Lovely post - glad you had a productive visit. I think the Chinese labs view the US ecosystem with deep respect / awe, so I'm sure they appreciated your respectful visit.
I think a lot of the cultural differences you described comes from the massive amounts of VC capital that has flooded into the industry. It has really perverted the growth curve.
On the plus side, it's precisely this massive flood of (risk-tolerant) capital which has attracted massive amounts of talent and capex, to enable the awesome build outs across the entire stack.
On the negative side, VC ethos has focused on building monopolies, moats, and really "owning" super intelligence. For VCs and the startups they support, it really is seen as a zero sum game - their closed source model HAS to achieve absolute dominance in order to justify the capital / valuations we are seeing.
As someone currently working at one of the labs you mentioned, I want to offer a few counterpoints.
First, tribalism between Chinese labs is real. It’s probably less dramatic than what we currently see in some US labs, but that doesn’t mean it doesn’t exist. I think you may not have seen it simply because people did not want to show it to a visitor.
Second, I want to question the engineering culture you described so positively. I agree that China has an extremely strong engineering culture, and that this helps with execution and fast iteration. But I’m not convinced that this culture alone will help China build a superintelligence.
Building superintelligence is not only an engineering problem. It requires people who are mature enough to take responsibility for intelligence. It requires multidisciplinary thinking across science, society, politics, and morality. In practice, though, these broader perspectives are often overwhelmed by a pure engineering mindset and also by power.
China’s AI sector, just like America’s, is mostly driven by tech firms and research institutes — especially universities focused on computer science. Most people working in this field are STEM majors. Compared with the US, Chinese colleges offer far less general education and cross-disciplinary communication. People tend to only care about their own specialized fields, and simply leave all other social and ethical issues to experts in those areas.
It’s not that no one in China is talking heatedly about AI’s ethics, economic impacts and social challenges. Plenty of scholars are already writing and debating a lot on these topics. The problem is, real cross-disciplinary dialogue is still pretty rare here.
If China wants its AI and large models to gain bigger influence across the globe, it’s absolutely necessary to have more cross-field discussions, build wide consensus, and turn these shared values into basic rules for how AI is built and developed.
With regards to Chinese Lex/Dwarkesh, Zhang Xiaojun 張小珺 is excellent. Like a more intelligent Lex. Not sure if there's a convenient way to listen in English.
Great read. I visited Beijing, Shanghai, and Shenzhen late last year, and had a very similar observation around how competitive intensity coexists with genuine collaboration. Most entrepreneurs I met named other Chinese startups as their primary competitors, not American ones, yet there was real respect across the ecosystem. I keep wondering what makes that possible and if there's something other ecosystems can learn from it. Would be curious to hear your or anyone else's thoughts on that.
Great post, love the comparison you made between the two. It's true that the majority of China's top AI creatives comes from pure engineering and scientific background. They believe all problems are solvable through technical means, focusing less on he boundaries of rights or procedural justice. The Chinese research environment has also long been driven by pragmatism. The proportion of high-level government officials with STEM backgrounds is significantly higher than in the U.S., thus, steering policies toward a pragmatic direction.
This is a very valuable set of field notes because it captures something that is often missed in the Western AI debate.
The usual question is whether China can catch up with the strongest U.S. frontier labs under chip constraints. That question matters, but it is too narrow. What your observations show is a different layer of competition: young engineering talent, low organizational ego, open-model feedback loops, build-not-buy data practices, practical technology ownership, and large technology companies trying to control their own AI stacks.
That is why I think China’s AI trajectory is difficult to understand through a purely frontier-model lens. The Nvidia constraint is real. But constraint does not simply stop the system. It also pushes labs and companies toward domestic inference adaptation, internal training environments, open diffusion, and parallel-stack formation.
In that sense, China’s AI strength may not only be about producing the single best model at the frontier. It may also be about building a system that can absorb more developers, more enterprises, more industrial scenarios, and more non-Western markets into deployable AI.
The U.S. may continue to dominate the frontier ceiling. China may build diffusion power. Both forms of power matter, but they are not the same.
Parallel-stack formation is the right frame, and the US side is doing the same thing under different pressure. ARM built its first chip in 35 years, Intel bought back the fab it sold two years ago, Nvidia took equity in its fourth supply-chain partner. Vertical integration happens when inference value exceeds licensing fees. Constraint just makes China's version more legible.
The cadence gap is bigger than the benchmark gap. DeepSeek, Qwen, and MiniMax have shipped frontier-class models on tight iteration cycles while several US labs spent 2025 absorbing talent shuffles and political reorgs. Org friction compounds. Architecture catches up faster than culture does.
I'm a little concerned that the Chinese researchers don't see it as their role to think about the larger implications of what they are building. Is that because they are used to the government making those decisions, rather than the private sector? I agree that it would be more productive if American researchers could focus on building more than politics but given the huge amount of fear and anxiety that LLMs are provoking in the U.S., it would be naive not be concerned about that. Otherwise, you get more misfires like Grok and the whole deepfake-nudity blowup.
Your intuition is right to a certain extent. That’s exactly why I suggest Chinese AI researchers engage more with scholars from other disciplines to reach a consensus. We should allow cross-disciplinary experts play a bigger role.
It’s not just about politics. Chinese people naturally tend not to speak out boldly on topics outside their expertise. In a way, it’s a humble virtue. But amid the huge societal changes brought by AI technology today, cross-disciplinary cooperation is no longer optional—it’s a must.
LOL, China doesn't have the same issues of a Grok or deepfake-nudity blowup as we have here in the US. It is difficult, maybe impossible, to have an anonymous social media account in China. One of the positives of having state censorship is that they are on top of abuses. It's not perfect, but the state and the culture do not put up with that sort of thing.
The organisational differences you describe map directly to which part of the stack each ecosystem treats as scarce. High participation. Open first. Rapid iteration.
Western labs organise around model secrecy because they believe the model is the moat. China's ecosystem organises around deployment speed because open model access is assumed. Same starting point. Very different binding constraints.
The consequence is that China's bottleneck naturally migrates toward deployment infrastructure and verification faster. The West's bottleneck stays stuck at model access longer. Which ecosystem has the structural advantage depends entirely on where you think value settles when models commoditise.
Great post. I believe China’s approach to AI will ultimately capture a significant share of the global market. The Chinese tech sector maintains a proactive attitude toward AI development; they are highly dedicated, and they bring unique perspectives to the open-source community. Furthermore, the Chinese AI model is both more economically sustainable and developer-friendly.They have a healthier attitude towards AI.
"To Nathan, the domestic LLM scene is basically paradise—everyone is respectful, and people keep things civil even when they don’t see eye to eye. By contrast, he can’t even bear to look at the 'Big Three' across the pond right now. It’s just non-stop, heated trash-talking, like some kind of primitive tribal warfare."
Lovely post - glad you had a productive visit. I think the Chinese labs view the US ecosystem with deep respect / awe, so I'm sure they appreciated your respectful visit.
I think a lot of the cultural differences you described comes from the massive amounts of VC capital that has flooded into the industry. It has really perverted the growth curve.
On the plus side, it's precisely this massive flood of (risk-tolerant) capital which has attracted massive amounts of talent and capex, to enable the awesome build outs across the entire stack.
On the negative side, VC ethos has focused on building monopolies, moats, and really "owning" super intelligence. For VCs and the startups they support, it really is seen as a zero sum game - their closed source model HAS to achieve absolute dominance in order to justify the capital / valuations we are seeing.
Hey Nathan, thanks for sharing a fantastic piece.
As someone currently working at one of the labs you mentioned, I want to offer a few counterpoints.
First, tribalism between Chinese labs is real. It’s probably less dramatic than what we currently see in some US labs, but that doesn’t mean it doesn’t exist. I think you may not have seen it simply because people did not want to show it to a visitor.
Second, I want to question the engineering culture you described so positively. I agree that China has an extremely strong engineering culture, and that this helps with execution and fast iteration. But I’m not convinced that this culture alone will help China build a superintelligence.
Building superintelligence is not only an engineering problem. It requires people who are mature enough to take responsibility for intelligence. It requires multidisciplinary thinking across science, society, politics, and morality. In practice, though, these broader perspectives are often overwhelmed by a pure engineering mindset and also by power.
China’s AI sector, just like America’s, is mostly driven by tech firms and research institutes — especially universities focused on computer science. Most people working in this field are STEM majors. Compared with the US, Chinese colleges offer far less general education and cross-disciplinary communication. People tend to only care about their own specialized fields, and simply leave all other social and ethical issues to experts in those areas.
It’s not that no one in China is talking heatedly about AI’s ethics, economic impacts and social challenges. Plenty of scholars are already writing and debating a lot on these topics. The problem is, real cross-disciplinary dialogue is still pretty rare here.
If China wants its AI and large models to gain bigger influence across the globe, it’s absolutely necessary to have more cross-field discussions, build wide consensus, and turn these shared values into basic rules for how AI is built and developed.
Thank you for this clear and thoughtful piece!
With regards to Chinese Lex/Dwarkesh, Zhang Xiaojun 張小珺 is excellent. Like a more intelligent Lex. Not sure if there's a convenient way to listen in English.
This is an awesome post -- amazing how much you can learn from the Chinese AI ecosystem in one trip. Without Mandarin at that.
An errata: bullet 5 refers to the above bullet but meant to refer to bullet 3 I believe.
30% of our group spoke Mandarin, so I had a lot of help :)
Great trip and such an insightful and unique report you're not going to find anywhere else on the internet
Great read. I visited Beijing, Shanghai, and Shenzhen late last year, and had a very similar observation around how competitive intensity coexists with genuine collaboration. Most entrepreneurs I met named other Chinese startups as their primary competitors, not American ones, yet there was real respect across the ecosystem. I keep wondering what makes that possible and if there's something other ecosystems can learn from it. Would be curious to hear your or anyone else's thoughts on that.
On a related note, I wrote up my own observations from a slightly different lens (more startup ecosystem, hardware, and infrastructure than AI labs): https://doctaignorance.substack.com/p/notes-from-my-trip-to-china
Great post, love the comparison you made between the two. It's true that the majority of China's top AI creatives comes from pure engineering and scientific background. They believe all problems are solvable through technical means, focusing less on he boundaries of rights or procedural justice. The Chinese research environment has also long been driven by pragmatism. The proportion of high-level government officials with STEM backgrounds is significantly higher than in the U.S., thus, steering policies toward a pragmatic direction.
This is a very valuable set of field notes because it captures something that is often missed in the Western AI debate.
The usual question is whether China can catch up with the strongest U.S. frontier labs under chip constraints. That question matters, but it is too narrow. What your observations show is a different layer of competition: young engineering talent, low organizational ego, open-model feedback loops, build-not-buy data practices, practical technology ownership, and large technology companies trying to control their own AI stacks.
That is why I think China’s AI trajectory is difficult to understand through a purely frontier-model lens. The Nvidia constraint is real. But constraint does not simply stop the system. It also pushes labs and companies toward domestic inference adaptation, internal training environments, open diffusion, and parallel-stack formation.
In that sense, China’s AI strength may not only be about producing the single best model at the frontier. It may also be about building a system that can absorb more developers, more enterprises, more industrial scenarios, and more non-Western markets into deployable AI.
The U.S. may continue to dominate the frontier ceiling. China may build diffusion power. Both forms of power matter, but they are not the same.
Parallel-stack formation is the right frame, and the US side is doing the same thing under different pressure. ARM built its first chip in 35 years, Intel bought back the fab it sold two years ago, Nvidia took equity in its fourth supply-chain partner. Vertical integration happens when inference value exceeds licensing fees. Constraint just makes China's version more legible.
The cadence gap is bigger than the benchmark gap. DeepSeek, Qwen, and MiniMax have shipped frontier-class models on tight iteration cycles while several US labs spent 2025 absorbing talent shuffles and political reorgs. Org friction compounds. Architecture catches up faster than culture does.
感谢你平实的描述，非常感谢
I'm a little concerned that the Chinese researchers don't see it as their role to think about the larger implications of what they are building. Is that because they are used to the government making those decisions, rather than the private sector? I agree that it would be more productive if American researchers could focus on building more than politics but given the huge amount of fear and anxiety that LLMs are provoking in the U.S., it would be naive not be concerned about that. Otherwise, you get more misfires like Grok and the whole deepfake-nudity blowup.
Your intuition is right to a certain extent. That’s exactly why I suggest Chinese AI researchers engage more with scholars from other disciplines to reach a consensus. We should allow cross-disciplinary experts play a bigger role.
It’s not just about politics. Chinese people naturally tend not to speak out boldly on topics outside their expertise. In a way, it’s a humble virtue. But amid the huge societal changes brought by AI technology today, cross-disciplinary cooperation is no longer optional—it’s a must.
LOL, China doesn't have the same issues of a Grok or deepfake-nudity blowup as we have here in the US. It is difficult, maybe impossible, to have an anonymous social media account in China. One of the positives of having state censorship is that they are on top of abuses. It's not perfect, but the state and the culture do not put up with that sort of thing.
Yeah, I couldn't agree more.
The organisational differences you describe map directly to which part of the stack each ecosystem treats as scarce. High participation. Open first. Rapid iteration.
Western labs organise around model secrecy because they believe the model is the moat. China's ecosystem organises around deployment speed because open model access is assumed. Same starting point. Very different binding constraints.
The consequence is that China's bottleneck naturally migrates toward deployment infrastructure and verification faster. The West's bottleneck stays stuck at model access longer. Which ecosystem has the structural advantage depends entirely on where you think value settles when models commoditise.
Different ways Kung Fu it starts there ..
Great post. I believe China’s approach to AI will ultimately capture a significant share of the global market. The Chinese tech sector maintains a proactive attitude toward AI development; they are highly dedicated, and they bring unique perspectives to the open-source community. Furthermore, the Chinese AI model is both more economically sustainable and developer-friendly.They have a healthier attitude towards AI.
Now every Chinese knows this trip: https://mp.weixin.qq.com/s/G2O42ck4vej97Qg1osAKLg:
"To Nathan, the domestic LLM scene is basically paradise—everyone is respectful, and people keep things civil even when they don’t see eye to eye. By contrast, he can’t even bear to look at the 'Big Three' across the pond right now. It’s just non-stop, heated trash-talking, like some kind of primitive tribal warfare."
Hilarious that this article got banned
Hmmm I can’t seem to access this link. Is it from some 公众号 that I could look up?
From 量子位, one of china’s most well-read AI 公众号！