China’s today’s mainstream sentiment is:
This is the current mainstream. I would say China is a post-industrial party society now.
Oh, good, good. I’ll feel more at home there because I’m a good old Gen X slacker, so I know all about avoiding work.
I mean, it’s interesting to me because I feel like I agree. There used to be a period where one side of that divide was absolutely treated as:
and the other was just written off. But yeah, I’m glad to see this swinging back.
Before we get into what Lingao represents, I think it’s worth situating it a bit. For listeners outside China, it’s almost completely unknown, as you said. How widely known is it inside China, especially among communities that care about:
Is this like a cult classic or is it something closer to shared cultural infrastructure?
I don’t think it’s widely known as a popular cultural product like a movie or Journey to the West. This is basically the most common vernacular day-to-day language.
But I think Lingao is very popular, very influential in a niche community. This community itself is what I would say the elite class of technologists, the STEM people who see themselves as pillars of China’s urbanization and industrialization, and predominantly male.
So Lingao, to be completely honest, strikes me as a semi-misogynist, misogynist novel because a lot of plots imply many things towards women.
But Lingao is a cult fetish. It is a Bible for the industrial party, this loosely connected intellectual group in China.
Yeah, I definitely want to ask you about the gendered nature of this book and about science fiction more broadly. I remember reading Senti, Three Body Problem and just being shocked. There’s stuff you could not get away with in America today, just the level of misogyny that was in there.
But how did you come across it yourself? You told me that you heard about it from Crypto Bros in the Valley, right?
Yes. Chinese Crypto Bros. I heard it from Chinese Crypto Bros.
Yeah, that’s hysterical.
What finally got you to read the thing? I mean, what did it just keep coming? And before you actually surfed over to it, what did you think it was? What kind of reputation did it have in your mind before you actually read a page?
I actually didn’t know anything except it is like a DAO experiment. It is a crowdsourced sci-fi lore. And like, to be honest, when I read anything that’s Chinese internet native, I tend to have lower expectations because I know some of the products, some of, especially those fiction writing stuff, is almost like Harry Potter fanfic, right?
It’s not written by J.K. Rowling herself, but written by the fans who just spend an afternoon, put a lot of scrappy plots together, and then you have a fanfic. So I tend to treat Ling Gao as an interesting phenomenon, like a part of the deeper corner of the Chinese intellect as a lore instead of as a serious science fiction. So I kind of had a lower expectation entering this novel.
And it turned out to be, yes, it is very scrappy. It was written by so many people to the point they started collectively writing it since 2006. And then people just keep writing and piling up and piling up.
A few years later, people were like:
“Okay, now we have too many things. Like the plots are going to multiple directions. We need to sort of come up with a kinetical plot together.”
So someone came up and compiled the storylines together, which creates the sort of, quote-unquote, kinetical Ling Gao timeline as we see today. But you can guess the nature of collaboration is:
There are thousands and thousands of different branches. When I was reading it, I couldn’t really tell which part is the kinetical story and which part is the fanfic back then. But because it’s well written, it’s sort of merged into the kinetical.
Is it because they’re all very put together and scrappy? It doesn’t read a thoughtful thing but reads like a collective stream of consciousness. There are these people who did the actual organizing, who actually decided what is canon and what is sort of peripheral.
What do we know about these people, about these principal writers? Who are they? What kinds of backgrounds do they come from?
They all use pseudonyms online, but we know some phenomenon writers sort of emerged out of the Ling Gao scene, later became the influencers or the writers for Guancha Zhe Wang (观察者网). And Guan Cha Zheu Wang is inseparable from Ling Gao’s collective writing.
Give us a sense of what Guan Cha Zheu Wang is. I mean, they have a certain political slant, a certain reputation. Why don’t you explain what Guan Cha Zheu is?
Okay, so in the Wired piece, I told the readers that Guan Cha Zheu Wang is almost like Chinese breadboard, but I think it’s less like breadboard because it doesn’t punch up. It kind of only punches west.
So Guancha isn’t that up though?
Yeah, so it is, I would argue, a more thoughtful patriotic or nationalistic collective online magazine delivering a lot of pro-industrial policy, pro-state opinion pieces, and some of the pieces are quite persuasive.
You know, I used to be a reader of Guancha when I was in college. Guancha reached its peak in the early 2010s. The founder himself, Eric Lee, I think he studied at UC Berkeley.
Yeah, I think he was the same year as me, in fact.
I see, I see. Yeah, he studied at UC Berkeley. It seems like he made a lot of money and he sort of diverted his money into this collective intellectual body building and started Guan Cha Zheu Wang.
It’s like a think tank and online publication, but it really represents a cohort of writers who, just like Ling Gao, have a strong stamp background, very pro-China, very pro-industrialization, and very anti-West.
And early on, a lot of their pieces are similar to a little bit like today’s narrative on how to establish a strong national, industrial national identity, and unapologetically loving China and being patriotic.
Yeah, so it’s very, how to say, very rad, very internet native. I would argue they’re very internet native because all of them know how to talk. They’re actually… Really good writers. You know, I mean, Eric X. Lee, the person who’s really sort of at the heart of it, as you say, you see Berkeley graduate and a venture capitalist of some success, very, very wealthy guy. He, in fact, is very well-spoken and quite persuasive in some quarters. You know, he has this famous TED talk in English. I agree. You know, he gets this gigantic standing ovation from him. I’ve described him before as sort of the first sword of China apology. He’s very gifted, I mean, in that sense. Yeah.
Let’s get back to the 共业党. Yeah. You know, it’s often spoken of in juxtaposition to the so-called 秦华党, which I’ve seen translated variously as the sentiment or the sentimental party. Does this 秦华党 actually have a representative online novel or a body of literature associated with it, you know, like we see with Lin Gao or is this just a straw man? Is it a real thing even?
I think 秦华党, if I understand correctly, is like the basically the civic space existed once on Chinese internet and I would say they no longer exist. I can say Chai Jing would be seen as a 秦华党 by industrial party standard because Chai Jing is this Chinese journalist who would make a documentary about air pollution, you know, she would Under the Dome. Yeah, Under the Dome. Like she would make a lot of influential documentary or journalistic pieces to remind people that the human cost of China’s rapid development load, you know, she would care about the migrant workers’ rights. She would care about the people who are dislocated because of the deformation of the city, because of the reconstruction of the city. Xi would, you know, care about air quality, right? Yeah.
So anything that’s been negatively affected or left behind by China’s headlong rush toward industrialization, right? Yeah, I would say that both the party and industrial party, I mean, industrial party doesn’t have the power to purge the sentimental party or, you know, the humanistic, the free journalist, the China’s civil civic space, but industrial party justifies for the state to marginalize and purge what they call sentimental party.
But I think sentimental party is actually a core part of my formative experience because I was growing up in China where the internet was a place to discuss real things from political reform to rule of laws to freedom expression to many things. I remember reading a lot of absolutely brilliant investigative reports about
I remember reading so many great stories about the one-child policy, about how this one town in China has forged some ties to systematically trade female babies to have them adopt in the U.S. A lot of the stories like this couldn’t exist in today’s China because of the demise of a sentimental party, because of the state’s effort of eradicating them.
So the industrial party in a sense doesn’t have any real political power, but I think they are a collective unconsciousness of the regime, of what CCP really prioritized or really think about.
Just to be clear, when you eradicated them, it’s not like they were rounded up and locked up. You’re talking about censorship, you’re talking about all sorts of different lawfare efforts, pressures to, yeah.
Yeah, I mean, when I was in high school, when I was in middle school, I could go on Weibo and read about Han Han’s pro-democracy essays, and those are really bold, quite fundamentally radical essays, if you see them now. I would be reading Chatter 08, written by Liu Xiaobo. I would be reading a lot about Arab Spring. I mean, a lot of the content sort of existed inside of the Great Firewall. It’s a beautifully diverse, chaotic, steaming, intellectual space. I kind of grew up in this internet.
People in those internet forums seriously talk about civic stuff, seriously talking about can China have a political reformation in the future? Because those possibilities were so real back then.
I think when you talk about the industrial party, you need to sort of dial your clock back to the 2000s and 2010s, it was because the tension between sentiment party versus industrial party were really real. I wrote in the piece, I think the signature event was the Wenzhou high-speed rail crash, the train wreck. I still remember vividly where I was that day and how I felt because I was about to board the high-speed rail from Beijing because the high-speed rail finally because I grew up in Shanxi.
Shanxi is an economically backward… Province, so I was really excited to see Taiyuan finally had a high-speed rail connecting to Beijing.
Instead of staying in the old train to take an overnight train to go to Beijing, you can actually spend only three hours to go to Beijing now.
Beijing as a cosmopolitan city, in my mind back then as a high schooler, it’s so close by to me, I can just go there. I was so excited and then the story burst out about the terrible train wreck in Wenzhou.
I remember there was a huge debate online about who was guilty, right? Like, where is the weakest link in this? If you dig way back in the Seneca archive to July or maybe August of 2011, you’ll find the show that we did about that.
Yeah, so I remember back then, all the public intellectuals were still active, their other accounts are still not banned. So a lot of people online writing lengthy articles or posting online about the liability of the authority that didn’t have a proper monitoring system.
And so basically, the thing is because a certain signal was missed, two high-speed rail trains basically crashed face-to-face. It was basically a pure human mistake. It was because a certain message didn’t send to the other side, so the tragedy happened. It was pure human mistake.
But anyways, I remember so many people writing about it online and there’s this one piece basically crying for China to slow down. And it was like,
“slow down China, wait for people.”
Implies don’t let such bloody train crashes happen again because this is a price we cannot afford just to aimlessly progress.
And this is a moment when industrial party people came and then they took the stage. They organized a systematic rebuttal against the humanistic sort of pro-slowdown discussion.
Because the industrial party intellectuals have a lot of advantages for knowing so much industrial knowledge because they are the ones building a lot of Chinese infrastructure. For example, I featured this one intellectual, his name is Ma Qianzu, one of the authors.
So the industrial party people organized a big rebuttal and they systematically published many articles to not justify this accident but saying we should take this accident seriously, but this shouldn’t be the reason for China to slow down its building on the high-speed railway infrastructure.
And yes, I think the industrial party and the development logic won in the debate and so the result is China didn’t slow down.
I mean, like I think retrospectively if it slowed down maybe China wouldn’t have such a convenient, vast, amazing network of high-speed rail today. But I think back then if China should develop was a real and very visceral and painful question to confront.
A lot of people’s idea is no, we really shouldn’t progress like this:
Like, are we, like, why are we allowing ourselves to be the colony for development?
But I think right now we’re basically sitting in the future to meditate on the dispute and one could say of course development is China’s thing, is what China always wanted.
But no, like, you know, there are people strongly against a lot of things the government proposed. There are people interesting to ponder that alternative.
Yeah, but that’s what this itself is — it’s a, you know, Lingao itself is pondering an alternative.
Now I haven’t read it myself, not one bit of it, I’m probably not going to, but I’m hoping you can give us kind of a controlled spoiler.
So a wormhole opens to 1628 from our present, or from the present of the time you know, 10 years ago when they were writing this.
So how does the alternate timeline then unfold? What kind of society do these guys end up building in Hainan? How different does it end up looking from our own history? How much do they change history in this project, in the book?
Yes, so okay. So reading this book is very interesting because the plot evolves as the people who write the story evolve. So like, and also a lot of the writers would write themselves in. The story features a captain—like a captain of the ship that would transport the 500 time travelers back to the Ming dynasty. The captain himself, his real name and real-life nickname, became known as Captain as well. At a certain point, the boundaries between past and present, fictional and reality, kind of blurred.
The same happens with Ma Qian Zu himself. He is one of the main people in the novel. So, it’s Qian Zu and Qian Zu—they almost spell the same in Pinyin, but one is Ma Qian Zu and the other implies humbleness. Ma Qianzu means you stand next to the horse to serve, but fictional Ma Qianzu is arrogant. You are Qianzu: you can see a thousand miles away. Zu means seeing.
I think things like this are very interesting. The basic premise starts from a simple thought experiment: what if you can travel back to the Ming dynasty with modern knowledge and equipment? People started writing about it without character building or discussion. The first 30 chapters are all about people getting together to think about what equipment they should bring to the Ming dynasty.
You will see this laborious preparation list, almost like the list a very organized person writes when packing for a long trip. People spend 30 chapters to prepare for this list.
Then, around chapter 37, people finally get together to board the ship that will take them to Ling Gao. You also see this immense obsession with details:
They conduct serious, detailed risk assessments. It’s really first principles thinking—almost like an action manual. If you really had a wormhole to travel to the Ming dynasty, you could simply follow it.
This is because a lot of the knowledge is factual. Professional people research and fact-check it themselves and each other in a peer review process ensuring scientific accuracy. People are thinking about how to bootstrap an industrial revolution on this island—what do we need?
But, let’s get to my question: how far do they take it? Are we talking decades of institutional development, or does it mostly stay in an early building and consolidation phase? Do they change history profoundly? Do we even know what history looks like now as a result of the changes they make?
The story kind of progresses as the current time progresses, I would say. Everything stays in the Ming Dynasty—there is no fast forward to the Qing Dynasty or the Republic period. The time flow of the Ming Dynasty basically matches today’s pace.
Because the story has been written for about 20 years, a lot has changed:
There are also plots, like some fanfic, which are not part of the main story:
There is also a story plot that diverges from the main story that… people ended up colonizing Australia, and they formed a huge sort of empire, almost like a British empire. In the 19th century, they forged a huge Linggao Australian empire across Australia, New Zealand, and Southeast Asia. The north part would be like Linggao county, like Hainan and Taiwan. Yeah, so like crazy stuff, really crazy stuff.
But what really strikes me about this, as you’ve described it, is this is an alternate history that doesn’t imagine salvation through new ideas, or a moral awakening, or the scientific revolution necessarily, but actually just through kind of competence, and very specifically through technological, technical competence.
There’s this like obsessive attention to getting the tech tree right, like:
It’s just like precise accumulation step by step, as you’ve described. But alongside that, there’s also, I guess, as you talked about in your piece, this kind of unglamorous work of building institutions that can sustain these capabilities over time.
They’ve thought through a lot it seems, seen that way. Linggao feels less like escape fiction and more like a thought experiment about governance and about why technocratic instincts have such appeal to China.
Let me ask about this because there’s this framework explicitly in academic terms. We usually talk about the Needham question and about things like Ken Pomerantz in his book The Great Divergence. These are ways of explaining why industrialization took off in Europe rather than in China when China seemed quite ready for it in some measures, by the Song dynasty.
We had the capability to do mass mechanized production in some ways. But again, I haven’t read it, but reading your essay and the broader discourse around the industrial party, it feels like this community has its own implicit theory of history.
How would you characterize the industrial party’s answer to the Needham question? What do they seem to think actually mattered in producing the divergence that we saw, that Pomerantz describes?
I think, I got educated in China. I think the sort of national scar, the hundred years of humiliation that China left behind, didn’t modernize until the European powers kicked your ass. Then China started industrialization.
This part of history has always been a sort of a collective scar, a wound, a true wound basically, among everyone that I know who received primary education in China.
Alternatively envisioning a China that started to modernize, started to industrialize at the pace of the European counterparts has always been, I think, a psychological comforting thought experiment.
I also noticed that sometimes the national consciousness in Linggao’s plot is really weak. Of course, it is a big part of almost like a salvation porn or like salvation.
“That’s a good way to talk about it: salvation porn.”
Part of it is salvation porn, but I realized a big part of it is the joy of meticulously planning everything itself. To the engineers, building itself is very joyful; it is beautiful, it is satisfying.
Because I observed this among not just Chinese engineers but among a lot of the western engineers that I’m friends with. They love YouTube channels like Primitive Technology. It is literally an Australian man using mud to build all sorts of tools from zero.
It seems like engineers really enjoy this sheer ability to transform the surroundings with the scientific knowledge they possess in their mind. It seems like totally my dad.
Like homo sapiens seems to us like we’re homo sapiens. We seem to really enjoy thinking about our ability to transform our surroundings.
I mentioned Robinson Crusoe. I think Robinson Crusoe is like the 18th-century Primitive Technology YouTube channel.
“For sure, for sure.”
When I was a kid, I obsessed with this book because I constantly imagined myself being this all-powerful human being, like going into a savage island and humanizing and civilized a place by my sheer intelligence, by the modern advanced knowledge I possess. And I think, thinking about this, it’s not just Li Gong Dan; it’s also Western engineers. I know, it’s also you, it’s also me. Very interesting, right? It’s like reading this makes you happy.
Seeing the primitive technology YouTube videos makes me calm. Like, I think as a hunter-gatherer, like offspring of a hunter-gatherer society, human being, I found this psychologically safe.
So I think a big part of Ling Gao’s dopamine hit comes from writing about technology and planning itself, writing about building the civilization itself, other than national, yeah, right, for sure.
I get that. I get that for sure. There’s something about—I mean, it’s a flex, you know? They get to show, look how I understand the very fundaments of the technologies that I deal with. But there’s also something like this kind of inherited historical vulnerability at work here.
You know what you talked about, this century of humiliation thing. I mean, not a grievance in a narrow sense, but just kind of a memory of how badly things can go, you know, when state capacity falters.
So I wonder, in addition to this satisfying kind of, you know, tech just tech qua tech, there is—I wonder if there’s this kind of implicit never again embedded in the discourse, you know? Not just about foreign domination but also about chaos, about fragmentation, about, you know, loss of national agency, right? I mean, that’s in there too. I wonder if that appeals to you as well.
I agree. I agree. I think this, we should memorize that engineering and industrialization and urbanization are the true things that truly gave the Chinese nation power. Like we shall engrave this in our bowl.
I think this is part of the message the Ling Gao Qiming Morning Star of Ling Gao has been sort of projecting. And it reminds me of—so there’s a scholar whose name is Wang Xiaodong, you’re probably familiar with, yeah, of course, who wrote, I think in 2009, China is Unhappy. I remember it was a big intellectual sensation.
Like he is the one who coined the term industrial party. So in this article that he coined the term industrial party, he stated very clearly that—I actually want to read this—he stated really clearly that:
“We must never envy the finance Hollywood, the Grammys, and NBA of the West. We would rather forge iron and smelt copper and let the Americans sing and dance for us because forging iron and smelting copper is the true—this is where true power lies.”
And I think this is a big—this basically crystallizes industrial party’s salvation arc, which is it is the industrial capability that made China powerful so other people couldn’t kick our ass again.
The true power, the true international strength that European countries wouldn’t bully us, like America and Japan wouldn’t bully us, is because now we can forge iron and melt copper. It is not because we can sing or dance, it’s not because we care about social welfare, it’s because we can build stuff.
I think industrial party has such a clarity about the importance of engineering and industrial knowledge.
I want to quickly shout out Fred Gall, who actually wrote another essay right after yours came out, and it happened that the very day that I read yours right away suddenly in my inbox there was Fred’s Substack. And he had actually written about it as well, and you know he definitely helped me to get oriented with this.
But what you’ve just described, it’s engineering then becomes an act of patriotism, right? It becomes synonymous with patriotism. Building is loving your country, and that connection seems to be quite explicit in the whole industrial party discourse.
I mean, building itself becomes a moral act. It takes on moral weight, which is a really interesting worldview.
Fred also frames this though in his writing as a generational revolt, especially against earlier, maybe more literary or humanistic modes of thinking about China, the China that you maybe described when we worried about the cost, we worried about the human cost.
I mean, it doesn’t describe this hostility exactly, but a sense of that those ways of talking had become just kind of unmoored from material reality.
So there is this tension between the When Yi Ching Nian phase and the Li Gong Nan dominance phase.
And, but I want to get to this gendered layer here that feels really important for me to acknowledge—that this industrial party worldview, this whole emphasis on engineering on… Discipline on technical mastery that to me feels very gendered in terms of who speaks with authority, what kinds of traits are valorized.
You’re somebody who identifies as a feminist and you work very fluently across technical and cultural domains. How do you read that gendered dimension to that, who gets to imagine the future in these narratives?
I think first of all Ling Gao Qiming itself is a piece of historical record because I think the collective writing process peaked maybe during 2011 to 2015, and this is the internet before China’s feministic awakening. So I would say certain feministic consciousness hasn’t arrived in China yet.
So Ling Gao Qiming is in a sense a product of its time — a pre-feminist cultural product — and people just really don’t have a lot of tools or instruments or frameworks to criticize it.
Just like a lot of women writers would participate in writing, they would probably feel extremely uncomfortable but they couldn’t name why they feel uncomfortable. But now, retrospectively looking at this text, looking at these primary sources, it is very much misogynistic.
It’s just like how much Liu Zixing’s Three-Body Problem feels extremely misogynistic when you’re reading in Chinese.
I mean Ken Liu did a great job in removing a lot of the poorly written female parts, it’s still in there, yeah, yeah. But you know like there’s definitely some plots in Liu Zixing’s work that would be like:
“Oh, you’re a woman but how can you listen to Bach, this German composer, like because Bach is such a representation of rationality, a rational music. How can women appreciate this beautiful, high class, high broad rational music?”
You know, such plots permit Lin Gao and the first 500 pioneers — like a very small group of them are women, predominantly men. And I think the made revolution is the part which is really fascinating because Lin Gao basically operates in the semi-military structure where the resource needs to be centrally planned and allocated to people.
It is a techno-authoritarian society where it’s also a little bit like plutocracy. I would say people who possess the most engineering knowledge have a better social status.
So at the time, there is this distribution of:
These people are very unhappy. I mean, they’re all fictional plots by the way, and those plots are the incels of Linggao — the single people from Linggao.
In the sense of domestic servants are also, you know, sex slaves, which is not being explicitly said but later you will see this Linggao society operating as a semi-feudal but techno-authoritarian style political structure.
Later, they recognize that:
This is all part of the modernization process.
China’s modernization success depends on female workers in the factory, so Linggao is like:
“Okay, if we’re rational enough to truly industrialize Hainan, to truly industrialize Ming dynasty, we shall truly give the female servants proper treatment, so we can properly…”
So it’s basically all rational, not like:
It’s not moral—it’s rational.
So it’s rational for the Linggao community to progress to a female-male equality scenario, and then this is basically a historical fatalistic direction instead of out of, you know, humanitarian concern or out of cuteness or moral goodness.
Wow, there’s just so much to plumb here, and it’s sort of the theory of history that underpins this that I’m particularly interested in. Maybe I will at some point take a crack at this thing. I’ll be good for my Chinese anyway.
So let me shift a little way away from Linggao here.
I do want to bring it back in frame but this book Breakneck, by Dan Wong, which is one of the most talked about books of 2025. Dan, of course, as you know, describes China as an engineering state.
I mean, listening to you talk about Linggao and the industrial party, that phrase starts to feel less like an abstraction and more like an actual lived… Worldview, right? Does that framing resonate with how you understand what Linggao is imagining, or does it miss something important?
You have this book club where you have been talking about, reflecting on Chinese language discussions of Breakneck. You know, it’s called What? Reading Breakneck in China.
Yeah, reading Breakneck from China.
Right, right.
One thing that struck me in your book club reflections—I’ll link to that because you’ve written about it on your Substack—is that Chinese language discussions about that book seemed less surprised by that framing than English language ones.
So, I mean, did the idea of an engineering state feel like any kind of a revelation to Chinese readers, or more like seeing something familiar finally given a name?
I really appreciate Dan’s framing. I think Dan’s framing is at least to better capture certain reality in China. I honestly think the democratic versus autocratic binary is not helpful anymore. Like, if you look at the US, what’s democratic about the US, right?
I know a few Chinese, China-focused scholars who used to study the authoritarian regime of China and now all sort of pivot to study the US authoritarian term.
You know, I honestly think Dan’s framework can somehow better explain the reality and better get to the point. It’s really helpful, it’s really instrumentally helpful.
And then, according to Dan, he tends to be playful with this framework. He’s like not 100% serious about it, doesn’t want to challenge the status quo of democracy versus autocracy. But yeah, I’m going to borrow that cop-out from him.
I’m just being playful here, I’m not really—it’s a way to not commit completely, right? I mean, that playful is—it doesn’t have like, you know, we have generations of scholars studying authoritarian systems, right? But like in a sense, I don’t think Dan wants to challenge that.
I think he comes up with this framework just to better explain today’s China and today’s US.
Yeah, I think I do appreciate this framework, and I think the engineering state captures a lot of the developmental, the knee-jerking intuition for the Chinese society as well as the party’s industrial policy.
I think the industrial party ideology is reflected by the CCP itself as well, and I would argue this industrial development is the priority spirit, is a collective unconsciousness among so many powerful people, so many decision makers in China.
For example, Xi Jinping mentioned the new productivity force. I think new quality forces of production is very industrial party coded—it’s because this implies that China’s economy is stagnating; the growth is that as we don’t have the prosperity like the growth like before, how do we solve this problem?
Okay, let’s shift to this magical new productivity, new quality productivity force. Let’s do more engineering, let’s upgrade our engineering so problems could be solved.
I think there is this industrial party-coded naivety or innocence in it, and then I think a big part of the CCP’s decision makers still think they can engineer a lot of problems away. But in reality, it’s not true anymore because the industrial party itself has a lot of intellectuals start to have their own reckoning on a lot of China’s problems, and then they realize that a lot of problems couldn’t be engineered away.
So, Dan Wong’s book, do you feel like it hits differently between English and Chinese audiences when it comes down to their different lived experiences? How would you, if you had to sum up the difference between how your Chinese friends—many of them have maybe not spent time in the West—how that hits differently?
A lot of people are overly obsessed with if China is a real engineering state. For example, they would argue:
So a lot of the Chinese language readers who are living in China would be dissatisfied with Dan’s engineering state verdict, because they would argue like, you know:
I tend to think it’s very useful to accept it in a sort of provisional and playful sense. But there’s this irony I keep coming back to: it feels like it’s only just in the last year or so that many Americans have really fully become aware of the scale of China’s industrial might or industrial power in China.
Yet, in our conversation, it sounds like the industrial party worldview—the whole framework that helped articulate and legitimize this push from within China, this crazy breakneck, engineering-driven mentality—is already losing some of its explanatory force in China. It’s weird that Americans are only starting to believe this is the case at the moment when the industrial party logic has lost or is losing its grip.
I don’t think the industrial party logic has lost its grip in China. I’m pretty sure a lot of the industrial policy decision makers still very much adhere to the industrial party logic:
But the intellectuals who were part of the industrial party movement in the early 2010s, I think they’re starting to suffer from China’s declining economy and, say, COVID. For example, Ma Qianzhu himself, an influencer in China with two million followers on Bilibili, is a very articulate writer. But his account was banned because he voiced certain issues during COVID. Ma Qianzhu himself got cancelled by the state even though he used to support everything for the state.
This brings us to the irony where the industrial party people, the engineers themselves, are very smart and aware of certain societal issues like:
These issues don’t necessarily yield to the logic of industrialism.
I’m curious about Fred Gao—I don’t know if you know him personally, but I’ve met him in Beijing. He’s a really nice guy and has been explicit about moving away from the industrial party orbit over time.
I wonder if this is a personal evolution or symptomatic of a broader shift in discourse. I think for many industrial party intellectuals, it feels like a personal evolution. They have kind of grown out of the industrial party phase. I would say they lost their innocence in believing engineering could solve everything. It’s not a magic potion.
Mai Tienzo himself definitely took some hits in life to realize that his youth was starry-eyed and innocent about many things. It’s called growing up. A lot of people I know had that kind of super faith in technology early on, and anything that didn’t surrender to the hard logic of mathematics and engineering was just worthless. They’d ask, “Why bother reading novels? You should be reading that kind of thing.”
People grow up, right?
It’s really funny because within the crypto community, I also met a lot of rational engineers—people who hang out in the rationalist forum community. I see them growing up as well, starting to learn that:
I see them also sort of grow out of this obsessive, almost purity phase.
It’s funny like my tensile right now, he speaks out. A lot about the child supply, and he speaks out about local government debts and certain central-local relations. He also has an absolutely descending voice during COVID. Well, I mean, it’s comforting to know that it’s still possible for people to change.
Yeah, let’s go for one final question just to wrap this all up about what Lingao tells us about China today. If someone wants to understand contemporary China—not the politics necessarily, or the policies, or the political imagination—what should they take away from the Lingao phenomenon? What does it tell us about how China thinks about:
What’s your big bottom line takeaway?
A big thing that tells us is maybe stories like Lingao are worth more attention. In a sense, it’s a more grassroots Senti—a Three Body Problem that’s more widely accessible. In a sense, it’s an egalitarian Liu Cixin collective building process. Like, you know, Three Body Problem’s Liu Cixin is representative, but I think Senti maybe speaks more to the unpolished, the authentic, the grassroots, the organic aspect of these things.
For me, reading Lingao is such a journey. It introduced me to knowledge I never really thought about. Part of the Industrial Party I constantly laughed about during the peak of their debate in the early 2010s: they constantly laugh at this humanistic journalist who would complain about the suffocating urban life and want to escape to the forest. As long as this journalist can take a hot shower and have access to the internet, the Industrial Party would laugh at this fantasy.
This escapist imagination ignored the infrastructure it needs to have a hot shower and wifi connection. The Industrial Party deeply advocates for the invisible wires buried in the ground. They advocate for the pipes that transmute the hot water to this escapist little Eden garden. This humanistic journalist would imagine oneself to be like this, but Industrial Party people are really making a lot of the invisible stuff visible to me.
In the process of US re-industrialization, such knowledge is revealing because I used to take hot water and electricity for granted. Then I learned that’s not true. China’s electricity supply is top of the world right now—the high voltage grids, convenient industrial basis—everything to fuel China’s innovation.
Yeah, I think Industrial Party really gives me certain knowledge that humbles me because I could be that ignorant humanistic journalist complaining about urban life. I want to take a hot shower in the forest and don’t reply emails, but I still want wifi. I could completely ignore the infrastructures—that’s like the iceberg under the ocean.
Yeah, I think, in a sense, Lingao is a textbook for me to learn about the industrial process at its very first principle. It’s not fun to read but also fun to read. That’s really an interesting take. I gotta wonder what these guys today would think of Li Ziqi.
I mean, you know, for those of you who know, Li Ziqi is a very, very popular video blogger, huge on YouTube and stuff like that. This woman is very attractive, who left her life in the city to go home and take care of her aging parents or grandmother in the countryside in Sichuan, and has made this enormous following because she’s so good. On the one hand, she sounds so far like that kind of journalist who wanted to flee as long as there were hot showers and internet.
But this woman also has mad skills. I mean, she crafts, she does, she’s a good asset on, you know, Hainan Island in 1628 for these guys because she knows how to build stuff, how to make stuff, and all these traditional crafts. I wonder what they would make of somebody like her.
She embodies, on the one hand, both what they don’t like and what they very desperately need.
Oh yeah, I think if I were a Lingao writer, if I were part of the Engineering Party, I would salute Li Ziqi because if I were them, I would meticulously break down the amount of planning for her to do in order to create. A seamlessly beautiful video like that—if I were an industrial party member, I would appreciate the engineering part of her production. I would be like,
“Oh my god, it’s because you did so much invisible infrastructural production work.”
So the 20 minutes—the visible time of you showing up on the screen—can look so effortless and seamless. I think, yeah, I generally think the Ling Gao people would appreciate her engineering skills in a sense—like production engineering and resource management skills for sure. Fantastic!
What a fun conversation this has been, and the time has just flown by. Afra, let’s move on now.
First of all, thank you for spending so much time speaking with me, and again, everyone’s got to go and read your piece if you haven’t done it already. It’s just a wonderful piece of writing. For me, I think it’s one of those things where this little slice, as you say, just this artifact of Chinese culture, made me think so much about the contemporary Chinese condition. It made me think so much about, you know, the mindset that really does—in so many ways—just sort of inform and shape the world that we inhabit today.
It’s become—it’s not just ideology, it’s more like infrastructure, right? The whole mentality, in many ways, has come to define the modern polity.
But let’s move on and talk about this segment that I call paying it forward. If you’ve got a young colleague or a friend or somebody whose work you want to call attention to, now is the time to do.
I think one thing I need to shout out is—I mentioned in a piece that there’s no English translation for Lingo, which is not true. So, two months ago, obviously a group of people took it as a passion project and translated the canonical version into English and made it a website.
They also basically have a GitHub commit about the tools they use to translate the piece. They use the GMLI 2.5 to translate everything.
Yeah, I’m just really glad that people are spending effort systematically translating Lingo into English, so I would recommend reading that. I think that’s the first recommendation.
Second is, unfortunately, if you’re not a Chinese language speaker or don’t listen to Chinese, you won’t get the great content. Baihua is this podcast incubator actually started by my friend Izzy. We’re all like sort of the founding members of Baihua, and we’re trying to incubate more Chinese language podcasts.
One of the podcasts I really like and really appreciate is called Xin Xin Renlei. I can also send the link.
“Please do.”
Xin Xin Renlei is a podcast hosted by three tech journalists who are also, like me, really bilingual and understand the tech world on both sides. They find some very interesting niche topics to discuss. For example, they would talk about:
Yeah, so highly recommend Xin Xin Renlei. The English name is Pixels Perfect.
Pixels Perfect, Xin Xin, Xin Xin, Xin Xin.
Okay, well excellent, excellent—that’s fantastic. Now, I don’t know whether that was your paying it forward recommendation or your actual recommendation recommendation. I distinguish between them, but did you have a book or something that you wanted to recommend?
Yes, I actually read voraciously. I do have a lot of books I would recommend. One would be, I think, it’s edited and written by Carrie Brown. It’s called
China from European's Eyes: 100 Years of History
I think that book, to me, is—
You know, like we always talk about how China is the foil and mural for the West’s imagination, and people’s obsession about China—the way people project China as a beacon for technological advancement today—is actually a sense of otherness, right? Like other in China.
So this book illustrated that this phenomenon is not new. It has been existing for 800 years. You know, many European intellectuals have been portraying China as the otherness projection—like it’s elderly, alien, different—but it… It could be either really beautiful or really ugly. It could be elderly powerful or elderly powerless. The reason why China couldn’t develop modern technology and modern systems, Hegel would argue, was because the Chinese language, the characters, are so laid back.
Basically, Cary Brown, as a historian, compiled 16 or 18 permanent European intellectuals on their takes of China. So the people from like Voltaire to Hegel. Yeah, so I think it’s a fascinating intellectual genealogy. I would recommend it.
Yeah, that sounds great. I mean, I have all the time in the world for Cary Brown. I think he’s wonderful, brilliant, and a fantastic writer. I don’t understand how he writes so much—like he’s gotten a new book every six months.
Oh, I have another one I really must say is Ilin Liu’s upcoming new book. It’s called The War Dancers. It’s coming out, I think, at the end of February, and this is a book about the history of the Chinese internet in the past 30 years. I think you’re going to be interviewing her.
Yeah, I read it. It’s absolutely such a craft—it’s a beautiful craft, so well written. She’s a great writer. Oh, she’s such a great writer. Honestly, as her friend, I really admire her craft. Such a role model.
Yeah, we know each other socially as well, and I am going to have her on the show to talk about the book. So yeah, I mean, it’s great because the book is really well written. I read that book—it’s called The War Dancers. I couldn’t remember the full title, but I have it. So I’ll make sure to put the title in there, and it’s an excellent recommendation.
Related to your recommendation of Cary Brown, just to remind people, I recommended this book ages ago. But it’s a very similar approach, although it’s not just China; it’s all of Asia. It’s Jürgen Osterhammel’s book Unfabling the East: The Enlightenment’s Encounter with Asia, which is something that I haven’t recommended before, and yeah, it’s absolutely great.
My recommendation for this week actually has something in common with that. It’s Tami Mansari, who I’ve recommended another of his books before. He’s an Afghan American writer and journalist, and he wrote a book called Destiny Disrupted: A History of the World Through Islamic Eyes.
It’s a real deep dive into the history of Islam as understood by Muslims themselves, from the time of the prophet in the 7th century all the way up to September 11th, viewed through the eyes of Muslims themselves. I think it’s a very useful exercise in building cognitive empathy and understanding the Islamic worldview—not that there’s one single monolithic worldview, but it’s a great book.
It also reminds me of another book written by Kim Stanley Robinson, who also likes to write about hard science like Liu Cixin and the Industrial Party. He has a book called The Years of Rice and Salt. I was just talking about that book the other day with a friend of mine. It’s a great book.
I have recommended that one on the show years and years ago. It’s an alternative history, which I really like. Even since we’re talking about alternative histories here, not a time travel one, but the premise is that the Black Plague actually ends up killing 99% of people in Europe. It starts with Tamerlane’s troops coming up to the Bosporus and then deciding, “Nope, we’re not going over there,” because they were planning on conquering Europe. But no need—the plague has already killed everyone.
Fascinating, yeah, fascinating book. It also has a lot of Buddhist touches, like reincarnation. The interstitial chapters are like the Bardo chapters.
Yeah, I really hope China has someone like Kim Stanley Robinson. I think he could be both spiritual and insanely technical, like Red Mars and Gray Mars, which are very detail-oriented in terms of Mars terraforming.
But a lot of his work is also deeply humanistic. Of course, there’s this cli-fi classic Ministry for the Future as well. So yeah. I would love to meet him one day. He seems like such a wonderfully interesting man. I know, I know, I love his recent preservation of Sierra, it is almost like he’s the embodiment of California spirit—both technologically aware but also deeply drawn to the mountains.
I don’t know, I think something fascinating about this guy, I really like him. Yeah, yeah, yeah.
All right, hey, well thank you so much, what an enjoyable conversation. I think we could go on recommending books to one another for several more hours, but we will call a stop to it.
I look forward to meeting you in person one day. I’m going to be in England at the end of the month of February, but I don’t know if you’ll be around. I think so.
If it’s London, yeah, I’ll be around. Yeah, it’s such a fun recording of a podcast with you.
Okay sir, thank you for inviting me. Yeah, yeah, what a great time.
You’ve been listening to the Seneca Podcast. The show is produced, recorded, engineered, edited, and mastered by me, Kaiser Guo. Support the show through Substack at www.synicapodcast.com, where you will find a growing offering of terrific original China-related writing and audio.
Email me at synicapod@gmail.com if you’ve got ideas on how you can help out with the show or if you just want to say hi. Don’t forget to leave a review on Apple Podcasts.
Enormous gratitude to the University of Wisconsin-Madison’s Center for East Asian Studies for supporting the show. Huge thanks to my guest Afro Wong. Thanks for listening and we’ll see you next week. Take care.
I earned my degree online at Arizona State University. I chose to get my degree at ASU because I knew that I’d get a quality education. They were recognized for excellence and I would be prepared for the workforce upon graduating.
To be associated with ASU both as a student and alum makes me extremely proud. Having experienced the program, I know now that I’m set up for success.
Learn more at:
ASU Online
asu.edu
Hello, and welcome back to the Cognitive Revolution. The presenting sponsor of today’s episode is Granola. Regular listeners have heard me describe the blind spot finder recipe that I’m using on Granola to look back at my recent calls and help me identify angles and issues I might be neglecting.
I love that concept, but it’s also worth highlighting how Granola can help raise your team’s level of execution by supporting follow-through on a day-to-day basis. This morning, for example, I had two very practical calls in which I committed to a number of things. In the past, to be honest, there’s a good chance I’d have forgotten at least a couple of the things I said I’d do. But with Granola, I can easily run a to-do finder recipe and get a comprehensive list of everything I owe my teammates.
This is the sort of bread and butter use case that has driven Granola’s growth and inspired investment from execution-obsessed CEOs, including past guests Guillermo Rauch of Vercel and Amjad Massad of Replit.
See the link in our show notes to try my blind spot finder recipe and explore all of the ways that Granola can make your raw meeting notes awesome.
Now, today, my guests are Vlad Tenev and Tudor Achim, co-founders of Harmonic, an AI research lab dedicated to building mathematical superintelligence, and also the creators of Aristotle, an AI system that achieved gold medal-level performance at the 2025 International Mathematical Olympiad.
While OpenAI and Google DeepMind achieved similar performance by scaling reasoning in chain of thought, Harmonic stands out for their commitment to formally verifiable methods. This is because it generates candidate proofs in Lean, a programming language that serves as a proof-checking assistant by using a trusted kernel to confirm that every single step of reasoning follows from a few explicit premises and accepted logical rules.
Aristotle’s work can be automatically validated, and its performance is in principle limited only by the scale of compute available for reinforcement learning.
In an effort to better ground my own intuitions for mathematical superintelligence, we begin with a metaphysical discussion about:
From there, we turn to the Aristotle architecture that delivered IMO Gold performance. It consists of:
- A large transformer model that uses a Monte Carlo tree search strategy, reminiscent of systems like AlphaGo, to discover valid paths from point A to point B in mathematical reasoning space.
- A lemma guessing module that helps manage context and keep things on track by generating candidate waypoints between a given starting point and a potentially distant end goal.
- A specialized geometry module modeled on DeepMind's alpha geometry.
We also discuss the Aristotle API’s informal mode, which attempts to auto-formalize whatever the user asks it to prove.
We discuss what its responses to my admittedly silly requests imply about the boundary between statements that could in principle be mathematically proved, and those which are sufficiently factual or philosophical in nature so as to fall outside the scope of the system.
Examples include propositions like:
“all is love”
and
“Epstein didn’t kill himself”
In the final section, we discuss:
On this last point, I have to say, with so many grandiose AI promises flying around these days — from a country of geniuses in a data center, to a century of progress in five years, to curing all diseases in our natural lifetimes — it is rare that I am genuinely taken aback by a company’s vision for the future.
And yet, as you’ll hear, Tudor did manage to leave me at least momentarily speechless when he described a future of theoretical abundance in which all physical phenomena we observe have multiple competing coherent explanations, which can only be separated by increasingly exotic experiments.
If you’re like me, you’ll find this episode a useful opportunity to:
Vlad Tenev and Tudor Achim, co-founders of Harmonic, makers of Aristotle, and winners with an asterisk of the IMO gold in 2025.
Welcome to the Cognitive Revolution.
Thanks for having us.
Greetings and salutations.
Thank you.
So this is going to be, I think, a fascinating conversation. It’s probably going to be more metaphysical than most of our episodes, but also there’s a lot of practicality because what you guys are doing certainly has aspirations to go beyond the pursuit of mathematical superintelligence.
Maybe just for starters, how do you guys understand what math is? That was something I was really wrestling with in preparing for this. And then, you know, that’s obviously very metaphysical. To make that a little bit more practical, what would you say are the core cognitive skills that people that are good at math really develop and excel at? And how do those skills do when we look at the performance of like the frontier large language models that all of our listeners are familiar with today?
“Yeah. Well, look, first, thanks for having us. It’s really great to be here.”
You know, when you ask, what is math? What is it useful for? What are the core cognitive skills? it gets like one of the core theses of our company, which is that mathematics is reasoning.
So a lot of people think of mathematics as this really esoteric thing. You know, you’re thinking maybe group theory, stuff you’ve seen in movies like Good Will Hunting, but mathematics at its core is the process by which humans understand the world by breaking their understanding down into small sequences of logical steps that other people can understand and verify for themselves.
So when you’re solving a physics problem or doing your taxes or thinking about what happened at the beginning of the universe, ultimately you have to have an explanation that is
And so when we talk about what it takes to be good at math, the question is what it does take to be good at reasoning. And so that’s, again, that ability to break this down into steps.
It turns out math is really useful for understanding the universe and building lots of engineering things, but ultimately it’s just about reasoning.
I watched your podcast that you did with Sequoia maybe 16 months ago or so now. And I recall Vlad’s story of like, basically,
“I thought that if I got good at math and I’d probably be good at other things and it sort of worked for me.”
So that’s like one way to, in a very practical sense, unpack the idea that math is reasoning. It certainly seems to help people generalize to at least related domains and be really effective, for example, in entrepreneurship.
But I’m not entirely clear still on like, are you making a more almost platonic claim there? It seems like there’s the very simple notion that like, okay, I should teach my kid a lot of math because then they’ll be smart generally. And again, that works for humans.
But is there something that you see as like a more fundamental law of the universe, sort of correspondence between what we are doing in math and what we are doing in these other domains? Because it doesn’t seem like we have the same sort of like verifiability in almost anything else.
We do have it a little bit in computer science, but even in physics, right? We’ve got like still very fundamental questions about
“I don’t think that stuff is at all agreed upon.”
So maybe you guys throw up your hands at this mystery too, or maybe you feel like you have kind of an intuition for what the answer is.
“Yeah, I can give you my perspective.”
I got into math through physics. So when I first came to Stanford as an undergrad, I had read Brian Greene’s The Elegant Universe, which was sort of like the first popular string theory book.
And when I was a kid, one of the earliest memories, one of the first full English books that I read was A Brief History of Time by Stephen Hawking. So I’ve always been interested in kind of the big questions, right?
Cause you know, back in the day, that was not obvious. You know, we thought electricity was separate from magnetism and it was just like a really big… I probably think one of the greatest achievements of science is figuring out that these two are actually two sides of the same coin really. And then, and then the big question is like, well, what’s going on with gravity? Is it, is it the same? Right.
And, in the middle of this, we found out that the weak force and the electromagnetic force were also splintered off of one electro-weak force. So it kind of feels like there was just one thing at the beginning and we have to understand what that thing is.
And what I found when I became a physics major at Stanford, and I started asking all of these questions, eventually they’d send me over to the math department. And they’re like,
“Well, in order to understand string theory, you have to understand all of these other things. Right. And if you want to understand general relativity, you’ve got to get into differential geometry.”
And so that’s how I became a pure math major and ended up doing a PhD. The impetus was actually trying to understand the real world through physics.
If you think about what’s the usefulness of physics, I mean, all of the big inventions that humanity has that really push us forward are kind of like physics inventions, really. I mean, when you think about:
They’re physics things.
So the real reason to do math is math is interesting and beautiful. There’s an art aspect of it, but it helps you. It helps you understand physics. Physics helps you understand engineering. And then you can create things that have huge value.
You were asking, how does math work in other fields where things are not as precise? I think math shows up just maybe a little more subtly than people think.
So there is this physicist, Eugene Wigner, who wrote a famous essay called the unreasonable effectiveness of mathematics, which was commenting on a really interesting phenomenon.
So Vlad mentioned differential geometry and special relativity. It turns out that when Einstein was creating that theory, he relied on these thought experiments from the 19th century around how to think about certain manifolds and their properties.
And that was actually the key tool that we use to explain what special relativity is, and then develop it for general relativity.
That’s a perfectly representative case because those thought experiments in the 19th century were almost preposterous. It made no sense to think about them because
“How could you possibly apply these concepts to the real three dimensional world?”
And then it turns out that it’s very useful for understanding the four dimensional world when you include time and curvature.
There are myriad examples like this.
If you consider number theory for a long time, that was really seen as an incredibly esoteric branch of math with no practical implications. But people pushed on that theory for a long time. And then it turns out that that’s the key tool you need to create a secure digital economy.
So now essentially all of human civilization has a digital economy, which is based on this branch of math.
So I think it’s almost the wrong question to ask,
“Well, I don’t know, there’s a lot of math out there. How is it useful?”
The point is you just do the math and then eventually some of it, not all of it, will be more useful than you possibly could have imagined.
So the investment in math is: it’s not just to build a really smart system. It’s to create a lot of new math that we can then figure out ways to apply later.
One interesting thing that the conversation reminded me of when you first asked,
“What is math? What does it look like?”
I think one of the reasons we got excited about applying AI to this domain is there are lots of different things that mathematicians do.
And I think we’re very excited about the prospect of AI accelerating the former. We think that’ll happen.
But the latter is something that AI is already really, really good at today and was good to some degree when we got the idea for Harmonic. You know, you look at GPT-4, which had just come out when we started and it excelled at just pulling information, doing these types of needle-in-a-haystack tasks of, can you just really quickly go through all the literature and pull things that might be relevant.
And I would say even if you can be an amazing mathematician, you’re in that category.
I think a lot of the work could be accelerated if you just knew all the math that was being done and could pick out the relevant things to an unsolved problem that you have at hand.
So I think the problem itself lends itself really well to what AI is already good at.
Hey, we’ll continue our interview in a moment after a word from our sponsors.
One of the best pieces of advice I can give to anyone who wants to stay on top of AI capabilities is to develop your own personal private benchmarks, challenging but familiar tasks that allow you to quickly evaluate new models.
For me, drafting the intro essays for this podcast has long been such a test.
I give models a PDF containing 50 intro essays that I previously wrote, plus a transcript of the current episode and a simple prompt.
And wouldn’t you know it, Claude has held the number one spot on my personal leaderboard for 99% of the days over the last couple of years, saving me countless hours.
But, as you’ve probably heard, Claude is the AI for minds that “don’t stop at good enough.”
It’s the collaborator that actually understands your entire workflow and thinks with you.
Whether you’re debugging code at midnight or strategizing your next business move, Claude extends your thinking to tackle the problems that matter.
And with Claude Code, I’m now taking writing support to a whole new level.
Claude has coded up its own tools to export, store, and index the last five years of my digital history — from the podcast, and from sources including Gmail, Slack, and iMessage.
And the result is that I can now ask Claude to draft just about anything for me.
For the recent live show, I gave it 20 names of possible guests, and asked it to:
Based on those, I asked it to draft a dozen personalized email invitations.
And to promote the show, I asked it to draft a thread, in my style, featuring prominent tweets from the six guests that booked a slot.
I do rewrite Claude’s drafts, not because they’re bad, but because it’s important to me to be able to fully stand behind everything I publish.
But still, this process, which took just a couple of prompts once I had the initial setup complete, easily saved me a full day’s worth of tedious information-gathering work and allowed me to focus on understanding our guests’ recent contributions and preparing for a meaningful conversation.
Truly amazing stuff.
Are you ready to tackle bigger problems?
Get started with Claude today at Claude.ai/TCR.
That’s Claude.ai/TCR.
And check out Claude Pro, which includes access to all of the features mentioned in today’s episode.
Once more, that’s Claude.ai/TCR.
AI agents may be revolutionizing software development, but most product teams are still nowhere near clearing their backlogs.
Until that changes, if it ever does, designers and marketers need a way to move at the pace of the market without waiting for engineers.
That’s where Framer comes in.
Framer is an enterprise-grade website builder that works like your team’s favorite design tool, giving business teams full ownership of your .com.
With Framer’s AI Wireframer and AI Workshop features, anyone can create page scaffolding and custom components without code in seconds.
And with:
it’s no wonder that speed, design, and data-obsessed companies like Perplexity, Miro, and Mixpanel run their websites on Framer.
Learn how you can get more from your .com from a Framer specialist or get started building for free today at Framer.com/Cognitive and get 30% off a Framer Pro annual plan.
That’s Framer.com/Cognitive for 30% off. Rules and restrictions may apply.
Okay, that’s quite helpful. I think, coming into this, I had focused my own mind on sort of two modes of math, I guess.
One being the kind of Einstein-like — obviously that’s a high-level example of a kind of eureka moment of having some insight that,
“hey, this highly abstract and, you know, seemingly perhaps like very esoteric formalism can actually unlock like major understanding.”
That’s kind of amazing. Very amazing.
And then there’s also this sort of grind-it-out, like I’ve got this thing that I want to prove, and I’m going to kind of, perhaps stumble my way even through the space of possible logical moves until I finally chart a path there. And then you’re adding another, a third layer, which is like problem selection in the first place, which I guess is pretty related to the Einstein thing, but certainly distinct in some ways.
Let’s take a minute before we get into the Aristotle system and how it works and how you’ve trained it and all that stuff to just talk about Lean.
Lean is basically a programming language that does this kind of very bit-by-bit logical maneuvering, right? Where you have certain assumptions coming in, you’re going to take these various steps, and the goal is to get to a certain outcome.
Tell us, because I’m just learning about this, in the context of preparing for this and a couple other podcasts, and I think most people don’t know anything about it.
So maybe give us a little bit of a more intuitive understanding of what Lean is. And I’d be keen to understand it on a little bit of a practical level to like:
So Lean, in my view, is the best programming language ever created.
In Lean, you can write any program you would write in Python or C or C++, but you can also express essentially any logical concept.
So if we’re okay getting into a bit of the details, it is a dependently typed programming language, which means that at compile time, you can express very complicated properties of the program that you can check before ever running it.
So on the one hand, you have on one end of the spectrum, you have something like JavaScript, where you can check basically nothing. And then on the other end, you have Lean.
But the really cool thing is you asked about axioms. So when Aristotle produces any output, it’s produced as annotated Lean code.
So there’s the programming language Lean, we write theorems, we write programs, we prove things. And there’s a lot of comments explaining to the person reading it what it’s doing.
But when we talk about proving things, you end up relying on three axioms, in addition to just the basic concept of the calculus of constructions, which is what the programming language is based on.
And just as an example to show what an axiom means: the axiom of choice —
“It’s not saying anything that would be controversial, it’s saying if you have a non-empty set, it’s possible to choose an element from it.”
And so from these three extremely basic axioms, it turns out you can build:
It’s all based on this core set of axioms.
And so the goal of a system that outputs Lean is to find interesting statements and programs then prove things that just depend on these axioms. And that’s really where the difficulty lies.
As you alluded to, sometimes you have to make big logical leaps, sometimes you have to grind through a lot of math, but both of those are essential. So you can’t really skip one of those steps.
But the Lean itself is just incredible. You can express so many ideas in it, you can prove so many things, and you can use it as a programming language too.
So it’s really up there for me in programming languages.
I started playing with Lean when Tudor and I started making a plan for this business, and we had a pretty early decision about whether we wanted to go formal and informal.
One thing that struck me about it is, as a former mathematician, I barely used the computer when I was doing math.
I was in my PhD in the late 2000s, and the only time you’d really be using a computer when doing math is when you wanted to type up your homework or your research paper or something.
But all the thinking about it would happen on a chalkboard or a whiteboard. All the collaboration about it would happen in person at these conferences or on a chalkboard in one’s office.
For a while, it was just like maybe mathematics would always be this pure thing that would just be kind of untouched by technology.
But what Lean has done is it transformed the mathematics from kind of like chalkboard and couch to now it’s in VS Code.
You know, you can do it in Cursive. You’re putting your math on GitHub, where now you can run these large collaboration projects.
So even when you subtract out AI, I think the Lean by itself without AI changed how people do mathematics, because now you’re seeing extremely prolific, famous mathematicians running these large projects where they’re collaborating with dozens of people around the world trying to do things like formalize research or formalize the proof of Fermat’s Last Theorem. And more and more and more of the folks are adopting Lean as like an accelerant.
So I think it’s changing how mathematics is being done and actually accelerates collaboration and accelerates progress and sort of like removes this notion of peer review.
If you’re a mathematician, if you’re a mathematician and you want to prove something, a big part is getting someone to read it and actually spend the time to tell you if it’s correct.
And so, you have the proof of Fermat’s last theorem, which took many, many years to be proved.
What happened was sort of this collection of people got together and when they all agreed that the proof was complete, it was sort of like ordained that the thing was proven.
And I think another thing formal does is it makes it so that that’s unnecessary.
Like if the proof checks and there’s no caveat that there’s no bug in the Lean kernel or how you’ve set up the statement, you obviate the need for manual human verification.
And the implications of that are pretty interesting too, right?
You have all of these potential citizen mathematicians who now with AI can solve unsolved problems and they don’t need to get anyone at a, you know, PhD program, a lean institution interested in their problem in order to tell that it’s correct.
They just have to have the Lean certificate and the proof is correct.
So, yeah, I think that’s a powerful thing.
If you think about journals, journals and math exist for this: it’s like the prestige of the review board tells you whether you should read something or trust it.
So I do that. The notion of trust is really changed fundamentally with tools like Lean.
Yeah. And I think that the open source software community has really solved this problem a long time ago.
So if you go on GitHub, one can simply open a pull request on some repository.
So now you’ve contributed.
That element of trust is not so present, you can just run the tests.
Also, when you talk about impact and prestige, you can look at the number of stars you have.
So if a repository is very popular, it gets forked a lot, it gets a lot of stars.
So you’ve disintermediated essentially any gatekeeper here, it’s totally open source, there’s no morning trust required, and there’s a measure of impact.
And so I think math is going to start going the same way.
Previously, mathematicians relied on their social networks to figure out:
But with Lean, you can have a big math project, anybody can come and contribute a proof.
And if Lean accepts it, then it’s right.
If a lot of other mathematicians start to depend on that result, we’re going to notice:
And so then you start to measure the prestige that way.
So it would be very interesting if Lean is the one tool that allows you to go from kind of the cathedral style of development where very closed networks, et cetera, to more bazaar style development where it’s kind of wild west.
But Lean is like the computational certificate that everything is correct.
I wish I understood a little bit better, had a more intuitive sense for what exactly is going on with Lean still.
This is going to be hard, I think.
But in doing my kind of research, one thing that stands out is the kernel is really small.
So, in terms of what you need to trust, it’s a pretty small amount of core code that has been thoroughly vetted many times by many people.
So there’s kind of that level of understanding.
I think I would still love to have a little bit better sense because when you mentioned the three axioms, for example, it’s a little weird for people outside the field to be like,
“Oh, there’s two that are kind of bizarre and technical. And then there’s this one that’s like if you have a non-empty set, you can choose an element from it.”
And I’m like, that seems like common sense, but why was that ever controversial?
Is there a way to describe the sort of space of legal moves in math or in Lean in sort of— I don’t usually like analogies.
I often try to set this up as an analogy-free zone, but because I’m—I think I and a lot of others are going to struggle with the very literal understanding, maybe this is a time for an exception to my no analogies rule.
Is there sort of like a— I don’t know, like a chess analogy or something where you could say, like, here’s the pieces and here are the legal moves that you can make to kind of give people a little bit of a better sense of what it actually means to move through these spaces?
I think the chess example is perfect. So a theorem in Lean is something like, given this starting configuration of a chessboard, it is possible to get to this configuration. And a proof of this theorem would be listing the sequence of moves. And what the kernel is doing in Lean is saying for every single move that you claim is valid, it’s checking, “hey, does this rule exist in my rulebook?”
So the theorem says you can get from A to B. The sequence of moves is, okay, here’s the sequence. And the kernel is just saying, “yes, this step is right, this step is right, this step is right.” And now I’ve confirmed that I’ve ended up in a target state. So Lean is doing that, but of course, the individual steps are different, they’re mathematical steps, and they depend on one or more of these three axioms.
The three axioms, although they’re technical, they’re very short. So if you write them down as mathematical statements, they’re under, I think, each of them is under a tweet in length. Like the axiom of choice definition in Lean is maybe 10 characters, and the other ones are maybe 100. So they’re not very complicated, they’re just a little bit annoying to write in math.
And then people say, okay, well, if we assume these axioms are true, and they’re also common sense, just like a bit more complicated. And we’ve checked every single step against those axioms, then we say the whole proof is correct.
Could you give like a few examples maybe of like the pieces and the moves? Obviously, we can’t come anywhere close to being exhaustive, but what are the primitives in terms of the…
I’ll give a mathematical but simpler example of a primitive. So let’s consider first-order logic.
So the deduction rules you have are:
So let’s say you have a proof that says: if I have A, and I know if A then B, and if B then C, the theorem says C is true.
And the proof of that says:
A is true,
I have if A then B, which means B is true,
And then I have the step B is true,
I know that if B then C,
And then I can conclude that C is true.
So this is a first-order logic, so it’s not quite the same as what we’re talking about in Lean. You can do more advanced types of logical statements there. But ultimately, that’s what’s happening.
I think it’s going to be hard to…
Essentially, the next step beyond that is just getting to Lean and the calculus of constructions and these axioms.
So there is one thing when I learned it. There’s actually…
People are also exploring use of Lean to teach math. And I think now it’s sort of like practical at the high school level, but you could see a world where it extends to like middle school and maybe even younger if someone’s precocious enough. But I think mathematics education will go from sort of like the chalkboard to the computer lab.
So there’s this thing called the natural number game where you learn Lean by deducing properties of like multiplication and addition basically. So for example, the commutative law, which is basically that
A plus B equals B plus A, right?
Or the distributive law, right?
A times quantity B plus C equals A times B plus B times C.
So you can sort of like discover and prove these fairly basic facts just using the core axioms and the Lean language.
So that’s a good way, you know, if anyone just wants to like, all right, what is this Lean thing? Why is it useful? But I’m not a research mathematician. Dip your feet into it. I think I would recommend that.
And that’s been extended to harder things too. I think there’s like the real analysis game now, which is if you want to learn real analysis, it’s very proof based. And it’s basically the foundation of calculus.
You can start with like basic facts about:
And then you can kind of keep proving more and more complex things.
That’s a great tip. I’m definitely going to bookmark the real numbers game and see if I can get my soon to be seven-year-old into it.
Hey, we’ll continue our interview in a moment after a word from our sponsors.
Want to accelerate software development by 500%? Meet Blitzy, the only autonomous code generation platform with infinite code context, purpose built for large, complex enterprise scale code bases.
While other AI coding tools provide snippets of code and struggle with context, Blitzy ingests millions of lines of code and orchestrates thousands of agents that reason for hours to map every line-level dependency.
With a complete contextual understanding of your code base, Blitzy is ready to be deployed at the beginning of every sprint, creating a bespoke agent plan and then autonomously generating enterprise grade premium quality code.
Grounded in a deep understanding of your existing code base, services and standards, Blitzy’s orchestration layer of cooperative agents thinks for hours to days, autonomously planning, building, improving and validating code. It executes spec and test driven development done at the speed of compute. The platform completes more than 80% of the work autonomously, typically weeks to months of work, while providing a clear action plan for the remaining human development.
Used for both large scale feature additions and modernization work, Blitzy is the secret weapon for Fortune 500 companies globally. Unlocking 5x engineering velocity and delivering months of engineering work in a matter of days.
You can hear directly about Blitzy from other Fortune 500 CTOs on the Modern CTO or CIO Classified podcasts. Or meet directly with the Blitzy team by visiting Blitzy.com. That’s B-L-I-T-Z-Y dot com.
Schedule a meeting with their AI solutions consultants to discuss enabling an AI native SDLC in your organization today.
The worst thing about automation is how often it breaks. You build a structured workflow, carefully map every field from step to step, and it works in testing. But when real data hits or something unexpected happens, the whole thing fails. What started as a time-saver is now a fire you have to put out.
Tasklet is different. It’s an AI agent that runs 24/7. Just describe what you want in plain English, send a daily briefing, triage support emails, or update your CRM. And whatever it is, Tasklet figures out how to make it happen.
Tasklet connects to more than 3,000 business tools out of the box, plus any API or MCP server. It can even use a computer to handle anything that can’t be done programmatically. Unlike ChatGPT, Tasklet actually does the work for you.
And unlike traditional automation software, it just works. No flowcharts, no tedious setup, no knowledge silos where only one person understands how it works.
Listen to my full interview with Tasklet founder and CEO, Andrew Lee. Try Tasklet for free at Tasklet.ai. And use code COGREV to get 50% off your first month of any paid plan. That’s code COGREV at Tasklet.ai.
And we haven’t really talked about Mathlib, but the Lean kernel is quite small. There’s an open source project called Mathlib, which you can kind of think of as the largest digital repository of mathematical knowledge.
So all of the, a lot of the famous theorems and results can be found in Mathlib, and those give you almost like additional complex moves or algorithms to prove your thing. So you can apply a theorem, and it’s almost like applying a function from a library. That can help you get to the goal.
Yeah, I think that people can understand what it is better. Just think of it like every math textbook in the world merged into one in a self-consistent way. So eventually, all of mathematical knowledge will be in this one repository.
And if you hit build on your computer, you’re going to be able to check it all from the foundations. If you have any question about any math concept, you just search for it, you click on go to definition, you can jump around. It’s really going to be the new foundation for math in the future. It’s pretty exciting.
I think mathematics is certainly going to change fundamentally—like how it’s done, how fast it moves. And I think to a large degree, it already has. And AI is just going to accelerate it.
The great thing about our timing is Harmonic really started when both of these things matured to a level of capability where you could start doing interesting stuff.
I think both of these matured to the level where you can start putting them together and doing really cool things. And I think we were just the first to see that. That’s how we came up with this concept of mathematical superintelligence, which really means the combination of formal verification and formal tools with artificial intelligence.
Funny story, as I was using Aristotle a little bit to try to wrap my head around all of this, I don’t have the sophistication to pose any really interesting problems. So one challenge that I gave it was to prove that two plus two equals four. And then I had to laugh when it came back, just citing something from Mathlib that was like, “this is already proved in Mathlib” for the, the theorem is literally like the two plus two equals four theorem. So I was like, it’s done. And I was like, yeah, that’s not exactly what I was looking for, but I guess I kind of got what I deserve there for asking it such a basic question.
Did you use the, were you using the web interface or the terminal UI? I started by having cloud code installed the terminal and then was using that a little bit. And then somehow it tipped me off to the fact that there was a web interface. And so I, then I, after that I’ve moved over to the web interface. Yeah, that came to the last week and they’re probably a little bit more appropriate for those types of questions.
I think we wanted to roll it out for on terminal, because I think it makes it a little bit more clear what the tool is great at. I mean, lots of things can answer two plus two equals four, but even I can answer that in the calculator.
Yeah.
And I think, I think for a while we were talking about like, how do we describe this, this, like what Aristotle is? I mean, it’s, it’s kind of like an amazing calculator where you can imagine you could just talk to your calculator. So it has:
You know, something like ChatGPT or Claude are very expressive, but sometimes you have to double check its work because it doesn’t always, you know, it doesn’t have the verification. But really the intent is to put those together.
And it turns out that the things, the first things that people really want to be sure about and to verify are like more complicated things. So I think the, you probably found this out, but the complicated things I think is where you really start to have aha moments when you’re using it.
Yeah, let’s get into Aristotle. And I appreciate the time spent in remedial education. I think it’s beneficial, not just for me, but hopefully everybody will now be able to kind of grok what we’re about to get into much better with that foundation that we’ve laid.
So Aristotle has three core parts. I’ll just kind of sketch them and then you can, you know, give me the double click on them.
First, there is this Monte Carlo tree search type thing. I kind of think of that as sort of an AlphaGo-like structure where we are systematically exploring the space of moves. I guess that’s where I got the chess analogy, right? Is that I kind of was making this equivalence between Aristotle, at least that part of Aristotle, and AlphaGo.
And so it’s kind of maybe I can make this move. And then there’s this learned scoring function that’s like,
“Okay, does that move seem promising? Does this path of, you know, does this branch of all possible moves that I could make, does it seem promising? Do I seem like I’m getting closer to my goal?”
And with that, you can kind of grind things out, run deep tree search, right?
The second part in some ways to me jumped out as even more interesting and kind of, I really want to dig into the metaphysics of it a bit, because this is the lemma-based informal reasoning system, which I take to be sort of saying,
“Okay, if I have some really big mountain to climb, and it’s maybe so big that I can’t just grind my way… it’s maybe becomes impractical to grind my way through like all these small localized steps.”
It’s sort of guessing like what’s the base camps that I would want to get to along the way that are like the really good way points such that if I can get there, then I know I’ve like, I’ve made it somewhere.
But that’s really interesting because it sort of strikes me a little bit more like a, like it behaves, it seems like a little bit more like a language model where it’s kind of guessing and not so formal. I mean, it says in the, in the technical report that it is an informal reasoning system.
And then there’s a third part, which we maybe don’t have time to go as deep on, which is specifically dedicated to geometry. Again, in the technical report, you described that as being like AlphaGeometry, which I think DeepMind developed.
So correct any misconceptions that I have there and give me the double click on what, like what more I should understand about how this thing works.
Sure. I think, I think you covered the components pretty accurately. So one thing I have to say is that, you know, we revamp our systems pretty often here. So I think Aristotle now looks quite different than Aristotle for the IMO. You know, I think a lot of things are consolidated and improved.
I think that you made this point about the Monte Carlo tree search being more of a grinder. I wouldn’t quite characterize it that way.
So the Monte Carlo tree search is actually doing a lot of inference on its own about high-level steps. So the levels that we’re talking about, they’re much closer to solving a challenging math problem than they are to prove that two squared equals four. So there’s a lot of reasoning that goes into them.
In some sense, it’s grinding once you get low enough in the search tree because you’re just closing out cases or easy subproblems. But it’s really solving harder problems on its own.
And so when we combined it with the informal reasoning system, you could almost think of it as a form of context management, actually. So ultimately, you need to end up with a lean proof, and that’s going to involve big steps and small steps. And it’s helpful when you’re focusing on the smaller steps to not have to remember the entire context of the bigger steps.
And so it turns out the informal reasoning system itself actually makes enormous quantities of mistakes. So one should not think of it as,
“oh, it’s a really smart human that’s laying out the steps to base camp.”
It’s more like a system that can propose lots of things that are wrong and don’t have to be formalizable or even correct. And you kind of try to assemble things from that.
So you can think of both of them as kind of doing the same thing, just at slightly different scales and complementing each other. And they’re actually all LLMs. So as we described in the tech report, the tree search itself is driven by language models.
Part of the language model is proposing steps. Part of it is scoring steps. But they work in concert to solve the lemmas and then eventually the full problems.
And as you mentioned, alpha geometry, it’s a slightly different system. We’re exploring kind of high-level steps and then trying to use an algorithm to grind through the rest of it. I think if we’re talking about systems grinding through a lot of math, I would say alpha geometry in the deductive reasoning system is really a grinder. So it’s really trying to find every possible conclusion of a geometry diagram.
I would say there’s not too much pattern recognition intelligence going on there. And that’s because geometry, if you think about it, is more constrained. You basically have points. You can basically start with points.
If you have three points, there’s only so many angles involved. Obviously, if you go to like 10 or 15 points, things blow up pretty quickly. But it also then becomes hard for humans to solve. And I think that’s why geometry was among the first class of competition problems to fall to AI and automation.
I think there’s also a couple of other components that might seem simple but are non-trivial that Aristotle, the system, does and are independently improving.
One is auto formalization: taking input that you provide in natural language and faithfully translating it into Lean in the best possible way. And I think, relative to our competitors, at least, I’m not aware of anything that’s as good at that as we are.
And also theory building: sometimes in the way of solving something, you have to create new theories and new structures that might not exist in mathlib. Aristotle has the capability of actually building that on the fly and incorporating that into the proving process.
Another funny anecdote, so that you’re referring to what I discovered is informal mode, right? Where I can provide—I think real users would not do this—but you can provide anything, any natural language input, just something that the system will then try to prove.
I asked it to prove all is love. And it came back and said,
“this is a philosophical statement and outside the scope of the Lean formalizer’s ability to prove.”
I also asked it to prove
“Epstein did not kill himself.”
And it came back and said,
“this is a statement about current events. And again, it’s sort of outside the Lean formalizer’s ability to prove.”
But yeah, I think this kind of gets back to this sort of metaphysical question that I find so perplexing around that translation from the messy real world of human affairs and intuitions to the formal definitions of,
“okay, this is actually the thing that we would want to prove.”
I did find that very, very interesting that you had such a thing at all. And I guess, well, do you have a sense for— I also do want to get into a little bit more details of just like technically how you created the models and all that stuff. But, you know, on my spectrum from 2 plus 2 equals 4 to all is love, is there, how do you think about the intuition for what the boundary is of what is inside the, what, because I, because I, again, when in listening to your previous interview with Sequoia folks, it seemed like you had the sense that eventually as the system and systems like this get capable enough that more and more things that are of interest every day people will start to become the sorts of things that they can do.
So like, how do you think about that boundary and how does that boundary expand over time?
I think the, the ultimate boundary of a system like Aristotle is in reasoning through any problem where people can also agree on what it means to be a valid sequence of reasoning steps. So right now you have math. That’s one obvious one. When we talk about mathematics being the same as reasoning, that chess example you gave is a perfect one. So you can express the logic of a chess game and then check it, right, and then reason about it.
I think one area that’s really going to touch a lot of people’s lives is it turns out you can use the same reasoning approaches to think about software. So when people write software, they write these things called:
And it’s kind of having the computer just run the program and check the output against what they expect. But that’s what they do after they’ve written the code.
It turns out that when engineers are writing code, they’re thinking logically:
“Okay, if I have this range in my input, I can think okay, as I go to this for loop at these if statements, it implies certain things about the output.”
And that itself is logical and mathematical reasoning. So we’re starting to see API users reason about programs in the same way that they can reason about math. People are writing cryptography implementations and then checking:
I think the same kind of input that will go to software will help take us to a bug-free software future.
Now, Vlad and I disagree a little bit. It’s not clear to me if we’ll be writing history essays or something — maybe there is a way to value them objectively. But I think the boundary is really in anything that’s quantitative and logical in nature.
Yeah, I think in the first version of Aristotle, it would actually formalize and build a theory for your all is love example. And it would give you a correct proof that it’s probably true.
I think it surprised us. People were asking all sorts of questions. We had people asking:
And Tudor mentioned computer science. So I think it’s actually surprised us how broad of a set of things it can successfully create a theory around and formalize.
I think the constraints we put were just, you know, when you’re building a product, you want to make sure that you deliver value. At this point, I don’t think we provide the most value if you want to write a history essay.
So we’re trying to nudge people to the point where they can discover what Aristotle is really, really good at as quickly and simply as possible. I think over time, you should expect that the surface area increases.
We start formalizing things. And I don’t think it’s inconceivable that at some point it pulls current events and news from the internet, puts out the axioms, and can sort of fact-check and make conclusions based on real-world events.
Not our focus right now, but I don’t think it’s a crazy thought.
I mean, I ask a question sometimes. I’m interested in astronomy, right? And I wanted to know:
“When’s the next full solar eclipse that I can see from within 50 miles of Palo Alto, California?”
The models usually struggle with this type of stuff because nobody’s asked that identical question out on the internet, so they can’t pull it. You actually have to do some math.
So you can imagine there’s a spectrum and there are questions like this that a model that can reason actually from first principles is going to be way better at.
Okay, let’s talk about just how you created this thing a little bit and how your experience, lessons learned, et cetera, kind of relate to some of the live questions more broadly in the AI space. I think you can take on faith that folks listening to this show will be familiar with things like reinforcement learning from verifiable rewards and stuff like that and certainly understand kind of how the ability to generate synthetic data feeds into a system like that and that’s, I’m sure, part of what you’re doing.
What more can you tell us in terms of like, would it make sense to start training something like this from some off-the-shelf pre-trained model or does that messiness that those, you know, LLMs start with corrupt or pollute your, the purity of the mathematical reasoning too much? Can you tell us anything about size of models, which could be parameters, could be tokens, whatever?
I’m interested in things like also, is there any role for taste in this process? Obviously, like mathematics, mathematicians are very interested in correct proofs, but they’re also interested in these eureka moments and the sort of sense of elegance of the proof, right? There’s a sense of the beauty that, you know, matters as much, I think, to many people as the correctness or maybe not as much, but, you know, it’s certainly heavily weighted.
And then I also noticed there’s test time training that’s part of this, and I think that’s, you know, a huge trend that I’m kind of watching in general.
So, you know, you can swing or take any of those pitches, but what do you think are kind of the most interesting next level of depth that people can use to inform their own AI worldview with?
Well, first, I have to say that if your audience knows about reinforcement learning from verifiable rewards, you’ve got a great audience.
“That’s not betting data.”
Yeah, that’s not. So, I think that is a safe assumption. Nobody was talking about that stuff, right? It was like science fiction almost, but it’s cool to see it entering the popular consciousness.
I want to address the taste question, because that actually, you know, strikes at a key thing that, you know, companies can decide on.
So, we get gold performance at the IMO, we have a very powerful system, and it was obvious we had to give it to people.
And there’s two ways you can do it.
That’s one way of expressing taste in the research map.
The other way, which we ultimately decided to do, and we think it’s been great for the community, is we said, well, we’re not going to be the ones to decide what’s important in math. We’re going to make Aristotle accessible to everyone.
And so, we opened up the API, the web interface, there’s a lot of great features coming.
And then, in this scenario, taste is expressed by the community by the revealed preference of what they submit to the API.
So, we don’t choose what kind of math they do.
We’re not saying, hey, Navier-Stokes is more important than P versus NP.
It’s the mathematicians that have the credits on the API to say, well, we care about X or some other thing.
And that’s why we’ve seen so much interest in:
And for a while, there are people doing a lot of interesting conjectures in graph theory on the platform.
And I think that that’s actually the right way for companies to engage with the community.
You know, you open the system and you let the people decide, you know, where they want to allocate those computer resources.
So, I think that’s an important decision. We’ve come on one side of it, but I think that’s the right long-term approach.
I think there’s a philosophical question there, too, which is, are we headed for a future where the AI labs themselves are going to generate all the discoveries?
Will the cure for cancer or diabetes look like a giant AI lab with a two gigawatt data center just churning on this problem? And then, you know, it comes out and they capture all the value?
Or does it look more like millions of people empowered with these tools working independently and collaborating and, you know, in that world, they’ll get the credit and the value will largely accrue to them?
And I think we believe that the second world is more interesting and it’s probably the one that’s more likely.
The first one is rather dystopian and less likely.
And I think we noticed that because when we rolled out Aristotle, you know, we had one view of what people would use it for, but then we started getting all of these, you know, Erdős problem results and things like that.
And it’s like, we’re not going to run on all the Erdős problems. We’re not going to do like computational learning theory, formalizations in house.
So I think the amount of cool things being done with it just explodes if you put it, if you make it generally available. So I think it’s not only right from a business strategy standpoint, but also like, I think that the world that we built, assuming this path, is a better world that I would like to live in.
So that speaks to taste in terms of problem selection.
But I was also just thinking in terms of, as you’re training the model, you’ve got the correctness signal, but maybe one sort of heuristic for elegance would be like just brevity.
Which is maybe one kind of way of trying to send an elegance-like signal through a deterministic mechanism. But I would be very interested to know if there is like a panel of mathematicians that you guys have reviewing solutions for elegance to try to make sure that this thing is not just a pure grinder long-term, but really has a more eureka flavor to it.
Well, brevity—if brevity is the definition of elegance—then our two plus two equals four proof probably takes the cake, right?
“I can’t get any shorter than that.”
I would feel bad for any mathematician’s job of us to compare AI proofs. That’s certainly not the job I’d want.
So we, we have never. It’s a big business these days across all domains:
Yeah, we have done essentially zero of that in the two years we’ve been around.
I think the metric we optimize for is the net present value of future proofs or computational costs of future proofs. And so that guards very naturally against certain phenomena.
When you’re solving easy problems early on in reinforcement learning, you absolutely can solve them with grinding. So you can say,
Let me just do brute force.
But you know that if you do that, it’s going to cause issues later because you haven’t learned how to do more complicated things.
In contrast, if you’re given two proofs that are not grinding, but one is drastically longer and more inefficient than the other, you prefer the more efficient one.
So there’s a tension there because you can get more efficient by grinding, but that messes you up in the future. So it’s a balance that our AI researchers strike based on their intuitions about what’ll be helpful long-term.
But we have never had panels of mathematicians do testing on proofs or anything like that. Really, you want to give your system as few priors as possible and just run reinforcement learning at scale.
There’s a famous essay called The Bitter Lesson, which I’m sure your viewers are familiar with. We really believe in that at Harmonic.
To get to your question about how we started: sometimes we’ll start from pre-trained models. Ultimately, you want to do whatever optimizes that net present value of future cost of proof. So pre-trained models are great for that.
I think at some point you might ask the question,
“Is that going to bias you too much towards how humans do math?”
And so you want to mix in reasoning systems that are not trained from human knowledge, right? They have more entropy and more complementary knowledge.
That kind of thing we always play with, but it hasn’t really been the living factor so far. I think that pre-trained models are a great starting point.
Cool. I guess one thing: Goodfire just announced today that they raised a bunch of money at a unicorn valuation. I was a very small-scale supporter of theirs, and it got me thinking.
This also connects to Vlad’s comment where you said the system can sort of invent new theory.
Obviously, one big thing people have said AIs can’t do, or AIs can never do—which is always a dangerous position to take—is that they can’t come up with new abstractions.
Sure, they can learn from what we have done and what we’ve encoded into language, but will they ever come up with their own abstractions? I think that’s not a very strong, increasingly hard position to defend.
But what is so interesting with Goodfire is they’re now starting to look at model internals and unlock new kinds of understanding based on looking at what the model has learned.
The famous one they just put out is like new markers of Alzheimer’s that people didn’t know about, but the model was able to figure out, and they were able to figure out what the model had learned by looking internally.
I’m kind of wondering:
I mean, one of the things that I’m very excited about is eventually Aristotle powering a spacecraft, right? Much like HAL 9000, but a benevolent one, one that doesn’t go crazy. So, yeah, I think eventually you’ll see it expanding into more real-world things.
I think the… I don’t know if you’re as excited about that. A safe HAL 9000. A safe HAL 9000, I think, would be very valuable.
You know, to your question on interpretability, I think that interpretability is often used as a proxy for trustworthiness. So, a lot of the reason that people explore interpretability technology is that they can make sure that the system does the right thing or aligns with the user’s intent.
So, when it comes to trustworthiness, we made the explicit decision at the very beginning of the company to focus on Lean. By outputting our reasoning in a formally verified way, that is the most interpretable possible output. So, the computer can check it. If the human wants to understand how the proof works, they just keep hitting “go to definition.”
It’s almost like navigating through a code base. There’s no more interpretable way to output math than in Lean, really. That’s the maximal version.
So, now the question is, okay, well, how interpretable is the model? I think, in the context of the bitter lesson, we just focus on letting the system do whatever it can to optimize for computationally cheap proofs of more and more complex things, with a caveat that it has to output in a way that’s verifiable.
I think down the road, we’re very curious, how does it do math? How is it so smart? And we’ll look into that. But for us, we’ve solved the trustworthiness question upfront by focusing on formally verified output.
Yeah. Okay. That’s quite interesting.
I do sort of feel like, I have this one kind of mental—mathematicians are famous for visualizing things—my kind of visualization of what is happening in a large model is sort of like shrink-wrapping reality.
Like, you’ve wrapped in plastic all of, you know, all internet data or all the kind of whatever domain it is that you’re trying to learn at scale, and you’re just sucking all the air out of it and gradually shrinking down to whatever, hopefully, is kind of the true structure.
And it strikes me that in math in particular, that structure might be amazingly simple. Or, there might be really interesting things to learn by running that process and then kind of cracking it open and seeing what is inside.
I would expect it to be maybe a lot more interpretable internally than something that has had to learn all internet data and can recite Wikipedia and all that sort of stuff.
I actually think that what these models are doing is interesting because they’re smashing together all of the techniques that all mathematicians have done before.
And so, while I haven’t seen the spark of superintelligence yet where it’s some breakthrough eureka idea that’s incomprehensible, I’d say that if you push it in, learning how the models do things, you kind of ask it to solve more and more complex problems and just see, like,
I think that’ll be a lot more interpretable and comprehensible than trying to dig through the way it’s structured—I might be wrong, but that’s probably where I’d start to interpret how it does things.
Yeah.
So does that mean maybe we can kind of look at different levels of difficulty of problem?
We’ve got the Erdős problems.
There’s definitely a phenomenon happening right now where people are using either Aristotle by itself, or—I’ve also seen a lot of examples, not that many, but increasingly more, of GPT 5.2 Pro to sort of generate a proof in token space, then bring it over to Aristotle for formalization.
Then there’s, of course, the IMO.
If I understand correctly, everybody who—and I think it was just three, right?—that you guys, OpenAI and DeepMind, got the gold level performance. I think everybody missed the same one question, which is really interesting to me.
I’d be interested in your thoughts on,
“why that—why so consistent?”
And then, of course, we’ve got these extreme problems where you would need this sort of move 37-like moment to solve them.
So maybe kind of sketch out,
I mean, I think – so on the IMO, the three labs that announced gold medal performance—us, DeepMind, and OpenAI—all missed question six. And I think that it wasn’t super surprising to us because question six is probably, I don’t know, 5x harder even for humans, right? It’s just a more complex question with lots of steps, and it requires this type of spatial reasoning that right now is more difficult to encode in formal systems.
We were running our system on it quite a bit, and we felt like we saw signs of life. So it’s definitely not inconceivable that before too long, question six is going to fall and be gobbled up just like the other questions. I mean, even one year before, questions three and five would have probably been well beyond reach for most of the models. So I think it does appear to be more or less a smooth exponential.
Yeah, I agree with that. I want to highlight that there’s two aspects of this.
So I think years ago, if you had asked someone, “Hey, could you automatically formalize a number theory paper in Lean or Rock or Isabel, these other languages?” you’d have been laughed out of any room of mathematicians you’d be in. And today, we are seeing people upload the full text of a math paper and run Aristotle a few times. We’re thinking of adding a Ralph button to just keep going, keep going, keep going. And then you get a formal version of it.
I think that phase transition has essentially come and gone now because of Aristotle. So in the next couple of years, as AI keeps improving, the fact that we can now formalize the AI’s arguments obviates the need for the humans to just be the verifiers, right, just sitting there and checking if some output is correct, to ones being the tastemakers. So we’re the ones setting what problems to work on if we’re happy with the techniques used. So that, I think, is the interesting transition that’s happened. So smooth exponential capabilities, but I think we’ve gone zero to one on verification.
I think that’s such a great point because I think there was some debate about this at the beginning.
And in a way, if you look at DeepMind, they started with formal, with AlphaProof, which was the silver medal-winning model back in 2024. It was a great result at that time, and that was a formal model. And then they went back to informal for Gemini this year, and I’m sure they ran AlphaProof. Maybe it was just that AlphaProof didn’t do as well. OpenAI, obviously, informal.
But if you think about, okay, let’s say we go to a world five years from now, and the autonomous math being done by AIs increases. Instead of five to ten-page proofs, you’re starting to produce 5,000-page proofs, which you should assume, right, as these models can autonomously reason more and get more efficient, they’ll produce longer and longer output per unit time. It’s going to be a proxy for complexity.
Who’s going to review that? Nobody’s reading a 5,000-page math proof. So I think it’s becoming even more clear that the future is formal because you have this problem of someone having to validate it and check it. And we want to make sure that the time to validate it and check doesn’t actually grow linearly with the complexity of the proof.
Yeah, that was really the founding thought experiment of harmonics.
So we asked ourselves in 2023:
- These models can do high school math poorly, but they could do elementary school math poorly a year ago.
- What happens in 10 years if we ask it to prove the Riemann hypothesis?
Any model will make an attempt at it and give you 100,000 pages of output, which you might as well throw in the trash for two reasons:
No, you just can’t wrap your head around what is going on in that proof.
And so there were two hypotheses, both of which have been proven out:
If you compare the resourcing we’ve had compared to the big labs, we’re punching well above our weight at the IMO. So I think, in our view, the debate on formal versus informal is settled. I mean, clearly, it’s going to be formal.
One can debate, okay, what’s the most efficient way to train a model? There’s some aspects to informal that are helpful, but I don’t think we’re ever going back to a world where we’re like, “oh, it’s just going to be informal from here on out.”
I think the interesting question, though, is to extend this to software, right? Because the same things actually hold for software that hold for math.
Let’s say AIs are getting to the point where they can autonomously work and create a software project over a period of a week or multiple weeks. You know, who was it? The cursor team ran this and generated like a Chromium-compatible browser, right? It was something like one and a half million lines of code. It was incredible.
So who’s going to read that code and find all the security vulnerabilities and the bugs? And is that code in the future that’s generated by AIs going to be in Python and Java anymore? Like, why would it be in Python and Java? Those are just languages optimized for human readability.
And, you know, if the answer, we think, to humans reading and trusting something or even an AI that the model is collaborating with checking something is the same. You want to make the cost of verification as low as possible. And that makes us believe that the future of software is formal as well. And more and more software will be written in formally verifiable languages.
Yeah. And I think, you know, Lean is our favorite language. It would be amazing if everyone can write in Lean. I think that as AI writes more and more code, it will be easier for people to accept that. But we’ll see.
And I’ll start with mission-critical, important stuff where bugs are much more serious and much more costly. And there’s a bunch of domains that already are doing formal verification for software, but they’re doing it in a very artisanal way.
You know, they’re hiring Lean or Rock or Isabel experts and kind of painstakingly formalizing stuff. So I think you’ll start to see it accelerating the work of those people first, but then it’ll just diffuse and you’ll see, like, formal vibe coding before too long.
Yeah, I love the term vibe-proving, by the way. Yeah, I think that vision is an incredibly compelling one. And, you know, it’s also one that I’m still kind of wrapping my head around.
For listeners who haven’t already heard it, I did one episode with Kathleen Fisher, who was at DARPA, and I think now has just moved to ARIA in the U.K. to lead their whole operation. And Byron Cook, who’s like a legend of the formal methods field at AWS. And, yeah, they’re kind of right there with you, you know, envisioning this world of basically totally verified, bug-free software, starting with mission-critical stuff, but potentially extending to everything over time.
I guess one – so I think that is super compelling.
The one kind of nagging – I don’t know if it’s a worry that I have or what exactly, but I’ll just frame it as a question – is, like, if we are training an AI to be superhuman at formal reasoning, within the formal reasoning system that we have,
how do we get new abstractions from that or how do we get a sort of Einstein kind of moment where, you know, like, it seems that at some point we all sort of thought the world was just naturally 3D and that was, like, obviously intuitive.
And then it’s kind of come to light, obviously now, that, like, well, that was an adaptive understanding of the world that served us well as monkeys, you know, and allowed us to survive. But it was at – at the end of the day, we now know that it’s, like, a lossy approximation of true physics.
And so I’m kind of like, do we have any room for doubt or worry that the math that we have now, as sophisticated as it has become, might also at some point prove to be not quite the right paradigm? And is there any way – if you’re training in this, like, purely formal way, is there any way sort of to punch your way out of the box as an Einstein did, right? He seems to have –
“The fourth wall.”
So he broke the fourth wall conceptually, but the key thing to remember is that he was able to describe his theory rigorously and formally in the framework of differential geometry.
So the point I was making earlier about math being reasoning is the point I’ll appeal to now, which is to say that no matter what complicated theory somebody might come up with to explain how the universe works in the future,
If it’s going to be based on a series of logical deductions that can be explained to someone else and checked independently, that is itself a logic that can be encoded with Lean or other languages like Lean and then verified. So, again, the axioms that Lean is based on are so minimal and just expressing just the most basic possible common sense about how reasoning should happen, like, one thing might fall from another, or if two things look the same, they are the same.
That’s the level of axiom we’re talking about. So I really don’t think there’s any conflict here. I think that one should just think about formal reasoning as an especially detailed version of informal reasoning that a computer can check automatically. There’s no limitation to it. Sometimes it might be a little more verbose than you’d want, right? So you want to write tactics and things to cut down on that, but there’s really no fundamental tension to turn into.
And I think there, you also, you know, might be thinking about Gödel’s incompleteness, like the fact that in any sort of axiomatic system, there’s statements that are true and unprovable. And there’s also statements that are undecidable, right? And independent. So there’s sort of like a bunch of edge cases here, but I think it doesn’t prevent us from making a lot of progress and proving actually the lion’s share of useful things. I mean, there could be things that are unprovable but true that are very, very useful to know as well. But, yeah, no way to know unless you explore the frontiers.
Do you think there’s always going to be a role for entropy of some sort in these systems? I mean, I think hallucinations are a key part of a reasoning system. Hallucinations are what allow a model to explore something that has never been encoded by a human before.
So, you know, when we run Aristotle, whether it was at the IMO or noun, it makes a lot of mistakes. It tries a lot of paths that don’t work. But that exploration is the very thing that lets you get to the right answer after enough attempts. So entropy is crucial. I think this whole notion of seeking fundamentally hallucination-free LMs doesn’t really make much sense.
Now, of course, you want to pair them with a system like Aristotle that can verify things in 10. But, no, I think entropy hallucinations are a key part of the training process for models like this. You’ve got to be able to pose false statements in order to prove that they’re false. Learn like humans. You know, you try a lot of room for humans. Some of the most creative humans are the ones that hallucinate the most.
So what’s kind of the latest progress on the path to superintelligence? You said you, and I think this is true of all good frontier AI companies, whether, you know, at the application layer or the model layer or anything, any hybrid of those, you know, you’re updating your systems frequently. It sounds like there’s kind of a convergence of some sort going on between the tree search part and the informal lemma guesser that you described in the technical report. What can you tell us about kind of what the trends are right now?
I think a lot of the—well, just to review the progress, right? So we started in 2023 and then in 2025 goal performance, the IMO, we topped out this Verena benchmark at the end of the year with our public API users started solving Airdish problems, right?
So I think there’s a very clear trend, right? And, and capabilities. I think the phase transition I mentioned has also happened.
So I think what’s next for harmonic and for the field at large is, you know, a couple of things.
Well, we can expect math live to grow. So math live is the, think of it like the Wikipedia for math that’s computationally certified. So as Aristotle makes it possible to auto formalize a lot of math, you can expect that users will start contributing a lot of pull requests to math live. And that makes it possible to solve more and more problems on top of that base.
I think when we look at how mathematicians are using our API, certainly people are starting to work on more important unsolved conjectures that a lot of people would care about.
So you can kind of think about conjectures as like,
“Okay, there’s a conjecture that’s technically been open, but nobody really cares about it.”
So it’s not like people are trying all the time, but now you might have some conjectures that, yeah, like a mathematician might try it once or twice a year, just take a shot at it. Maybe a hundred mathematicians would.
And then eventually, but the millennium prize problems where, you know, any mathematician would be happy to spend years on it if they might be able to solve it. So I think what you can expect from Aristotle and other systems is, you know, more and more problems get picked off. So it becomes easier to use it extends to software, as I mentioned.
So we have users using it to check, say, decretable software, whether in Lean or other languages.
And overall, if I had to pick out just one trend, it’s really just that formal reasoning goes more and more mainstream. So as more stuff is produced with AI, I think you’ll see complementarily more formal reasoning to kind of verify all of it.
And I think on the product side, we’ve gotten a lot of feedback coming in from the folks using it. Obviously, whenever you’ve got customers that are using a technology like this, they’re very passionate.
So there’s lots of ways in which they’re still complaining about things and improving the ergonomics of it, making it so that people don’t have to hop between so many different tools. And we could just solve their problem as simply as possible and at the lowest possible cost. You should see that continue to improve.
There have been updates to the system pretty much on a daily basis. Maybe you’ve seen some of them just as you’ve been kind of experimenting yourself. But that is going to continue. And you should expect that it gets exponentially more useful over time.
So maybe a good place to close is kind of the vision for what that looks like as you succeed. I mean, obviously, one thing is solving Millennium Prize problems. But I’d love to get a little bit more of kind of an intuitive understanding than that.
I mean, one dichotomy that kind of comes to mind is this very formal reasoning-based paradigm versus what I think of as intuitive physics. It does seem like models are very good at developing intuitive physics in kind of any number of spaces.
Right. Like folding a protein with a model is not something that’s done in a formal way. It’s just kind of something where whatever kind of mess of heuristics they learn, they can do a protein fold orders of magnitude faster than we would be able to do it.
And if we were going to do it through a sort of physics-based simulation approach, when we think of no limit to math and what a mathematical superintelligence looks like, I also think Eliezer, once famously—or at least famous to me—said:
“A real superintelligence in his mind could look at one still image and deduce all of physics from just the information contained in that one still image.”
That kind of also connects, I guess, to test time training.
What is your vision? You can bounce off any of those concepts, but what is your vision of how this thing evolves? Is it an ever bigger tower of formal statements? Is there some role of new kinds of intuition, new abstractions that emerge out of that that aren’t so strictly defined but potentially useful?
You know, what is this thing doing in 2030 once all the Millennium Prize problems are solved?
I think that by 2030, we will have theoretical explanations for everything, basically.
I mean, if you look at the history of science, there’s leaps of intellect and leaps of data:
Right now, there’s really been a shortage of people that are able to reason logically at the highest level.
So when you think about unifying general relativity and quantum mechanics, it’s just a very hard thing to do.
I think what you’ll see is really like anything that can be posed mathematically, which is what underlies all of science, we’re just going to get theories for everything that are self-consistent and make sense.
I think we’ll then go back into a regime where we’re data limited. So, we might have maybe five theories that unify QM and GR, and we’ll have to run very high energy experiments to figure out which one is right.
We’ll have to wait a while to build those colliders. But at the very least, we’re not going to be bottlenecked anymore on wondering, “Can we explain something?”
We’ll have a system that can explain anything perfectly correctly. So it really will be a renaissance of science. You just remove the intellectual bottleneck in everything.
So do I understand that correctly? Basically, you’re envisioning:
Yeah, because AI is not omniscient. Whether it’s our model or others, they’ll be able to reason about anything they can kind of ground in their own logical deduction rules.
But ultimately, there are aspects of the universe where you just have to run the experiment and find out how it really works.
Wow.
Just to be clear, I think there’s a lot of utility before you get there. If I have to analyze asymptotically where we get to that, that’s my point. Well, I mean, that’s, we’ve heard about centuries of scientific progress collapse into five years. That sounds like more like a few thousand years, perhaps, of scientific progress.
Also, that’s left will happen, and then you just have to get more data. But you’ll have a superintelligent system that can help you. Wow. Okay. That’s about as grand of a vision as I’ve heard anywhere.
Do you guys worry about the safety of these systems? It sounds like we haven’t talked about that really at all in this context, but I’ve done many explorations of different safety concerns.
You know, Eliezer, when he described the model, whatever AI he was kind of envisioning, when he described it, understanding all of physics from a single image, he also thought that was going to be super dangerous because it would be so powerful.
How do you guys think about that aspect of this whole, I mean, we’re talking about a lot of stuff in the next five years.
I mean, I think right now we’re not so worried about it because the outputs of our system are constrained.
I think that you’re likely to see, like, the first dangers will probably look a lot like cybersecurity incidents, right? Because, you know, you have the models that are making API calls and running autonomously, interacting with other systems.
So that both creates API level cybersecurity holes and the mechanisms to exploit those. So I think you’re likely to see a lot of those.
I think for our model, since it’s basically just the interface to the outside world is tightly constrained, and it’s not just going to fire off a request to your Gmail account or the iMessage APIs, we’re a little bit further away from that. But, you know, you can imagine we’re going to have to start taking that much more seriously when we do get to a point where we’re connecting the model to the outside world and it’s, you know, speaking in the interfaces are not just sort of like lean files being outputted.
Yeah, I do think constrained action space is certainly one of my favorite paradigms for keeping things under control. But I mean, there’s a full like molt book, molt bot thing that has been fascinating to watch. And, you know, I think we’re entering a strange new world for sure.
And I think the benefit is we’re probably not at the danger frontier. So we’ll have the opportunity to learn from others’ mistakes, and hopefully they don’t screw up too badly in order for us to learn.
Yeah, okay, fascinating stuff. This has been fascinating stuff, guys. I really think the approach is really interesting.
The vision for how far we can expect, or even somewhat entertain the possibility of being in 2030 is arresting, and both inspiring and for me, a little bit scary.
Anything else you want to leave people with before we break?
I think for me, and you kind of see this in the values that we put on our website of what we care about:
You know, that’s what we believe in and in the world that we’re helping — the future that we’re helping bring to life.
Yeah. And I think just to add to that, for me, when I started using Aristotle, it was very different to have an experience where the output’s always correct. And so I think if people haven’t experienced that before, they should just try it out. It’s a free to sign on for.
Cool.
Well, there’s, I’m sure there’ll be plenty of ways to monetize mathematical superintelligence when the time comes. We might do ads, you know.
Yeah. I can’t wait for that.
All right. We’ll do those anthropic ads to life.
Fascinating stuff, guys. I really look forward to watching your progress. Thanks for both the remedial education and a grand vision today. It’s really extraordinary. What a time to be alive.
Vlad Tenev and Tudor Akeem, co-founders of Harmonic. Thank you both for being part of the cognitive revolution.
Thanks for having me. Pleasure to be with you.
If you’re finding value in the show, we’d appreciate it if you’d take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube.
Of course, we always welcome your feedback, guests and topic suggestions, and sponsorship inquiries, either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. The cognitive revolution is part of the Turpentine Network, a network of podcasts, which is now part of a16z where experts talk technology, business, economics, geopolitics, culture, and more.
We’re produced by AI Podcasting. If you’re looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at AI podcast.ing.
And thank you to everyone who listens for being part of the cognitive revolution.
]]>Why would anyone create a new programming language today if AI can already write most of your code?
Andrey Breslav has an interesting answer.
Andrey Breslav is the creator of Kotlin, a language that runs on billions of Android devices and is one of the fastest growing languages in the world. Today we cover how Andrey designed Kotlin by deliberately borrowing ideas from Scala, C Sharp, and Groovy, and why he considers leaving out the ternary operator one of his biggest regrets.
We also discuss why making Kotlin interoperate seamlessly with Java was a gigantic undertaking, and what it took to get it done. Kotlin adoption went through the roof after Google announced it as the official language for Android, in a move that even took Andrey and the Kotlin team by surprise.
Andrey’s new project, CodeSpeak, is a new programming language built on English, designed for an era where AI writes most of the code. If you’re interested in the future of programming languages from someone who built one of the most loved languages of today, then this episode is for you.
This episode is presented by Statsig, the unified platform for flags, analytics, experiments, and more. Check out the show notes to learn more about them and our other season sponsors, Sonar and WorkOS.
Andrey, welcome to the podcast.
Hello.
Thank you for having me.
It is not often that I meet someone who designed such an influential language across mobile and backend. So let’s start with: how did it all start?
Okay, so that was a little messy because I went to school back in St. Petersburg, studied computer science, and I didn’t really know exactly what kind of programmer I wanted to become. I knew I wanted to be a programmer. At some point, while I was still at the university, I started teaching programming in school. It was a big, passionate hobby of mine.
At some point, I got a job with Borland and worked in some developer tools. That was awesome. Borland was a very big name, though they went under pretty soon after I joined. I hope it wasn’t because of me.
I worked at the tail end of the UML era, doing developer tools in the UML space. That was very interesting. I learned a lot. But then Borland went under, and I went back to teaching full-time. Then I started PhD school. All that was kind of not really planned out.
In my PhD, I was working on domain-specific languages (DSLs), and generally, I was interested in languages. I was curious about typed languages specifically. I was always curious about how these things worked, but never really serious. When I started looking into DSLs, it was slightly more serious. Although my PhD was a mess and I never defended because of that.
At some point, someone reached out — he was actually a person who was in charge of Borland’s office in St. Petersburg. By that time, he was already at JetBrains. He reached out to me while I was in Tartu, Estonia, where I was a visiting PhD student for a year. It was a lovely time.
He invited me, during my next visit to St. Petersburg, to visit the JetBrains office and talk about something related to languages.
What I thought was that it was about this project called MPS (Metaprogramming System) that JetBrains had. I knew about it. It’s about DSLs. I worked on DSLs; it was plausible they wanted to talk about something like that.
But I was completely wrong.
What they wanted was to start a new programming language.
I was completely unprepared for that. I had never thought about doing something like this. My first reaction was:
“You don’t do new language. You don’t need it.”
The basic pitch was that the Java ecosystem needs a new language. Java is outdated, so on and so forth. We can talk more about this.
It was 2010, I think. I said, “but there are other languages. Everybody’s doing fine. Why do you need to do that?”
Then this conversation was actually very insightful because the guys at JetBrains explained how things actually were. It was a big problem by that time.
So Java didn’t really evolve and hadn’t been for a long time.
What was the reason behind this? Can you take us back for those of us who are not in the ins and outs?
Yeah. So the last major version of Java by 2010 was Java 5, released in 2004 — a six-year-old language. Since then, there were updates. Java 6 made no changes to the language at all. Java 7 made minor changes. In parallel, other languages — especially C Sharp — were progressing very well. And by 2010, C# had all the nice things. There already were lambdas, like header functions and all that nice stuff. There were getters and setters and many other things that made the language much nicer. And Java was felt like it was standing still. There was a project to work on lambdas for Java, but that was in the works and had been in the works for a long time and only came out in 2014. So that was the situation.
And, you know, the ecosystem didn’t stand still in the sense that other people were building languages. And there was Scala, there was Groovy. And, of course, people at JetBrains knew both Scala and Groovy. They built tools for them.
It’s traditional to build your tools in the language you’re building the tools for. So the Scala plugin was built in Scala. And there was a lot of Groovy used in JetBrains as well. So they knew what the issues were with the language. And both languages are very interesting and very good in their own ways.
But they saw an opportunity in the market because basically Groovy was too dynamic and too far from, you know, hardcore, mainstream, large-scale production. Because dynamic languages are not for that, basically.
What are dynamic languages for? What are their strengths and best use cases? The trade-off, I guess, if you look at a statically-typed language like Java, Kotlin, and Scala, for example, versus dynamic languages like Python, Ruby, JavaScript, and Groovy:
And this may be changing nowadays a little bit. And this is in part what I’m working on now. But back in the day, it was completely true. The whole art of making a good language was to restrict the user in a good way.
Yeah, but in any case, the situation with dynamic languages is that they are much more user-friendly in the beginning. But then when the project scales, you’ll have trouble making large refactorings. You have trouble making sure that everything works together. You need to do a lot more testing and rely on other things like that.
As opposed to static languages where you have precise refactoring tools and other things that can make sure that at least a certain class of problems just doesn’t happen. And, you know, this is why, at least in our mind back then, it was absolutely clear that if we’re building a language for large projects, big teams, so on and so forth, it has to be a static one.
So with Groovy, that was a big issue of performance as well, because Groovy was building a dynamic language on top of a very static runtime. So there was quite a bit of tension there.
That wasn’t the Groovy side and the Scala side. Scala is a wonderful static language and incredibly powerful and with tons and tons of good ideas. But it had its own problems. It relied very heavily on implicits, for example. And I have a history of debugging one line of Scala for an hour to try and figure out what it does. Just because it was pretty complicated.
Also, the compiler was very slow and there were issues of stability, and many, many things were just not accessible enough for a lot of engineers. So from the experience of using Scala, JetBrains, my colleagues basically understood that it’s not what’s going to change the industry. Although Scala got a lot of adoption.
And again, like Martin Odersky, he is a great language designer. And I think one of the biggest use cases was old Twitter. A lot of it was built on Scala and they scaled to massive scale, etc. And I think LinkedIn as well.
So in any case, these were, you know, it’s always very nice when other languages kind of pioneer things. And then you can build on top of their successes and failures. And we were in that position, basically.
So the argument that people at JetBrains were making was basically that there is a window of opportunity. People need this language. We, JetBrains, are the company who can actually put out a language and make it successful because:
- We have access to the users.
- We have their trust.
- We can make good tools.
And it was another issue with Scala, for example. It was very difficult to build tools for Scala back then. Now Scala 3 is more tooling-friendly, but back then it was a nightmare.
Like, I said that, you know, if you have a static language, you can’t have precise refactorings if the language is too complex. And some languages are particularly challenging. So Scala back then and C++ were incredibly challenging to make precise tools for.
So, and that was the basic pitch. And I quickly understood that, yeah, they were right. And this was something that was worth a shot in the sense that it was not completely hopeless, not completely dead in the water. I had no idea if we could pull it off.
It’s, it was then when we actually sketched some initial features on the whiteboard.
Just because JetBrains is genuinely run by engineers? Hold that thought from Android on how JetBrains is genuinely run by engineers. This is because I happen to know another company also run by engineers: Sonar, our seasoned sponsor.
If there’s a time when we need true engineers, it’s now. As AI coding assistants change how we build software, code is generated faster than before. But engineering basics remain important. We still need to verify all this new AI-generated code for quality, security, reliability, and maintainability.
A question that is tricky to answer:
How do we get the speed of AI without inheriting a mountain of risk?
Sonar, the makers of SonarQube, has a really clear way of framing this:
Vibe, then verify.
The vibe part is about giving your teams the freedom to use these AI tools to innovate and build quickly. The verify part is the essential automated guardrail. It’s the independent verification that checks all code, human- and AI-generated, against your quality and security standards.
Helping developers and organizational leaders get the most out of AI, while still keeping quality, security, and maintainability high, is one of the main themes of the upcoming Sonar Summit.
It’s not just a user conference. It’s where devs, platform engineers, and engineering leaders are coming together to share practical strategies for this new era. I’m excited to share that I’ll be speaking there as well.
If you’re trying to figure out how to adopt AI without sacrificing code quality, join us at the Sonar Summit. To see the agenda and register for the event on March the 3rd, head to:
sonarsource.com/pragmatic/sonarsummit
So everybody I talked with was deeply in the weeds with IDEs and everything in new programming languages very well. We had a very technical discussion.
I don’t remember exactly all of the features we were talking about, but the current syntax for extensions in Kotlin was already there. I don’t remember why exactly we focused on extensions, but it was there.
So, from day one, we’re basically building on top of ideas from other languages, like extensions obviously came from C#.
Yeah, so it was a very exciting conversation, but I didn’t make a decision then because I was in Tartu and I needed to finish there. It took me a few months to finish.
Then I came to St. Petersburg for one month because after that I had an internship scheduled with Microsoft Research in Redmond. I was going to Seattle to stay there for about three and a half months.
I said, “Okay, guys, I have this month. I can work in the office and we can try to sketch things, but then I’ll go into Microsoft and then I will decide whether I commit or not.” Which in hindsight, I made the right decision in the end.
I had a great time for this month or so. I worked with the guys in the office — it was mostly Max Shafirov we were working with and it was incredible. We had such great discussions and I actually saw Max this morning and it was like, it was great time.
So then I went to Seattle, did something completely different. There are Microsoft researchers, some really great researchers working there, actually was exposed to the top notch level of academia for the first time — was very insightful.
But after that, I kind of realized what the question was: whether I want to try to pursue an academic career, which I didn’t feel like I was really built for and was not sure whether I can be a good researcher on my own or I’ll have to follow in somebody else’s footsteps.
So for those of us engineers, which will be the majority who have not built a language from scratch, how do you start with it? Like, speaking for myself, I know how to:
How does a language start?
In our case, we basically talked a lot for a few months. I think not everyone is like that, but I think the best when I’m talking to people.
This was the ideal environment because we were basically discussing things with the Macs constantly for many months. There were a few internal presentations that I made at JetBrains and some of the slides survived.
I can see, including my spelling mistakes in the slides — my English wasn’t as good then — and you can see some of the evolution through those slides. I think there’s a recording of one of those presentations.
So we were basically doing whiteboard design for some time. And the great thing about doing this at JetBrains was that there were a lot of people with opinions about not so much how to make a language, but what problems do programmers face and what they like and don’t like in other languages. So I had tons and tons of input from other people and very good people. So that helped. And I really, I don’t think I realized how special that environment was back then. Like I was 26, to be clear. And I had no idea how things were done in general. But somehow these people just trusted me. I’m not sure it was very rational on their part. It worked out. But I’m not sure I would recommend anyone to do this.
And so in the first few months, I understand that you kind of whiteboarded and wrote down how you want this language to evolve. You kind of, you know, like wrote out like,
“We’re going to have these features. Or how can we imagine?”
So I guess the easiest way to explain this would be like this. It basically went off what the pains were with Java. And there were quite a few. And there was a lot of experience of using Java across the community and inside JetBrains. And we kept making lists of things we wanted to fix.
I came up with some ideas and some other people suggested other ideas about how things can be fixed, what is an actual problem, and what we don’t care about, and so on and so forth. For some time, I was just, you know, pieces of the puzzle basically laid out on a table without fitting together. And then at some point, we started fitting them together. I was just doing a lot of that in my head, which is not the best way. But this is how I knew how to do it.
There were also some crazy ideas that we thought were important back then. For example, I wanted to implement multiple inheritance, fully-fledged multiple inheritance, which was a dumb idea. And multiple inheritance meaning that a class can inherit from like several classes, and you have to take care of like conflict resolution and all sorts of edge cases. Right? Yeah.
The actual challenge is not so much conflict resolution in terms of methods, but initialization of state. Constructors are really hard. And it was actually someone outside of Gibbons who explained to me that was a very bad idea. And I’m very grateful to them. Yeah. So, you know, there were crazy ideas as well. And some of them just fall off over time as we were discussing or prototyping.
I think I started writing code maybe six months in or something like that. Maybe a little earlier than that. I started with a parser. And it was actually a very unique way to start a language because the idea was to start not with a compiler, but with an IDE plugin. I have it in the editor first, which is, you know, an IDE plugin shares a lot with the front end of the compiler, so it’s not absolutely crazy. But I was just relying a lot on the infrastructure that was available in IntelliJ IDEA.
All the parsing infrastructure, and it was awesome. Like, the parsing infrastructure in IntelliJ IDEA is better than anything else in the world because it’s the heart of the IDE. It has to be incredibly fast and very robust and so on and so forth. But then later, someone who knew the infrastructure a lot better than I do had to factor that bit out to make the Kotlin compiler autonomous. And it was Dmitry Zemirov who did that. And he’s an awesome engineer. Like, he’s probably one of the best people to refactor a large code base and, like, take this one bit out of something that was already 10 plus years old back then.
So we started with this IDE plugin. I think Max wrote the scaffolds and I actually plugged in the parser and everything. And that was an interesting start because it was very interactive. So I could show off the language as if it existed because it had some tooling. But I couldn’t compile anything in the very beginning. And that was actually a very good way to experiment with the syntax.
But then soon after, I started working on a full-fledged front-end and on some translation. And Dmitry and Alex Kachman were working on the back-end. Everybody was part-time.
When you say you work on front-end, and they work on back-end, in a language context, what does that mean?
It’s slightly different in different languages.
Basically, the front-end is what deals with the:
And the back-end is what translates to the executable code.
In our case:
Front-end:
- reading the text
- parsing
- doing types
- all that
Back-end:
- generates Java bytecode
And Kotlin has multiple back-ends for different target languages:
At that time, nobody was full-time working on this project. Even I was part-time, a PhD student, part-time Kotlin developer. And it was the very early days.
Then, at some point, I gave up my PhD and focused 100%. Which was also, like, isn’t it a weird decision to start a new language part-time? Yeah. Looking back, I was young and stupid.
There’s a saying that we didn’t do it because it was easy. We did it because we thought it was easy. Absolutely that. I didn’t realize how hard the problem was. I also had an unreasonable amount of hubris. I just thought I knew how to do everything. I didn’t. But it worked out in the end.
So, when the language started, what did you call it internally? There’s always internal code names, right? Right, yeah.
So, I don’t think there was a discussion of this first name at all. It was generally understood that the language will be named Jet. And it was logical. We had all the code base using the name Jet. We had:
Then someone realized that the name was trademarked by someone else. It was actually people we know there in Novosibirsk in Russia doing something. It’s not a language, but it was a compiler, and we couldn’t use it.
This is when we started looking for another name. It was very painful — looking for names. Guys, this is so bad. It’s one of the worst things because you never know what name will work unless you want to do an extensive study.
And then all the good names are taken, of course. Then some of the names that are not taken are not taken because they’re not really Google-able.
Some people are just very brave. People who named their language Go. This is why people now call it Golang because otherwise you can’t identify it. It’s a verb in English, a very common word.
Yeah, so we had weird options. In one of my old presentations, I found a list of early names:
And those weren’t great.
By that time, other languages were popping up. One of the alternative languages was called Ceylon. The logic was: Java was the island of coffee. And Ceylon was an island of tea.
Dmitry Jemerov basically looked out of the window and said,
“OK, we have an island here in St. Petersburg. In the Gulf of Finland, there’s a big island called Kotlin.”
And it’s a good name in the sense that it’s very Google-able. Nobody uses it for anything. It’s very recognizable. It’s not super smooth for many languages, but it’s kind of OK.
Nobody was in love with that name and we were kind of hesitant.
You know, “Kot” means a bad thing in German. Also, there is like some negative connotation in Mandarin, I was told, or something like that. You know, it’s always some language has some nasty association with any word.
We basically were super hesitant. So when we announced, and we had this deadline, that we were basically putting this off, when we announced, we were still not sure.
So we called it, we decided it would be a code name. We called it Project Kotlin to have wiggle room to later replace the name — but it stuck.
The first thing we did was put out basically a Confluence page with a description of the language. It was just a bunch of wiki pages and there was no compiler available then, I think.
There, the word Kotlin appeared many, many times. I was like,
“My God, this thing doesn’t, like, I can’t do search and replace and then change the name everywhere.”
So the workaround that I came up with was to create an empty page called Kotlin. And so it had a name. And then everywhere else, you mention it as a page. When you rename a page, it gets renamed everywhere.
This is why there was an empty page called Kotlin in that documentation. But yeah, the name stuck and it turns out to be not a bad name.
So, when it started, what were the main differences with Kotlin compared to Java? Because Java was, what was the big one? How did you explain to developers who initially started onboard or wanted to give it a go?
Yeah, I guess there were a few major selling points. Then there were other things on top of that. When we started, like in the very beginning, we didn’t have null safety in mind. Null safety came a little later.
After one of the internal presentations, it was Max Shafirov who invited Roman Elizarov, who later was the project lead for Kotlin. Roman came and listened to the presentation, gave some feedback, and said something like,
“Guys, if you want to do something really big for enterprise developers, figure out null safety.”
And we did. It took a while.
So in the very beginning, it was the general idea of what makes Java feel so outdated. There were a bunch of things. Lambdas were very big. The general, like, the general feeling from Java back then was it was very verbose. It was called the ceremony language. A lot of people were grumpy about too many keywords, like public static void main is something everybody was really grumpy about.
But also, there were getters and setters for every property. There were constructors and overloads and all that stuff that looks like boilerplate because it is. Yeah. It’s super annoying to type out.
The problem with boilerplate is, on the one hand, it’s annoying to type out. But tools can generate it for you and fold it and so on and so forth. But the bigger problem is always readability. So reading is more important. Reading code is more important than writing code. We do a lot more of that.
And with boilerplate, it’s terrible because if some tiny thing is different in the middle of completely standard boilerplate code, you’ll miss it. You’ll become blind to it and you can debug for days not seeing that. So, you know, that was the point of sort of modernizing Java, making Java programs be more about what they do and less about the ceremony of making the compiler happy, basically.
And, you know, type inference was also a big thing because Java was repeating types a lot and many other things like that were, like, semicolons. The modern languages of the time already got rid of semicolons. And so in Kotlin you also got rid of it?
Yeah. So we got rid, basically, in terms of syntax, we got rid of semicolons and duplicated types. And that was a lot of noise across the code.
What does it mean that Java had duplicated types?
So in that version of Java, when you declare, say, a local variable, you say it’s a list of string called strings equals new array list of string.
Oh, yes. I remember this one.
Yes, yes. You need to type it out twice. And if you get one of them wrong, compiler, et cetera.
Right. So, and at best, you could omit the second mention of string by using a diamond operator, but that only came later, you know. Basically, it was very verbose, especially if your types are long.
So, and a bunch of things like that were really annoying to a lot of people, especially compared to C# or Scala.
So, we did all of that. And then, on top of that, there were other value-add features and null safety was a big thing that we spent multiple years actually on implementing. And I think it’s one of the main differentiating factors now for Kotlin alongside of with extensions and other things. But null safety is one of the core features.
And can we just spell out why null safety is so big?
I mean, I just today I came across a bug on the, I couldn’t send a package because in JavaScript on the Dutch post website, there’s a null issue happening in production.
But, you know, before Kotlin and a lot of languages, why is it such a big problem?
It is.
Yeah. So, dealing with null references is a big hassle in most languages. And I think it was Tony Hoare who called it the “billion-dollar mistake” at some point because, like, introducing, I think it was about introducing null pointers to C or something.
So, basically, when we look at all the runtime errors that we have in Java code, I think null pointer exceptions will be at the top. So, you know, the type system of the language is supposed to protect you from those unexpected errors.
So, there are errors you’re designed for and maybe errors that are not even your fault, like a file system error or something like that. But there are also errors that should be prevented by the compiler. So, for example, class cast exception or missing method error are things that the compiler is trying to protect you for. It’s trying to make sure that this never happens in your program unless you switch off the check by making an enforced cast or something.
And with nulls, it’s not a thing in Java. Like, anything can be null, and if it’s null, it will just fail. Yeah. It throws an exception and the program dies. So, it’s a very common thing.
So, a lot of people are kind of used to it, and there are different ways of being disciplined about it and so on and so forth. But, basically, this is a plague across any code. You know, there are different approaches to this.
And in Kotlin, we took the approach of:
- A: enforcing it in the type system,
- but also making it free at runtime.
What does that mean, that you made it free?
So, one very common way of dealing with nulls is to use something like an option type, where you have a box, which might be empty, or might have an object in it.
No. And that box is not free. Like, you have to allocate it, you have to carry it around everywhere. And, this easily creates a lot of objects in the old generation for the garbage collector, so it can be challenging. What we did was just have a direct reference at runtime; our nullable or not null reference is the same as Java’s reference.
All we do is compile-time checking and some runtime checking when we cross the boundary. But that’s a lot cheaper than allocating objects. Although the runtime is getting better, and they can optimize some of those objects away, it’s still an overhead.
What are features that you took in from Kotlin that were inspired by other languages that you admired?
A lot of them. I have an entire talk about this. It’s called Shoulders of Giants. We really learned from lots and lots of languages. And it was always the point. Andre just mentioned how Kotlin was built on top of the shoulders of giants, taking good ideas that existed, not reinventing them. This was one of the reasons Kotlin succeeded as much as it did.
But jumping forward from 2010 to 2026, one thing that is totally different today is the speed of things. AI is allowing nimble teams to build faster than ever before. Companies that used to take years to move into the enterprise are doing it in months.
This speed creates a new problem:
These show up almost immediately. This is where WorkOS, our seasoned sponsor, comes in.
WorkOS is the infrastructure layer that helps AI companies handle that complexity without slowing down.
Features include:
- SSO for enterprise buyers
- MCP offer agenda workflows
- Protection against free trial abuse with Radar
Teams like OpenAI, Cursor, Perplexity, and Vercel rely on WorkOS to power identity and security as they scale. If you’re building AI software and want to move fast and meet enterprise expectations, check out WorkOS.com.
With this, let’s get back to Andre and how Kotlin was standing on the shoulders of giants.
So the slogan for Kotlin was “pragmatic language for industry.” The pragmatic bit, which is a nice rhyme with your podcast, was kind of coming from the experience with Scala being called an academic language. A lot of people had trouble getting their heads around many of the very smart tricks in the design.
And so our idea was:
“We’re not doing academic research here. We’re not trying to invent anything. If we don’t get to invent anything, it’s a good thing, not a bad thing.”
From the engineering perspective, it’s generally a good idea to do this. Usually, you end up making something new, but most of what you’re doing shouldn’t be very new because you want familiarity. You want people to easily grasp what you’re doing. This has to be familiar from other languages.
Also, if you’re building on top of the ideas of other languages, you benefit from them having tried it already. You can look at their designs, their community’s reactions, and the implications all over the place. That gives you a huge benefit.
So we did a lot of that.
I think the language that influenced Kotlin the most is, of course, Java. Because the entire runtime of Kotlin is the JVM, and we depend on that.
Apart from that, Scala had a huge influence. We used many ideas from Scala, including:
vals and varsIt’s a huge pity that this didn’t make it into Java design. It was flipped at the very end of the design process to what Java has now. The Martin Odersky idea was much better.
We had to fix this problem on the Java boundary and figure that out.
There were many ideas we took from Scala, and that was very helpful. We usually transformed those ideas a little bit to adapt to our setting and to build on the knowledge of how it actually works in practice. We left some things out. We simplified some things.
For example, Scala had traits. Traits are a very powerful construct, like an interface where:
What you couldn’t have were constructor arguments. You always have a default constructor and can initialize all your fields.
It’s not as bad as multiple inheritance in C++, but it’s still a little complicated when it comes to the order of calling constructors. We decided we don’t want to deal with that. It’s a complex algorithm and hard to explain. Let’s just get rid of the state in interfaces and only have method bodies. And I think it was a good compromise. Especially given that Java ended up in the same place. It was easier to integrate.
Yeah, so Scala was a big influence. C Sharp was a very big influence. Extensions, of course. And we learned quite a lot from how C Sharp compilers do things.
There, there was also one particular trick that makes Kotlin syntax a lot nicer, nicer than Java’s and nicer than Scala’s, that we’ll learn from C Sharp. And it was actually my colleague who worked on the C Sharp IDE who told me about this, which is basically a super pragmatic thing they do in C Sharp.
There is like, when you call generic functions, you use angle brackets inside an expression. But the thing is that there is no such thing as angle brackets. There is less and greater. Right? And the parser can easily get confused and think that this expression, since we’re not in a type context, it’s an expression context. This expression is a comparison. It’s not an inequality, right? It’s not a call. And this is mathematically unresolvable. It’s an ambiguous grammar.
Yeah, look, you can do anything about it. And the way other languages handle this is:
collections.<Type>functionName().And we did the same or something very similar, and it just works. And the syntax is very familiar and very intuitive, and we’re very happy about that.
Yeah, because when you read it, as a person, I never get confused. Like, this is not a smaller sign. Like, I know it’s a generic. Yeah. Yeah.
Okay. Wow. Most of the time, it’s not a practical problem. And there is a way to disambiguate, if you like. So C Sharp was a big influence.
Groovy was a big influence as well. JetBrains used Groovy for build scripts. And there were incredibly useful patterns in the Groovy syntax that they call builders, which is not about building programs, but, you know, building objects.
And this is what inspired something fairly novel that we did in Kotlin, which was typed builders, where we had the same syntactic flexibility, or almost the same syntactic flexibility, as Groovy, but it was all typed. And we could make sure that all the arguments matched and so on and so forth.
So all that side basically was inspired by how Groovy people did this and reworked into a typed setting. And this is why we have, for example, extension function types. And this is why we have dangling lambdas and other things that are actually very nice syntactic constructs.
So, yeah, many, many things came from different languages.
A less known language called Gosu, I think it was what inspired us to do smart casts.
What are smart casts? Oh, yeah. So, I think smart casts are one of the nicest things a compiler can do to a developer. Because it’s a very common situation when you say:
If x is a string (so you do an instanceof check), then do something with x.
The annoying thing is that in a lot of languages, you have to cast x to string again. Like, you’ve done the check. After you’ve done the if, you know it’s a string, but then you need to write it out again.
Yeah, so you’ve just done the check, but you have to say string again to make the compiler happy.
So, smart casts basically get rid of that. So, that cast gets figured out automatically. So, if that’s a string and then inside the bracket, you can now use it because it’s a string. Yeah, you can use it as a string.
And isn’t it an easy thing, right? So nice. Yeah, it’s a very nice thing.
Yeah, it’s a pretty complicated algorithm. Because, you know, variables can change values and the check that you’ve just made can go stale. And, you know, there’s a bunch of algorithmic trickery around this.
And you can’t do a smart cast on any expression. It has to be a certain type of expression that can be stable enough and so on and so forth. But, you know, it’s a very nice thing. And you can get rid of so much noise in the code because, like, all the code in the world is riddled with this instanceof cast. instanceof cast.
So, we wanted to get rid of that. And it worked. And it was fun to implement.
What were things that you looked at other languages, you considered, maybe we should bring it in. But you, after debate, you’re like:
“No, let’s just leave this out.”
Like, not all of them, obviously, but some of the big ones that kind of came close. We had a design for pattern matching in Kotlin that was inspired by functional languages like Scala and Haskell and others. But at some point, early on when I was still working on the parser, I just realized that this is a huge feature.
So, when I was sketching it out on a piece of paper, it looked like a very useful thing, just another feature in the language. But then when I started working on the parser, I realized it’s an entire language in size. Like, you have to create a parallel universe in syntax for pattern matching. And I was like, okay, this will be a lot of work. Let’s postpone it.
Later on, when we were doing review for 1.0 or maybe a little earlier than that, I just realized that smart casts plus we have something called destructuring together give us like 80% of all the good things pattern matching can do to normal developers. Then there is another group of developers that can be very vocal, mostly compiler developers and people super into functional programming. They have a point, but that point is only relevant to them, and there are not very many, so we decided to not have pattern matching back then.
And, you know, maybe there comes a day that pattern matching gets added to Kotlin. And pattern matching is, is it in the case? Yeah, it’s the… So you can have, like, a lot nicer case statements, a lot more expressive ones, right? Yeah.
Generally, Kotlin has this compromise where you have our version of switch case, which is called when, and you can have smart casts there. So you can say:
That kind of gives you a lot of the niceties of pattern matching, but some things you can’t express like that. And that was, I think, a good compromise because it’s a really big feature. It’s hard to design well. There would be a lot of work on the tooling side. But maybe it gets in the roadmap one day. I’m not sure.
Java is trying to get towards pattern matching, so we’ll see. Maybe they kind of make it more mainstream.
Why did you omit the infamous ternary operator, which is when you write out something with the question mark and the colon, and it confuses new developers every single time if you’ve not seen it before? Yeah. Was it for readable reasons?
This is the saddest story I think in the design of Kotlin. I didn’t realize how much people liked it. The reason was, Kotlin used this principle from functional languages that everything we can make an expression is an expression. So if is not a statement, and the ternary operator is sort of a patch on the design of C and other C-like languages that makes an if expression, basically.
The logic was:
okay, we have if as an expression already,
can we just get rid of this extra syntax construct,
especially given that it's using very precious characters?
Like, there is a question mark and a colon, and we might find some other use for that. So we decided to not have it. We used question marks for nullable things and the colons for types and so on.
But it turned out that if as an expression is pretty verbose; people don’t like it. I resisted for some time, and then by the time I agreed, it was too late because you can’t retrofit the ternary operator into the current syntax in Kotlin—it just doesn’t agree with how other operators have been designed.
So you’re actually sad about it not being there a little bit? Yeah, I think in retrospect, it was a mistake because pragmatically, it’s more use than harm to have it. But we just can’t retrofit it.
What are some other interesting features that you like about the language that you added that we could explain for those who are not familiar?
Okay, so the good ones, there’s quite a lot of them. One feature that is not a traditional kind of language feature is Java interoperability. That’s probably the single thing we spent the most time on. And I always say that if someone offers you a job to create a system that interoperates transparently with another huge system you don’t control, ask for a lot of money. It’s a very tricky deal to figure this out.
Interoperability means that from Kotlin, you can invoke Java, and from Java, you can invoke Kotlin. You do a bunch of work there, but it just works in the end as a developer. You don’t need to think about it.
The idea is whenever you have a Java library somewhere in the world, you can always use it from Kotlin. It was a big selling point because if you start as just a language in a vacuum and you don’t have any libraries, that’s not a good start.
In this direction, definitely, it was an absolute requirement for Kotlin. But also, we had the requirement to go the other direction. In an existing project, you could just rewrite parts of your code from Java to Kotlin, and everything keeps working. And some libraries actually did that. Many projects started using Kotlin bit by bit.
A lot of people started with just writing tests. But then, you start adding things in Kotlin, new things, for example. And all the Java code around that has to transparently use the Kotlin code. So we put a lot of effort into that. And that was fun.
Can you explain to us as engineers, like, it sounds like it was a friggin’ big project. What is the work, right? Because from the outside, again, I’m just being your average developer, where I’m invoking a Java class.
And things I can think of are:
What is hard? Tell me, tell me. I’m dying to know.
So one thing to note here is that we don’t control the Java compiler. We somehow need to make it work so that in your Java code, you make a call into something that only exists in the Kotlin source. And the Java compiler somehow agrees to call it to begin with. It’s not a Java file. It doesn’t know it exists.
So the way it actually works is: when we build a mixed project, what we do is we first compile all the Kotlin code. That can depend on the Java sources in the project. So we have a Java frontend baked into the Kotlin compiler so we can resolve everything in the Java code. Then we produce class files, binaries for the JVM that the Java compiler can read. So when Java compiles, it takes Kotlin sources as binaries. And this is how it works.
We would have to implement a Java compiler otherwise. Fortunately, Java has separate compilation, so this works.
This trick means that whenever you have tooling, like in your IDE, for example, when you navigate from Java sources to Kotlin sources, it has to be a special trick. Someone needs to go and teach the Java world to know about the Kotlin world.
Of course, the IDE doesn’t do the compilation to navigate. But at compilation time, we don’t control the compiler. So we did our own IDE. This way, we could do something about the Java tooling, but we couldn’t do anything about the Java compiler. So that’s trick number one.
Then, when it comes to incremental compilation, it becomes even funnier because Java incremental compilation is a complex algorithm on its own. Now we are incrementally compiling two languages at once. And that’s fun.
Incremental compilation algorithms are generally a very messy, very complicated heuristic with tons of corner cases. So, that’s like one example.
But then you start making interesting new things in Kotlin. You need to expose them to Java. You need to make sure that whatever fancy thing you have, Java can actually interoperate with that.
One example would be Kotlin’s approach to making Java collections nicer in Kotlin without rewriting the collections using the same library. Java collections are what’s called invariant because they’re all read-write. So if you have a list, it always has a set method.
That’s a little bit of a problem because whenever you have a list of objects, you cannot assign a list of strings to that. That’s annoying because you want to be able to represent a list of anything, and for that, you need to play with question marks, wildcards, and stuff like that.
It would be very nice if we had a read-only list interface that doesn’t have a set method. Then there is no problem in assigning a list of subclasses to a list of superclasses. But this interface doesn’t exist at runtime, right? We can’t just invent it. Or can we?
So we actually can. No.
In the Kotlin compiler, we have this layer of trickery specifically for Java collections where Kotlin always sees Java collections. If they come from the Java world, they are read-write, mutable collections, we call them. But mutable, right? Yeah.
So the Java collections are always mutable or platform mutable. I’ll talk about that later. But when you do it in Kotlin, you can actually distinguish between read-only and mutable collections, and it’s all very nice on the Kotlin side.
But then when Java sees the Kotlin collections, they are normal again. When we expose them through binaries, the Java world always sees them as normal collections; they’re mutable for Java, and it’s all right.
Okay, I’m starting to see why you said you need a lot of money for this because this is just one of many things. But this itself sounds like, I don’t know how you solve that.
Yeah, so just to add a little bit of detail to this. So the nice thing about those read-only collections is that you can pass a list of string for a list of object, right?
Wouldn’t it be nice if a Kotlin method that takes a list of any could accept a list of string in Java? But aren’t we erasing all the Kotlin nice stuff? We are, but we know that this list is actually what’s called covariant. So we can expose it to Java as a list of question mark extends and not just list of objects. So, you know, it becomes covariant for the Java world as well. And that’s like one hack that makes it a little more transparent.
And there’s a bunch of that. So, you know, so that’s another thing that we had to play with. But the biggest thing is, of course, nullable types. And actually, we handle nullable types and these things with collections kind of similarly, which makes the whole typing layer of the interop quite interesting.
But basically, so Java doesn’t know anything about nulls, right? Well, it knows about nulls, but not about nullable types. It does not exist. Yeah, Java doesn’t know about nulls at compile time. So in terms of types, it’s just not represented. So technically, every Java type is a nullable type.
And this is where we started. We said, okay, so Kotlin types can be not null and it’s very convenient. And when you have a not null type, you can just call a method on it normally, right? But if something is nullable, you can’t just dereference it. You have to first check for null and then use it, right? Or if there is a safe call operator, well, just propagate null is on the left-hand side.
So we started with saying,
“Okay, all Java types are nullable, which is a conservative, like very mathematical way of treating it.”
This is correct, right? Yeah, you’re not going to be wrong with that. Yeah. And we implemented that and we started using it inside JetBrains. And the feedback was horrible. Like your code is plagued with those null checks and you know that they shouldn’t be there because you can’t express anything on the Java side the right way.
And there were like, we had some annotations for the Java side. It was also brittle and not always worked because, you know, there can be long chains and stuff. And some libraries just don’t have the annotations. And we struggled with that for a long time.
And basically we realized that this assumption that everything in Java has to be treated as nullable just doesn’t work. This was a turning point where we sat down and reimagined the whole thing.
And we worked with a great type theory type practice, I would say, guy from, I think it was back then he was in Cornell, Ross Tate. So Ross helped me figure out the sort of mathematical side of how you can represent those types that come from Java and should be, like we should be aware of that they are from Java and can possibly be nullable.
But we shouldn’t treat them as nullable because it was very inconvenient. And Ross put together a very nice sort of calculus about those.
And when we started implementing it, all the nice things are gone. The mathematical beauty is completely gone from all that. And I think we took the general idea of sort of splitting a type in two and everything else is just very messy industrial kind of thing. That’s not sound, but it works well.
Okay. And interoperatively sounds like it was a journey, but a necessary one.
How long did it take? Can you give me just a sense of like how many people working on it? How much, because I think in traditional projects we can get a sense, but I have no idea with the language. How does this work? And how long did you think it would take versus how much it took?
Yeah. So let’s start with that.
Yeah. So, so I had no idea. And I always said like, okay, a year from now feels far enough. We’ll probably be done by then.
In practice, we started in 2010, yeah, autumn of 2010, basically. And we released in February 2016. So, you know, it was a long time, five-ish years. And that, you know, in part was just because I didn’t know how to manage projects.
And my initial team, the people who worked full-time on the project, I looked up on GitHub to verify that. Everybody who, almost everybody, who joined JetBrains to work on Kotlin was a fresh graduate. Because I used to teach and I had some good students and I knew how to work with students. And so basically everybody on the team was a student, apart from a few veterans from JetBrains who were helping, not all of them even full-time.
So we started getting experienced engineers on the team a bit later. And, you know, to be fair, a lot of those people, people who are following Kotlin know those names. People who are core contributors, who built out, like, absolutely foundational parts of Kotlin, joined as fresh graduates. And they became great engineers.
But I think I overdid it a little bit. So it’s great to have, you know, younger people have no fear. And that’s wonderful. But, you know, the balance was not right.
And how big was the team initially and then towards the release?
So we started out basically with four people part-time. And, yeah, we went like that for maybe a year or something. So the initial prototype was built like that. And then people started joining in. By the time we released, I think it was around 25 people or something.
And the team grew quite a bit. So by the time I left in 2020, it was about 100 people on the team, 70 of them engineers. So it became a pretty big undertaking.
Can you tell us about the development process inside language?
I think a lot of us are used to building, you know, like services, backend services or products or mobile apps, etc. They typically have a release process. How does this work inside a language? Like, what is your release process and what is the, I guess, best practices?
Like, do you even do code reviews or, you know, like how can we imagine? Because, again, it feels such a rare project. There are people building languages, but not many of them.
Yeah, so one peculiar thing about building languages is what’s called bootstrapping when you write your compiler in your language.
Oh, nice.
Which means that, you know, to compile your code, you need a previous version of your compiler. And you better agree with your colleagues which version it is. It can be really tricky, especially when you do things about the binary format. And there is, like, quite a lot of bootstrapping magic going on.
And I don’t think you can reproduce the Kotlin builds from scratch. Because, you know, if you just take a snapshot of the Kotlin repo, you can only build that with a Kotlin compiler. And I don’t think we kept all the bootstrapped versions. So it might not be really possible without a lot of manual intervention to rebuild all the sources from the very beginning and reproduce all the versions.
Because sometimes, you know, we had to, like, commit a hack into a branch and use that branch as a bootstrap compiler for the next build and then throw the branch away. So that was, like, a one-off compiler used to facilitate some change in the binary format or syntax or something. So that’s a separate kind of fun.
But generally, I mean, many practices are very similar. Like, we had code reviews pretty early on. It’s my personal quirk, again, that I like to talk to people. So in code reviews, I often just sat together with someone and either they reviewed my code or I reviewed theirs. But this is, you know, I can’t argue that it’s much better or worse. It’s just how I prefer it because I like talking to people.
So code reviews, yes. And, of course, we had an issue tracker like everybody else. Ours was always open. So everybody can submit bugs to the Kotlin bug tracker, which was very helpful. It’s hard to manage because there will be, like, with usage, there will be a lot of bugs and a lot of feature requests and all kinds of stuff. But it’s worth it. You have a communication channel.
Release cadence is a very difficult thing to figure out for such projects. Because one big consideration you have for languages is backwards compatibility.
In part, this is what delayed 1.0 because we wanted to be reasonably sure we can maintain compatibility as soon as we call it 1.0. In part, because it was the expectation, especially Java is incredibly stable and very good with that until Java 9 came about. And also, Scala had a lot of trouble because they were breaking compatibility a lot. And the community was struggling, really. So we really didn’t want to repeat that.
But, you know, it turns out you can even break compatibility Python 2 to Python 3 and survive.
Barely. Barely survive.
They’re doing very well. Now they’re doing well, yes.
Yeah.
So we were really serious about that. But basically what it means is you start doing interesting things like deprecation cycles. So we actually invented an entire tool set for compatibility management.
So before 1.0, we tried to help people migrate. So we had those milestone builds. Embarrassingly, we had 13 of those.
And, you know, when we broke the language in major ways, we tried to provide tools for automatic migration.
That’s nice of you.
Which was, I don’t think, a standard practice in the industry back then. Now people are doing it more. So I’m very happy to have sort of popularized this idea. And then when we were preparing for 1.0, we did a major review of everything and took a year to sort of review all the design.
What we’re doing is basically trying to anticipate what changes we might want to make or what new features will require. And to basically prohibit things that might block that. So we tried to make sure that the changes we were planning were guarded well by compiler errors to make sure that users don’t accidentally write anything that looks like a new feature. And that was fine.
We had design meetings, I think, every day at some point—basically working on that, like, “okay, let’s outlaw this. Let’s prohibit that.” And we prohibited a lot of stuff correctly and some stuff incorrectly. But, you know, generally worked out. So this compatibility thing was a big deal.
But there’s also a lot of stuff that we didn’t anticipate. So we had to figure out ways to manage this. And there is something in Kotlin compiler called “message from the future,” which is basically when in a newer version of a compiler, you introduce something that the old compiler doesn’t understand.
We have different options. And one option a lot of languages go for is:
But it’s a little hard for people then to manage their versions because new libraries, new versions of libraries come with new compiler expectations and you have to migrate your entire project to do that. It’s a little annoying. And if what you’re adding is like one method that basically invalidates the whole library for an old compiler, that’s not great.
So what we’re doing is a newer compiler can write something into the binary that tells the old compiler, “okay, this method is what you can’t understand, but everything else is fine.”
Wow, that’s smart.
Yeah.
So we call this a message from the future and like it can provide some details. So there’s that.
And there’s also the discipline of experimental features, which is incredibly helpful. And I am very happy to see other languages doing it now. And even Java does experimental features now, which is wonderful.
Andrei just talked about experimental features in programming languages and how that used to be rare back in the 2010s. What this reminded me is that running experiments in production used to also be rare. Not because teams did not want to do it, but because doing it meant building a lot of internal tooling around it:
Assignment, rollouts, measurements, dashboard, debugging, the whole thing.
For a long time, only a handful of companies really pulled this off at scale. Companies like Meta and Uber.
Which brings me to Statsig.
Statsig is our presenting partner for the season. Statsig gives engineering teams the tooling for experimentation and feature flagging that used to require years of internal work to build.
Here’s what it looks like in practice:
And the key is that the measurement is part of the workflow. You’re not switching between three different tools and trying to match up segments and dashboards after the fact. Feature flags, experiments, and analytics are in one place, using the same underlying user assignments and data.
This is why teams and companies like Notion, Brex, and Atlassian use Statsig. Statsig has a generous free tier to get started, and pro pricing for teams starts at $150 per month.
To learn more and get a 30-day enterprise trial, go to Statsig.com/pragmatic.
And with this, let’s get back to Andre and experimental features in Kotlin.
So we did quite a lot of work when you’re doing something experimental. This is something that’s supposed to break, and you want to emphasize this to make sure that the user is aware that:
“this is something we are not promising to keep compatible. This is something we’re going to break.”
We used to put the word experimental in package names for people to understand that this is going to be renamed. And warnings when you use language features, and we require compiler keys to enable language features and stuff like that. It kind of helps. So we did quite a lot of that.
All this is an extra layer. And unlike a SaaS system, for example, a compiler leaves behind, but not behind, creates a lot of artifacts that pin down its history in the world. There is source out there and there are binaries out there, and you’re guaranteed to encounter them every time anyone hopes that
“this is an obscure case. Nobody will ever hit that.”
With enough users, you hit every freaking case. And this is so surprising.
I discovered this fairly early on. I think before 1.0, when we had a few thousand users, I realized that
“if something’s possible, some person out there will actually do it.”
So you got 1.0 out. Can you tell me how Kotlin grew in popularity? When you released it, what was your target audience? And then how did Android happen?
Okay, so that’s a complicated story. Let’s try to not get off track, because this has a lot of sidetracks to it.
When we started Kotlin, we were not really very aware of Android. And I mean, we knew that that was a thing called Android.
Kind of ironic.
Yeah.
From now, message from the future.
Right.
Yeah.
So basically in 2010, we were focused on the majority of Java developers that was all about the server side.
Clear.
Yeah.
So the most money IntelliJ was making was on Spring users. And, you know, everybody knew that this was what the Java platform was about by then. So we were targeting server side developers, basically.
And also desktop developers, because JetBrains had the, probably the last desktop application written in Java, or at least in Swing.
So that was the target. It was initially not even a plan to do Android.
Kotlin got some usage for the server side. And, you know, it’s still there and it’s growing there, not as fast as on Android, but still has quite some representation on the server side.
But then a few years in, some person on the Internet asked us whether Kotlin works on Android. And I was like, I heard Android uses Java, so Kotlin should work. We’ll never try. Go and try.
I think it was either the same user or a different user who came back and said
“the toolchain crashes.”
And it wasn’t even Kotlin toolchain. It was the Android toolchain that crashed. And, you know, we looked into it and it turns out that some tool in the Android toolchain that’s written in C just fails with a core dump. And it’s not very clear what’s going on.
We later figured it out. It turned out that the Android developers and the people who built the Android platform actually read the spec of the JVM, unlike the people who implemented the Hotspot VM. Because the Hotspot VM, I suspect, came before the spec. So it was the reference implementation, but it was actually specified after it was built.
The Hotspot VM was super lenient to weird things. Like, there would be, if we put a flag on a class file that was not allowed for classes, Hotspot wouldn’t care. And we ran everything on Hotspot. So we thought everything was fine.
But then on the Android side, those were the people who actually read the spec and implemented it. Yeah, they would complain about everything.
This is why we used the Android toolchain as a testing environment basically, because
“this is how we could get rid of stupid things in our bytecode.”
They helped us a lot with validating everything. But, you know, there were some gotchas there. Some legacy stuff nobody cares about in mainstream Java just were faithfully implemented on the Android platform.
That was fun.
So, you know, at some point, pretty early on, I had this realization that Android was a growing platform. Which, to me then, I didn’t have much understanding of the dynamics of markets, but it meant that there would be a lot of new applications.
And it’s much easier to start completely anew with a new language.
So, I made sure, at some point, that we worked well on Android. It was already after the lawsuit.
So, the big context to all this was that when Oracle acquired Sun Microsystems, they sued Google for billions of dollars for using Java.
And I think that is settled.
It was settled in some way, yeah.
And then everyone could go on their own way.
Right.
But yeah, it took years and years to settle.
Back then, it was very much a thing. And, you know, that dispute was somewhere in the background.
But yeah, so basically, we saw that a lot of people on Android really liked Kotlin. They loved it.
Yeah.
As soon as it was stable, pretty much. I mean, I think for all the things that you mentioned: it was just so much nicer than Java. Easier to write, easier to read, lots of nice features.
So, you know, you use Android as a way to actually make sure that Kotlin compiled correctly.
And then, why did it take off on Android?
Yeah, so the situation in Android was pretty interesting because unlike Java server side that is kind of under control of the teams that develop on it. In the case of Android, there are devices in the pockets of people, right? And when you have billions of those devices, and those devices don’t always update the virtual machine.
So, people on Android were basically stuck with old Java. And even when Java started progressing, and, for example, Java 8 came out in 2014, it was very difficult to roll out this new version of Java across the entire Android ecosystem because it required updates to the virtual machine.
There were workarounds, and Retro Lambda really helped, and so on and so forth. But there was still a lot of people stuck with really old Java. So, Java wasn’t on par with Kotlin or C Sharp in 2014. But it still was much better, and solved the major problem. But it was not available to the Android people.
So, there was a lot more frustration with Java in the Android community.
And also, there was Swift on iOS. Where it was a real example of a big ecosystem transitioning from a really dated language to something really nice.
I think compounding these two things were the major factors. Also, we made sure that Kotlin worked well on Android.
Very fortunately, at some point, Google switched the developer tooling from the Eclipse platform to the IntelliJ platform when IntelliJ was open-sourced back in, I don’t remember, 2013, I think.
So, we had a nice plug-in because everything worked on the IntelliJ platform, and the same plug-in worked for Android. Many other things were just very smooth. Well, very smooth—there were a lot of bugs, but reasonably smooth.
So, it felt like a very good match, and a lot of people appreciated that.
We really wanted to somehow draw the attention of the team at Google to maybe talk about it or something, but it just didn’t happen.
We released in 2016, and there was some communication with Google in general, but there was no interest on that side. They were like, okay, we, I guess we’ll just keep going as we do.
Some people were already building Android applications, and some people were building production applications in Kotlin before we released 1.0.
Kudos to the brave people because they gave us indelible feedback. But you guys are too brave.
So, it just grew organically.
When we started, in the very beginning, I set this internal goal to myself, that if we get to 100,000 users, it’s a success.
I’ve done well enough if it gets to 100,000. Of course, it’s hard to tell how many users the language has, but you can kind of estimate that.
I think we were on track to get to 100,000 users during 2016 because it was growing, it was in the tens of thousands, it looked good.
Then, some people from Google reached out and said they wanted to chat.
It turned out they wanted to chat about announcing official support for Kotlin at Google I/O 2017, that would be in like three months from the time of that conversation.
They said, “yeah, sure, let’s do it. What do we need to do?”
It turned out we had to figure out quite a few things, but we managed.
I think it was a heroic effort on the side of the Google team. They did amazing, impossible things.
I have good friends among them now.
It was really, really close. Like, we could have missed the deadline, but we figured it out.
On our side, we had to make many things work and figure out how we interoperate with Android Studio better, and how to set up processes and everything.
But there was a big legal thing around it. This is when the Kotlin Foundation was invented. We had to design the protocols for decision-making in the Kotlin Foundation.
Google owned the trademark for Kotlin for one year because of legal things. It was basically a guarantee from the JetBrains side until the foundation was set up.
You can look up the public record:
Google was in possession of the Kotlin trademark for a year.
But then the foundation was set up and the trademark was transferred to the foundation.
It was fun. It was a pretty crazy time.
But it was amazing to see how happy people were at Google I/O when the announcement happened.
Then usage must have skyrocketed. You probably blew past 100,000 pretty quickly.
Yes, I think we went into millions that year.
So this was basically the moment happening.
I knew many years before that the easiest way for a language to succeed is to be part of a platform.
For example:
And I knew that Kotlin had no platform. So it was supposed to be a much tougher time for Kotlin than for some other languages. But, yeah, the platform came along somehow.
Jumping forward to a lot more closer today, you left Kotlin in 2020. Later, you left JetBrains. What are you doing right now?
Yeah, so I’m also working on a language right now. But it’s sort of a different kind of language because the times have changed. And, you know, you can look at it from a similar perspective. Like, in Kotlin, we wanted to get rid of boilerplate. We wanted to make programs more to the point. And less of a ceremony.
And I think this is where we, today, we have a great opportunity to do the same thing at a different level. Because of AI, right? Because of AI. It’s all because of AI.
Yes. AI is great because many things that are obvious to humans are obvious to LLMs as well, which closes this gap between what the machine can understand and what a human can understand quite a lot. Which means we might not need to write dumb code anymore. That would be very nice.
So, on the one hand, you know, the entire history of programming languages is going from lower to higher levels of abstraction. We started with machine code. And then assembly was a step up, actually.
Teams could grow and you didn’t have to be a super competent programmer to build working software. And then, you know, things like Kotlin built on top of that success. And we raised level instructions some more, but now we can do even better in the same domain.
So, you can imagine a normal program, some application code. A lot of the things in this code are obvious to you and to me. So, if you ask me to write this code, you don’t spell everything out. You explain what the program needs to do and I can implement it. And it will work the way you want.
There are, you know, it depends on how detailed the specification is. But you can tell me a lot less than you would have to tell a compiler.
And so, this is the point with Codespeak. We want to basically shrink the amount of information a programmer needs to tell the computer to make the program work. From my current anecdotal experience, you can shrink a lot of the code about 10x.
Which means that a lot of projects out there can be a lot smaller. And it will be a lot easier for humans to deal with that and a lot easier to read — and reading is the most important bit — and a lot easier to navigate.
It becomes, you know, the essence of software engineering. When you are not dealing with a stupid compiler, you’re not restricted by that anymore. What you’re expressing is what only you know about what needs to happen because everything else, the machine knows as well.
So, can you tell me a bit more on what Codespeak is or what this language is? Is it designing an actual, like, in a formal language, just simpler? Is it using, of course, we know that AI and LLMs and agents can do all the funky stuff. Where is this? What is this?
Okay, yeah, so I’ll try to explain this.
So, I think the best way of thinking about Codespeak is it’s a programming language that’s based on English. It’s not a formal language or not an entirely formal language. But it’s a programming language. It’s a language that’s supposed to be used by engineers. But it uses LLMs heavily.
And this is like the way new languages will be. Because, you know, you can think about the ultimate language of today as a normal programming language that uses an LLM as a library.
You know, there was a time where NPM was wonderful because, you know, it’s a huge repository of all kinds of JavaScript libraries. It’s the node packet manager, one of the biggest package managers in the world, right?
Right, yeah.
So, you have:
- a huge library out there that you can call,
- but now you have an even better NPM,
- The LLM that has seen all the code in the world,
- and if you're inventive enough, you can fish this code out of the LLM.
Yeah. You need to know how to prompt.
Right.
And the trick is, like, it would be really nice to have a programming language that has the entire LLM as a library or as a bag of libraries, right?
The trick is to take anything out of an LLM, you have to use natural language. So, the query language to this incredible database of all the knowledge is informal. And there is no way, at least known today, that you can make it formal.
So, inherently, this ultimate language of today has to be, at least in part, informal. And this is what we’re working on.
So, it’s still in the air, like, how formal can we make it? And, you know, it’s not the goal to make it super restricted. But the goal is to leverage all the power and support the user. You know, we need to rule out stupid mistakes and things like that. And we’re still working on that. But the basic idea is, if you, instead of spelling out every line of code and every bit of your algorithm, you can basically communicate intent the same way I can communicate it to you, you will just get there much faster.
So, one question that I asked Chris Lattner, which I’m going to ask you as well, you’re talking about designing a language for software engineers to build software more efficiently, maybe more concise, in a new way, and it sounds super exciting. But going to the other side, we have LLMs. Do you think there is a need to design a new type of programming language for LLMs to use more efficiently?
That’s a very interesting question. And I had a few discussions about this. My position is it’s probably misguided because of a number of things.
So, one, to get an LLM to understand some language well, you need a huge training set. And with the new language, that training set is not there. You can try to synthesize it and so on and so forth, but it’s not going to be as good as other languages. Like, for example, right now, the newer languages are just harder for LLMs than the more established ones.
And, you know, there are ways around it. I think the later models added some more Kotlin into the RL sets and it’s getting better. But still, it’s pretty hard. And so that’s challenge number one.
Also, challenge number two, I don’t think there necessarily has to exist a language that makes it better because LLMs are trained on human language. Their knowledge of programming languages is part of that. Their power is in having been exposed to all the code in the world and its existing code. And inventing a new language for that, I don’t know how promising that can be.
You can do another thing, which is an interesting research project. You can sort of extract a language from an LLM because, internally, it has some intermediate representations of what’s going on during inference. And maybe you can sort of extract the optimal prompting language.
It’s not guaranteed to be intelligible to humans. And there are some experiments that show that you can create completely unintelligible prompts that give the same results as normal human prompts, but they will be shorter.
You maybe can do something like this. I don’t know if it will help a lot. But what we’re doing in code speak as part of working in this language, we need to really nail down this query language capacity.
What we’re doing now is we are looking at existing code, and we’re trying to find the shortest English descriptions for this code that can generate equivalent implementations—not necessarily character to character, but they have to work the same way.
That’s an interesting exercise because you need to figure out how to represent the ideas in the code in a way that:
But also, this code you represent evolves over time, right? So you have a commit history on top of this version. Going forward in time, you need to be able to represent all the changes in your code speak version.
You need to make sure that when it’s a small change in the original code, the change in the spec is smaller. That’s an interesting challenge. So in this way, we’re sort of discovering code speak as a language, or at least parts of it, and not really designing that bit of it.
You know, it’s a very new world in the sense that, nowadays, if you work with AI, everything is a machine learning problem. That means, back in the day, if you had a very smart algorithm on paper, you could just implement it and make sure it works. Nowadays, whatever algorithms you have in mind, you need the dataset.
First of all, like if you don’t know how to collect a dataset, don’t even start. And, yeah, this is what we’re doing.
So just taking a look at, you are using these tools day in, day out. I mean, you’re building with them. How do you think programming as a whole, or I’ll say software engineering, is being changed by AI? And how do you think the future is starting to look? Especially thinking about software engineers. You’re a software engineer yourself. You’ve written so much code in your life. And are you still writing code?
Yeah, I’m writing some code, yeah.
Sorry, typing or prompting?
I’m doing both. Sometimes I’m just typing. More often, I’m typing with cursor tab completion. I’m doing quite a lot of prompting as well. And that’s a combination of all this. But cursor’s completion is really a step up from traditional IDEs. And I think the IntelliJ side has something similar now. So it’s like a lot of coding, but in a very different kind of mindset and a different tool set.
Yeah, so in terms of what’s happening to programming, I think we are in the early days of the new era. So, you know, it’s only last year that we figured out that coding agents are good. No. Cloud code and cursor agent and so on and so forth. And I think this is a very early step.
Right now we are in this phase where a lot of people are in love with agents and it can be very useful and I use them every day. But I think there are inherent problems with the model, with how you interact with a coding agent because it’s a one-on-one chat. And as a human, I talk to the agent in human language. So I’m communicating my intent on a high level.
And that intent gets translated into code and it’s the code that I commit to the repo and it’s the code that my teammates will see. So my chat history is lost. Big problem.
Yeah, so it turns out I’m talking to a machine in human language. But the way I communicate with my team is the machine language. That’s kind of backwards.
So, yeah, so what we’re trying to do in the Codespeak is to elevate everything to the human language level. So this is where we start. We say, okay, we have this incredible tool. We can prompt agents to implement code for us. And we are just picking it up. So I think a lot of teams haven’t yet realized how difficult it is to review the code.
And I’ve talked to people who are like,
“Maybe we can just not review this code.”
I’m like, yeah, I mean, you can for a couple of days and then it just collapses. And I think another big theme of today is that we’ll be doing a lot of testing.
And like, you may not need to review the code if your tests are really good. You need to verify it, right? Yeah. That’s what you’re saying is verifying might not mean reviewing. Right. Or it could not mean. Yeah, depending on the domain. Of course, of course.
You might get by without reviewing the code as much, but being sure somehow either reviewing the tests or somehow else, making sure that your tests are good. That’s a trend. And we are putting a lot of effort at Codespeak into automated testing and making sure the tests actually check the right things and that they check all the code and all that stuff.
It’s very interesting computer science. And also, it’s now a question of, especially in the case of Codespeak, and I think for other agents as well, like, yeah, reviewing code can be too much, but can we present the tests we generated to the user in a way that actually verifies that we did what was to be done?
It’s tricky. Some tests will be just very long and tedious to read and, you know, but we’re working on that. And that’s where we are.
And I think we’ll see a lot of development in terms of power of the models and we’ll get some quote unquote obvious things implemented in agents. For example, the agents are just starting to use like language servers and basically all the stuff that we’ve always had for code is not very utilized.
And, you know, if you compare like IDE-integrated agents like Cursor or Juni at JetBrains, you have a lot of like code navigation capability and, you know, databases of code is indexed and you can navigate it very quickly. You can find things very quickly.
When you run cloud code, for example, it might not have that and use grep and it will be as successful, but take a lot longer and burn a lot more tokens.
So, you know, I’m sure this year all these tools come to most agents and we’ll have a lot more sophisticated scaffolding around the models.
So that’s one thing. But then, you know, my question is always what’s going to happen in the endgame or in further future. And there it’s very hard to predict. And we can assume that models will become much smarter. But an important thing is that humans will not.
So one thing I know about the future and it’s hard to know the future, but this thing I do know about the future, humans will be as smart or as dumb as they are today. And if we have incredibly smart models, what we will be doing is constrained by how humans are and this is one of the reasons why I’m working on Codespeak because Codespeak is a tool for humans, not for models.
Yeah. And humans, I know I can build a tool for them.
I guess an important footnote is that many people will say things like,
“If we have smart enough models, they can review the code themselves and they can test the code themselves.”
But then my question would be like, who’s making the decisions here?
You know, if all the software engineering work is done by models, it means humans don’t have any say in that. And this has a name. It’s called technological singularity.
Yeah. When humans are not making decisions, it means we’re not in charge.
Yep. So this is not the future I’m building Codespeak for. Nobody should build any projects for that future. In that future, we’re gone. Your projects don’t matter.
But so my assumption when I’m talking about the future is that the technological singularity is not happening. And so the basic assumption is humans are in charge.
And if humans are in charge, it’s their job to communicate intent. So we have to say what kind of software we need to build. And when we’re talking about serious software, it’s always complex. There’s no way there’s some very simple thing that will make a difference.
And when we talk about this complexity, this is what our jobs will be, like dealing, managing this complexity, figuring out what we actually need to do. And this is absolutely engineering. There is no way someone can tackle huge amounts of complexity without an engineering mindset. It can be called software engineering, can be called something else, but you will have to do it. You will have to navigate this complexity, organize this complexity, figure it out.
And I’m not talking about the complexity of many, many layers of implementation. Maybe not. Maybe that is what’s called accidental complexity, something that happens or arises from how we implement systems. But there is also essential complexity. How we want it to behave is complex enough that we need to figure it out.
And this is why I believe there will be teams of engineers working on systems like today. Maybe they will be a lot more powerful teams. Maybe fewer people can deliver a lot more software. Yes, but still teams of people working on organizing complexity.
And this is what Codespeak is for.
Going back to where we are today with what the models can do today, what do you see with developer tools? It feels a little bit of a wild, wild west right now, very much so. I mean, there’s a lot of, obviously with Cloud Code, with Curse or with others.
But what are areas that you think we will see, we will have to see new, different, better tools to actually just catch up with how we can generate? And what parts feel the most messy and the most interesting? Especially because at Kotlin, you have, and the team has built so many tools for developers.
Right. So I think, as I already mentioned, this year will be the year of making developer tools available to agents.
There are some technical challenges, but you can’t figure it out. The people will be doing that.
There’s also a surprising advantage to using a good UI for your agent. It’s very nice to have everything in your terminal, in one sense. But then you can have a lot better user experience if it’s a dedicated environment.
The terminal tools, especially Cloud Code, are amazing. And it’s a complete breakthrough of what you can do in a terminal. But generally, you can do better in a specialized environment.
So I think we’ll see more of this integration into development environments or just new development environments built from the ground up to work with agents primarily. So that is an important thing.
Since we are putting a lot more emphasis on review, there should be new tools for review. And I think we can do better than what we’re doing now in many respects.
I don’t expect many breakthroughs in testing this year because it’s hard. I’m doing it right now. It’s hard. It’s not going to happen this year. But maybe some advances will arrive this year.
But generally, I think the big lesson of the last couple of years is that all the things that were, quote unquote, obviously needed and, you know, the idea of connecting agents to developer tools was absolutely the trivial thing to think of two years ago. But they take a long time to happen because it’s hard.
And, you know, nobody in this industry is lazy. Like everybody’s working their asses off. But it just takes time. You need to figure out the basics before you can do advanced things. So, you know, all the straightforward ideas will get implemented at some point.
I think there’s been this massive jump with AI, especially over the winter break, where the coding agents, the CLIs, have become a lot more capable.
And I know so many developers who are actually just prompting most of their code, if not all of it. It’s just a massive, massive jump. I don’t think we’ve seen anything this fast.
I see a lot of engineers scared because it can shake you to the bone. You know, it took 10 years to get really good at coding. And the writing the code part feels that it’s kind of going out, you know, the trash can.
You yourself have coded for a longer time. What would your advice be for developers who are feeling like this, that they’re feeling, you know, it is scary.
I think we, and I talk with some folks, a lot of people message me as well. How are you thinking about this specifically these last few months? It’s really hard to give advice.
There are a few ideas I can share. So one thing is there’s a lot of hype and a lot of it gets to the management and a lot of people make suboptimal decisions. But that will go away.
So, you know, there’s more and more news about people not hiring junior developers, for example.
And a lot of other things can be really stressful in the moment, but some of them will be rolled back. So that’s one thing.
Another thing, it’s absolutely worth it to invest your time into learning these tools and getting good at it. There’s a lot of skepticism around in the developer community about how useful it actually is. And, you know, I tried it on my project and it’s no good.
There is quite a bit of skill to using these tools. Unfortunately, it’s not super formalizable. At least so far, nobody figured out a really good, clear way of communicating how to do it well. But there are people who can do it much better than others. They not always can’t articulate why their prompts work better. But, you know, you can learn it. You can get a lot better at it.
And, you know, not necessarily believing everyone on Twitter. Some people claim crazy things, but you can be very productive with these things when you use them well. And it’s absolutely worth investing into that.
And yeah, so as I mentioned before, in the future, it will still be engineers building complex systems. So keep that in mind. It’s not like we all go to nothing.
And for new grads, people coming out of university, what would your advice be for them who are like determined, like, “all right, I actually want to be a standout engineer. Maybe with these tools, I can do it faster.” What would you advise them to focus on either skills or experiences to get?
I guess it’s a matter of what your inclinations are.
So, you know, generally, if you have any inclination in looking under the hood and figuring out how things work, go as deep as you can. As a younger person, you have a lot of mental capacity for that. And this helps a lot. You become a very good expert in very wide fields, just through drilling down on many things.
That’s closing. I just wanted to do some rapid questions. I just ask and you shoot what comes next.
What is a favorite tool that you have? It can be digital. It doesn’t have to be digital.
Well, I love my AirPods. They’re incredibly convenient. They fit under my earmuffs.
Well, another tool would be earmuffs.
Earmuffs. Incredibly good.
Yeah, I saw you wearing it. I’ll take that one, Earmuff.
And what’s a book recommendation that you recommend and why?
There is this classic that’s been recommended across the tech community for many years. It’s called Zen and the Art of Motorcycle Maintenance.
I heard that recommended.
Yeah, it’s a very good book. I mean, there is a part of it that’s about technology and how to deal with the real systems and others, but it’s also a very good novel. I really like it.
Well, Andrew, thank you so much. This was very interesting and I think inspiring as well.
Thank you very much. It was great to chat.
It was great. Thank you.
The thing that struck me most from this conversation with Andrey was his observation about how we work with AI coding agents today. You talk to an agent and play in English. It generates code. You commit the code. But that conversation, your actual intent, it disappears. You communicate with the machine in human language, but with your teammates in code, in machine language.
Whether or not CodeSpeak becomes the answer, what is sure is that we’re missing an intent layer. And someone is going to figure out how to preserve it.
If you enjoyed this episode, please do share it with a colleague who’s been thinking about where programming is headed. And if you’re not subscribed yet, now’s a good time. We have more conversations like this one coming.
Thank you and see you in the next one.
]]>Welcome to the Sinica Podcast, the weekly discussion of current affairs in China. In this program, we look at books, ideas, new research, intellectual currents, and cultural trends that can help us better understand what’s happening in China’s politics, foreign relations, economics, and society.
Join me each week for in-depth conversations that shed more light and bring less heat to how we think and talk about China. I’m Kaiser Guo, coming to you this week from my home in Chapel Hill, North Carolina.
Sinica is supported this year by the Center for East Asian Studies at the University of Wisconsin-Madison, a national resource center for the study of East Asia. The Sinica Podcast is and will remain free.
But if you work for an organization that believes in what I’m doing with the show and with the newsletter, please consider lending your support. I know you think of this as boilerplate by now, but seriously, I am looking for new institutional support.
The lines are open and you can reach me at:
cynicapod@gmail.com
or just my first.last name at gmail.
Listeners, please support my work by becoming a paying subscriber at cynicapodcast.com. Seriously, help me out.
I know there are a lot of Substacks out there and they start to add up. Yes, but I think this one delivers some serious value. You get my stuff, the China Global South podcast, the fantastic content from Trivium, including not only their excellent podcast but also their super useful weekly recap. You get James Carter’s “This Week in China’s History” column. You get Andrew Methvin’s Chinese Phrase of the Week.
I am really trying to deliver value for your hard-earned dollars, so please do sign up. Things are tough, I get it, but consider help now. Tough for me too.
As we move into the second year of Donald Trump’s seemingly interminable second presidency, U.S.-China relations have once again defied easy characterization.
What began as a return to tariff escalation and hardball trade tactics has somewhat unexpectedly given way to a period of relative strategic calm marked by:
Even in the national security strategy and the national defense strategy that was just released.
The once dominant language of great power competition has definitely receded, and many of the most vocal China hawks who shaped Washington’s approach for the past decade appear to have been sidelined.
In their place, we’ve seen a policy posture that reflects Trump’s highly personalistic approach to foreign affairs and emphasis on leader-to-leader rapport.
“Xi Jinping’s my friend,” deal-making over doctrine, and a willingness to bracket or at least downplay ideological disputes in favor of transactional progress on trade, technology, and risk reduction.
Trump’s repeated praise for Xi Jinping, his apparent sensitivity to certain of Beijing’s red lines, including on Taiwan, and his apparent comfort at treating China as a peer rather than a civilizational rival mark a sharp departure from recent bipartisan orthodoxy in Washington, if you indeed believe that it was a bipartisan consensus.
Supporters argue that… This shift has lowered the risk of conflict and delivered tangible gains. Critics, though, counter that the United States is conceding leverage without securing durable returns. Either way, the result is a relationship that feels less confrontational for now.
In my private communications with certain among my more panda-hugging friends, there’s this sort of bewilderment. It’s like, we kind of agree that Trump is awful for this country but not so bad for U.S.-China relations, right? But beneath the surface calm lie unresolved structural tensions, deep mutual dependencies, of course, that can be weaponized, and parallel efforts in both capitals to reduce those vulnerabilities.
So, what comes next? Are we headed toward a genuine lasting stabilization or a familiar snapback to the acrimony that once dominated, once our expectations collide with reality? Or a more ambiguous middle path, one in which both sides buy time, avoid escalation, and quietly work to insulate themselves against future shocks?
Well, to help us think through all these questions, I am joined by Ryan Haas, director of the John L. Thornton China Center at Brookings, and one of the most clear-eyed analysts of the U.S.-China relationship working today. Ryan has just published an essay on the Brookings website laying out three plausible pathways for the relationship under Trump scenarios ranging from:
He joins me from D.C. And Ryan, welcome back to Sinica, man.
Thank you, Kaiser. It’s wonderful to be back with you.
So Ryan, like I said, you’re joining us from Washington. Let me start there. One of the strengths of your piece is that it treats leaders as not free agents but constrained actors. From where you sit in D.C., what are the most powerful domestic forces that are shaping the U.S.-China policy right now? And which of them do you think actually matter to President Trump?
Well, it’s a really interesting question. I have to say, sitting in Washington, D.C., one thing that is very palpable is a hope, a wish among many inside the beltway that we will soon snap back to the way things were before—that this one to two-year window is just sort of a brief pause from the long-term trajectory of intensifying competition and confrontation.
I’m a little less confident of that. In fact, I’m fairly skeptical that’s where things are headed, but that’s certainly a palpable sense of mood within the beltway.
To your question, I actually think that President Trump is fairly unconstrained in terms of his approach to China. I believe he is pursuing the approach that he thinks will yield the best benefit for him personally and politically, but also for the country. The basic contours of it, to the extent that you can assign strategy to what President Trump is doing, are:
- Trying to lower the temperature of the U.S.-China relationship through direct engagement with President Xi.
- Showing tremendous respect to President Xi and, by extension, China in service of that effort.
- Building deterrence in Asia militarily.
- Reducing dependence upon China for critical inputs to the U.S. economy.
- In his own way, trying to rebalance the U.S.-China economy.
That’s the direction he is trying to take things. I don’t think he surveys the landscape of the U.S. political class and finds too many threats to his vision and approach to the relationship. But he’s thinking about midterms, he is thinking about 2028, and he’s thinking about affordability and things like that.
I mean, is that part of the logic that’s driving him to soften things with China right now—to hit pause?
Yeah. I think that there are a few things causing him to do that. First, he believes that China has us over a barrel in terms of their control over earth and other critical inputs. Until we get out from under the sword of Damocles that the Chinese have above our head, I don’t think he sees much value in taking the U.S.-China relationship toward head-on collision.
He also recognizes that he’s managing a lot of other problems around the world simultaneously. Adding to that list with intensifying confrontation with China may not be wise or prudent.
But I think he also recognizes that there isn’t a ton of appetite in the United States among the body politic for head-on confrontation.
This is something, Kaiser, you have written about and talked about—the vibe shift in the United States. President Trump, one of his unique strengths is… His reptilian feel for the mood of the American people. And in this regard, I think that the president reflects what he can sense from the American people in terms of what their expectations are for the U.S.-China relationship today.
Well, that’s comforting. The other questions, industrial policy coalitions used to be, at various times, a ballast for stability or even an active force for improved relations with China. Are they acting on him today? Is there business pressure somewhere? Is Jensen Huang a major force in his thought these days?
Well, I think that President Trump operates much differently than traditional U.S. presidents, in the sense that he is not sitting in the Oval Office waiting for his staff to bring him options for him to decide upon as it relates to China. As we’ve talked about before in Berkeley and elsewhere, he is his own China desk officer. He takes his own responsibility for calling the shots and setting the direction of U.S. policy towards China.
And in doing so, he is not informed by stale, turgid intelligence briefings that stone-faced people deliver to him early in the morning. He is talking to a range of people in and outside of government. He’s talking to people he treats as peers and considers as peers, including Jensen Huang, but not just Jensen Huang. He is basing judgments upon the body of inputs he’s receiving, which are far broader than a traditional U.S. president would.
So if he is so unconstrained and if his policy toward China, as with all things, is such a function of his just idiosyncratic whims and his character, is this current pivot away from ideology credit where it’s due? It’s something that I’m really happy to see. Is this something that could survive Trump or is it inseparable from his personal instincts and his incentives?
Well, I’ll try to take this in two parts. The first is that I think Trump is in a category of one amongst the U.S. political class in his willingness and tolerance to affect the change in America’s overall orientation towards China. And you noted this very articulately in your introduction, that he has moved the United States away from sort of an emphasis and a framing of great power competition as the sole lens through which to view the U.S.-China relationship to something that’s much broader.
I think of it as sort of non-conflictual coexistence, a more pragmatic, realistic appraisal of the nature of the U.S.-China relationship than preceded President Trump. But it does raise the question, I think a very legitimate question that you’re asking, which is, is this just something that will perish when President Trump departs office?
I can’t tell you. I honestly don’t know. But my instinct would be that no, this has the potential to outlast President Trump. However, for it to do so, a few things will need to happen:
First, President Trump will need to demonstrate return on investment. Over the next couple of years, he will need to demonstrate that this less harsh approach to the U.S.-China relationship yields tangible benefits for the American people and American workers.
Secondly, whoever succeeds him, whether Democrat or Republican in 2029, will need to be able to make a case for what America’s national goals are and how China relates to them.
It’s impossible to know how those two variables will play out, but it is certainly a possibility that we could see an elongation of this period beyond just Donald Trump.
The ball then is sort of in Beijing’s court. They need to pay a return on that investment, and I think if they want it to endure beyond Trump.
But speaking of Beijing, let’s flip the lens to Beijing. Is Xi similarly unconstrained? Is he a sort of singular determinant of Chinese policy toward the U.S., or does he have domestic determinants of China’s policy toward the United States at this point?
I mean, and if they are, is it like economic stabilization in the post-COVID period? There’s plenty of things that bedevil the Chinese economy right now.
Is it:
What are Xi’s considerations as far as you can tell?
Well, one of the unique aspects of this moment is that we are in a situation where the two countries are driven by very personalistic leadership styles. There are some, for me, uncomfortable similarities now in the way that the two countries are sort of operating.
I don’t think that Xi is perfectly unconstrained. I’ve never subscribed to the view that he has a monopoly on power in China and that he alone can determine the outcomes for 1.4 billion people. But I do think that there are certain things that… He is very invested in and that his brand is associated with, his political brand. One of them is making progress towards greater self-reliance and less dependence upon the United States and the West for China’s future growth, innovation and technological breakthroughs. And this period of relative calm in the relationship, I think serves that purpose. It gives breathing room and space for China to make progress down the path of greater self-reliance.
The second is being able to give proof to the narrative that time is on China’s side, that China has “winded its back” and that it’s the United States that on a relative basis is declining. And I think there are plenty of proof points that President Xi and those around him can point to, to build that case persuasively inside China today, which I think also gives some momentum to the current direction that we’re in.
I mean, I know it’s hard to say with any certainty, but is it your sense that there’s debate within the Chinese system about how hard or soft to lean into this current period of calm? Is this something that, you know, is he facing opposition? In other words, are there people who are saying,
“Hey, America’s showing weakness, time to press our strength,”
or does it seem to be, you know, Xi’s calling the shot in that case?
Ryan Hass: You know, it’s a good question. My latest sort of touch for that is a bit dated. I was last in Beijing and Shanghai in December. So I’m a month plus out from my last contact with people who are in policy circles in China.
But based upon that last round of conversations, my view is that many people recognize that this moment is serving China’s interest well, that China’s goal is to try to relieve pressure and sort of unblock the path to China’s continued rise.
To the extent that President Trump is willing to play a role in that by relaxing pressure upon China, whether it be through:
I think those are all sort of indicators that this is working to China’s long-term benefit.
Kaiser Guo: So Ryan, a central claim or assumption in your essay is that both sides, Beijing and Washington, are behaving less out of mutual trust than out of mutual sense of vulnerability. That, I think, isn’t a claim that many people would challenge, actually, and I wouldn’t.
To what extent do you think that policymakers in both capitals genuinely understand this as kind of a negative sum dynamic? And to what extent are they simply discovering through painful trial and error that they are mutually vulnerable and that they need to chill out?
Ryan Hass: Well, I have a pretty high degree of conviction around this point, but I don’t have some smoking gun evidence that I can point to to prove it.
My sense is that both leaders and those around them have come over the past year to recognize that the other side is capable of doing immense harm to itself.
And I think that this has been a revelation, more so on the US side than the Chinese side. The Chinese side has been well aware for a long time that the United States is capable of being a dangerous superpower that can do immense harm to China.
But when President Trump and Secretary of Treasury Besant and others entered office last year, they entered office with a certain degree of bravado and hubris. Secretary Besant famously said that
“China is holding a pair of twos in terms of, you know, the cards it has in its hand and the lack of leverage it has over the United States.”
No one is talking like that anymore.
Through painful trial and error, both sides have come to realize that they are each capable of doing harm to the other. And that if one side initiates action against the other, it should expect painful retaliation response.
And so I don’t think that President Trump and President Xi over the past year have developed some like brotherly friendship where they decided not to do harm to each other.
I think they both come to recognize that if they take actions that are harmful to the other, that they will get hit back in response. And that it will hurt.
And that was the whole lesson in 2025 leading up to Busan, right?
Kaiser Guo: And you know, your trip may have been a couple of months ago, but that was still in the post-Busan era. So I think you have a probably quite accurate read of how they’re feeling right now. Not much has changed since then, so.
Ryan Hass: Right.
Right.
Yeah. There haven’t been many major ruptures or fluctuations from then till now. Except the rupture that, you know, Mark Carney spoke of.
But so Ryan, let’s jump in with your first scenario, the soft landing. In this pathway, both leaders:
I mean, you know, they’re really talking about investment right now. We’ve got Ford talking about working with Xiaomi possibly, according to the FT, at least on a battery plant, right?
Yeah, you’re absolutely right. I think for this scenario, the soft landing scenario to take root, a couple of things would need to happen.
So that’s the first prerequisite.
I think one of the things that some people point to who are advocates of this approach would be some type of grand bargain.
So we know that President Trump is planning to travel to China in April. If that visit were to yield a sort of significant breakthrough on a contentious issue, most people would identify Taiwan as the candidate, Taiwan combined with some type of transactional benefit for the United States and its workers. Then that would give momentum or solidity to the idea that we could travel down this path.
But short of that, I think it’s hard to imagine both sides really sort of believing and acting in ways that both leaders believe they can sustainably improve over the long term of the nature of the relationship.
What makes that costly from the American side?
So you say that it would require both sides to send costly signals. What sorts of signals are we talking about from Beijing, and what would be costly about those? How hard would they be to deliver domestically in Beijing?
It’s a great question. I think in the case of China, there is a certain degree of skepticism about whether the Chinese leadership would be comfortable seeing some of its companies and crown jewels invest or produce outside of China. We see this in particular with Meta’s efforts to acquire a Chinese-origin AI company that relocated to Singapore.
Meta’s, yeah.
Another area, in the Taiwan context, would be if President Trump were to alter longstanding declaratory policy toward Taiwan, would China reciprocate by:
These are the types of questions that sort of point to costly signals that each side would expect the other to give if they were to give it themselves.
I have trouble seeing that is costly to China compared to the electoral costliness of signals from America. So it feels like China can ram this through; Trump faces electoral pressure.
Yeah, he might. But let’s keep in mind, he’s never going to be on a ballot again for the rest of his life.
That’s true.
And so, President Trump has never shown a lot of conviction about election outcomes that don’t involve his name on the ballot.
Ryan, looking back over recent U.S.-China history, is there a precedent that you can point to for restraint for restraint actually holding for any decent length of time?
I can’t think of anything off the top of my head right now that would give a lot of confidence to the notion that restraint for restraint is a time-tested and well-established trend. This is the critique that I think people of the soft landing approach would make, is that the soft landing would The discussion involves the United States making concessions to China without receiving reciprocal benefits in return. There’s a pretty calloused skepticism that has built up over years, including within the Trump administration, as a consequence of the underperformance of China in the phase one trade deal.
Obviously, you floated this possibility that something like a fourth joint communique on Taiwan could anchor the sort of soft landing you’re talking about, the grand bargain.
What problem would such a document actually be trying to solve? What would be the content of a fourth communique? And is Taiwan ultimately the issue that makes this scenario maybe politically untenable, even if both leaders are inclined toward restraint? I mean, is Taiwan going to flummox this?
I think it will be very difficult. The idea would be that the last time that the United States and China had a communique was in 1982. A lot has happened in the last 40 plus years. A new framework that sets out a baseline of understanding for how both sides will approach cross regulations may be a useful stabilizing mechanism.
I’m on the more skeptical end of the spectrum on this question. I don’t think that the challenge is a lack of understanding about the nature of cross-strait issues. I think that there are just competing interests involved that need to be managed.
In Washington, it’s treated as sort of a foregone conclusion that Beijing is desperately seeking a fourth communique or some type of new understanding related to Taiwan with President Trump. There are a few factors that may mitigate against that as a foregone conclusion:
Yeah, yeah, yeah. Just to remind everyone, this is your most optimistic scenario. And in this most optimistic version, there is still a sense that the soft landing would be kind of inherently provisional, something closer still to a pause than to a full reset.
I am ineradicably optimistic but still have trouble seeing either polity really arriving at some kind of durable modus vivendi right now. There’s just no trust. There are many deeply entrenched habits of mind on both sides.
But there are other scenarios that you posit here. The second scenario is the one I sincerely hope to avoid: a hard split.
You frame this as a familiar arc: Trump starts conciliatory, grows very frustrated, and then swings really hard. We’ve seen this many times. What are the most plausible triggers that could push the relationship down this kind of path toward a hard split?
Well, there are a few ways we could get here:
1. There could just be a misunderstanding on what each side agrees to. President Trump comes to the conclusion that the Chinese are under-delivering on their promises. He grows frustrated, angry, and we find ourselves back following the same cycle as we did during the first term, where:
- The first three years focused on negotiating a phase one trade deal.
- The fourth year focused on letting it rip because the president was so angry and frustrated that COVID had spread and undercut his reelection prospects.
2. China takes actions against American allies that involve **use of force** and puts the United States in a very difficult position of deciding whether or not to employ force against China to come to the defense of their allies and uphold **Article 5 commitments** or traditional understandings of security commitments.
Examples of such allies include The Philippines or Japan. Right. Right. Right. And then what some people in Washington would say is that as the midterms get closer, the political incentive for President Trump to become harder and harder towards China will grow, and that the political imperatives of President Trump wanting to hold off Democrats gaining control of the House and relaunching impeachment probes against him will compel him to grow tough.
This is the hope, I think, of a lot of people in Washington who want us to get back into the business of great power competition. And I’ll just offer just a quick caution, Kaiser, as to why I’m not yet convinced that this is the natural course of events that we’re going to find ourselves in.
First, you know, the president has demonstrated that he is very sensitive about America’s dependence on rare earths. That dependence is not going to change in the next 12 months, 18 months, even two years.
Yes. The second is that President Trump just genuinely is not activated by the military threat or the ideological nature of competition between the United States and China. But he’s much more focused on economic and tech issues. He wants to make deals that he can point to and tout his as successes and breakthroughs.
And having a hostile relationship with China would sort of move against that objective.
I also think that President Trump is pretty comfortable with the status quo right now. He doesn’t face immense political pressure at home for where the US-China relationship stands. He also likes to brag privately with his colleagues and counterparts about how much tariff money he believes that the United States is generating from tariffs on China, never mind the fact that it’s US importers that pay the tariffs.
And then lastly, I think that President Trump is very focused on legacy and blowing up relations, burning down the house with China is not a legacy enhancing exercise. Putting the relationship on a new plane potentially could be.
So, I mean, the fear of a blue wave in 2026 in the midterms, I mean, I get that. But part of him also has somebody’s got to be showing him these polls that say,
“there’s just not a lot of appetite right now among voters for tough on China. It’s not a winning campaign strategy right now.”
I mean, poll after poll after poll is showing that that is fundamentally weakened vibe shift once again.
Right.
So, I mean, hopefully that’s a mitigating force.
Yeah.
And traditionally, midterm elections are not animated by China or by foreign affairs. I mean, there really isn’t any empirical evidence that going tough on China improves the odds of House and Senate candidates getting elected.
So, from Beijing’s perspective, I mean, it’s pretty easy for us to think of what kinds of U.S. actions would collapse strategic calm and force Beijing to take a harder line that would be reciprocated by Washington. I mean, all sorts of triggers, right?
But where do you think miscalculation is especially dangerous? What are the areas where you think that crossed wires and signals misinterpreted are particularly dangerous?
I would suggest, just as a hypothetical scenario, if the United States became more aggressive with other countries about urging them, insisting that they adopt America’s AI tech stack—
Right.
—and conditioning security support for them doing so. That could be an example of how things could go off track.
And if there were further actions like we saw last fall where the Department of Commerce rolls out something in an uncoordinated fashion, the 50% rule, the affiliates rule.
Right.
Something along those lines that the Chinese perceive as violating the truce, the understanding that was reached between both leaders—that could compel the Chinese to reciprocate and retaliate.
Well, that problem may be solved. Trump has apparently neutered BIS, right? So we’ll see.
One thing that struck me is how much this scenario depends on momentum, on anger compounding on anger. Once the relationship starts moving in this direction, how easy is it to reverse?
I ask because this isn’t the first time either Beijing or Washington has seen things go sideways. And you’d think that both sides might have learned something about how to manage that sort of crisis. And at least sometimes they’ve managed to get the relationship back on track.
And we saw that with the taco meal that resulted in Busan.
Has there been any learning? I mean, do you think that there’s enough sort of wisdom on either side to avoid that kind of scenario?
Well, I think that the key to avoiding that scenario is the two leaders. When things begin to veer off track, it’s the two leaders that usually put things back on track. And the challenge, the structural challenge, is that the Chinese traditionally, historically, are pretty reticent about requesting calls from President Xi to President Trump.
So if there is an incident that is an unplanned encounter between naval vessels or whatever it may be, and things begin to sort of go off the rails, pressure builds. We have a spy balloon-like dynamic emerge inside the United States where there is just boiling angst and anger about something that China has done that violates American airspace or hurts American sailors or whatever it may be.
When the Chinese do not appear to be reaching out to President Trump personally, we could find ourselves in a tough spot. And if the Chinese are perceived to be the instigator of this downward spiral and they don’t communicate directly with President Trump but try to operate through intermediaries, I think that President Trump could find himself both humiliated and offended in ways that could sort of compound the initial problem.
So that’s scenario two: one where there’s a hard split, not an optimal outcome at all, obviously.
You, fortunately, ultimately judged scenario three, which is about buying time and building insulation, as the most likely path. I would certainly concur. But what, in your mind, makes this outcome more resilient than the other two? I mean, because it seems sort of inherently unstable, right? It’s provisional. It’s about sort of just playing for time. And so it feels very impermanent.
But why do you think this is maybe more durable than the other two possible outcomes?
To me, Kaiser, and this is unscientific, this is just sort of a feel, it feels like the most realistic scenario. I don’t think that either of the two leaders is prepared to sort of make significant lasting concessions to the other. I don’t think that either country is prepared to accept a subordinate status to the other.
I think that both countries, in their own way, are able to tell themselves a story that time is on their side. And if they just regenerate or strengthen themselves, that they will be able to outlast and outpace the other.
And so this third scenario of sort of buying time and building insulation, it’s most appealing to me because it works for both leaders and how they describe their intentions and their goals.
This scenario allows him to make directional progress on all those goals.
Similarly, for President Xi, I think that there’s a fairly mirrored set of objectives.
President Xi is very committed to strengthening China’s self-reliance and moving down that path. He certainly, in my mind at least, does not seek a confrontation or conflict with the United States. But he also isn’t interested in making any significant gestures or major concessions to the United States either.
I think that the Chinese believe that they have momentum behind them. And the wave of leaders that have come to Beijing over recent weeks to visit President Xi, I think, have reinforced that perception.
So a core insight of your piece, Ryan, is that both sides are constrained by deep mutual dependencies. I think most people who are listening are aware of some of these and can rattle them off:
But what do you see as underappreciated vulnerabilities on each side that might reinforce this uneasy equilibrium? Are there things that we’re not talking about enough where there is mutual dependence?
Well, I’ll offer a few.
When I was in China last December, I was discomforted to be reminded in almost every meeting about America’s dependence upon active pharmaceutical ingredients from China, APIs. And I don’t think that that was just sort of a stream of consciousness idea that bubbled into the minds of everyone we were sitting down with. It was a reminder that rare earths aren’t the only source of American dependence upon China.
Similarly, I think for China, they are painfully aware of their dependence upon the United States and the West for:
But also at a more intangible level, access to America’s higher educational system. This is something both from the students themselves and their future contributions to Chinese society, but also Chinese leaders’ ability to keep that door open for students, the children of their peers, is critically important. And if the relationship were to deteriorate, we’ve already seen that this is something that the Trump administration has considered using as a retaliatory tool.
Now he’s talking about 600,000 Chinese students in America. I guess maybe he thinks about them as a service export rather than as human beings who contribute to the flourishing of our academic community.
But whatever the case, I think that having Chinese students in the United States enhancing the education of classrooms that they’re a part of is a net benefit for the American people.
So, Ryan, in this scenario, you kind of suggest that the way we score this is by measuring who reduces dependence faster. I mean, if we look out five, ten years from now, which side do you think is better positioned to actually succeed in reducing those dependencies? I mean, who’s working hard at this?
But what’s your assessment of this?
Well, we have a tendency to swing from one extreme to the other in the way that we talk about this in Washington. A few years ago, Kaiser, you and I were talking about peak China, whether it’s a serious thing, how should we think about it? Everyone was focused on all of China’s weaknesses, vulnerabilities, and soft spots.
In recent months, it feels like the pendulum has swung to the other extreme where China can make everything. China can do anything. Ten foot tall again, right?
The world is sort of gravitating towards China. The United States is in dire straits. I’m uncomfortable with either of those extremes.
I think that China does have profound challenges, but it also has immense strengths. Neither of those are going to go away anytime soon. We have to get comfortable to be able to look at both of those side by side.
And the same can be said of the United States.
I will just make one observation that I hope is in service of answering your question, which is that I am deeply uncomfortable with the direction that our country is headed in certain respects. I think that right now the social fabric of the country is tearing, and national unity is the foundation of national strength.
No country can be stronger on the world stage than it is at home.
What we are watching in Minnesota and elsewhere is deeply troubling, both for me from a spiritual standpoint, but also just from a civic standpoint, and also in a measure of national power.
Secondly, I worry very much about America’s alliance network fraying and unraveling. Alliances traditionally have been a force multiplier of American influence on the world stage. Now, I think that our alliance network exists more in name than function.
This is going to be a long-term cost that the United States is going to pay for the moment that we find ourselves in.
But more fundamentally, and this I think, speaks most directly to the question that you’re asking, I worry that America’s economic competitiveness is eroding somewhat.
I just feel like at a certain level, President Trump is pursuing a 19th century strategy of assuming the control of natural resources will be the source of national power. We find ourselves in a different world today.
I think that his resource obsession is a strategic distraction.
For me, the goal needs to be to stimulate growth.
Growth comes from productivity. Innovation and diffusion come from:
- Talent
- Ideas
- Efficient allocation of capital
- A transparent and predictable legal system
This is how America gains strength.
The further we turn from that, the more that I fear we will lose our ability to achieve the sort of escape from dependence that your question was anchored in. Yeah, I mean, it’s so frustrating to be, this is a man whose favorite metaphor is cards, but, you know, he’s talking about who’s got the stronger hand, you know, who holds more cards.
It feels like somebody’s got to be able to convince him that what he’s been doing by, like you say, turning away talent at the border, by destroying those things like predictability, rule of law, alliances, all these things, you know, that act as force multipliers for us.
He’s plucking valuable cards out of his hand and, you know, lighting them on fire to light his cigars. It’s just bizarre.
I mean, I feel like at this point, Beijing must look at, you know, the hands that each side holds and conclude that there’s some very pronounced asymmetry here.
I feel also like that could really make this equilibrium that you described in scenario three more fragile. I mean, if one side succeeds faster than the other in reducing vulnerability, and right now it looks like China’s succeeding faster in reducing vulnerability, that actually seems like it would destabilize this equilibrium.
I agree with you if the equilibrium is measured in bilateral terms only.
And I thought that Adam Tooze made a very important point in the interview that you flagged to his with Ezra Klein after Davos, which is that if we are thinking about the world as undergoing a power transition from the United States to China, it is going to trigger all the anxieties, insecurities, and antibodies in the United States about China’s rise and compel us to try to suppress it.
And if we rather think about what’s going on in the world, not as a power transition, but as a power diffusion, where the United States is not significantly declining, but power is growing much more diffuse in the international system. The international system is splintering. It’s growing more disordered.
Then the nature of the challenge shifts, and the way that we think about and address and respond to it also evolves.
I am much more inclined to the latter view, that we’re seeing a splintering and a diffusion of power rather than a transition in power. But this is going to be, I think, sort of a core aspect of the debate that will be underway about the way that America relates to the world for the next couple of years.
Yeah, it’s interesting. I seized on that metaphor that Tooze used, too.
And I started thinking about that kind of moral panic securitization that we’ve seen in this country as an autoimmune response.
“You’ve got to take some goddamn antihistamines and chill.”
I agree with you that this scenario, this third scenario that you describe, is probably the most likely.
Does this framework, just stepping back, suggest that we’ve entered a phase right now where U.S.-China relations are less about, you know, trying to build trust or establish shared norms and more just about engineering resilience under assumed conditions of enduring mistrust?
I mean, where each side, you know, we’ve got a hand on the other’s choke points,
It’s, you know, I guess it’s structurally analogous to, obviously not identical to, kind of, you know, mutual assured destruction during the Cold War.
If that’s right, how should it change the way policymakers even think about stability?
Well, it’s a great question. I am inclined to your second scenario that you just described. I do think that we’re both sort of holding each other’s oxygen tubes to a certain extent.
I don’t think that there’s any outbreak of goodwill or warm, fuzzy feelings towards each other right now. And I also think that we’re in a pretty fraught moment. Both countries believe that they are gaining a certain degree of advantage over the other or that they can do immense harm to the other.
But on top of that, if you look at, you know, social science work and some public polling data,
And so this isn’t the time. We are not at a moment where there’s going to be some grand breakthrough in the relationship.
I think that if we manage it well through this coming period, we will have done a service as stewards of a long-term relationship rather than as authors of some concluding chapter to it.
Well put. Beautiful.
A final question to you. I mean, if listeners wanted to just cut through the rhetoric and only watch for just a handful of real concrete indicators over the next, say, 12 to 18 months, what would you tell them to focus on to assess which scenario we’re actually in or which we’re careening toward?
I would encourage people to watch the frequency of interaction between the two leaders,
- how often they talk on the phone,
- how often they acknowledge exchanging views through each other as ambassadors or intermediaries.
I would pay attention to the degree to which both sides are preparing for engagements, direct face-to-face summits between the two leaders, whether this is a professional process or just sort of a slapdash trip across the ocean. I would watch to see how well the United States is doing in terms of building or stockpiles, reducing its sort of vulnerability to shocks in the industrial supply chain system from China.
And similarly, I would watch to see the degree to which China is sort of making progress and innovating around some of the export controls and other obstacles that the United States has put in its development path.
So how important are atmospherics going to be around the April Trump visit to Beijing? Well, I think it’ll be significant.
You know, it’s somewhat ironic, Kaiser, because traditionally, the United States trades form for substance. You know, we decide to negotiate away different sort of bells and whistles of a Chinese leader’s visit to the United States in exchange for substance. Because we know that the Chinese leader cares deeply about the imagery that comes out of such engagements because
it bestows respect and gives people inside China pride that their leader is being treated with dignity on the world stage.
Now, I think we’re in a moment where sort of the roles are reversing, where it’s President Trump will be committed to the trappings of dignity and respect, and we’ll want something grander and more dramatic than what he experienced with the state visit plus in 2017 or 18. 17 it was, yeah. I expect that he will probably go to a second city this time as part of his trip.
And so how he is received by the public, but also, you know, the imagery that comes out of that will be important to him. But ultimately, I think that the measure will be to what extent has his travel to China benefited the American worker and the American people. And, you know, we’ll have to see.
Well, I will be there on the ground in Beijing in April. I’m leaving very soon. In fact, just two weeks from now. And I will report faithfully. I’ll do a couple of shows about, you know, preparations for the Trump visit and see how that plays out. Because I think that is a very, very telling indicator.
And I think you’re absolutely right. We are in this world right now where the Trump presidency cares very much about all the symbolism, the pageantry, all the sort of etiquette and the formalism of it. And I think Beijing knows that. Beijing knew that before November 2017 when he went. They sort of turned up the flatterometer to very, very high. They know how to do this.
Well, I will be listening carefully to your reporting from on the ground, Kaiser.
Well, thank you, Ryan. Make sure to read the piece. It’s on the Brookings website and everything else that Ryan writes because it’s all super, super good.
Ryan, thank you so much for taking the time to chat with me. Let’s move on to paying it forward. Do you have a younger colleague or somebody who you’ve been working with who deserves a shout out here on the show?
I do this selfishly because, you know, I’m looking to cultivate, you know, new guests to bring on. I would point to Audrey Wong, who is an incredibly thoughtful, talented researcher, writer, public intellectual, who is doing tremendous work explaining China’s economic orientation to the world.
Fantastic.
And we can find her stuff on Brookings?
Audrey, I believe she’s at USC right now.
Oh, okay. Cool. Excellent. Audrey Wong. I will look out for her.
And what about recommendations? As you know, we do a recommendation every week. What do you have for us? You got a book or a film or some music, a travel destination, something that you want to recommend?
You know, Kaiser, I wish that I had something super cool to share. I’m going to just default to a book recommendation from Robert Sutton. He wrote The Conscience of the Party, the biography of Hu Yaobang.
And it’s as much just a gripping human story about Hu Yaobang, the last reformer in China, as it is a sort of an x-ray of the Chinese Communist Party system and the way that it operates and how it operates. So it’s for anyone who’s sort of interested in the functions of the party. I think that Robert’s book is a tremendous starting point.
That’s been on my list for a while. I really need to finally get around to reading it.
That’s an excellent recommendation. Thanks, Ryan.
So I’ve got a book as well, as well as a couple of China-related things. But my book is just for fun. I’ve been reading the long-lost final book that Alexander Dumas wrote. The English translation that I have is called The Last Cavalier, but it’s also known as The Knight of Saint-Hermain. The French title is Le Chevalier de Saint-Hermain.
But either way, it is a really fun bit of Napoleonic-era historical fiction in which actually Napoleon himself is a major character. And Dumas gives him a really kind of believable personality. I mean, much better than Ridley Scott gave him in that lamentable film, which I hope none of you had to suffer through.
But there are loads of fascinating characters. Many of them are historical. It sent me skirting to Wikipedia many a time just to sort of look these people up. But it’s also just got a ton of historical material mixed in. It’s got letters and decrees and courtroom proceedings, all kind of jumbled into the fictional stuff.
I mean, the story, the plot is a bit of a shaggy dog. It’s maybe, you know, 40% fewer total tangential plot lines might have made this book a little more sort of readable. But it’s still worthwhile if you’re interested.
Dumas actually writes himself or his father. I mean, he does this sort of breaking the fourth wall thing where he suddenly starts talking to the first person and then talks about his father, who was this Napoleonic general, who’s also Alexander Dumas.
It’s anyway, great stuff to take your mind off the world as it is. But still, you kind of get to scratch this itch for, you know, political turmoil and intrigue. If you’re listening to this show, you probably have such a niche.
For a couple of quick China-related recommendations, some really good sense-making of the Chinese economy has dropped just in the last couple of days for the day we’re taping. Check out the Asia Society conversation led by Lizzie Lee, who listeners will know, of course, from her many appearances on the show.
She’s joined by two of my faves:
It’s about the challenges of rebalancing the Chinese economy, but it goes way beyond that. It goes, you know, into the – obviously, you know, the problems of the property market and much else. It’s as good as you would expect with these three all taking part.
Related to that is the latest outstanding Trivium China podcast, of course, which you can find on the Sinica Network. It’s hosted by Andrew Polk, and it is just a banger of an episode.
Joe Peisal, who heads macro research at Trivium, is the guest for the first half, and they do this thing that they’re going to be doing every month or so, just looking at the macro numbers. But this one sort of looks at just – not just macro numbers for Q4, but for the year. And it’s a great survey.
The second half, though, features Danny McMahon and Corey Combs, who are both absolutely brilliant.
They are both really brilliant. It’s on, you know, why China is facing headwinds on boosting capital expenditure, which, if you follow the Chinese economy, you’ve probably heard, dropped really, really precipitously in the last quarter. So check out those shows.
I’m a neophyte soul when it comes to the Chinese economy, but I’m always interested in learning. So these guys have taught me just enormous amounts.
Anyway, Ryan, great to have you on again, man. And this is going to be a very Brookings-heavy month because I’m going to be talking to your colleagues, Kyle Chan and Patty Kim about the work of theirs recently.
“Delighted to hear it, and thanks for having me on, Kaiser.”
Thank you. You’ve been listening to the Sinica Podcast. The show is produced, recorded, engineered, edited, and mastered by me, Kaiser Guo. Support the show through Substack at SinicaPodcast.com, where you will find a growing offering of terrific original China-related writing and audio.
Email me at SinicaPod@gmail.com if you’ve got ideas on how you can help out with the show. Do not forget to leave a review on Apple Podcasts.
Enormous gratitude to the University of Wisconsin-Madison Center for East Asian Studies for supporting the show this year. And, of course, huge thanks to my fabulous guest, Ryan Haas, who is always a favorite, fan favorite, my favorite.
I’m really – thank you, Ryan, once again. Thank you, guys. Thanks for listening. We’ll see you next week. Take care.
]]>The China Global South podcast is supported in part by our subscribers and Patreon supporters. If you’d like to join a global community of readers for daily news and exclusive analysis about Chinese engagement in Asia, Africa, and throughout the developing world, go to ChinaGlobalSouth.com/subscribe.
Hello and welcome to another edition of the China Global South podcast, a proud member of the Sinica podcast network. I’m Eric Olander. Today, we’re going to get an update on the state of the Belt and Road Initiative.
Now, for the past several years, we’ve been hearing that the BRI is spent. The Chinese have run out of money and Global South countries that were the destination for so much of that investment simply can’t afford to take on more debt. And even the Chinese themselves have tried to change the narrative to make way for what was supposed to be a new, more austere era. Remember all of that talk about small yet beautiful? In Chinese, it’s called Xiao Er Mei (小而美). That was the line that they told everyone about smaller, more affordable, less risky BRI projects around the world.
Well, the data tells a very different story. BRI engagements last year actually reached an all-time high of more than $200 billion. Construction projects increased by 81% and investments surged by 61% compared to 2024. Energy engagements, especially in the fossil fuel sector, were very, very hot in 2025. And while the U.S. may have soured on Africa, Chinese investors haven’t. The continent was the top destination for BRI engagements anywhere in the world last year.
All of this comes from a new report published by the Green Belt and Road Center at Fudan University in Shanghai and Griffith University in Australia. Our old friend, Christophe Nedepil, is the man in charge of the project and joins us today from his office at Griffith University.
Good morning, Christophe, and welcome back to the show.
“Good morning to you, Eric. Great to be here and good to see you.”
It’s wonderful to see you and to get these surprising numbers because, again, we had heard that the BRI was all but done. Even the Chinese themselves were trying to brace us for a much more austere era. Your data says otherwise. What do you explain for this big surge of construction and investment by Chinese stakeholders?
“Yeah, I think that was a very big surprise already. When we were tracking the data, this level of commitment that we’ve seen in 2025, it’s something that we hadn’t expected and obviously hadn’t seen before.”
We’re at levels of BRI engagement and engagement, again, there’s construction contracts and investments—construction contracts where Chinese construction companies, like particularly state-owned enterprises, take the lead in implementing a large project. And investments are much more where the Chinese are investing their own money through equity investments, so they take ownership.
Now, these levels are more than double from the COVID years. So it is quite impressive. And you mentioned the energy engagement over $90 billion, and that over $90 billion is more than we’ve seen. Just kind of only the energy engagement is higher than we’ve seen during the COVID years. So this level of engagement is really something that is quite surprising to us.
I think there are a couple of explanations, of course. In the COVID years, 2020, 2021, and 2022, there was this whole idea of Xiao Armei, like small, yet beautiful, which was, I think, very logical. There was a lot of global risks. It was difficult to make deals. It was difficult to travel. And so the projects overall, the project volume decreased.
And now, really, this uptake in a still very volatile world, but with massive deals, scales more than $10 billion for single engagements.
So I think the largest one, also outside the BRI, is $37 billion by TikTok in Brazil. But we’re also tracking outside BRI. So we’re not just tracking BRI. We’re reporting on BRI, but in all the massive engagements:
And this $10+ billion engagement, we’ve not seen before. This is a new level of BRI engagement that I think is quite interesting to observe, and we’ll see whether that continues over the years to come.
So a lot of us were surprised, not only because of the size of the numbers but also the timing of it. Coming in 2025, when Donald Trump comes back into power, the international system goes into disarray, is there any connection that you can see in the data between the events that have been happening, say, within the new Trump era—that is the disruptions? Do they see an opportunity to move as the United States is pulling back from the world? Or are these just more coincidental in terms of the timing?
“I think there’s both. These projects take a while, particularly large-scale projects, take a while to negotiate.” This is not something that the Chinese are able to do with their partner countries in months. This is usually maybe a year in the making. So not everything that we have seen in 2025 was agreed to in 2025.
Now, what we know, of course, that over the last years, and this is not just the Trump era, this is also Obama era. And there was supply chain diversification, supply chain de-risking, with manufacturing plants being constructed in countries outside of China in order to reduce the tariff burdens from exporting directly from China, so rather exporting from other countries.
And there, of course, then came Liberation Day in April 2025, and with the massive increase of tariffs around the world to the U.S. And again, some of the Chinese companies have actually reacted quite quickly and also, for example, scrapped investment decisions for manufacturing from Vietnam and brought them to Morocco or other countries that have lower tariffs. So there’s still a lot of movement around in terms of the investment decisions, and that is also driven, of course, by geopolitics.
Let’s go back to energy. You mentioned that that was one of the major investment surges of last year. In fact, it was the highest of any period since the BRI’s inception at $94 billion, more than double what it was the previous year back in 2024. Give us the profile of these energy investments, because we had heard that the surge in Chinese investment overseas was in solar panels and new energy. But it seems to get these numbers at $94 billion, you’re going to have some of the older energy modes in there as well. Tell us a little bit about what happened in the energy sector.
Yeah, so I think kind of one of the quotes that have been picked up, it’s like
“2025 was the dirtiest and the greenest year in terms of energy engagement.”
And that’s true in absolute terms. So the overall engagement, as you said, increased quite a bit. It is particularly driven by oil and gas related engagement. These, for example, in Nigeria, the gas energy industrial park, there are a number of other fossil fuel engagements across the region.
So fossil fuel engagement actually has taken by far the majority, I think 75% of the total engagement is related to fossil fuel. And that’s a very high emitting energy engagement. And this is, so I remember in 2020, we celebrated that green energy or renewable energy has broken the 50% mark of the total energy engagement. And we’ve been backsliding since then in terms of the share. So that’s a worrying trend in some ways, particularly if we want to talk about a green Belt and Road Initiative and China’s green engagement.
Now, at the same time, the green energy engagement also increased to record levels. So that’s why we can also say it’s been the greenest year. And so that’s particularly in solar construction, but also in solar and wind construction, as well as in battery storage. As a broader kind of engagement portfolio that the Chinese have compared to previous years.
What’s important to note here, and I think we’ve also discussed this previously, Eric, is that we’re not looking at exports. So China’s green energy-related exports, solar panels and wind, whatever Pakistan, for example, imports 19 gigawatt of rooftop solar, this is not captured in the data. But because this is just pure export, we’re not capturing export. We’re capturing construction engagement and investment. And again, in the export space, China’s green-related exports, of course, are also increasing. And there’s great other reports out there that look at that.
Do you get a sense that in the fossil fuel sector the Chinese are building infrastructure and connectivity for exports from other countries to China? Or is this building coal, gas and oil infrastructure for these countries to use themselves or a mix of both? How does that break down?
So we don’t know exactly what it is used for each single project. What we see is that a lot of the fossil fuel engagement is indeed through construction contracts. So where Chinese construction companies just have a very strong expertise in
where Chinese companies potentially, either through a government-to-government contract or even through open bidding, have offered the most competitive price and therefore get chosen to lead this implementation.
And it might come with some Chinese financing, but it also might just come with local financing. What’s interesting for the Chinese construction companies is that a lot of these projects are very well-financed because you have the fossil fuel that in the end generates revenue. So you can be pretty sure that you’re going to get paid back for the construction that you do. And that’s different, for example, probably we’re going to talk about it in road infrastructure, which is public infrastructure, where there’s not such a strong revenue model. And therefore, the risks for the Chinese construction companies are much higher.
Again, fossil fuel, very clear. You’re going to sell the fossil fuel. You’re going to make money. And then you can pay back the Chinese construction companies. And so it’s a very lucrative business also for the Chinese.
That seems to be one of the trends that you’ve been following over several years now: the types of infrastructure that the Chinese are financing and building that used to be railroads, roads, things that we would call public goods, are less prominent today as opposed to telecommunications networks, fossil fuels — things that the moment you turn on, revenue starts coming in.
So that debt sustainability issue becomes paramount in what the Chinese are funding because, obviously, a lot of the countries where they’re doing these activities are having debt issues. So they’re looking for projects that are revenue generating right from the start. Is that a fair assessment?
I think that’s a very fair assessment.
So in 2019, the Chinese published the debt sustainability guidelines. That means for companies to evaluate whether they’re going to give a loan or work with a country to build an infrastructure project and look at the country’s profile, whether they are able to pay back the debt and whether they’re actually exacerbating the debt issues of that country.
Since 2019, this debt sustainability framework exists, and so that was before COVID. Then we saw during COVID that a lot of the global South countries were subject to a lot of sovereign debt issues. That impacts Chinese construction companies quite severely because, in the end, it’s the construction companies, if they took out loans for building, let’s say, a coal-fired power plant or a road project, and whatever country — Pakistan, for example — does not pay back the loan or does not pay back the loan in time, who’s going to be paying the loan?
And in the end, it is actually often the construction companies that have to shoulder some of the banks necessarily, but it’s the construction companies that have to shoulder the risk.
So there’s a very clear risk management necessity to understand:
I was surprised that Africa turned out to be the top destination last year for BRI engagement, $61.2 billion, an increase of 283%, largely by this big project in Nigeria that you referenced, $24.6 billion.
Just to be clear, is that project in Nigeria:
Because sometimes it’s not clear.
It is so true, Eric. So we are trying our best in our data to distinguish project-level commitments. I think every database is running into the same issues. We don’t track money. We track announcements of projects by two independent sources where possible, or a stock market announcement. So we try to be as rigorous with our methodology as possible.
There are different levels of commitments that we track. This one is, I think, more than an MOU. It’s agreed. We have the location. We have the amount. We have both sides’ agreement.
In the end, I believe this project will change and will evolve. It’s not kind of… The design phase is definitely not finished from what we can see. So there’s a lot more work that needs to come to make this project actually real. But the commitment is quite explicit from both sides and confirmed. And so that’s why we were willing to include it in the database.
Yeah, the report also said that part of the reason for this surge in Chinese engagement in Africa was because of potentially, again, just a theory, because of the lower U.S. tariffs that African countries received traditionally through the African Growth and Opportunity Act (AGOA), which is now in the process of being renewed through Congress.
By the way, something very interesting on the renewal of the African Growth and Opportunity Act.
But you said that there might be some connection between lower U.S. tariffs and the surge of Chinese BRI engagement. Tell us a little bit more about that. If a Chinese company wants to export to the U.S. and is in the process of making an investment decision, the logic is, of course, that the Chinese company will look at countries that have a tax regime that is favorable or a tariff regime that is favorable to them to be competitive against other competitors. It might be sitting in a country with a high tariff regime. So these investment decisions are just normal. I don’t think that any country or any company would not make those.
What’s interesting with the Chinese, and I think there’s an upside and a downside to that, is that China’s speed, making quick decisions, being able to build factories very quickly, and to churn out the products very quickly, is an opportunity, I think, also for host countries, for BRI countries, to attract specific types of investment.
The downside is that, and the downside is that, once the regime changes, and maybe there’s some issues, and maybe another opportunity for the same Chinese company, there’s also a risk that the facility will be abandoned very quickly.
Now, early on in the BRI, back in the 2013-2014 era, it was a lot of Chinese state-owned enterprises backed by Chinese policy bank loans that were going out and doing these big deals, these huge projects. We saw that run-up of lending that peaked in 2016, and that’s gone down. And the Chinese private sector back then played a secondary role.
Over the past couple of years, as we’ve talked to you, one of the things that we’ve noticed is that
Are you seeing that in the data for 2025 as well?
Yeah, so definitely for the investment side, it’s mostly private companies that are leading the fray, and it’s interestingly also a lot of these new tech companies that are both in kind of the IT tech, like TikTok and Alibaba, as well as in the green tech space, like Jinko Solar and other green tech companies that are leading the way.
These are private companies that are interested in:
- being closer to their customers
- diversifying their supply chains
- de-risking their supply chains
- going abroad
And that’s also now the capacity, management capacity, and technological leadership to be actually a really, really attractive partner in host countries to set up factories.
And that’s not just in BRI countries, that’s also in a lot of developed countries, and those are trying to attract battery manufacturers from China, because this is state-of-the-art technology, very different from the early phases of the BRI, where such technological leadership just did not exist.
From that perspective, I’m always very, very impressed, and I think the rapid emergence of these technological leaders in China over the past couple of years has very much flipped kind of our logic, what type of investment we want to attract.
It was at the beginning, of course, the Chinese wanted to attract Western technology, and now it is often the case that everybody wants to, particularly in the green space, attract the Chinese technological leaders to set up shop.
In the construction engagement, it’s still a lot of state-owned enterprises that are very engaged abroad. So, these are real leaders in driving these construction engagements.
What’s, I think, also clear is that state-owned enterprises have a different mandate, particularly of spending their own money. Now, construction engagement really brings them in money. That’s just revenue. You’re a service provider.
For investment, you have to, of course, use your own money, and Chinese state-owned enterprises might have a mandate to also invest domestically to create jobs. And it also, kind of, their financing modalities are quite different. Their approval processes are quite different from private companies. And so, their ability to invest abroad has also changed over the last year.
So, that’s, I think, why we’re also seeing less state-owned enterprise investments compared to the previous years.
As we look forward to 2026, which obviously is now underway, some of the trends that we should watch out for are probably:
And mining is something we didn’t talk about, but that was one of the areas that was also showing a lot of activity.
Do you expect Africa to continue to be a main focus, or will the Chinese look elsewhere to spread out some of those investments?
Man, that’s always the golden question, looking into the future. Obviously, we don’t know. I always start with that.
The trends that we’ve seen over the last years, I think, can continue. So, I think we’ll see even more tech-related engagement.
And we’ve seen this tech-related engagement, not just in developed countries, but really in emerging economies. I believe that this will continue. There are a lot of opportunities for the Chinese to set up shop. There is a lot more capacity for the Chinese in the management skills to do so, to manage all the local staff. So, I think this is really a learning. And therefore, I think this trend will continue.
In terms of the mining, there’s a very clear engagement across the world to kind of own more mines, to kind of also use this accelerated need for a lot of the transition minerals to utilize on this trend. And it’s not just the Chinese. It’s also, of course, the Australians and other countries that are trying to get their mines and their processing in order.
In terms of regional engagement, there will be a lot of kind of up and down. So, I always believe that kind of one year does not give a trend. And so, this year, of course, we saw a lot of Africa engagement. In the previous year, we often saw a lot of Southeast Asia engagement.
I think the only one that has been very constant over the last years is actually the Middle East, where there has been just a very strong engagement across a number of different sectors. It includes:
So, I think this Middle East engagement has been very strong. Also, in countries that are often not seeing a lot of Western engagement, it includes Iraq, Afghanistan, where the Chinese had some good engagement.
Again, Africa, Southeast Asia, I think this is always kind of up and down. I’m not able to see a very clear trend where it will go over the next year.
Okay. The report is the China Belt and Road Initiative (BRI) Investment Report 2025. It is by far the most authoritative report on the trends related to the BRI, where the money is going, what they’re doing with it, and who is actually engaged.
It was prepared by Christoph Nedepil, who is the director of the Griffith Asia Institute at Griffith University in Australia, and the acting director of the Green Finance and Development Center that’s part of the School of Finance at Fudan University in Shanghai.
Thank you so much, Christoph, for letting us know about everything that’s going on. We’re looking forward to talking to you later in the year to get an update on how things are going in the first half.
You do these reports every, I think, two or three times a year, correct?
“Every six months.”
“Every six months.”
So we’ll talk to you over the summer to get an update on how the first half of 2026 is going.
Thank you so much for joining us.
“What a pleasure to be here again, Eric.”
“Always a pleasure to see you.”
Thank you, Christoph, and thank you, everybody, for joining us today.
We’ll be back again next week with another edition of the China Global South Podcast on behalf of everyone around the world at the China Global South Project.
Thank you so much for listening and for watching.
The discussion continues online. Follow the China Global South Project on Blue Sky and X at China GS Project or on YouTube at China Global South and share your thoughts on today’s show.
Or head over to our website at ChinaGlobalSouth.com where you can subscribe to receive full access to more than 5,000 articles and podcasts.
Once again, that’s China Global South.
Once again, that’s ChinaGlobalSouth.com.
Thank you.
]]>The China Global South podcast is supported in part by our subscribers and Patreon supporters. If you’d like to join a global community of readers for daily news and exclusive analysis about Chinese engagement in Asia, Africa, and throughout the developing world, go to ChinaGlobalSouth.com/subscribe.
Hello and welcome to the show. I’m Eric Olander. Today, the fallout from Canadian Prime Minister Mark Carney’s speech at Davos is still reverberating.
If you recall, he declared that the old U.S.-led international order is dead and called for middle-power states to work together to form a new coalition.
Not surprisingly, his remarks were not well-received in the United States, but they sparked a lot of conversation in wealthier middle-power countries like:
We’re reading a lot about that right now in international media coverage.
But we haven’t heard much at all about what all this means in smaller, lesser-developed countries in Asia, Africa, and Latin America. The dynamics are very different and oftentimes because they are much more vulnerable due to their size. Oftentimes, it’s poverty or weak governance that are factors that play into all of this.
So we’re going to focus on a fascinating report that came out last October focusing on how small states in South Asia are navigating this new multipolar world that we’re in.
What’s interesting is that the dynamics of what’s happening over here in Asia are very similar in many ways to the challenges that smaller countries and other developing regions are also confronting. There’s an opportunity here to apply learnings from one region to another. But of course, not in all cases, and there are a lot of differences.
The report I mentioned looks beyond just the U.S.-China competition, but also includes India in the mix as well. And that’s something important in certain parts of the world.
I’m thrilled to have two of the lead authors of the report join me today for our discussion.
Zagar and Mandakini, thank you so much for taking the time to join us today.
Thank you for having us.
It’s great to have you today, and what a great time to have this. When you wrote the report back in October, you could never have foreseen where we are today.
Before we get started looking into the report, I’d like to get both of your perspectives, both from India and from Nepal, on the Carney speech. Whether you think the message he signaled is as important where you are as it’s being discussed in Europe and parts of industrial Asia.
Sagar, let’s start with you, and then Mandakini, I’d like to get your take on that.
Yeah, so it’s like that moment when somebody suddenly screams from the sides, you know, the emperor has no clothes, right?
And so, in that sense, the existence of the U.S. hegemony was well understood at all levels, at political levels, at sort of in financial domain and otherwise.
And the average Nepali cannot buy or sell anything without first changing their currency into dollars. And so, the presence of the dollar is quite overwhelming everywhere.
But for the immediate stakeholders, which is the foreign policy establishment in Nepal and those who keep track of the issues like these, it was like, well, we all knew. It’s just that there is an open admission.
And in that sense, even in that speech, the precursor was that, well, we all knew, but, you know, at the same time, we never quite mentioned it or openly confronted the U.S. in this fashion.
And so, there was some, let’s just say, a quiet celebration that the truth is out, right, from that angle.
But for countries like Nepal, you know, which is right in the middle of India and China, it’s got only two neighbors, China to the north and India to the south. Both are emerging giants, disproportionately larger than what Nepal is.
And so, therefore, it lived in a different geopolitical setting where the U.S. mattered, of course, because it overwhelmingly matters everywhere, and to a certain extent, particularly as a sort of developing partners, and Europeans also mattered.
But beyond that, Nepal has always a predominant concern about what happens in China and India rather than elsewhere.
Mandakini, the India reaction has been very interesting in part because India has seen this dynamic play out before as well. India, during the Cold War, very skillfully played both sides.
And so, I’m wondering if the reaction in New Delhi was similar to what Zagar was hearing in Nepal.
Well, I think for one thing, I’m not sure that it actually made the frontline news. I think it was buried somewhere in the newspaper. And, of course, I heard about it and was very curious to hear what he had said. And when I heard it back, I was actually a little underwhelmed. Underwhelmed in the sense that what he was saying was not really new.
I think countries, developing countries, middle-income countries, countries which are kind of small island countries, have been talking about the structural inequalities that they have been seeing in these international processes, whether it’s the WTO, the World Trade Organization, or on trade, or financing for decades. And I’ve been calling it out quite vociferously.
And I think India has been one of those countries, South Africa, Brazil. You know, the Prime Minister of Barbados, if nobody heard, is absolutely fabulous. I mean, I think she calls out…
Hypocrisy, she calls out quite a bit. She calls out the hypocrisy.
I think what was interesting was the fact that, as Sagar said, for the first time, you had a Western democratic leader actually calling it out and saying that,
“Oh, you know, the post-World War institutional structures, this rules-based international order that has been shoved down the throats of many countries is unfair in many ways.”
And that larger, more powerful, more financially powerful countries for years have been pursuing their own foreign policy or diplomatic economic imperatives with a lot of impunity. And it’s been the kind of hush-hush secret that everybody has kind of gone along with.
So I think it was a bit of… Yeah, I think that the reaction, I think, just not only from India, but many countries in the global South was,
So I think, for example, just to give an example, with the rise of China, as Sagar mentioned, the concerns about China’s expanding footprint across the world has been…
It was such big news for the last decade. It led to the Indo-Pacific becoming a new geographic construct. The Quad alignment between India, USA, Japan, and Australia came as a result of that. All a bit focused around very much controlling China’s strategic rise. And in fact, even Canada came up with an Indo-Pacific strategy for China. And now you have Mark Carney saying, okay, you know, we’re willing to talk to China.
So, I think India very much on…
It took over to your point around the Cold War, which is, you know, when you had the US-Russia, the tensions rising, particularly of the last couple of years. Trump wanted India to stop buying Russian oil. He still wants us to stop buying Russian oil. And I think India has been more muted about it now. But the foreign policy position was like,
“Look, we’re going to exercise our strategic autonomy and buy oil from where we can, because we’ve got, you know, our economy needs to grow.”
And India has actually done a lot to respond to Trump’s demands as well. But yet now we have some of the highest tariffs being imposed of all the countries in South Asia. So I think calling out that kind of double standard is something that countries have experienced for a long time. But now that it’s coming to bite the West, I think there is more open acknowledgement target.
Yeah. And he even acknowledged that they knew that this was a flawed system, but went along with it.
Just very quickly before we get on, I mean, you were being very polite that it was saying it’s from a wealthy or G7 country. Is the fact that this is coming from a white man different?
Yes. Because we don’t, I mean, the whiteness matters here.
Oh, it does. I was trying, I was wondering, should I say it or not? But yeah, it does matter.
No, no, no. Let’s kind of be, take, you know, be as direct as you can.
I mean, 100% it matters. I think, you know, the fact that a white person who is, you know, the leader of a G7 country saying,
“Oh, you know, it’s unfair and it’s unfair to Canadian people.”
You’re like, well, what about the millions of people south of the equator who have been saying it’s been unfair for generations?
So I think there is definitely a factor. And I suspect it would not have made such mainstream headline news had it not been a white leader who had said it, a white male leader. I mean, if Modi said it or if Modi or she said it, people would have been like, yeah. And they have. Modi, she, the prime minister, Barbados, the BRICS countries, all of them.
If you Google it, ChatGPT, you will find statements from them going back decades, which would have said something to the effect that the existing world order is not fair.
You know, there’s a similar phenomenon going on in the United States where white people are shocked, shocked that the police are abusive and that even video recordings of police brutality… Against white protesters in places like Minneapolis and killings now of white people and brown and black people, many have been saying this for decades, for centuries actually, that the police have been impartial. So again, this is a reckoning happening both inside and outside the U.S.
As much as I’d like to continue that line of our conversation, I want to get back to the report that you guys worked on last year. Now, it focused on three countries in particular: Bangladesh, Nepal, and Sri Lanka. You also had some insights included in it from Bhutan and Maldives as well.
Sagar, let’s talk a little bit about the understanding that a lot of countries have where we hear the top line, which is they don’t want to take sides between the various powers. And as you pointed out, in Nepal, we cannot make this only about the U.S. and China. Obviously, India plays a very important role.
You also wrote in the report that they don’t follow the textbook strategies for hedging because there’s the impact of domestic politics, there’s regime survival, all sorts of other factors. Let’s start at the high level about how these three countries in particular are managing these rivalries and what we should take away from it.
What we are essentially bringing out in that paper is that, look, countries are—it’s difficult to say countries are rational actors because countries are only as rational as their ruling establishments are rational, right? And it’s like what you see in the U.S. right now.
Like you can’t call the U.S. behaving rationally or irrationally. It’s more like Trump and his coterie behaving rationally and irrationally. So that happens in smaller states too.
You’ve got the ruling elites who have a particular interest. They would want to extend the legitimacy to rule as much as possible. And in that process, if China is a resource, if China’s influence is useful, then they would be more than happy to take it.
You see this in countries like Maldives, where Maldives has periodically, election after election, either become very close to India or very close to China. Other states have sort of, in some ways, tried to balance it.
But what we are arguing is this balancing act is really, really difficult because it’s never—the foreign policy positions are never derived from a broad, national, consensus-based interest determination.
These things happen at the will of the ruling elites, and it can go in any direction. That makes it all the more risky.
Mandakini, Sagar gave us a really nice kind of setup for this. One of the things that we’ve seen is that in Sri Lanka, Maldives, Bangladesh, and certainly Nepal, there’s been this flirtation with the major powers in the region to varying levels of success.
But again, talk to us about this question of the interests that Sagar brought up. Sometimes the ruling elites’ interests are not necessarily aligned with those of the population or the foreign policy. And as such, they don’t necessarily behave rationally.
So if we want to look at how these countries are managing these rivalries, give us a little bit of your insights of what you found on the research.
Well, I think it’s useful to think of it in sort of like an analogy, right?
I think we often forget how strongly that legacy of partition—both in terms of the division of land, people, and resources—has truly affected the way in which states in the region actually see each other and are able to engage with one another.
So it’s kind of like when you divide land amongst your, if you were to divide land amongst, you know, your five brothers of five men and women. And it’s been fundamentally unequal in some instances. Some geography was traded, some people got left to Bahrain on one side.
Those wounds, I don’t think, have ever really knitted.
So India has, and the region has a baggage which it carries, which I think very often plays very emotionally into foreign policy decision-making.
And very often, by the political parties in different countries, in particular moments of either political upheaval or economic hardship, it plays into decisions that they might take with respect to:
- Choosing a particular infrastructure project from India versus China
- Taking a particular line of credit or a particular loan
So what I’m trying to say is that engagement with India always comes with a certain degree of historical baggage, one of which also is this idea of it being a regional hegemon and behaving like a big brother. It’s something that India has been accused of for decades, and I think justifiably so.
But at the same time, it’s kind of like that big brother who you hate, but you love to hate. And we all know we love to hate him. But in a time of crisis, you know that big brother is the one that’s going to come.
So in the instance that during the COVID pandemic, when the whole world locked down, it was India that actually manufactured vaccines and was the first to provide them to a lot of countries in its neighborhood. But to be fair, until India shut down its own vaccine manufacturing, the rest of Asia could not get drugs from India. So there are limits to that. And that exposes the risks, though.
So we in Vietnam were counting on India to provide vaccines to us. Now, the West hoarded all the vaccines for themselves. But when India made a decision in its own interest, at the expense of everybody else, it exposed the asymmetries in these relationships.
Can you speak a little bit to the imbalances that exist in these great power rivalries?
When you’re sitting in Nepal and you’re relying on India, you’re up to the whims of what happens in New Delhi, and that’s it. Like the vaccine during COVID. I mean, I think it’s not just the vaccines, right?
And I think would be a rational policy choice for any government in that moment to diversify options.
But I think what the paper is also showing is that those decisions sometimes are genuinely reflective of what the country needs at that point of moment. Sometimes it’s to do with just servicing the interests of the ruling political elite, for example, right?
So that hedging sometimes works and sometimes it doesn’t. That balancing works sometimes, but it doesn’t.
The lesson I think for countries like India is that, you know, also the geography and South Asia has changed in the sense that, very often you’re looking at a population of 2 billion people. The median age in India is 20, not India, in the region is 27. That means that’s a young, very young population, all of whom are looking for jobs, all of whom have social media, and they’re seeing a lifestyle which they all aspire to.
So there’s a lot of pressure on local governments, on countries in the region to provide for their young voting elites and middle class a lifestyle that they aspire to. And the question is:
And India, unfortunately, has done a terrible job of opening up its markets to its neighbors. And so they will look for markets elsewhere. They will look to send their labor elsewhere because India, I mean, the region is famously called one of the least integrated regions in the world, right?
So unlike ASEAN, which is quite a well, you know, really well-functioning, to some extent, regional unit and political bloc, mobility is really hard in South Asia. You know, people can’t even visit relatives across the border.
And what you’re saying, I think, will resonate with a lot of people in Africa where mobility is also an issue and also a very young population that is looking to upgrade their lifestyles and certainly against what they see in TikTok, but also just in absolute terms as well.
Sagar, this question of the great powers, the U.S., China, and India, and how they’re being perceived. When we were in Indonesia a couple of weeks ago, we met with some senior stakeholders and they explained the relationship that they have with China as one where Indonesians don’t look at China as an ally or a threat, but an opportunity. And what they said was,
“This is basically a conditional relationship. The moment it ceases to be an opportunity, they will look somewhere else for opportunities.”
How do you think the smaller powers in South Asia, especially in places like Nepal, look at the major powers, all three of the major powers, in that same way as Indonesia? Or do they see it differently?
It is more or less the same as Indonesia. China is an emergent actor here. And then it comes with all these goodies. It’s an opportunity, right? But what the Chinese trajectory is of a kind that will probably not stop being an opportunity for some time to come, right? And that’s largely because how well it has established itself in the technology front, right? Like you, in the whole world as some anticipation that AI, for instance, would be part of their economic engine or a sort of a new window for innovation in all economies.
But look at how AI is developing, right? If in the entire world, there is this particular space in the US, in Silicon Valley, where seven companies have invested more than a trillion dollars in that technology. And for that technology to become ever affordable or for any other country to sort of think of coming up with their own AI ecosystems is completely impossible from cost-wise, talent-wise, and so on and so forth.
So while there is almost a preoccupation among the seven giants as to who beats who, China is quietly putting ecosystems, the entire AI ecosystems, that’s the hardware and the model and, you know, lock, stock, and barrel ready to be sold at much lower prices to any buyer in the global south, right?
So that’s what they did with the cell phone industry. That’s what they’ve done with the EV. So if you just look at these two products with which we already have prior experience, which is EV and cell phones, now think about if at a much lower cost, companies, governments, militaries across the world can buy Chinese-produced almost break the seal, open the package and start running kind of AI modules, right?
So there are opportunities like that, right? And even in terms of financing, right? So it’s easy to say we can live without the U.S. But the reality is, U.S.’s current annual budget is 1.5 times larger than the Indian economy, right? So you can’t escape the influence that comes from that kind of money, right?
From that angle, and then for most South Asian countries, India’s market is as in, why is the EU in India today? And Canada as well was there as well. Canada as well. So they’re in India. And that fact is not unnoticed by India’s neighbors. Like, what about us, the little guys here, right?
So from all of those considerations, I don’t think China is, at least not foreseeably, going to weaken in terms of what goodies it has to offer. And from that angle, the balancing, edging, sort of thinking about what lies ahead in future will continue to make geopolitical calculations difficult.
Mandakini, the points that Sagar raised on AI and the goodies relate to oftentimes infrastructure. And infrastructure becomes a very important part of the dynamics of great power management in these parts of the world.
The U.S. has sought to become a bigger player with its various initiatives that it’s brought out over the years with the DFC and others to counter the Belt and Road. India is a big infrastructure builder in the small states that you guys covered in your report. And of course, China with its Belt and Road initiative, particularly in places like Bangladesh and Sri Lanka.
Talk to us about the importance of infrastructure as a vector of the great power competition in this part of the world.
I think infrastructure is a really big one. And of course, India cannot hope to compete with China in terms of the scale and the number of projects with the BRI, the Belt and Road Initiative. Of course, India is very concerned about things like CPEC, you know, the China-Pakistan Economic Corridor, because that has a certain, you know, and the border roads construction happening around India’s northwest and western eastern frontier.
But I think when it comes to the small states that we looked at, obviously, whether if you’re a small island state like the Maldives, which is basically a bunch of island atolls, which are very inaccessible, I mean, inaccessible either by flight or by boat, right? So, infrastructure for them is a real need.
And I think it’s interesting to see how the Maldives has been very effective in kind of extracting infrastructure contracts or getting infrastructure investment from both China and India, and also by successive governments.
So a few years ago, the government of, I think it was Mohamed Yameen, had investment that he brought in from China. And then subsequently, the following prime minister brought in, president brought in investment from India.
So I think also you have this flip-flop very often between competing opposition political parties where, you know, one is openly pro-China, while they’re in government, they’ll bring in Chinese investment, that the person in opposition will be like,
“No, no, China out, India in.” And when they come into power, they bring India in.
And of course, a recent president of the Maldives came into power on a very anti-Indian stance. He wanted India’s defense support to the Maldives. We had some troops stationed there for them to leave. He came into power, the troops left, and then the following year, he came to India seeking investment.
So also a lot of these decisions are politically expedient and demonstrate certain optics to your domestic constituency, which is also important. So verity is very important to small states. The optics of being seen as being neutral, non-aligned, not pro-one party or one power or the other is actually strategically very important to them.
So I think to the point around infrastructure, I just want to make one point, which is, I think it goes without saying that, if you go to Sri Lanka, for example, China has built the most amazing fall-in highways.
So I think India needs to do better if it hopes to compete with China. But it is in many ways it can’t because of the scale, the sheer proficiency with which China has been building roads and infrastructure around the world. Africa is a good example: I had a friend who was posted in Sierra Leone, six-lane highway in like a couple of months. It’s very impressive.
I’d like to close our discussion looking forward a little bit. You wrote this report back in October of last year. And again, the world has changed dramatically since October. We see a breakdown of the international system and also of the institutions themselves. The United States has all but quit the United Nations. The United Nations is doing significant layoffs now of its staff.
What does it mean for these kinds of countries when the institutions and the systems that have been in place for 70, 80 years are not there anymore? It’s obviously a risk, but is it also an opportunity?
So for the small states, it’s a risk. It’s a risk because the number one issue comes from the fact that small states as such couldn’t or never did have much of a voice in actually making these rules in the rules-based order. But anything that promises to treat everybody equally is always good when you are a geopolitically weak actor.
And so there is a natural leaning towards a rules-based system in small states. And that being shaken is a serious problem. Because now the middle powers jostle. In the sense that when the Canadian prime minister spoke about it, it sounded good. But then there is internal competition between the middle powers.
So there’ll be a lot of jostling. And then the smaller states have a more heightened risk of being squished in one direction or the other.
The third thing about the upheavals that we’ve seen is this whole jeopardy on development financing stream. America withdrew lock, stock, and barrel. Europe, because of its own war in the backyard and failing economies and now that it has an issue with tariffs with the U.S., its biggest trading partner, the European outlook economically isn’t good.
So whatever they were able to do through EU or at a bilateral level, particularly U.K., Germany, France—France in Africa, others elsewhere—that development financing stream is also in some ways being compromised. And then now the latest news is Japan is being shaky. Japanese bonds being as cheap as they were, borrowing from Japan was a great advantage for very many developing countries in Asia where Japan has some degree of focus:
- India has borrowed heavily.
- Sri Lanka has borrowed heavily.
- Nepal has borrowed heavily.
- Bangladesh has borrowed heavily.
That’s because the interest rates were so low. Now the Japanese interest rates are growing very rapidly high.
Because of all of these changes, it’s like just because the dominoes started falling from the U.S., it has sort of taken the whole world in a sweep. And so all of those development prospects, financing and so on and so forth has become a problem for smaller states. Mandakini, what do you think?
I think it’s, you know, it’s sort of like you may, we all may have known that the rules of the game were not fair, but at least we knew what the rules were. I think now when you’ve thrown the rules out of the window, it is a situation of, it’s an unknown situation of just not knowing what will happen, what you’re going to wake up to and read in the papers tomorrow, right?
I mean, for a large country like India, yes, certainly it’s a concern. You never know whether the tariffs will go up or down tomorrow, what Trump will tweet about overnight.
And I think for small states, the existential anxiety will probably be even more. And I think one underestimates the power of a single vote in the UN, right? So even a small island state, like a small like Nauru or Kiribati or one small little island in the Caribbean, that vote really mattered in the UN.
So if the devaluation of that UN vote, I think is significant. Equally, the fact that, you know, the UN has been passing all these resolutions on whether it’s Ukraine or on Gaza, and none of them have been backed. You know, if a country like Ukraine or, you know, a large political, a big political conflict like Gaza, no one is going to come, essentially, the message is:
“No one is coming to our rescue,” right? And no one is listening.
And I think that’s very disconcerting.
I think in terms of an opportunity, you know, with this whole Don Roe or Monroe doctrine of America wanting to kind of withdraw and create its own sphere of influence in the West, that means it’s going to create a vacuum, right? Now, who is going to fill that vacuum?
It’s a player, but not in the same way. It doesn’t have that kind of military capability. But I would suspect that there’s a lot of head scratching and thinking going on in countries in South Asia, whether they are small or big about, you know:
I think we’ll see the rise of more minilaterals or trilaterals, you know, triumph groups of two or three countries trying to come together. But as Sagar said, you know, economics matters, and they will be looking at how do they shore up their economy so that you don’t see the kind of domestic political upheaval you’ve seen in Bangladesh, Sri Lanka and Nepal, right?
So it’s going to be a very tough balancing act and also maintaining your own strategic integrity as a country, you know?
Yeah. And we also didn’t touch on it, but there’s going to be bottom-up pressure as well from Gen Z where if a cigar, I mean, we can talk about that at some other future time, but, you know, Nepal was ground zero for one of the most violent uprisings of Gen Z that expressed their frustration.
So these governments are going to be facing:
And we saw these same pressures in Nairobi and in Jakarta and in other parts of the world as well.
Absolutely fascinating to start thinking about this because we’re in a whole new world now. And it is, as you pointed out, maybe this is something that, you know, many of these countries expected because they’ve seen the hypocrisies for so long, but actually talking about them now is so important given that it’s being discussed in Berlin and London and Brussels and Washington, but it’s not necessarily being discussed as much elsewhere.
So we’re happy that you both were able to join us.
Mandakini Suri and Sagar Prasai are both independent development consultants who’ve been in this business for a very, very long time. They did some fascinating research on how small powers in South Asia are dealing with this new world that we’re in.
Now, again, they wrote it last year. The new world is even more new this year. And so we’re happy that you were able to join us.
Thank you both for taking your time today to share some of your insights. We really appreciate it.
Thank you so much.
Thank you, Eric.
Thanks, Eric.
And we want to thank everybody for joining us today for another episode of the show. We’ll be back again next week with another edition. And we hope that in the meantime, you’ll check out everything that the team at CGSP is doing around the world in French and Spanish and English.
Go to ChinaGlobalSouth.com and you can get all of the coverage that really is the only place in the world that we’re doing it at this level. So we hope that you’ll give us a look and check us out.
So for the entire team around the world, thank you so much for joining me today and for watching and for listening.
The discussion continues online. Follow the China Global South Project on Blue Sky and X @ChinaGSProject or on YouTube. At China Global South, share your thoughts on today’s show or head over to our website at ChinaGlobalSouth.com, where you can subscribe to receive full access to more than 5,000 articles and podcasts.
Once again, that’s ChinaGlobalSouth.com.
Thank you.
]]>So I think that we have moved on from this idea that the Chinese cannot innovate. I think that idea is now decisively buried, and I am glad that we have buried that idea.
We’ve moved on to another idea, which is that, okay, the Chinese are much better at scaling, going from 1 to 100, whereas the Americans are still good at going from 0 to 1. And I want to suggest, no, let’s bury this idea, too, that the Chinese are able to both innovate as well as scale, that numbers are continuous. And so I think this idea of going from 0 to 1 as just exclusively the remit of the West is not empirical at this point.
And there’s also an idea of, you know, what does it matter if you go from 0 to 1 if you cannot go from 1 to 100?
Welcome to Manifold. It’s my pleasure to be here with Dan Wang at his home institution. We’re here at the Hoover Institution.
Dan, as you know, is the author of Breakneck: China’s Quest to Engineer the Future. This was a huge success. The book was listed as one of the best books of 2025 by the Financial Times.
I think it will become one of the go-to books for everyone in the United States or in the English-speaking world that wants to learn more about China and the competition between the U.S. and China.
Dan, welcome to the podcast.
Steve, welcome to the Hoover Institution. We are sitting in a seminar room at the library just to showcase our intellectual content. I’ve placed behind me a copy of the U.S. Industrial Outlook from 1991. This is the intellectual caliber you’ve reached, Steve.
A classic, classic. So I want to congratulate you for the success of your book.
Thank you.
Now, for the audience, I’m not going to try to do justice to the book because it’s quite a lengthy book, and Dan has been interviewed something like 70 times about his book.
Is that the right number?
Yes.
So I won’t try to rehash all the details of his book. I’m going to drill down on certain topics that I think my audience, the manifold audience, which is very interested in the U.S.-China competition and the development of China in the last generation — I think that audience already has a fair amount of background on this.
Dan’s book does a great job of introducing someone who isn’t an expert in this topic to the situation. But we’re going to try to drill down on a few topics that I think my audience is, and I myself am, particularly interested in. So I hope that’s okay, Dan.
Sounds great.
Okay, so one of the things, one of the themes in your book is contrasting the so-called engineering state of China versus the lawyerly society of the United States. And I think that’s a brilliant formulation, and it’s gone viral.
So just the other day, I was listening to a keynote talk that Adam Tooze, the historian, was giving for the London Review of Books, and there was a pretty big audience. It was a keynote address, and he spent a fair amount of time discussing — I don’t know if he mentioned you, but he used that exact terminology.
So you’re already affecting some of the leading thinkers in our society.
Affecting or infecting, Steve?
Affecting.
I would like to hear a little about your reflections on the book tour. So what were the things that surprised you? Were there any particular questions where people challenged you or changed your thinking on the topic?
I think the first thing I should acknowledge is that this idea of lawyers and engineers has become a memetic idea, but it hasn’t exactly been original to me. And I want to first acknowledge that this is sort of an idea that’s been more or less in the air for quite a long time.
We’ve had Bill Clinton quip in something like 1996 when he was in China to say,
“Well, we are governed by so many lawyers, you’re governed by so many engineers, we should have a swap.”
So I think what I’ve really tried to do was to take this concept and weaponize it and really try to create a bit of a framework in terms of thinking through China and write it in a way that fits my strengths as a writer, which is to write these very long essays in the form of annual letters, really to try to explain what works and doesn’t work in China.
And I think one of the things I’ve really wanted to do is to try to capture this fact that I would say that China, modern China today, is defined by two central facts, two central trends:
And so I think that is sort of the two facts I really want to get across. And as part of this book tour, I think what surprised me is the extent to which people are now quite curious about China. I think China is in the water, so to speak. Everybody has to have a view about China, whether that is something very social media driven like home prices in Chongqing—that’s been a weird mimetic trend to take off. You know, whatever aspect of weird industrial chemicals, that was a TikTok meme very briefly.
So the youths are on to this China thing. At the same time, the older elites are also very on to this thing. And so I think what I’ve been heartened by is that my book isn’t just being read by folks in the Bay Area where we’re speaking, not just folks in Washington, D.C., not just folks in New York, but also folks in Michigan, Ohio, Indiana, the broader Midwest, as well as the rest of the world. Because I think it is really important for all of us to be at least more curious about China.
I think Trump really did you a favor by starting this trade war and ensuring that every day on the front page of The New York Times and Wall Street Journal, there was some headline about China or U.S.-China competition. So definitely the timing of your book was exquisite.
Perhaps. And I think that one might have to say, “thank you, President Trump.” And maybe more of us will be compelled to say, thank you, President Trump, more on an ongoing basis now.
But I think there was something about the year 2025 in which we started the year with news about DeepSeek, moved on to news about electric vehicles, and then had more of the trade war, the 14th five-year plan. But I think that, you know, every day now there is going to be more and more China news. And frankly, that’s probably a good thing because
What is more important to the United States now than these two big trends, namely the rise of China in a more adversarial relationship, as well as the rise of technology, which has been something that I’ve been thinking about for the past decade.
One of the things I maybe not everybody in the audience is aware of is that your life really prepared you exquisitely to write this book and even prepared you for that specific formulation of lawyers versus engineers.
So when you were working in China, was it roughly 2017 to 2023?
You were working for Gavekal, which is an investment analysis firm. I mean, you were often analyzing companies in the chip industry, semiconductor industry. So you were very familiar with the engineering state, the technological development of China, the competitiveness of the products and companies.
But then I think when you were writing the book, you were in residence at Yale Law School.
So you were surrounded by the top legal minds in our country. And so you had both juxtaposed right before you, and that must have helped you formulate the ideas in the book.
So I think that that is absolutely right, that I’ve been thinking about China and technology, working for Gavekal Economics, working for my rabbi, Arthur Kroeber, thinking very extensively about China’s developments in
living in Beijing, as well as Hong Kong, as well as Shanghai, throughout the entirety of zero COVID as well.
And after zero COVID fell apart in China, I moved to the Yale Law School where I was a fellow. And that was really the contrast that set everything up, that I lived through zero COVID, in which the numbers are right there in the name, no ambiguity about what zero COVID could possibly mean.
Thinking through the history of the one-child policy, which was in part heavily influenced by a missile engineer who was one of China’s top cybernetics experts, mathematicians at the time.
And then sitting at the Yale Law School among what was, I think, the self-consciously grooming America’s ruling class at a time when this was in 23, 24, when the Biden administration recruited very heavily from the Yale Law School. We had folks like
who were all graduates of Yale Law. These were people who, I think, suborned the economists and tried to really assert their influence as lawyers to run a lot of policy.
And I think all of that came together very, very well to think that, well, you know, actually, the engineers and the lawyers, though it collapses quite a lot, no question, that is actually a pretty decent framework.
No less bad than socialist, capitalist, democratic, authoritarian to have another lens to think about this important relationship. Now, setting aside China entirely, as somebody who spends all his time talking to other scientists and talking to technologists in Silicon Valley, the idea that our nation should be run by lawyers is, to us, a shockingly bad outcome.
And, I mean, it’s something that people decry constantly in the circles that I move in. And, I don’t know, perhaps you disagree with me, but I’m curious how you think about that.
Yeah, well, there’s a brilliant politician who offered this quote that:
“Power isn’t something that is ever given to you. Power is something that you have to seize.”
Now, who said that? Was that Mussolini? No, that was someone far more ruthless. I mean, of course, Speaker Nancy Pelosi.
So, I think that, you know, if you are going to want to have power, no one is going to give it to you on a silver platter. And rather than, you know, decrying the lawyers for having seized all that power for themselves, I would say, why don’t we point the finger at the economists, let’s say, who had power tenuously, but then gave it up and they were really, it was really pried out of their fingers by the lawyers, I would say.
Why don’t we say to the engineers and scientists, can you not get better at organizing society? Can you not make a case better to the population and speak in a little bit more of a coherent and appealing manner than you presently do and try to get that power away from them?
So, that’s my challenge to the scientists. We can’t expect the lawyers to let go. You’ve got to seize it from them.
Yeah. It’s very interesting because we may have just entered an age where, if you think about it, so Trump is more of a business person, entrepreneur, although he certainly has had a lot of experience with the legal system. He’s not a legal mind himself.
The other people who are contending for power, like Elon Musk, they often control these huge platforms and can subtly influence the messages that are promulgated on those platforms.
So, we could be entering an era where vast wealth and control over these media platforms is what propels people to power, not Yale Law School.
Yeah. Well, I think the first thing to acknowledge is that Trump is an excellent, splendid product of the Lawyerly Society.
Yeah. He has abducted Maduro from Venezuela using, because this fellow is designated a narco-terrorist, charged by… And a machine gun owner.
And a machine gun owner.
And I saw some tweet about this. I’m not sure if it was actually true, but one of the charges against Maduro now is that he has been dispossessing native peoples, indigenous peoples, from land in the course of oil drilling.
So, you know, there is something, there is a legal pretense, a legal fiction upon a lot of things that Donald Trump tries to do.
I think a lot about some of these quotes. There was someone, there was a Latin American ruler who once said,
“Anything for my friends, for my enemies, the law.”
So, you know, law can be terrifying. And Donald Trump knows what lawfare is. He was schooled by Roy Cohn, who practiced lawfare extensively.
This is a man who, for whom lawsuits are absolutely central to his business career. He has sued totally everyone. He keeps suing people three times a day before breakfast. And he is very intent on, you know, flinging accusations left and right, intimidating people, and trying to establish guilt in the court of public opinion.
So the comment you made about the sort of trumped-up charges against Maduro is illustrative of why people who come from more of a science or techno-entrepreneurial background really hate the ideas, the idea of lawyers running society.
It’s because, for us, the number one thing is uncovering truth, uncovering how the world works.
So in the case of a scientist:
And for an entrepreneur, the truth is discovering:
So for us, it’s mainly that concern, not making things up or making arguments regardless of the true reality of the situation, which is the way that we think about lawyers. And so it just seems very repulsive for me.
The most jarring thing that has occurred to me in dealing with people in positions of high power in our government and other governments who are trained as lawyers is, as a scientist, I see they’re often just not very anchored in the reality of what is true or not true in the world. And they’re just making arguments on whatever side they want to make the argument in favor of.
They’re free to make that argument and often free of the actual kind of scientific or rigorous thinking that we’re used to.
Yeah, I get where you’re coming from. And I think it is a position that is easy to be sympathetic to. But I want to rebut it somewhat with this idea that what is truth? And if we can discover the truth, what follows is not necessarily something that scientists are able to comment on very well. You say that entrepreneurs are trying to figure out what consumers want.
Well, is that a truth of what consumers want? That sort of process of trying to determine all of that does not seem quite as straightforward as measuring gravity or something, right? That is, consumers are fickle. The economy is a web of relations. There’s no commandments dictating what people want to do. And so that process immediately becomes very, very complicated.
Now, what we could have are scientists asserting themselves to say that global warming is true. Okay. And I think that is a statement that we can more or less say is a factual statement now. But what we ought to do with it, I think that doesn’t necessarily follow.
Does that mean that we need to cease all economic activity and lock people up so that no one emits any carbon? That is something that lawyers, as well as economists and many other folks, as well as all sorts of humanists, have to get involved in to figure out how do we resolve these normative questions? How do we handle disputes within society?
Because the Chinese are very capable of following science to its logical conclusion, which in the case of zero COVID was to essentially lock up the residents of its biggest city of about 25 million people, or to say that, well, we have this population crisis. The solution is extremely simple, one child per couple, which ended up being enormously disruptive to everyone.
So I would say that what we want is pluralism. What we need to have are some scientists and the ruling elites. We need to have some economists, lawyers, humanists, et cetera.
Yeah. I think that in navigating a rules-based system, principled system for determining how society should react to some discovered conclusion about reality, I mean, that is where you need a sophisticated legal system, a sophisticated system, civil society.
I think all those things are quite necessary and beneficial. It’s the commitment to our knowledge of how the universe is uncertain and we have to be disciplined in practicing our discovery of what turns out to be true or what doesn’t turn out to be true.
I think good business people are like scientists in this way that they can’t come in and say,
“I am sure people want to buy X. I am not going to give them X.”
They have to sort of look at,
So there’s a process of discovery that they’re committed to, which is an empirical process of observing what’s true in the world and reasoning based on that.
And I just find a lawyer that is practiced in debate and will adopt either position and throw themselves fully into that position without concern about whether the position is actually fundamentally true or false. That to us is very disturbing. That finding of truth is our sort of most sacred activity.
Yeah, well, what you call arguing out of both sides of your mouth, I can call empathy and understanding what the other side thinks. This is something that I think the Communist Party, for example, is not very good at doing. They do not have a good sense of how other people think.
And, you know, the Lawyerly Society has created some astonishing companies. The West Coast is the only region in the world that has created several companies worth trillions of dollars.
And maybe that’s a weird thing. Maybe things are overvalued. In fact, we can be sure that they’re overvalued. But I think that, you know, it doesn’t seem to me that the Lawyerly Society has been awful at creating companies, creating products, even though lawyers can argue any side.
Yeah, I think characterizing the U.S. as the Lawyerly Society makes sense in contrast to China. But as far as the way that the Google founders had to operate or the way that Jensen Huang had to operate, the law is part of the system that they’re in. I don’t know if it’s their primary concern or the thing that made their company valuable and Yahoo not so valuable.
It wasn’t necessarily that,
“Oh, this is a Lawyerly Society and that dictated the outcome.”
But certainly, I think property rights, well-defined rules for how businesses have to operate in society — all those things I think are important and maybe lacking in China at this point.
I think that’s a fair point.
Yeah. So I want to turn to a concept called industrial maximalism, which is promulgated by something sometimes called the Industrial Party in China. And in your book, you spend a little bit of time talking about this, and I think you actually investigated, I think you went and read some of the original documents or essays that started this movement all the way back in the early 2010s and mainly online. It was an online movement. It wasn’t started by the central government. It somehow made its way into the halls of power, but started really as an Internet phenomenon.
Some people who are watching the tech competition with China very closely, both from the Silicon Valley side and from the Chinese side, might say that this industrial maximalism idea has actually won out in the Chinese government. So the Chinese government, some people would argue, is actually behaving as if they’ve now embraced this idea of industrial maximalism. And so I just want to discuss that a little bit with you.
So maybe just say a little bit for the audience about what you wrote in the book about the Industrial Party and its history.
Yeah, the Industrial Party is not a legitimate political party because there are only, aside from the Communist Party, I believe there are eight other tolerated political parties in China, but all of them must be subservient and loyal and obedient to the Communist Party. So the Industrial Party is more of an online meme movement. Maybe this is one of these early memes from the Chinese Internet when it was a relatively free space before the Great Firewall really managed to slam down.
These were people who were essentially advocating for technocratic rule in order to pursue science and technology. These were people who have understood that China’s weaknesses stem almost entirely from backwardness in science and technology, that it was invaded brutally by Japan as well as partially colonized by Western powers. And the solution is to pursue not just the bomb, not just the satellite, but all sorts of important science and technology, and really to organize society, the entirety of society, to pursue these sort of things.
Now, there are some important canonical texts about this. One of them is something like, the title is something like, study or wash dishes. One of the most interesting ones is something called the Morning Star of Linggao, in which there is this online fan fiction community that imagined sending a lot of people from the present into something like the Ming Dynasty to industrialize the island of Hainan. And so, you know, it’s a kind of an interesting little read there.
At various points, the Industrial Party has been censored. And so, some of these people have been quite interesting. They were empowered to be major voices on the Internet. And I think these creative people did not always follow the dictates of the party. And when they spoke up, sometimes the Morning Star of Linggao was interrupted based on that fact.
I think that actually the canonical text of the Industrial Party was The Three-Body Problem, which is one of China’s best-known cultural exports now. And I love The Three-Body Problem. I think that it is a remarkable piece of science fiction. One of the underlying themes was that Liu Cixin said that it created a scenario in which humanity bands together under technocratic rule, builds these enormous spaceships controlled by the Navy in order to confront an extraterrestrial threat. And that feels like the sort of prescription that the industrial maximalists are really interested in.
So, for our non-Chinese listeners, I think it’s important to give some cultural context here, which is that for thousands of years, China regarded itself as at least the preeminent civilization that it knew of on the planet. And the reading of the past couple hundred years is the idea that some barbarians with superior science and technology came and inflicted terrible humiliations on Chinese civilization. And we are now recovering from that period, and China is reassuming its rightful place as one of the leading, perhaps the leading nation in the world. That’s sort of the background to all this.
So, I think if you’re a scientist or engineer in China, it’s natural for you to attach to science and technology as the thing which the barbarians used to beat us a couple hundred years ago, and the thing which we have to perfect now to restore our civilization to its rightful place.
Do you think that’s fair? That’s in the mind of like almost every Chinese person on the planet?
Yes, I think that is a fair reading. And I think that that is instilled into the minds of many. But I would offer two remarks. By the way, I’m not endorsing it. I’m just saying that that is a story that every Chinese person is familiar with.
Yes, that is a story that even before the Communist Party, the Chinese rulers have created this story. And the Communist Party has indeed put it into the heads of everyone. And I would want to offer a little bit of nuance here. First, if these were barbarians, how do they get all this great science and technology? So maybe there is at least a little bit something else here going on that these British, Dutch, red-bearded people were able to invent a lot of great things.
So was there something deficient in China that it wasn’t able to do? Maybe, maybe not. And the other part that I would offer is that, yes, China fell behind in what it referred to as the century of humiliation, in which these Western barbarians came over and seized major parts of China.
It made the people and the government suffer various indignities. And much more serious was the Japanese invasion, which was a brutal invasion by a fascist power that really ravaged a lot of the country and pushed the state into these pretty interior remote areas in order to carry on the fight.
And what I would offer is, you know, that was the century of humiliation. Yeah, pretty bad. I agree.
But what about the quarter century of self-humiliation that the Communist Party inflicted right afterwards, after 1949, when tens of millions of people perished after various landlord struggles, after the Great Leap Forward in which a famine ravaged the land? We had the Cultural Revolution.
Afterwards, there was this other spasm of the Wanchau policy. And then there’s some things that we don’t even speak so much about. You know, the strike hard against crime campaign in the 1980s. There were various issues with the Inner Mongolia Separatist Party that we don’t even think about. That created incredible ravages.
And that China was a society, after the Communist Party took over, that suffered these extraordinarily violent convulsions throughout society.
How do we explain something like that? Was that something that the barbarians inflicted upon the Chinese? No, I would say that that was something that emerged organically from them.
And so it is fine and good to think about the century of humiliation. Let’s also consider a little bit about the quarter century of self-humiliation.
Yeah, I mean, I think that the apologists for the current Chinese government, for the Communist Party, you know, they would say,
“We had to break a bunch of eggs and we made a bunch of mistakes to get us to where we are now. But now we finally emerged in our rightful place among the leading nations.”
I think one should not minimize the mistakes that were made, the terrible suffering, all that sort of thing. But I think from a nationalist perspective or even an ethno-nationalist perspective, you know, mistakes made by your own people, which inflict suffering on your own people, somehow—
I’m not justifying this, but I think this is the perspective—are more tolerable than things inflicted on you by some alien group of people. Right? I think that’s the psychology that governs this.
So they’re willing to overlook:
A miracle did happen that in a period of one, between one and two generations, it went from absolute levels of poverty, similar to the poorest countries in the world, to being possibly the most advanced technological civilization on the planet.
And so they’re willing, these people who are nationalists, are willing to forgive all of that based on where the country has now arrived.
Yes, and I think they are able to justify this and self-justify this. And I think that the rest of us don’t have to give in to some of that. And we can question whether self-inflicted harms is somehow morally better than other inflicted harms. And do the ends justify the means, whatever, an ancient philosophical question.
They would say yes, but that is something that we can also be—we can also interrogate and be critical about.
Yeah, I think from—I mean, you and I were both raised in—
And so from our perspective, the idea is that the greater country is one where one can critique what the government has done in the past, criticize it, try to open it up for consideration, even if it is embarrassing for the current regime or current leaders. And that’s a sign of strength of the country that that can happen.
You can have dissidents. The dissidents are allowed to air their views and people consider those views. And the country is internally stronger because of that.
I think one argument that people would make on the other side is that we are rising, speaking as if I were an industrial maximalist or Chinese nationalist or something,
“We are rising from the ashes. These guys came, they fed us opium. Japanese came and killed so many people. MacArthur wanted to use nukes against Chinese cities during the Korean War. We had to overcome all that. And we’re still overcoming.”
The moment we showed some ability to move up the value chain, the Americans tried to completely crush us.
And so when this is all over and we are on top and we’re no longer threatened, Threatened by the West, then we can relax and create this beautiful vision that you have for how society should be.
But right now we’re in the middle of a war and you don’t necessarily see the war. The war is being conducted in a very serious way, not for the average person to perceive, but for people who are watching carefully, this is a war.
This is the third world war, but it’s being conducted in a more subtle way. And I think that is what some of my listeners who belong to this party, the industrial party, that’s how I think they would articulate their position. That is a view that I wonder whether they can allow for any sort of relaxation afterwards.
Do you want to believe that you are in a war and a sort of a silent war in which the West is trying to suppress you? I think that is a reasonable reading of the evidence so far, especially over the past decade when the U.S. government really weaponized China’s dependence on semiconductors.
I was covering all of these twists and turns of U.S. export controls. I’ve been on net fairly critical of U.S. export controls. I’ve written several essays in the New York Times and in Foreign Affairs saying that this will, in fact, over the longer run, stimulate China’s technological self-sufficiency.
But what I would like is for a little bit more self-confidence from the regime. If China is, as you posit, the world’s leading technological civilization, a claim that I would mostly sign on to, they’re just starting to enter that phase. They’re not fully in that phase, but they’re just starting to.
There’s some questions about:
But I think that it is a reasonable claim to say that China is, in many ways, technologically more self-sophisticated.
I just had an op-ed come out in the New York Times saying that China is winning the electrical age while Donald Trump is invading Venezuela for oil that the U.S. doesn’t need.
But, you know, in that case, why can it not gain a little bit more self-confidence and be a little bit more, let’s say, beneficent in a Confucian sense against the rest of the world?
If there is kind of this siege mentality, one of the things I think about as someone who thinks about not just authoritarian systems but Leninist systems in particular, in which I think that the Communist Party and its Leninist heritage—the leaders wake up every day feeling like this is a life-and-death struggle against the Western colonialists, capitalists, whatever it is.
And I wonder if they will ever feel like they can relax. I suspect not.
Let’s say that somehow the West is in disarray and fairly weak, that China can close the gates of the Celestial Empire and just be serenely untroubled. Do you really think that this Leninist regime can ever relax?
No, I think that these people are highly, highly paranoid. They’re trained to be paranoid. And what I want to do is to start relaxing now rather than some mythical end state that they will never feel comfortable with.
Yeah, I think this is the essential point. When China does achieve its full potential as a civilization—I mean, its economy, its technology—will they be able to transition to a different kind of government, a different sort of set of values for what’s allowed in society or not allowed in society?
You could have different views about how likely that is. Will they be so locked into this paranoid worldview that they won’t be able to get out?
And I want to say that this, by the way, this NATSEC (national security) perspective exists on both sides.
In the United States, we had this McCarthyist era, we were very worried about communism, we had the Sputnik scare.
So the question is, at what point do you get confident enough that you can sort of relax and stop locking up your dissidents and not think every visitor from this competing country is a spy?
Can they:
I think nobody really knows the answer to that. I think some people just assume China will fully recover from its bad few hundred years and then it can become just a normal country again, even though it is starting from what is currently a communist authoritarian system.
Other people would think, no. It’s very different. Like I was just listening to Dario Amadei at Davos talking with Demis Hassabis and saying:
“No, we cannot possibly give them NVIDIA chips because these people are so horrible. If they get AI first, it’s game over for the whole planet,” et cetera, et cetera. So I don’t claim to know the answer for which of those two scenarios is more plausible, but I think it’s the one, it’s what people should really be focused on, the future of a world in which China becomes possibly the most powerful country.
Yes. And I would say that, in general, the track record of America is that there have been these paranoid spasms. The Sputnik moment was a stimulus towards science in the US, but there’s been parts of the American berserk, in Philip Roth’s terms, that really deranged this country.
On the other hand, this is also a country that I think tried to build up Europe, tried to be kind or utilitarian towards Japan and Germany after it vanquished them.
That for the most part embraced, what we call the liberal international order by being open to globalization, hoping—now we say foolishly—that China will also grow into its own image, which was a bet that proved wrong.
I think we can say that now, but there, the episodes of paranoia also coexist with episodes of generosity in ways that I think the Trump administration calls “foolish generosity” at this point, given how globalization, in their view, has not worked out very well.
So coming to the current Chinese government and the nature of life in China, you’ve commented on the parity that’s experienced even by the elites.
So I guess the way I would describe it is: I think the average person in China is pretty happy with their government just because of the very strong growth and development that’s happened in the last generation.
I think if you question them carefully, they might say something like,
“Oh, there are certain topics it’s best not to talk about. We could get in trouble. Let’s not talk about that,”
and just change the subject. They have that awareness.
But in general, I think they think of their government as a good government and they have some confidence in it. I think I could be wrong about this, but this is my impression for sort of ordinary people.
I agree with that impression.
But then I think you make this insightful comment, both in some of your interviews and also in the book:
And that’s why you send your kids to Western universities, try to get as much of your capital out, buy a bunch of property in Vancouver and L.A.
So talk a little bit about what you think the world looks like for these relatively successful elites in China.
Do they not feel as comfortable about their future or secure in their future as, say, a wealthy person in San Francisco?
Who are Chinese elites?
Well, I think that in the state party context, these are people who are relatively high up within the Communist Party. These are relatively high up in the People’s Liberation Army.
Let’s check in: How are they feeling?
You know, right now, the Politburo has, I believe, now formally 23 members. The one spot traditionally reserved for a woman has been axed.
One of the two People’s Liberation Army Politburo members has been given the sack.
Potentially another Politburo member who is in charge of the weapons state-owned enterprises may also have fallen.
There’s this journal now that has tracked that the Communist Party has disciplined around a million officials last year, and this keeps going higher.
And, you know, the discipline process for a Communist Party official is extra legal. They have their own party system. It’s like being court-martialed in the military system here.
That’s kind of a grim fate.
And so, you know, that’s a million party officials disciplined out of a population of about 100 million party officials. That feels fairly substantial, both tigers as well as flies.
So, where do Chinese elites work outside of the party state?
Well, much like the U.S., some of them are working in:
How are they doing?
Among my friends who did attend Harvard, some of them knew her. So, if Harvard is not too good for Xi’s daughter, why should anyone else not send their kids to Stanford or Berkeley or Michigan or wherever else?
When people have the ability to acquire an American status, passports, university education, they generally tend to seize it. It is exceedingly rare for Americans to want to feel like they have to do the same thing.
I think it is still relatively rare among people in developing countries, given a choice between spending much more of your time in China or the U.S., that they still, on net, a little bit more, want to spend their lives in the U.S. That sentiment, I think, is easy to find.
If I were talking to some people who grew up in China but are working in the United States and I stated the case you just made, I think all of them would be familiar with it. I’m not sure how many of them would really strongly feel that they are better off, safer, or more secure in the United States than in China.
I know a lot of people who have the opposite view and feel that now is the time actually to go back to China because the companies there really are doing cutting-edge stuff. They wouldn’t sacrifice their intellectual development by being at Tsinghua versus being at Harvard or something like this.
I think it’s quite complicated. I don’t actually know the answer myself, but you could easily find people who don’t necessarily have the psychology you described, but there are many people who do have that psychology. I think that’s a key variable in what’s going to happen to the country.
So I think it’s something to monitor very carefully—the mood of the most able, dynamic, entrepreneurial people in China. Like, are they better off staying there and building that economy, or are they better off trying to get out?
Yeah, I think that’s a great observation. It’s hard to monitor them. We can find plenty of examples of people who’ve decided to move to China over the last few years.
You can also cite that, I believe it was Matt Sheehan who did this work, that 90% of the major AI researchers who published in big papers over the last couple of years are still in the U.S. and have declined to move elsewhere.
We can find numbers for both of these. But maybe, Steve, you should be the sharp tip of the spear in trafficking.
I try to spend a fair amount of time in China. Then inevitably, when I’m at one of the big labs here, I bump into a bunch of people who have China backgrounds but maybe got their PhDs here or something. I’m constantly interrogating this question, trying to see how people feel about things.
It may vary by age. It may vary by the family background of the individual.
A funny story I think I told before on this podcast is that I was visiting DeepMind. My host, who’s actually, I think, a Google fellow or has some pretty fancy title, was showing me around. I was there to talk about the use of AI in physics, using the current best AIs to help physicists do their research.
This guy was showing me around, and we encountered a group of three or four obviously Chinese AI researchers, very young people who were walking through the pavilion down at Google where they’ve put all the AI people in Palo Alto and Mountain View. Everybody’s young, and you hear a lot of Mandarin at all these labs.
My host, who is not Chinese, introduced me to these young people and said,
“Oh, this is Professor Xu. He’s helping us make our AIs more useful for physics.”
I was shocked because, first of all, I’d encountered this in Japan and Korea, but not in China. All of them bowed to me slightly, which I thought, Is that coming back? Like, students bow to teachers? Because I had not experienced that in China in the past when I visited.
But anyway, they all kind of bowed to me, and then one of them said very sincerely and naively, like a young nerd:
“Professor Xu, if you find any problem with our model, we will make it better.”
It was so naive and sincere, the reaction of this kid. So anyway, but it was like a cluster of Chinese. And these guys were involved in the most hardcore model training, pre-training of the models at DeepMind and Google.
So, like, I would love to pull them aside and say,
“Well, why are you here? Why are you not at Tsinghua? Like, what are your classmates doing?”
You know, so I’m trying to monitor that kind of situation as much as I can.
Well, I would love it if more of these, your students at Michigan State bowed to you much more actively, Steve.
What a nice idea.
So give us a sense of the buy check of the moment.
How many of these talented Chinese origin researchers that you’ve been speaking to on your big tour of Silicon Valley, how many of them want to move to Tsinghua?
You know, my best spy when it comes to this is a Chinese-American kid who grew up in California, but unlike me, actually mastered the Chinese language.
So he’s fully bilingual and after doing his undergrad degree in the UC system was asking me where he should do his PhD in AI. And so we were talking about the strengths and weaknesses of various programs, and he ended up going to Tsinghua.
So he’s actually a grad student in Tsinghua now, and he regularly reports to me about what the scene is like for these huge numbers of kids going through Beida and Tsinghua and other universities there and working with the companies.
So here it’s a little funny because the big labs aren’t that well integrated with academia, and academia is a little bit isolated right now from the frontier developments. But I think that’s less true in China.
So a lot of the academic groups are collaborating tightly with ByteDance and some of the other companies that are building frontier models.
So he reports to me that the scene in, partially for this reason, because he’s on the academic side, he’s doing his PhD, the scene in Beijing is more vibrant than the scene here on the Stanford campus.
Because most of the Stanford people are locked out. They’re not actually able to do this frontier-level work at Google or at OpenAI or at Anthropic.
But in Beijing, there’s a very fertile mix between these groups.
And so he keeps thanking me that I didn’t encourage him to go to Tsinghua, but I gave him what I thought was a realistic view of what he would find there. And he’s quite happy there.
So it’s only one data point. It’s a very important thing to just keep your finger on, to understand where is the talent.
Yeah, and the people that you’re speaking to here, how many of them want to keep actively going to China?
Yeah, it’s very interesting. Will they work for ByteDance?
Yeah, so after my stay here in the Bay Area ends, I’m actually going, I will be visiting Tsinghua for a while.
And my host is a young, new assistant professor at Tsinghua, whose prior position was as a postdoc in the U.S. at one of the top universities. And he’s quite happy to go back.
I don’t know. I mean, I think, of course, he’s super nerdy, so he probably doesn’t think at all about precarity or getting on the wrong side of the Communist Party.
He’s just thinking about,
But I don’t think it’s a completely zero-one kind of, you know, one-sided decision. Some people could decide in one way, and some people could decide in other ways, maybe even because of food.
Some people might just be like, “Wow, the food is so much better in Beijing than here,” or something.
So, and the research stuff is kind of comparable.
So, who knows? But I think it’s something we should all keep tabs on.
Let’s switch gears now.
I have a bunch of quotes, which I loved, that came from your, I think the ones I wrote down came from your annual letter.
And so, for my audience, maybe, I’d love to hear a comment on this.
Maybe one of the reasons Dan’s book was so successful is that he already had a huge following among all the people who seriously think about China and U.S.-China competition through the annual letters that he wrote.
And you always write these at the end of the year during the holidays, right?
And how many years were you writing that letter?
“Eight now.”
So, I mean, I’m sure at the beginning, maybe not that many people were reading it.
Only my mother.
Okay, only their mother.
But by the end, I know you skipped 2024, right?
But by, say, 2023 or 2025, like, everybody who is seriously thinking in this space is reading his annual letter.
They’re beautifully written.
They’re not short.
These are long, kind of discursive essays.
And one in 2025 I really enjoyed.
Do you think that helped your book when it came out?
People already knew who you were and already were familiar with your thinking?
I think that it was a slightly higher stakes letter to write because this was a, I have a book audience that is onto me now.
Yeah.
And so I’m speaking to a broader array of people who are not necessarily thinking about China.
And what I really tried to do was, for my book, was to reach the lawyer in Ohio or Indiana.
And so, now that these people, if these people are interested in my work, I wanted to hit them with a really big annual letter.
This is also an annual letter that is more driven by Silicon Valley, where I spend a lot of time now, and not just China.
Yeah.
So let me quote, I think these quotes, I think I got them from you, not from the book, but from the annual letter.
And I just want you, just feel free to riff on, you know, extend what you said in the quote, or make any comment that you think is appropriate.
So here’s a quote:
“Probably the most underrated part of the Chinese system is the ferocity of market competition. It’s excusable not to see that, given that the party espouses so much Marxism. I would argue that China embodies both greater capitalist competition and greater capitalist excess than America does today.”
So, beautiful.
Yeah, I think that, you know, there’s a lot of ways in which the tech companies in Silicon Valley feel they are very, it’s like a gentleman’s club of, you know, playing with, against each other.
Where, you know, Google’s got search, and then Amazon’s got e-commerce, and Facebook’s got everything else.
And so, they don’t really tread that hard against each other’s toes, whereas in China, everyone is fighting everyone over everything.
And you have these, I like this chart that Kyle Chan of High Capacity has created.
We got these big interlocking circles that are highly complicated.
And I think that just really shows how everybody is up in each other’s business all the time.
Whereas in the U.S., at least among the tech companies, it is much more gentle.
You know, we’ve got union protections here, where you don’t have unions, really, in China.
They arrest the Marxist organizers.
And so, you know, what sort of socialism is this?
So, you know, Peter Thiel was famous for saying that competition is for suckers, for losers.
“What you should try to do is get a monopoly, and that’s how you really create a trillion-dollar company.”
And one critique of our lawyerly society here, or our political system, is that business can co-opt government to the point where they can actually end up with a monopoly, and government doesn’t do anything about it because they have enough influence in Washington.
And whereas in China, that’s not going to happen, the government likes to see these companies beat each other up and, of course, drive the profit margins to zero, maybe to the benefit of the consumer.
Do you think that’s just a caricature, or do you think that’s a fair picture of what’s going on?
I think that’s pretty fair.
If we take a look at the Chinese equity market, I mean, that’s – this is not the whole reason, but part of the reason that, you know, the Shanghai Composite Index has mostly trended sideways for 20 years.
To my dismay.
To your dismay.
Well, I’m sorry about your losses.
But, you know, maybe it should have been –
Or lack of gains.
Investing in America, Steve.
I do both, but –
You do both.
So, I think that is part of the reason that, you know, profits are just much lower in China because it’s much more competitive and a bunch of losers, maybe.
But I think it’s great to be a consumer there.
Yeah.
That’s my – well, for Americans who’ve never lived in China or spent time there, buying power you have there is just unbelievable.
Yes.
Like, dollar for dollar, what you can buy, not just food, but electronic gadgets, cars, whatever.
Services.
Yeah.
It’s just insane.
It’s insane.
Which is another reason, actually, I won’t say his name, but one very prominent person who is in Beijing, big venture investor now, was a very prominent technologist in the United States for one of the big tech companies.
He just says, like,
“I can’t have this quality of life in the United States, even though I’m a billionaire — he’s a billionaire — even though I’m super wealthy, I can’t have this quality of life in the U.S. because I get on my phone and do this, and some really delicious food appears 20 minutes later delivered to my office.”
And, yeah, he says, I can’t have that.
Yeah.
Lower labor costs is a real thing.
Yeah.
Next quote:
“Beijing has been preparing for Cold War without eagerness for waging it, while the U.S. wants to wage a Cold War without preparing for it.”
Let me reply to your quote, or rather my quote, with another quote.
This is an apocryphal quote attributed to Viennese satirist Karl Krauss:
“In Berlin, the situation is serious but not hopeless. Whereas in Vienna, the situation is hopeless, but it is not serious.” And this is where I think a lot of the U.S. is, where they are just not a serious country in a lot of different ways.
Now, is China eager to wage Cold War?
I can say maybe they are. But I would also say that the balance of evidence has been that China grew very, very rich under an international system in which it became a major trading power without a large navy. That might be unprecedented.
You know, you didn’t, the U.K. was a major trading power within a large navy. So was Portugal. So was the Netherlands. So was the United States. China didn’t have to build a big navy in order to become a big maritime power.
The U.S. gave it to them for free, right? And so it prospered by any measure under the system. And that system has kind of turned against it. And so it is reacting. That’s all fine.
But China has mostly prospered under this system. And so, but it has also been quite cautious and preparing for energy self-sufficiency, food self-sufficiency. I spent some time in my book examining something called the Mayor’s Vegetable Basket Program, in which there’s incredible amounts of farmland around big cities like Beijing and Shanghai.
As soon as you take the high-speed rail out, you get to hit farmland really quickly. And that’s in part because the mayors, not the party secretaries, but the mayors, people responsible for logistics and operations, have to manage grains self-sufficiency, as well as various parts of vegetables and meat self-sufficiency as well.
And there’s obviously now a drive towards technological self-sufficiency.
Yes, the U.S. right now is obviously floundering to escape a system that it has very substantially built, that it is trying to figure out how to subdue China in various ways. Yeah, but it is unwilling to pay a lot of costs in order to get to a much better state.
And so, that’s why I think there’s parts of America that want to wage a Cold War without really doing enough to prepare for it.
Yeah, when I mentioned that a lot of my listeners are, in a way, members of the Industrial Party, I didn’t mean the Industrial Party of China. I meant they’re members of the Industrial Party of America. So, lots of entrepreneurs, Silicon Valley people, are astonished that we can’t seem to get our shit together.
We talk about reshoring manufacturing to the United States. The hard work of really doing it, the government doesn’t seem, or society doesn’t seem serious enough, really, to undertake that heavy lift. And it might take decades for us to get there. It’ll take decades of sustained effort. And I’m not sure, I’m not, as an American, as an American who grew up in the Midwest, I’m not sure that we still have it in us to do that. I hope we do.
Yes. Let me read a few things about innovation and manufacturing in China from you. I’m going to read a few of them and then you comment.
So, I think that we have moved on from this idea that the Chinese cannot innovate. I think that idea is now decisively buried. And I’m glad that we have buried that idea.
We’ve moved on to another idea, which is that, okay, the Chinese are much better at scaling, going from 1 to 100, where the Americans are still good at going from 0 to 1. And I want to suggest, you know, no, let’s bury this idea, too, that the Chinese are able to both innovate as well as scale, that, you know, numbers are continuous.
And so, I think, you know, this idea of going from 0 to 1 as just exclusively the remit of the West, I think, is not empirical at this point.
And there’s also an idea of, you know, what does it matter if you go from 0 to 1 if you cannot go from 1 to 100?
So, I cite the example of, you know, who was it who, you know, which American lab invented solar panels? Bell Labs, 1957 in New Jersey.
Where is Bell Labs today? Nowhere, because this thing is gone.
Who owns the entire solar industry today? It is the Chinese. And they own not only the final assembly of the modules, but also the polysilicon processing, as well as the equipment to produce all of the solar wafers and modules.
So, you know, this is something that I think that, you know, let’s not cheat ourselves out of the idea that Chinese cannot do this or cannot do that, that they’re fundamentally constrained because of the nature of their political system.
You know, the political system has gotten pretty far in terms of producing a lot of technology.
I believe that authoritarian powers can do incredible science and technology. We saw this with the Soviet Union. We saw this with Nazi Germany, in which totalitarian systems produce miracles in, let’s say, spaceflight, as well as many other things.
So, I think that the more that we can get our discourse and our level of competition to the level of sustained capacity building, which is what the Chinese have done, rather than thinking, “oh, well, they can do that, which is not that difficult, and we can do this, which is much harder,” let’s acknowledge that they can do a lot of difficult, important things, and that the United States should also get in this game in a much more active way.
If you take the normie boomer, sort of outdated view on China, okay, we could have different stages:
I think by now, people, even most normie boomers, have kind of set that one aside, and they say, “oh, well, they can do advanced manufacturing.” But the core idea, the zero-to-one, brilliant, genius breakthrough, that still has to come from the West.
And I think you’re saying, okay, even that misconception should be set aside. I’m curious, in your 70-odd interviews that you did, or addresses to big audiences of serious people, what fraction of the population is ready for that step? Are they already there in understanding that, or are they still resisting that step, or clinging to that illusion that the zero-to-one step can’t happen in China?
Maybe half-half. I think that there is still plenty of clinginess among the quote-unquote serious people. But I think that there is also a readiness.
I mean, it starts sounding a little bit absurd. “Oh, well, they can make all the toys and T-shirts, and they can make the large-capacity batteries for the EVs, but they can’t.” You know, these sort of things, statements, don’t really make sense to me.
I think that, you know, 2025 was an important year in terms of the DeepSeek moment. Maybe that scale of success was overstated, you know, there was this big market reaction to deep-seek, and that was a little bit hard to predict based on the release of a paper.
But, you know, EVs are definitely taking over, and, you know, we don’t have to over-index our expectations of the economy based on EVs alone. Transportation equipment is not a giant sector of the American economy, even though the auto sector is politically important.
But I think that people now are ready to hear this, especially given the moment we’re in with Donald Trump eroding a lot of alliances, making the world question American leadership. I think that it is good for there to be a little bit more worry about what is the American position and try to be a little bit more clear-eyed about what are American weaknesses.
This question of whether, in a less free political atmosphere, scientists can’t make big innovative leaps. I still hear that.
Yes. Do you still hear that?
Yes, and I think this is one of the sins of the lawyers, that they think that, “oh, we need so much free speech in order to produce scientific breakthroughs.”
In my book, I spent a bit of time thinking about the example of the Soviet Union, and there’s a particular historian of science from Harvard named Lauren Graham, who studied the Soviet science system. She pointed out that there were two big cataclysmic events for Soviet scientists:
Which event was worse for scientists? Actually, it was the dissolution of the Soviet Union, because that was when the money stopped. A lot of scientists had to close their labs and leave.
Where, yes, very strangely, throughout the 1930s, under Stalin, there are many, many examples of scientists who barely staggered out of gulags and then produced their Nobel Prize winning scientific innovations.
And so, you know, part of this is because Stalin funded science. And when you throw in a lot of money at science, you’re probably going to get a lot of science. At least this was true about 100 years ago. Maybe not so true today.
And there is also this view that, you know, why were the Soviets so good at pure mathematics, as well as chess, as well as a lot of science? Well, potentially, they treat doing science as an intellectual escape when they do not, cannot enjoy broader political freedoms.
So, I don’t think that, you know, authoritarianism is good for science. I definitely will not go that far.
What I will say is that money is good for science, for the most part. And right now, the Trump administration is cutting somewhat some scientific funding. University budgets are on the back foot. But the Chinese are funding a lot of science, and I think they’re going to get the science.
Yeah, I agree with you.
So, I finished my PhD in 1991, which coincided with a huge influx of former Soviet scientists to the West.
Right. The best of the best from the Soviet Union, former Soviet Union, came to the United States and competed with young Americans like me for the same jobs. So, it was a very, very vivid experience with me, and I spent many, many hours in so-called Russian seminars.
Our seminars are usually scheduled, oh, it’s an hour, and if a few people want to stay longer and talk to the speaker, they hang out, but everybody else leaves. The Soviet system, they talk for hours, and they fight until they get to the bottom of the matter. So, even the style of Soviet physics and mathematics I was very familiar with. When you talk to those individuals, you get to know those people after they become professors here and collaborate with them. They had very, very fond memories of the Soviet system.
It was a system. There were even cities, if they were adjacent to the weapons program, they would live in a secret city where they didn’t have to worry about food, they didn’t have to worry about salaries, everything was provided, and they could just focus on the science. And they had a chance to have an orchestra and play chess in their free time or write poetry. So, really did not interfere.
They knew there were certain things they were not supposed to do. They should not suddenly write a tract about why the wheat harvest was so terrible in Ukraine last year, right? Okay, avoid that topic, but if you want to learn about electrons in super lattices, no problem. So, they always felt very nostalgic for that system, and then what happened is when that system collapsed, just as you said, they literally couldn’t eat.
They had to grow potatoes in their front yard in order not to starve and just plead with Western colleagues to send them a few dollars. So, they could actually survive, you know, so they could actually survive, you know, in the Soviet Union. Eventually, they all left the Soviet Union.
So, this idea of connecting the political system to the productivity of the scientific effort is a very tenuous connection. I even wonder why that system took place. Like, how did we forget the Soviet Union produced a lot of amazing math and science? How do we forget about Nazi Germany?
I think most people don’t have direct experience with how scientific breakthroughs or technical breakthroughs are produced. So, it’s always a great remove. Like, you read about it in a textbook or you watch a movie about it. But for the people who really see how it’s done, the Aspie genius who’s going to improve your large language model, transformer model, that guy’s not potentially thinking about any political things. He just wants to actually sit in his office and do his own thing. And he could do that just as well in Beijing as in Mountain View.
Yes. But I do want to say that I think that Aspie genius would do even better in a system where he was allowed to criticize the Ukrainian harvest. We can state a minimal case in which authoritarianism does not have to defeat science. But I think we can still embrace a stronger case that freer societies can get better science because the American science ultimately was better in probably most ways, with some exceptions.
Now, there is a more extreme version of this hypothesis, which, for example, people from the French system and people who really experienced the Soviet system will advocate for it. I don’t necessarily hold this to you, but they will advocate for it.
For example, the French really outperformed the Americans in winning Fields Medals. It’s because they don’t have a system that glorifies money and becoming a billionaire. They have values. They have cultural values. Mathematics is one of their things that in their society is highly esteemed.
And a kid here who is really talented at math is thinking to themselves,
“Well, how can I make a lot of money in cryptocurrency when he’s 17 years old?”
That doesn’t happen so much in France and it didn’t happen so much in the Soviet Union. And so they claim that on a pound-for-pound basis, their math and physics was better because they were not corrupted. I’m not talking about intellectual freedom or political freedom. They were not corrupted by capitalism, by materialism. And that is an argument that some people, even French scientists today, will still make.
Well, France is interesting because they’ve had several leaders with mathematics degrees. Yeah. Like even modern senators. And that is interesting. Yeah, the prestige in France from having a mastery of mathematics is quite high still.
Here in the United States, it’s not so high. Although maybe it’s coming back with AI and crypto and stuff like that. I don’t know.
I doubt it, Steve. I don’t think mathematicians will be cool. I can only dream. Let me throw one more quote at you. I think we still have a little bit more time. This is from your letter.
I sometimes hear that the U.S. will save manufacturers through automation. The truth is that Chinese factories tend to be ahead on automation. That’s a big part of the reason that Chinese Tesla workers are more productive than California Tesla workers. China regularly installs as many robots as the rest of the world put together. They are also able to provide greater amounts of training data for AI.
We have to be careful not to let automation, like superintelligence, become an excuse for magical thinking rather than doing the hard work of capacity. There’s all these wonderful claims now that AI, AGI, superintelligence, whatever you want to call it, is going to solve all of our problems.
And, you know, I fear that before it cures cancer and rebuilds the industrial base, AI is simply going to derange us all. People are getting fooled quite easily on the Sora 2 with these video generation tools. How can you tell what is an AI video today? Like, are there even actually any watermarks here to say? Hard to say.
There’s teens with their AI companions. I feel like social media is already trolling me to insanity without this super intelligence hamming that up. So, I’m worried about AI. And I think that, you know, there’s all sorts of opportunity, but it seems like there’s also a lot of threats here that I don’t really know how to assess.
What’s your view on AI?
Steve, when you think, when you are in this, you’re doing your tour of Silicon Valley, you’re living in Berkeley this month. When engineers come up to you with these really apocalyptic claims about AI or pure utopia claims, how do you personally react? How do you assess it? How do you maintain epistemic hygiene and sanity? And I want to know your mental toolkit for dealing with these claims.
Yeah, I think this is the question of our era. And I think this is ground zero.
You know, on the one hand, on Saturday, I’ll be at a dinner, I’ll give a talk, and there’ll be a hackathon at AGI House, which is a huge mansion that’s been rented by these AI researchers. It has a view, it’s a twin to the Twin Peaks neighborhood of San Francisco, and has this unbelievable view of the bay looking to the east.
On the other hand, at Lighthaven in Berkeley, where there’s an assembled crew of rationalists who are all doomers and worried about AI killing us, they have a beautifully manicured campus where you can sit outside and discuss things in a philosophical tone.
So here we have both poles of the most optimistic, abundance-focused view of what AI is going to give us and the most doomerish existential risk view. And those two communities are living maybe 10 miles apart from each other. So this is the place to examine those views.
You know, for me, I would say we’re close to spending 2% of GDP now on CapEx investments for data centers and model training and NVIDIA chips and things like this. And back to your quote, the assumption that that will suddenly make us competitive in terms of industrial robots or automation in our ports or assembly lines where cars are made, things that China is way ahead of us on, I don’t see that the AI that we’re currently developing is necessarily going to close that gap automatically in some magical way.
I agree with you. It’s sort of magical thinking on the abundance side to think that these large language models are suddenly going to solve all those problems for us. So I think that’s pretty unrealistic. I think that’s a problem for U.S. strategy. If you’re sort of a hard power person thinking about U.S.-China competition, I don’t think these AI investments are going to, at least for a while, close these gaps. They’re very concerning.
As far as the long-run existential risk from AI, I think the people who are afraid are not wrong. I think in the long run, if we do create things which are more intelligent than us, more powerful thinkers than us, but also eventually have their own desires and wants and maybe even a kind of consciousness, that it will be hard eventually for them to displace us in some way.
I think there is a real long-term existential risk. I don’t think it’s as proximate as some of these people in Berkeley think. I think it’s going to take longer. I think there could actually be a period of human flourishing where we do create these great intelligences, and for a while, they are harnessed to our needs. They are well-designed. They are aligned in some sense.
That alignment, I think, can’t be guaranteed in any rigorous mathematical way, and so in the long run, they could diverge, and we could create godlike things that eventually don’t care about us and accidentally smush us. And I think that’s the argument that the Berkeley crowd would make.
What is the long run? Is that a matter of two decades or two years? Very hard to predict because the problem is that, and this is the key thing everyone’s on the lookout for, is when do the AIs become so good at AI research and software writing that they start improving themselves and at a pace where human engineers can’t quite follow the way in which they’re improving themselves?
I don’t think we’re close to that right now. I think I would guess we’re probably five or ten years away from that inflection point. Other people have much more aggressive timelines than I do. Yeah. It’s hard to know. It’s one of these things. It’s like, oh, what is it? You know, if China does succeed in becoming wealthy and fully developed, what will their government system evolve to? I literally don’t know the answer to that.
And similarly, I do think AI is definitely going to continue to advance rapidly, but I don’t know the timescale over which it will become threatening.
Do you believe that there will be a point in which it starts self-recursively improving?
Yes, I do.
Okay.
I mean, this is one of these hard things where, you know, it is really easy for, you know, once you look at all of these exponential curves and we see them and, oh, it’s like a log scale is a perfectly linear, you know, it becomes really, really difficult to think of anything but, you know, the extent of these curves.
And, you know, I don’t know if the saying to do, saying thing to do is to just rule out the idea of, you know, the idea that this will self-recursively improve because once we’re in that scenario, then all bets are off, right? And how can we predict anything?
It’s like the first order impact of AI is hard enough to understand. And then we have to like think about the 17th order impact. That’s just way too hard, right?
And so this is where I’m wondering whether, you know, this sort of totalizing aspect, it is occluding our view of what is really important. And maybe it is rational because there’s plenty of smart folks who are telling all of us that we’ll get to this point of recursiveness. And I don’t know, it feels like it is just going to be, it’s hard to come up with the epistemic tools really to deal with that argument.
So, I was in a war game held at the Tate Modern, this museum in London, and the war game was sponsored by people who are concerned about AI safety. The British government funded some of it, and some of the big AI companies funded it.
My team was the China team. I was a leader of China in this. It was literally a war game where every turn is like a year and you’re making decisions about:
The head of the U.S. team was a guy called Stanley McChrystal, who was a prominent former general. Very serious simulation.
The outcome was though very optimistic because what happened is that when we, in our respective countries, saw this recursive capability start to appear, we negotiated a slowdown and mutual inspections of our AI labs. So the American side could send AI scientists to see what was going on in the Chinese labs and vice versa. The game ended with us slowing down the development at a place where we might lose control over the AI.
Now, just before I came over here, I was listening to, as I mentioned earlier, an interview of Demis Hassabis and Dario Amodei at Davos, which I think is happening right now. They both said, they just kind of mentioned during their conversation, which is mostly not about this stuff, but a little bit about this cataclysmic type development.
They both mentioned that they’re in communication because they’re both on the lookout for this, and they are competing against each other and they’re competing against Chinese labs, but they are watching for this development. They do have in mind that they are going to start talking to each other when they start to see it. And hopefully I think it would just be wise that if we start to see that we slow down.
Now for the doomers, this doesn’t solve the problem because the doomers will say:
So it doesn’t completely solve the problem, but there is a plausible future where we delay the real catastrophe for a long time by being smart about this process.
Well, if we’re sure that the catastrophe is coming, why don’t we slow down now?
Those guys were actually interesting. So very, very interesting remark because both Dario and Demis said in this conversation, they would both be happy if they could slow down. They both said that.
So, but it’s a collective action problem. It’s a little bit like global warming. Like, can people actually put aside their narrow self-interest to cooperate to avert this catastrophe? I don’t know.
Well, global warming is challenging in part because there are so many countries that have to coordinate together who have to trade off against economic benefits, presumably with coal.
But why don’t we send General McChrystal to start inspecting those labs right now? Because we can do a lot of that right now.
When they designed this war game that I was part of, the people who designed it are actual specialists who designed war games for the Pentagon and also for the British defense ministry. They were honest.
So they designed the dynamics of the game to be as realistic as possible. However, the makers of the documentary were hoping that the ending would be an optimistic one where the competing sides could cooperate.
And the view is like, this provides a template because a movie version of it is very real for the ape brain to see humans enacting the story. It’s a narrative, right?
And the hope is that this will plant the seed in the Chinese leadership, the U.S. leadership, and the corporate leadership that yes, we do need to be ready to pause when things start to get dangerous.
So I think it is in everybody’s mind, whether it’ll, it could still be a tragedy. It could still be a runaway arms race, but I’m optimistic.
Actually, narratives matter. We are all ruled by ape brains. And so this is why the lawyers are in charge, and the scientists and engineers need to get better at narrative.
Very good. Well, that’s a good place for us to stop.
So thank you very much, Dan. It’s been a wonderful experience. Hope my audience enjoys this.
We’ll put some useful links in the show notes to some of these things like industrial maximalism and other topics that came up.
Thanks very much for your time.
“Thanks, Steve. This was wonderful.”
]]>MD was supposed to be the protein folding solution. There is a great counterexample. The counterfactual is basically a group called DESRES, D.E. Shaw Research.
They had similar funding to DeepMind, probably more actually. They tested the hypothesis to death that MD could fold proteins. They built their own silicon. They built their own clusters. They had them taped out all themselves. They burned into the silicon the algorithms to run MD. They ran MD at huge speeds, huge scales.
I remember David Shaw came to a conference once on MD, and he flew in by helicopter. He was a pretty famous guy, kind of rich. He gave an amazing presentation about the special computers and special room outside of Times Square and what they can do with it. It was beautiful, amazing.
I always thought that protein folding would be solved by them, but it would require a special machine. Maybe the government would buy five of these things, and we could fold maybe one protein a day or two proteins a day. And when AlphaFold came out, and it’s like, you can do it in Google CoLab or on a GPU or desktop, it was so mind-blowing.
I forget that protein folding was solved. I always thought that was inevitable. But the fact that it was solved and on your desktop, you can do it, was just completely floored. It changed everything.
This is the first episode of the new AI for Science podcast on the Latent Space Network. I’m Brandon. I work on RNA therapeutics using machine learning at Atomic AI. My name is RJ Haneke. I’m the co-founder of Mira Omics, where we build spatial transcriptomics AI models.
The point of this podcast is to bring together AI engineers and scientists, or bring together the two communities. These are two communities which have been developed independently for quite some time, but there’s been some attempt to combine them. And only now, after many years, are we starting to see some of the big developments start to play out in the real world and start to solve key scientific problems.
There’s no, like, one-size-fits-all solution. You need domain expertise. You need people on both sides of the aisle who can really talk to each other and really work together and understand both the modeling and all of the real subtleties of the system you’re actually trying to work on.
We hope that we can connect these communities and that we can provide a starting point for this new era of AI and science to move forward. So without further ado, let’s get started on the first podcast.
We’re really happy to have in the studio today, Andrew White, co-founder of Future House and newly formed startup Edison Scientific. Rather than introduce him, I’ll let him introduce himself.
“Hey, I’m Andrew from San Francisco, former professor, now running two startups, one that’s a non-profit research lab and one that’s a for-profit venture-backed company. And we’re trying to automate science.”
We’re going to get into all those points.
“Yeah, really happy to be here. Thanks for having me on.”
I want to know personally about jump from academia to industry and course, quasi-industry. So I would love to hear that story.
“Yes, I guess that’s the whole story, right? So I did my PhD at University of Washington and I worked in a group with, I think, 19 people doing experiments and like two people doing simulations.”
And I was working on a topic called molecular dynamics, which I think is actually suddenly becoming interesting again as everyone’s looking for ways to generate data from first principle simulation. And molecular dynamics, you know, covers basically everything that’s molecules moving around in dynamic systems, so like biology.
Of course, the complement in material sciences: things like density functional theory, where you can model chemical reactions in these like solids.
So I was working on that and we were working on biomaterials. And so the goal of my PhD was trying to find what are called non-fouling materials.
So in biological systems, whenever you put like a foreign object into the body, it will trigger a response. And that response called the foreign body response basically encapsulates it in like this layer of collagen.
This actually is exploited for some implants.
And this is great for pacemakers, but for like a glucose sensor or like a, you know, brain cognitive interface, BCI is what they call it now.
“Yeah.”
There, it’s not so great. And so that’s why some of those things have like a limited lifetime because eventually your body treats it like a wound and heals. Rejects it.
“Yeah. It’s kind of like some rejections like immune based.”
Okay. And so that’s where if the body can see anything on it, like if it can see some ligand that it combined with antibodies, then you get this inflammation, which is like a rejection response. You see this in organ transplants. But with materials, the body’s just like, oh, there’s just a wound or something here and it just covers it up.
I think the research in that field has gone on a long time since I left my PhD, and there were a lot of theories about its relation to the mechanical properties of the material. Like if it’s spongy, or if it’s trabecular, meaning it has a bunch of little pores in it.
We worked on the theory that it had to do with how hydrophilic the material was. But anyway, I was the only one working on computers in this group. I couldn’t figure out how to connect what’s on the computer with what’s done in the lab, because you can make a simulation of whatever - 10,000 particles, 10,000 atoms.
It’s like, well, this is not going to model the human body. It’s a lot more atoms involved.
So I had a good time. We did some cool stuff, some bioinformatic work. I learned a lot.
But then when I did my postdoc, I was like, okay, we’re going to try to merge experiments and simulations. So I worked on this theory called maximum entropy. It’s about how you take complex simulations and match them to limited observations.
It’s like the inverse of machine learning. Machine learning is like if you have simple models, then you go to a lot of data, whereas I had complicated models trying to fit very little data.
It was fine. It was great. We wrote some papers. It was useful.
Then I started my research group at the University of Rochester on applying these methods to model peptides. Yeah, I’m always a bit too early for things.
We studied peptides for, I don’t know, four or five years. It was a cool niche field, not that popular. Now peptides are like the hottest thing ever. I think there’s even a peptide rave I heard about a couple of weeks ago.
But when I was an assistant professor, nobody cared about peptides. So we worked a lot on different ways to combine them. We looked at different experimental methods, paired with molecular dynamic simulations of peptides.
Then in 2019, I was out on a sabbatical at UCLA. They have a place called the Institute of Pure and Applied Mathematics, which is this institute where people can go and do a sabbatical and learn new methods.
They happened to be doing machine learning for physics. I think the name of it was some kind of symmetric thing, like machine learning for physics and physics of machine learning. It’s a kind of cool concept.
Yann LeCun was there, and Frank Noé was there, a big figure in Europe in this field. Terence Tao even came by. It was a really great group, and everyone was kind of jamming.
It was 2019, so that was before the big hit of machine learning in non-computer science fields.
Then I came back from that and thought, well, I’ve got to teach a class on this. So I’m writing a book about how you can apply these methods in chemistry.
It was a very niche field because every machine learning class my PhD students could take at the time - this was when I was a professor at University of Rochester - would always end with something like:
But in chemistry, it’s all about graphs. It’s about representing graph structures. It’s about symmetry and geometry.
That was not a common thing. It was popular but not mainstream. You had Max Welling before - the godfather of geometric deep learning.
So I wrote this textbook about these methods, and there was a bunch of interesting mathematics to it. I had a good time.
Then I was following the news in the space, and when Codex, the original Codex, came out, I had been tinkering with transformers for a while.
We started trying them on some chemistry tasks, and we were really impressed. We wrote a benchmark.
This was around 2019 or maybe 2020. Ahead of the curve, a little ahead.
The task was like this:
Here's a function, a body of a function for a Markov chain Monte Carlo simulation.
It's missing some pieces.
Complete it.
Then we had a verifier that would see if it was a valid MCMC simulation.
Yeah, we wrote this paper. It ended up coming out, I think, in 2021, 2022, because it took a long time to bank enough questions. But I wrote an opinion piece about how transformers could change how we think about chemistry and things like this and how we teach it.
And then OpenAI, some people there, Lama was there. She saw this paper and they reached out and said, “Hey, we’re building this new model. And we think it’d be great to red team it to see what could happen with these models if they’re applied to chemistry or biology.”
So I was a red teamer for GPT-4, and I was using it like nine months or something before release. It was like August. So GPT-4 came out in March, and I was using it in August.
Then the ReAct and MRKL paper came out. I think Shunyu Yao wrote that paper, and I plugged it in with GPT-4 like in the fall, and I was like, “Wow, there’s so much stuff coming out with React.” Yeah, and it was really exciting.
When GPT-4 came out, I released this paper called ChemCrow. I worked with Philippe Schwaller in Switzerland and IBM. So that was like React applied to chemistry.
What we had was a cloud lab that IBM built in Switzerland, so we had GPT-4 operating the cloud lab.
Then I had written a literature research agent that did agentic RAG. At the time, nobody really knew what agentic RAG was. I think actually Harrison Chase had written a blog post about some ideas there, and so I stole some of those ideas-really smart guy.
Basically, we applied that and saw some really cool stuff. It was really exciting.
We wrote the paper, and it set off this crazy storm of anxiety about AI progress. I ended up visiting the White House. I guess my paper was the only time a preprint or peer-reviewed paper was presented to the president on their schedule for a 30-minute block.
The National Security Advisor at the time, Jake… I was confused. One of them was a talk show host, and one of them was the National Security Advisor. I forget which is which. That guy had a presentation about our paper, and they presented it because there was a big tech CEO summit at the time where they sent out Sam Altman and some other CEOs.
This was the future of chemistry as language or a different one? This is the ChemCrow paper. Oh, ChemCrow, that’s right. Sorry, I probably should name these things.
It was crazy. They had me go out there, and then I met a lot of three-letter agencies I didn’t really want to meet.
People from these agencies asked questions like:
I was like, “Guys, I don’t know for sure.”
It turned out that there weren’t many world experts on AI and science.
Right. So what’s the answer? Great question. We’ll come back to that.
In the end, I had a lot of energy and excitement about this area. So I took a sabbatical from the University of Rochester.
First time it was Sam Rodriguez, and Sam had been talking to Eric Schmidt and Tom Khalil (who was also a national security counsel at the Obama administration) about how to scale up these ideas.
Sam had this concept of focused research organizations-how do you do science not in academia or in a near-monopoly tech company but in a new, dedicated type of lab? He wanted to try this idea out.
I was like, “Hey, we should do this around agents for science or AI for science.” I love Sam. He pushes me to come up with really lofty ambitions.
So we decided to automate science as the goal instead of just seeing what fun stuff we could do with agents and science. But I think that was maybe the real mission. Of course, automating science is the long-term mission.
Yes. And so that was what led to Future House.
And that was a very long-winded story. No, no, that’s great.
But you chose to leave a tenure track position. I was on sabbatical, which is a beautiful concept, but then I did resign my tenure position when we co-founded Edison.
I had been on sabbatical for a very long period of time, so at a certain point, I just had to resign my tenure. I resigned tenure in June.
Oh, so that’s only recently. Yeah, only recently. You just felt like this is the direction of your career.
Yeah. I got tenure and I had these early career awards like the NSF Career Award. It was great. And I think academia is really exciting, but I just thought that right now, this kind of area, like ever sciences, is difficult to do in academia and be so exciting, but I think you can take bigger bets.
And I think having a tenured position and writing research grants is maybe not the biggest bet you can take on a field.
Yeah.
So now we have a venture-backed startup called Edison, which we spun out of Future House. And we took a lot of the ideas and we’re trying to do this at an even bigger scale right now.
Yeah.
And so Edison was always kind of the plan, like going back to Sam’s idea of a FRO or a, like, what’s a fundamental research organization. Like he always had this goal of, like, “let’s do fundamental research in this tightly scoped nonprofit, which can kind of explore.”
And then you have that as a natural arm for spinning off, you know, venture-backed.
Yeah. I think that’s right.
I think some things that make that not as clean these days are how expensive AI research is and how expensive GPUs are.
So I don’t think we can repeat it many times from Future House. It might be like an end of one thing right now. It just maybe not.
I don’t know if venture capital keeps growing, then maybe we can, but yeah, I think we took a lot of the ideas of Future House.
Another thing is, I think we expected it to be harder to automate science, and actually it’s really hard.
I’m always miscalibrated in this domain, but it’s always hard to predict progress.
And I think that I overestimate the speed of things on a month scale and underestimate things on a year scale.
So the two years from 2023 to 2025 was an enormous amount of progress.
Yeah.
And I always felt like things were not going as fast as I thought, but when you look back on it, wow, like there’s a lot of progress.
And so I think in Future House Center, Sam actually regrets us writing this, but in the original marketing or announcement it was like,
“It’s our 10 year mission, automate science.”
And now it’s like, okay, yeah.
Yeah.
So two years later we had Cosmos, and things are going so much faster.
And also, this is the kind of thing you notice in San Francisco, where it’s actually kind of hard to find problems which are:
We’re in this gray zone.
Actually, I feel like that’s where we are now: we can actually automate so much of the scientific method because it turns out, especially in a field like biology, which is very empirically limited,
the top 1% guesser of what they think will happen in an experiment and the top quintile or quartile are about equal.
Even if we wait 10 years and get even smarter models, I don’t think it will really change the fact that we’re ready to automate a lot of science with existing LLMs.
I mean, what do you mean by automate science?
That’s a pretty loaded statement. Science is lots of things. There are many ways to think about that.
So we try to draw a line between:
Groups trying to model something like the cell or how proteins fold, or how antibodies can be designed, or virtual cells (as an example). If they’re trying to use machine learning or AI to model a very specific system,
We’re trying to automate the cognitive process of scientific discovery:
We want to automate that sort of loop.
We thought that we would have to build a whole new organization from the ground up for agents.
So it means:
- Automated labs
- Putting all the papers in one spot
- Getting APIs wrapped around everything
But over time, the models have gotten better and better that we had to stop and rethink:
“Okay, we don’t actually have to hold their hands so much anymore, or they don’t necessarily need to have an automated lab.”
They can write an email to a CRO (Contract Research Organization), or tell you what experiments to do and you can take a video of you doing it and show it to the model.
Then they can say,
“Okay, well, this is what happened.”
So it’s been a really interesting experience of sometimes over-engineering. Things and sometimes actually basically just mostly over-engineer. So I always think about systems and scientific is a system like scientific processes, a system. I always think of it in terms of constraints, right. And like, what is a bottleneck in the system?
So that, so what is your hypothesis about this? Like in my mind, not knowing a ton, but in my mind, the constraint of the scientific process is the work you do in the lab. And that’s sort of notably missing from, well, not entirely. You said, you mentioned automating lab and whatever. So like, how are you thinking about this?
Yeah, I think you’re right is that basically the best model, whatever, Opus Seven or GPT-10, like, it really can only propose the first experiment, maybe slightly more clever, but at a certain point you just need information, right? Like some little calculations you can do that like there’s more atoms in the brain than you could ever simulate. Even if you had all the energy from the sun, right? Like I think it seems maybe a thousand brains in real time with all the energy in the sun because just too much information.
Yeah. So science really hits these bottlenecks where you just actually have to go measure things. Yeah. We definitely think about maybe like lab in the loop sort of situations. Like one of our papers, which was called Robin is that we like had one of our agents propose an experiment. We did the experiment and then we had our agent analyze the experiment that propose the next experiment. Yeah. And that kind of loop I think is where you want to get to.
So what is the bottleneck in that? I don’t think it’s like the intelligence of the first experiment. I think the bottleneck might be something like right now. I think the bottleneck is something silly, like knowing what’s the lead time on all the reagents that you need and what is available in the lab. Like, you know, I think whether GPT-5.2 codex max or Opus-4.5 is going to do better is probably doesn’t matter. It’s just a matter of like, which one’s going to have all the information about what’s in the lab and how much will it cost, how long will it take. Right.
Um, and also, I guess the kind of frontier that I think about for these models is taste, which is like a lot of science. I mean, of course we want to accelerate technology. We want to improve the economy. We want to improve people’s life expectancies. We want everyone to be happier. But a lot of what is done in science is based around like human preferences.
Like why do people study, I don’t know, a particular worm? Well, like there is a theory that by studying the worm, it has led to good medicines or it’s led to discovering new genes, but also people studied in the past, people’s careers depend on that worm and people want to write papers about that worm. And so there’s a human element to some of this. And I think that models don’t capture that so well about knowing what is an exciting result and what is a boring result.
I see. So I think that’s like a scientific taste. It’s like a broad category of all these things. How do you, like, do you try to quantify taste in any way? I mean, I know that I have some fun anecdotes about this, but maybe, yeah, just like hear what you.
Yeah, actually, we sat on this idea. We sat on it, but we like argued about it for a long time. Sam and I usually every Monday morning at eight o’clock in the morning, Sam and I meet and we’re both caffeinated and ready. And we argue about stuff like this. And we had a lot of Mondays where we talked about scientific taste.
And in the end we’re like,
“Okay, let’s just do the dumbest thing, which is to like have our agents make hypotheses and put them in front of humans and have them be like, I like this one or like that one.”
Right. So we just did like whatever, RLHF on hypotheses. And we learned a lot about how bad RLHF is with people. Just like people pay really attention to the tone, to the details, to like how many specific facts or figures on the hypothesis, right? Like actionability about if the experiment is feasible, but what people didn’t really pay attention to is like, I don’t know how to describe it, but like, if this hypothesis is true, how does it change the world? If the hypothesis is false, how does it change the world?
This like,
And that really didn’t come through from those things. So they were like, okay, well, this is maybe one strategy. And so we had to go back and think about it more.
And then we took a pause from that research and then we made Cosmos and then Cosmos has baked into it taste, right? Like at the end of the day, there will be some report and we’re working on generalizing this. Basically, at the end of the day, okay, I made these discoveries and a person would:
And that rolls up to some hypothesis that came earlier in the process. So we think we can get to end to end on this as opposed to human preferences.
So you mean the feedback loop is the click?
It could be the click. It could also be like, you know, we do an experiment sometimes in Cosmos, you could ask it to end an experiment and you can go see what experiment is success or failure or something like that.
But I guess like we brought it out of this kind of hard to quantify:
“Is this a good hypothesis or bad hypothesis?”
and into this, like, you can see some downstream consequences of the hypothesis.
So yeah, humans have, I think, a very strongly well-calibrated nose for science. Like, I mean, maybe you could argue there are sociological effects like across the community, but ultimately oftentimes people-really good scientists-know right off the bat:
“Is this going to be likely to be useful or not?”
How long, how many attempts did it take before you started to see results that to yourself seemed useful?
Like even working on this for, I guess, two years now, you know, I think when the co-scientist paper came out from Google, I think it was a really interesting idea to do this like tournament style or just pairwise ranking of hypotheses.
Right.
So I think co-science is very interesting.
Counterexample to what we built is that what we built is something with either lab in the loop or data analysis in the loop or literature research in the loop where you’re iterating on an idea.
I think co-scientists took a very different approach of like:
So co-scientists will produce these very long reports of like,
“Oh, we really tested this idea with lots of dialogue,”
and it was very interesting stuff. I was really impressed with the paper that came out.
And then we had this Robin paper.
One of the things that came out of the Robin paper is that the hypothesis that people thought was best was not the one that led to success in that paper.
Interesting.
It was in, um, age-related macular… ocular, uh, age-related macular, basically it’s like part of the eye where you’re going blind because you have this accumulation of debris in the eye and can’t clear it out. That’s one of the major causes of blindness in people over, yeah, 60.
Ollie, who works on the hill…
Yeah, he’ll cringe when he hears me say that, but something like that, something like that. Sorry, Ollie.
In that one, like, we went to optometrists or ophthalmologists who actually get confused on that as well. Sorry, Ollie.
Essentially, you know, ask them:
“What hypotheses do you think are good hypotheses? What do you think would lead to a good mechanism for treating dry MD?”
And, yeah, they agreed to be on the top 10, but beyond that it was kind of noise.
And then, you know, what we found was ripestoodle was a very good medicine and had a mechanism that I think is novel.
Although there was lots of debate on X because in, I think in 2012, there was a master’s thesis which proposed this mechanism on like page 38. I actually think it was a typo. I think they meant wet AMD.
Okay.
Anyway, I won’t belabor the point. I will concede that maybe there is one reported example of it in the past.
That was a really eye-opening experience for me because that was the first really serious test where we really went to the lab and we spent like four weeks on a battery of experiments to see what hypothesis led to a good mechanism and a good repurposed drug.
Right.
And it was not as correlated with human opinions as I expected.
And so since then, I think that I have a lot more faith in these like verifier in the loop kind of scenarios where you have either:
- Data analysis
- Literature search
- Running a unit test
- Running the experiment
Anything like that, I think is going to give you a higher signal than the sort of vagaries of like:
“Oh, this is a higher opinion or that we’d like this one better.”
Max Welling called it nature’s computer.
Yeah. It’s like you have this computational cycle you’re running and nature is part of that.
Yeah.
I’m curious. You said that there is a paper which maybe could propose where this molecule came from. But do you have some way of interpreting or understanding where that hypothesis originated in the absence of that? Like, is there a traceable thought train? Yeah. Yeah. Yeah.
Actually, this is something we pay really close attention to at Future House and at Edison - provenance of like information. So our first sort of agent was paper QA. Sorry about the name. Paper QA sounds like an email set, but it was an agent. It really does.
Paper QA was like every sentence that it outputs has a citation to a page. Right. So it’s a lot of provenance. And then we basically built a philosophy for everything.
So Robin, which is the name of this, I don’t know, workflow or something you can call it, led to this result on the repository being a good therapeutic for dry MD. It has data analysis that shows you which line of Python code led to the result here.
And then that is like, okay, then it goes to this other model, which says,
“Based on this literature finding and this result from the data analysis, I believe this is the right thing.”
But you know, where does the original idea come from? Like going after these rock inhibitors, which is the mechanism for the target, was basically enumeration.
And so this is like, if you can’t be smarter, you can try more times. And I think that was the theory of the Robin paper: we can put out a whole bunch of hypotheses and then we can filter them.
Just like, I think some of how Co-Scientists did is you go for a filtration process, but the difference is that in Co-Scientists, their filtration process was other LLMs sort of ranking it with rubrics or personas. Our filtration process was like literature search and data analysis.
And I think that’s the easy way to succeed in AI over humans - you can try more ideas faster.
Something I’ve heard people say, and maybe I’ve experienced this in my own life: sometimes hypotheses are kind of cheap, especially in biology. Yeah. In many ways, it’s actually easy to come up with what you think could be happening.
It seems like to me, verifying is oftentimes a big bottleneck, maybe the biggest bottleneck. Like if you have lots of hypotheses, and it costs one hundredth of your runway to test each one of them or something, you don’t have any shots on goal. Yeah.
So how do you make sure you are actually enriching for good hypotheses? Literature and data analysis, right? You know, yeah.
There was a time when we used something called tiling trees. Tiling trees is a literal brute force method invented by Ed Boyd and Sam’s PhD advisor.
Basically, the idea is:
I want to accomplish X.
I could try these methods.
Once you pick a method, you split into two paths: use this method or not use this method.
Then, I need some substrate: try substrate A, substrate B, or substrate C.
You can basically try to tile the space of all the ideas.
We tried some early experiments there and you’re right, you run into this thing where some of the hypotheses just don’t make any sense. You’re going to waste a ton of effort if you actually test them all.
Nowadays, I would argue that if you go to an LLM and ask it to evaluate hypotheses, including some garbage ones, it will probably do as good a job as an expert in the field filtering them out.
That’s not always the case. Yeah. I’ve actually seen that myself.
But there are a lot of gotchas, and people can miss those, but I think LLMs are pretty good. So I’m not as worried about hypotheses that can fail fast by an expert looking at them.
I think now the filtration process really happens in:
Yeah.
So with regards to existing data, another contrarian take is that the hardest part is just understanding the context of data - where it comes from and how to interpret it.
I can also think from my own life, multiple cases where the data in some sense was there, and you had two people who were both experts and very smart who looked at it and drew very different interpretations.
In fact, when we were interviewing Heather Kulik, she had some fun stories about using… LLMs and she would find that there would be raw data in a paper, which wouldn’t agree with the conclusions of the actual paper. And it’s straight from the paper. It’s not even like cross-paper talk or something.
And I’m going to be a really boring interviewer and be like, “yes, you’re right.” You know, this is a hard question. I think, to give you something concrete, we have a bioinformatics benchmark, we call it Bixbench.
Bixbench is like we put it out. We’ve updated a few times. It’s in some frontier alums when they release their system card, they’ll mention Bixbench is like one of the things they test on. And, you know, we’re getting to 60%, 70% correctness on Bixbench.
And we found that actually we’re at the point where humans disagree at this level, like humans only agree 70% of the analysis. So it’s true that when it comes to analyzing data, like humans do not agree a hundred percent of the time; there is a certain amount of choice that goes into it.
And, you know, we try to - so Edison is a for-profit company. We like trying to sell some of this stuff to the companies and we’ll go to some companies like,
Or, you know, whatever. And like, okay, well, we’ll have to change our ages so we don’t impute data for them. But then some of the companies are like, “oh yeah, we impute data. It makes everything easier.”
Right. And you want to know what the real modern dark arts are that like AI-resistant area of the world is like medicinal chemistry. That is like the spot where there’s so much superstition.
I feel like I was burnt out. But the religions never agree. Two medicinal chemists will have completely different viewpoints about a functional group. Yes, exactly.
I remember this as I talked to somebody who works at CRO and they’re like,
“Oh, whenever company X orders anything, we never put boron on any of the compounds because they hate boron, because there was one program that was killed because there was a boron somewhere in the core and it led to some toxic side effects.”
So no boron for this company. This company likes things to be fluorinated or something because they think it’s great for the AdMet properties.
Right. And so there’s all this stuff where you reach the point where you’re at, I don’t know, human bias level or human disagreement level. And I think we’re getting to that point in data analysis.
Of course, you will see then that if I take the raw data from a paper and I analyze it myself, I will get a different conclusion.
One of the cool tricks you can do - back to this brute force thing - is that I can go to our agent and run it a hundred times and take the consensus-like analysis. Or I can say,
And then you can - there’s even like words, it’s like epistemic versus aleatoric uncertainty, right?
Anyway, there’s a Donald Rumsfeld formulation of this as well. Like,
“No, no, no, no, it’s an aleatoric epistemic debate.”
Interesting. This kind of digging into your Cosmos model a little bit.
So I glanced at the paper and one of the things that jumps out is that there were certain classes of problems for which it was only 50-some percent accurate.
Oh yeah. And can you talk a little bit about that and how that, like, okay, so if I’m just raw getting 50% accurate answers and then I’m going into the wet lab and being like,
“Okay, try this.”
And then it’s like,
“Ah, like the stupid thing did tell me to do it.”
Well, how do you…?
I would say first of all, that 50% is actually pretty good because it’s rare that experiments in the lab are actually coin tosses. They’re usually a lot more outcomes than binary.
Yeah, sure. Okay.
But that particular number was a human agreement in the interpretation of the results.
So we asked people to evaluate different aspects of Cosmos:
That number that was 50%, that came from Cosmos’s interpretation of some of the analysis. So like it might go in literature and find this result and then it would say, “wow, this is super exciting. This is amazing.” Or it might do data analysis, but this is a novel discovery. We’re really excited about it.
And then people would disagree. “That’s actually not interesting” or “I don’t agree with the interpretation of it.” So it’s like picking bad problems, maybe. In the negative class. And so I think it’s like that, that 52 or 55, whatever it is, that’s interpretation. And so I agree. I think that’s where, like I was saying, I think the frontier right now is scientific taste.
And so that’s what we’re working on right now: how do you get that interpretation to match?
Could you step back and just introduce Cosmos from a high level?
Yeah. I would actually be even curious to hear starting from like Chem Crow and, you know, you have paper QA, Avery, E3, zero.
I’d like to hear a little bit of the lineage and how those different decisions were made. What were the key learnings and how did you get to where you are now?
Yeah. So I could retcon and tell a really great story about how we arrived at Cosmos, but I will say that, to a large extent, we just try a lot of stuff and sometimes it works and sometimes it doesn’t.
You know, I’ll say that we’re very-I’m a builder. I like to build things piece by piece. I’m probably some fancy word for it, but I’m like a Lego guy or something.
My vision was that we would make:
So we had like, Chem Crow, which is going to help us with setting up our medicinal chemistry work. We had Protein Crow, which we haven’t released. I don’t know if we will ever release, but Protein Crow was like designing proteins we might need for some part of our workflows.
Or we had a data analysis agent-that’s an agent, so an LN plus tools.
Okay. Or we had ether zero, which was like, okay, we noticed that the frontier models can’t work with molecules very well, so let’s make a model with intuition for medicinal chemistry. And that was what led to ether zero.
But then Sam actually really pushed on us to like, “let’s just see if we do the whole thing. Let’s just try to build an AI scientist. Let’s just try the whole thing.”
And that was what led to Robin. Robin was like, let’s just take these agents we already have and we’ll just put them in like a work basically. It was like you could express it in a concise Python file of like:
# Pseudocode representation
try_a_whole_bunch_of_ideas()
then_go_see_if_they_all_filter_through_literature_or_if_they've_been_disproven()
come_up_with_experiments_to_do_in_wet_lab()
analyze_all_the_data()
repeat_the_process()
Yeah. And this is our inventory list and then go analyze all the data, then go back and repeat the process. Right. So that’s like what Robin was.
And we came across Cosmos. We’re trying to understand what is the process that Robin is automating. And it came from this idea of a world model, which is that when we first started Edison, we were thinking like, what do we want to change about this? What is new here?
So we spent some time thinking about, well, the scientific process, like what is actually going on in my brain, which is that I have some understanding of the world or the phenomena I was studying. And that’s my world model.
Then a lot of the actions I take are about trying to update that world model. And it’s something that changes over time. This is like the ability to change over time, but it’s also something that is practical. Like I can use it to make predictions about, “I know from this experiment, this will happen.”
That’s why it’s like a model and not just like memory or a bunch of papers or something like that. It’s supposed to operate in Cosmos.
We tried this idea out and actually Ludo, who’s the first author on the paper, tried a whole bunch of ideas around world models. We kind of thought they weren’t really appropriate though.
Like, we tried a lot of different ways to do this:
And they’re okay. So we all just had to take a break. Ludo-his project didn’t work on trying to do this world model stuff. He’s like, “I’m going to keep trying it.” Ludo is a very stubborn person.
So he tried it for like a week or two weeks, then he was kind of like quietly, “Hey, can you guys come take a look at this?”
And we’re like, “Wow, this is actually really cool.” And then we started building on it and jamming really. And I think what Ludo figured out is that you have to get this experiment loop thing. You have to be able to let it, then the data analysis agent is what got us in the loop.
So if you put that in the loop of like, it can really update this world model because we were trying to build it around literature before. And when you build it around literature, there are just not really experiments you can do and then see the results for. That was like our surrogate - was literature. It just wasn’t working.
Data analysis actually really lets you explore ideas. And so that was what led to Cosmos.
In Cosmos, we basically had all the pieces sitting around. We were working on:
And then we were working on, you know, we built a platform for scientific agents. So we had things that can:
Then we put that all together. And like a world model was like sort of the glue that allowed it to fit together.
An analogy is like encoding agents - like GitHub is sort of the glue. There’s some shared repo and everyone works on the repo. And software engineers have spent lots of brain cycles thinking about what’s the way to coordinate and organize working on code together for a long time.
So the world model is actually like a memory system? Yeah, you can think of it as a memory system. We think about it as a model. So like it actually, you can put in input and it will output predictions. And we think about calibration. But really, it is a set of like a big bundle of information that we accumulate over time. It’s distilled in some way. And that is like what allows us to do this.
And I think you can think about like a GitHub repo - it’s a distillation, right? Like really, there’s a long graph of commits that lead up to it. And like the current file system in that GitHub repo, or keep saying GitHub, I’m such a corporate shill here. Your Git repo is like a distillation of all of the work that people put into the PRs, into the commits.
And so I think there’s a nice analogy between a Git repo and what a world model is. And I think that’s just sort of what allows us to automate scientific discovery so well.
Can you talk about kind of how you implement a world model? Or is that sort of like secret sauce? That’s our like secret sauce right now. Yeah, that’s fine. No, it’s fine. People have asked me.
So one thing that’s notably missing is the like simulation, right? Yeah. Molecular dynamics or like Boltz or… Yeah. I want to help you guys pump up your views here.
So I think molecular dynamics is overrated.
In fact, coming from someone - yes, that goes in the thumbnail, you know - and DFT is overrated. In fact, DFT may be even more overrated than molecular dynamics.
I think these methods -
Okay. And I can explain more about that.
Basically, MD and DFT have consumed an enormous number of PhDs and scientific careers at the altar of the beauty of the simulation.
Also, random interjection: once I did an estimate, I think pre-like ChatGPT, something like 20% of the world’s computing power just went to simulating water. Oh, my fucking God. Water.
Yeah. I had to deal with so many water simulations.
I did DFT simulations of water and they are so annoying. I use these big computers from the Department of Defense and I spent like, I don’t know, five months. And by the way, this is pre-LLM training days. Five months of compute is actually a really long time.
I simulated water with quantum effects with a Grotthuss mechanism for how a proton hops through water. And it’s on YouTube. It’s my number one YouTube video.
It represents like… until now. And it represents like, I don’t know, a million CPU hours of compute. It was, you know, one of the biggest computes that I… probably the biggest one I’ve done in my life so far. Maybe Ether Zero is bigger, but it took a lot more work.
Anyway. And what’s the point? Yeah, what’d you learn? What’d you learn?
All I learned was like, what set of hyperparameters reproduced some physical effects of water? But none of it was de novo, right?
And this is the issue with molecular dynamics and DFT - they don’t model the world correctly. And so we have to invent little stories we tell ourselves about we’re like making good inductive biases and then it models the world more correctly.
Like in DFT, you simulate water at 330 Kelvin when you want room temperature water. Is room temperature 330 Kelvin? No, it’s not. That’s a little too hot.
And so this is… The issue is that people just make up these things or, I don’t know, GGA or B-LIP or B-3-LIP, all these different methods people invent. They’re clearly empirical. And then they bolt it on to DFT and they say,
“Look, it’s a first principles method, right?”
But actually you made a whole bunch of choices and, you know, you overfit to the validation data to get this to work. And that’s, I think MD and DFT are like that.
Because if you go look at the catalysts, you know, what catalysts change the world, none of them are single crystal materials that are really well suited for DFT. They’re always like, they have grain boundaries, they have dopants, they’re complicated, right? And you’ll never capture DFT.
So I think this is one of the fundamental, I don’t know, dichotomies of the world is that:
And so that’s why I don’t do DFT and MD anymore.
What about somewhere like the machine learning stuff like AlphaFold and…
AlphaFold was trained on x-ray crystallography data. And I think, you know, this is the story of MD: MD was supposed to be the protein folding solution.
There is a great counterexample. There’s a, I don’t know what there’s a word, but the counterfactual is basically a group called DESRES, DE Shaw Research. They had, you know, similar funding to DeepMind, probably more actually. They tested the hypothesis to death that MD could fold proteins.
They built their own silicon. They built their own clusters. They had them taped out all themselves. They burned into the silicon the algorithms to run MD. They ran MD at huge speeds, huge scales.
Yeah. I remember David Shaw came to a conference once on MD and he flew in by helicopter and was like this pretty famous guy, kind of rich. And he gave an amazing presentation about the special computers and special room outside of Times Square and what they can do with it. It was beautiful, amazing.
I always thought that protein folding would be solved by them, but it would require a special machine. Maybe the government would buy like five of these things and we could fold, you know, maybe one protein a day or two proteins a day.
And when AlphaFold came out and it’s like:
“You can do it in Google CoLab, you know, or on a GPU or desktop,”
it was so mind-blowing. I forget like that protein folding was solved. I always thought that was inevitable. But the fact that it was solved and on like your desktop you can do it was just completely floored, changed everything. Like the bitter lesson on steroids.
Yeah. I don’t even know what it is. But it’s like imagine ChatGPT came out, but instead it was like,
“Oh, you can just run it on your phone or locally on your own desktop.”
Like that’s the level of shock that came out. And it gets down to this thing that humans are really bad at estimating problems that aren’t human-made problems.
Protein folding-we all thought it would require a huge amount of compute, a very challenging problem, the hardest problem in the world, right? And it turns out that you can actually do it on, I don’t know, I think the numbers are now like 10,000 GPU hours so you can train a good protein folding model. It’s actually turned out to be barely an inconvenience. Therefore, why not?
Oh, oh, therefore, protein folding was highly efficient based on experimental data. They took x-ray crystallography. That’s what DeepMind did; they took x-ray crystallography data.
DESRES tried the first principles method.
Yeah, yeah. And it’s like a nice head-to-head comparison.
Yeah, yeah. Two very well-resourced groups. They both tried different ideas and the machine learning on experimental data beat out first principle simulation by, you know, a very large margin.
And so why isn’t like Bolts or whatever inside of Cosmos? Like why isn’t there a tool that can run Bolts?
There’s all these companies that basically wrap a lot of these deep learning protein design tools or chemistry design tools. They wrap them in an API. You just give that to, give it to Cloud Code if you want. You can give it to Cosmos and you can be like,
“Hey, you know, if you want to design a protein for x, use these tools.”
Your mechanism, it sounds like, or one of the primary mechanisms that has been successful is like it,
Yeah. And so how do you think about serendipity and out of distribution thinking and getting there? How far have you gotten and what’s left?
Yeah, that’s a great question. I think, I guess the short answer is that there’s very, so this is the domain of CBRN. So chemical, biological, radiological, nuclear weapons or, I don’t know, safety. Yeah. This domain has been explored a lot in history by a lot of organizations. Yeah.
And I would say that there is a big question mark for us a few years ago was like, how much of this stuff is intellectually bottlenecked? Yeah. Like, how often are people like, “oh, wow, I want to cause harm, but I need to know like some facts?” And could LLMs make that easier or go faster or anything like that?
I think, you know, the first set of answers in 2023, I think was basically no, is that like, you know, you can go find the synthesis route for many dangerous compounds on Wikipedia. People know what are the targets in the human body that like are targeted by most biological weapons. It’s not really that much of a mystery. So I don’t think there was a lot of like, there’s a lot of new ground when LLMs first came about.
And then there’s a lot of concern about like laboratory protocols is that could agents or LLMs reveal some tacit knowledge that like maybe people couldn’t find on Wikipedia or like maybe for making something, there’s some technique that is acquired when you scale it up in size or something. Or maybe there’s like some way to get around like tracking lists by ordering different compounds.
So in that, I think, was really well tested by a few different labs. It’s not me, but there were some groups that spun up that started making like tests for this and labs pay attention to it. I think it’s really been put into process where LLMs will like kind of shut down or be filtered in those scenarios.
But I think that is actually an area where there is some risk. And so I think that’s something that people pay attention to for open source models. And there’s still, I think, some discussion there. But I think to a large extent, it’s not really been greatly accelerating in practice, or at least I haven’t seen much evidence of it.
And again, I think it comes down to the fact that it’s not really available. But like you, if you look hard enough, you can find most of the information you would need to get up to no good. Yeah. In the public domain already.
But then I think now is the next frontier is like, can it somehow help you with real-time protocols, troubleshooting, like more in the loop? And more, especially in the computational side of things. There are some scenarios that are now coming into focus that could be more dangerous or more intellectually bottlenecked. And so I think people are trying to pay attention to that.
To some extent, there was like a first wave that we thought this could unlock a lot of stuff. And I don’t think it came to pass. I think there’s now an emerging sort of second wave of like there are some actually new scenarios that were just too farfetched to consider two years ago that I think are now realistic. Some smart people are paying attention to it, but I don’t think it’s solved yet.
I don’t know. It’s very vague.
So I guess like one kind of differentiator, there’s a lot of talk about AI safety in like the modern LLM, you know, ASI space. And, you know, there it’s jokes about paper or pay-per-click maxing robots or something. But like the core threat here is more like a malicious actor using this as a tool to accelerate something dangerous.
And like kind of the first order hypothesis is that you basically already have to be an expert to effectively create a bioweapon or a chemical weapon. And a non-expert, an expert already know how to do this.
Yeah, I think, you know, so each of the categories in the CBRN (chemical, biological, radiological, nuclear), they’re all a little different. But I think to a large extent, it’s a lot of like pushing material around, you know, the classical example of nuclear is like it’s a lot of centrifugation, a lot of ultracentrifugation, a lot of high pressure or high RPMs. And so it’s just you can maybe get smarter about how to set up, you know, the economy of scale to do that with an LLM.
But to a large extent, I think you can call your friend and country X and they can tell you what are the steps. It’s not, I don’t think it’s that much of a secret. It’s just a lot of like moving material around. And I don’t think it’s accelerant, meaningfully accelerated.
Now, with that said, there are all kinds of like, you know, dumb dual use things of like, maybe you want to call a company that makes centrifuges and you want to make sure that they sell you them and they go through some KYC steps and maybe an LLM can get you through the KYC faster.
And that’s like a dumb thing that like, OK, like, yes, like, you know, email makes it so that you can order centrifuges off the Internet more easily.
Is email like a dual use technology? Like, yeah, to some extent it is. So I think there’s a lot of weird second order things that we don’t pay attention to in AI safety of like, does it make KYC easier? Does it make it easier for people to know like where to order this from? Or like, what is the expected price? Or like, what should you order first? Right.
All those like sort of simple logistical things, I think, are accelerated by AI just as like a consequence of AI being an accelerating technology. But certainly, I mean, shit, guys, there’s some scary stuff. And I try not to think about it too much.
Yeah. I don’t know. I guess I don’t want to get too political, but I do think that right now the United States government is maybe taking a slower, less intensive look at safety.
And but there’s definitely people, I think, in other spaces than the U.S. government thinking about it hard.
Do you think it’s a thing people need to spend more time on? I do get waves of angst about AI. I’m sure many people living in San Francisco do get a little bit of waves of it. And sometimes I think that there isn’t enough work being done on it.
And then sometimes I think,
“Wow, like I need to mellow out. And like, you know, we have lots of time to think about it.”
What is my opinion on it then? I don’t know. I think my opinion is not formed fully.
Yeah.
You and Sam have done a lot of thinking about funding science and future of science. You have been vocal about the reproducibility crisis and other things.
First question, why this focused research organization or FRO?
Yeah. FRO.
What does that get you that you don’t get from academia or, you know, big lab or whatever?
So I don’t think we’ve had as much friction there as you might expect. But yeah, this is all stuff that we - Sam and I - think about all the time. It’s like,
“How do you balance stuff like this? How do you balance the economics?”
You know, there are some venture-backed companies that are having cash salaries over a million dollars. And it’s like insane to me that you would use all of your cash from your equity financing in these insane salaries.
In terms of like total spend on GPUs, it can still be a total small fraction of your burn. So sometimes it kind of makes sense.
Yeah. Yeah. That’s one way to think about it.
So like you, this is a good lead into: you are automating science in some capacity.
Yeah.
So where does that leave scientists? So I think this is a Jevons paradox we can try here.
Yeah.
So where does that leave scientists? So I think this is a Jevons paradox we can try here.
So let me start with the contrast here:
So when you automate that, that’s like a scarcity thing - basically, you’re displacing jobs when you automate driving.
In science, I don’t think there is a finite appetite or a finite capacity for science. I don’t think science is like a scarcity thing.
Like there’s:
I think instead, actually, if we can make science go much, much faster, there will be no decrease in demand.
There will be actually, I think, an increase in demand that will match whatever automation amount we have.
And so my vision for what a scientist would be in the future is that they will be, I don’t know, like agent wranglers or Cosmos wranglers of like, OK, they’re exploring 100 ideas simultaneously.
Or they’re like working with systems like ours to make
10x discoveries,
100x discoveries,
because I think there’s an unlimited amount of scientific discoveries to be made. And so there’s no like scarcity set where basically we will displace them all.
Now, that’s kind of like, you know, this is what I would tell when I go talk to a first year PhD student.
Yeah, it’s going to be just fine.
You know, but then when it gets into the nuts and bolts, I do agree that this is going to be like a really hard thing where like if I am CEO of a company, that makes science like a pharma company or material science company or something like that, or an R&D arm at IBM, I think,
I might just choose to go with the AI scientist because, you know, to a large extent, like hiring people is hard, right?
And hiring an AI scientist is probably a little bit easier.
And so I think that there could be some friction. But another thing is like, science is in some ways closer to art, in the sense that there is a large number of people who just appreciate good science.
Like if you get published in Nature, it’s not because it’s really going to be world changing. Of course, that’s part of it. But it’s also because people are like, “wow, this is really interesting science.”
So I think that the enjoyers of science are also scientists. And so I think that it’s kind of hard to imagine a scenario when there aren’t scientists as the consumers of science.
And so I think if they’re going to be consumers of science, they’re also going to be some of the producers who are involved in the process by itself. I don’t know if that makes any sense.
Yeah, you’ve touched on this. The question in my mind is just what does a scientist do then?
There’s a great short story by Ted Chiang, I think in 2003 or something. And it’s about:
And the problem is that science is like, you have to translate science to make any impact. Science cannot exist by itself.
I do agree engineering can exist by itself. Like if you give some kind of system a goal of “making me a material that I can make a space elevator out of,” you could be not participating in the beginning, the process, or the middle of the process. And you just come by the end and be like, “okay, all of this recipe.”
Like science of:
That has to be hitting human eyes and human brains at some point. So I think a human has to be involved in the process.
Don’t want to be contrarian, but why does a human have to be involved? Why does a human have to be involved?
Well, a human has to be involved at least at some point to be like, “yes, this is good science or this is bad science.”
Okay. So it goes back to taste. Yeah.
But I don’t know. Maybe you’re right. Maybe there is no point for humans. Maybe we’ll be like, no. What is it? Sora. You know, like the AI slop app.
But I think in Sora, there’s still humans at the end clicking on the videos or something.
So the Sora analogy kind of brings up an interesting point.
Like, is it possible that due to the biases of AI science, if we really go full in science, there still is a market for kind of boutique human science? Like, there are still people who want to paint things the old fashioned way.
But more to the point, does it become even more important to have a human who is actively doing their own exploration because there will be:
because this is sort of baked in now due to your training data. And without a human, you’ll always get stuck in there. There will be a blind spot.
That will never…
Bio is a company in Oakland or Emeryville. They do really cool stuff with automation. I think they’re going to be testing this theory of like, “OK, maybe if that’s the bottleneck, we can see evidence of it” because they’re going to start doing really well.
It could be true.
I still want to say all of those in my mind are still sort of scoped in terms of R&D for pharma or bio, but none of them are attempting to answer big fundamental questions.
And maybe there’s different levels when I think about it, and you seem to be thinking the future or how the focus of future houses in Edison is much more towards:
But you know, I have some background in fundamental physics.
Yeah.
You know, it’s like, is there any thought about how do you take on dark matter candidates? And like, I just think the data to really give us a complete story is just not there yet.
You know what?
Like, I’m sure everybody at every company is the biggest critic of their own product.
Yeah.
So we think Cosmos is great, but there’s a very large amount of area for improvement.
So with Cosmos, there’s like an open, sort of access to everybody version.
Yeah.
Do you provide access to other labs that is less open?
We have a version of Cosmos that has bigger resources, like it can run for longer. It uses GPUs, so basically when it does data analysis, it’ll have a GPU.
So we use that for things like:
- machine learning experiments
You know, if you want to know this question about whether it’s better to pre-train first on noisy data or not.
Yeah.
We have pre-release models that are coming out and we try those. But yeah.
So I guess, yes, we do. And we do have research partnerships with companies where we build something specific for them. And that is something we think about.
Yeah.
But broadly, I would say Cosmos, that’s on the website, is pretty close to what is the best we have internally.
Yeah.
I have a question. So you previously have stated that you think that language is the natural language. What is it?
Language of chemistry. The future of chemistry is language.
Yeah.
Yeah.
Yeah.
Okay.
So I wonder, do you still believe that?
Good question.
I think I would say yes. I still believe that.
So in that article, the opinion article, my point was that, at the time when I wrote that article, which I think maybe three years ago now or something, maybe 2023, it was that we have models for predicting solubility of compounds.
We have data about our large populations and we have papers and we have code.
The only way to bridge all that information is natural language.
And the argument was that like humans, whenever we can’t bridge information, like if I can’t talk about my code or I can’t talk about some idea to you, I will invent words until I can get the point across.
“Humans are always innovating on language to make it represent all known observations and people innovate on language to represent whatever code pattern they have.”
This is the only shared activity we’ve been doing for this long - coming up with words to represent everything we know.
And so I think that for that reason, natural language is the only possible way to connect all the different pieces of data we need in biology, medicine, or any domain for that matter.
I think there are some caveats to this, like you can make an argument —
There are arguments against natural language, like maybe there’s something more, or perhaps natural language imposes limitations. You cannot exceed it because you’re stuck in this abstract space that was invented by humans and you can’t escape it until you can touch something.
Yeah.
I mean, it is an abstraction. And scientists basically work exclusively in abstractions to some degree.
I just find that interesting, because it seems like most scientists are right. When they explain things, they do so through language, but many conversations, maybe most, at some point result in people drawing diagrams.
For example:
These are oftentimes a language of graphs.
Bonds are abstractions, yes, but they are pretty good abstractions for many cases.
Or, think about the geometry of a protein.
It’s like that-that’s how people often like to think about things.
So I find it interesting that you are focusing primarily on language.
Have you thought about essentially a multimodal version of this?
Like where, when it comes along a SMILES string, it doesn’t just say, “Oh, this is a SMILES string,” but like, this is a graph. This is a representation of some higher, abstract object.
You’re absolutely right.
And the problem with these, this like Jacob’s ladder or something, whatever you want to call it, is:
And here’s the electron correlation. Like, well, you know, you can model electron correlation, but actually these things, when they’re in a solution, they have relativistic effects because there’s a whole bunch of stuff around it. So you really got to have the relativity in there.
And you’re like, well, you got the relativity and you have the electron correlation. You can have the bonds and you have the conformers, but you want to think about the cosmic radiation background because, like, it does actually impact everything. And there is some energy there.
Right. And before you know it, you’ve run out of, you know, compute or whatever resource you’re using to model this. And so I think you have to draw the line somewhere.
Natural language, like I said, is that humans have worked for a long time to make it be the, you know, what’s the word? Like the least abstract or it’s somewhere on the border of like, it’s still abstract enough that you don’t need to know all these details, but it’s still granular enough or concretized enough that you actually can make use of it.
There may be some other representation like multimodal that might turn out to be video or maybe there’s some other fusion you can make. I like natural language because we all work really hard to make it right at that boundary.
And I do agree. Sometimes ideas slip and they can’t be in language. You have to get out the whiteboard or ideas slip and you have to wave your hands around, or maybe then you need that degree of freedom to communicate.
Just digging in on this a little bit more: famously, quantum mechanics is like undescribable, right? There’s an argument that you cannot understand quantum mechanics with words or with our preconceived understanding of the physical world because it doesn’t behave like the macroscopic world. And so the only way to understand it is through mathematics, right?
I largely see language as the joint key of science as well, but I wonder if that’s not true for many domains. Quantum mechanics is just the one that hits you in the face.
I mean, I don’t know. I actually think there are like seven principles of quantum mechanics or five or something like this that you can actually express pretty concisely in language. I agree that you need to actually look at the consequences of them. You need some mathematics.
I don’t know. This is like a challenge. I think you could actually describe a lot of quantum mechanics in language. Sure, sure. But I see your point.
Yeah, I guess I’m a realist. When I talk to my kids, maybe I will be like, “Okay, let me draw for you.” I don’t make sure in our house everything is described with natural language.
So I agree with you there. I think maybe we can be a little flexible with natural language and include equations and SMILES strings in it. And I think we can get a little bit farther. So maybe that’s okay.
But some people, I think, like optionality. Like, Oh, it could be this, or it could be that. I’m somebody who likes to take strong opinions and see how much farther they can get me.
I think in my career, it’s actually been better for me to take strong opinions which, in my deepest heart, I know are maybe not correct or not fully correct. But once you take these strong opinions, you can move many steps down the road.
Once you take these strong opinions.
And like, for example, at Future House, we took the opinion that scientific agents are the future. And that allowed us to skip a lot of steps because a lot of other people were like, “We need to build a foundation model for X,” and we just skipped all that.
Right. And I think if you were also unopinionated and you had optionality, like I can think of a famous example of a different company that liked optionality and they wasted a lot of time on foundation models or something. Then I think you get stuck.
So that’s one of my strong opinions: that natural language is a way to join all these different domains. It may not be a correct opinion. It may be more subtle or more complicated, but it has allowed me to get very far.
I’ll drop it someday and find a new one. Not yet though. That’s my meta opinion.
The Ether Zero story on your blog, I find hilarious and kind of awesome.
You know, when I was a kid, I loved the genie / monkey paw concept of “be careful what you wish for, but you just might get it.” Maybe just quick story, can you talk about that? That was just really fun.
Ether Zero was a hell of a project because conceptually it was a very short project of… Hey, people have made a lot of progress and verifiable rewards in math and in computer and code. Let’s say we can do it in chemistry.
So chemistry is like not a verifiable field, right? Of course, you can go test something in the lab, but then we had to think about all these ways that we can make chemistry verifiable.
And one of the ones we settled on was like, make a molecule that has:
And we thought that was a pretty verifiable question. But every time we would train a model, it would find some new, insanely weird trick to generate these molecules.
I’ll just tell you one of the examples: it would make these molecules, and we would do some checks to make sure it had the right bonds, right number of electrons or atoms. But it would just solve the problem in any way possible.
So, like, it would just put all the nitrogens over here, put all the oxygens over here, just things that don’t look good.
We started coming up with these rules, like, “Oh, let’s check to make sure it followed these good practices.” And we found ourselves into this, like, opposite of the bitter lesson, where you try to make everything custom.
One of the things it kept doing was putting these nitrogens in a row: one nitrogen, two nitrogens, three nitrogens, all in a chain. This is like, if you have three nitrogens, it’s explosive; two nitrogens is bad, and four nitrogens you can’t make.
I kept telling everyone it would make these six-nitrogen compounds, and they’re just literally impossible and not possible. Many of the people on the team were computer scientists.
One of them sent me this:
“This is on the cover of Nature today on Nature’s website. Somebody made a six-nitrogen compound.”
This is like somebody’s career to deliver this compound because this is the most unstable, insane compound you can make. Some ridiculous setup. The spectroscopy to get that proven was very difficult. It was an amazing accomplishment.
I told Andrew, “Look, it’s not actually impossible.” It was so funny to me that our model was spitting out these six-nitrogen compounds in 2024 or 2025, and the paper just happened to come out that year that mankind had finally made a six-nitrogen compound.
Do you think those were actually synthesizable even under these extreme circumstances?
No. Our model was just reward hacking. The model was so creative in ways to reward hack.
Another one we did was: we wanted to make sure that, when it would propose a reaction to make this compound, all the reagents were purchasable - not made up.
The reason we came up with that is that originally we’d just take the end compound, remove one atom, say “here, you buy this,” then put the atom on it. It’s like, well, I wish it was like that.
So they had to be purchasable. We thought it might be hard if all reagents were purchasable because sometimes you order things custom. So we’d make sure at least one was purchasable.
The first thing it did was putting nitrogen in there because nitrogen is purchasable and has no participation in the reaction.
“Oh my God!”
So then, it had to be purchasable and participate in the reaction. It started putting acid-base chemistry: just putting an acid here, because acids are purchasable, and it would move one atom.
Then the constraints got tighter: everything has to be purchasable.
I found myself sitting one day building a ridiculous catalog of purchasable compounds and a bloom filter so we could go fast enough in our training loop. And I was like, “Why am I doing this? How did I get here?”
It was really funny because pre-training or training transformers on just data, like supervised training where you have inputs and outputs directly, is very nice and relaxing. Things are always robust; things go pretty smoothly.
When we do these verifiable rewards where you have to write a bulletproof verifier, it is really difficult. And we had so many models trained only to find out they were hacking some other like random thing in our setup. It’s really hard.
And I, and I, I don’t envy the frontier labs that have to do this at a very massive scale because we had a lot of adventures in ether zero and you guys should read the blog post. It’s very fun. It’s a great read.
GRPO. We did make some modifications to GRPO. Yeah. I actually, I used to know all the names of these modifications, but I think it’s like a DAPO is one modification and like the clipping we did was special and we explored a lot of that stuff. Yeah.
And it was also one of these things where like, you think the hypers are wrong, the algorithm is wrong. And then you find out it’s just because like you had somehow sorted the reagents when you made your training data, but when you made your test data, you didn’t sort them alphabetically and the model was just like barfing because its whole strategy was to exploit something in the way you sort of things.
So yeah, we explored a lot of different methods and it was, um, I learned a lot about chemistry, a lot about nomenclature.
Um, and actually there’s a, I learned a lot about medicinal chemistry as well, more than I ever wanted to.
Awesome. If you want to do some like engineering, just check out Edison Scientific and they have, you know, I think a lot, they’re hiring with lots of like interesting things, everything from scientists to, you know, infrastructure engineer.
Yeah. Thanks Andrew again.
]]>“Thank you very much for, for joining us.”