Differentiated Understanding

Nathan Lambert Reflects on China’s AI Labs: DeepSeek, Open Models, and the 'Race' with the U.S.

Joining me today is Nathan Lambert [https://substack.com/profile/10472909-nathan-lambert], author of Interconnects AI [null] and a post-training lead at the Allen Institute for AI. Nathan recently returned from a major tour of China’s leading AI labs, where he met with researchers and teams building some of the most impressive open models in the world. In this conversation, we discuss what Nathan saw on the ground: how Chinese AI labs differ from their U.S. counterparts, why open models have become such an important part of China’s AI strategy, and how labs like DeepSeek, Alibaba, ByteDance, Kimi, Z.ai, MiniMax, and others are navigating compute constraints, data access, and commercialization. We also dig into some of the most debated questions in AI today: Are Chinese labs really 6-9 months behind U.S. frontier labs? How meaningful are distillation accusations? Can domestic chips like Huawei’s make up for restricted access to Nvidia GPUs? And is China’s AI ecosystem actually government-directed, or is the reality more fragmented and commercially driven? Ultimately, this episode is a more nuanced look at China’s AI ecosystem that looks beyond simplistic narratives about subsidies, copying, or geopolitics, and instead examines the technical, cultural, and economic forces shaping the future of open models. Check out his two recent articles here: * Notes from inside China’s AI labs [https://www.interconnects.ai/p/notes-from-inside-chinas-ai-labs] * How open model ecosystems compound [https://www.interconnects.ai/p/how-open-model-ecosystems-compound] To find the previous episodes of Differentiated Understanding, see here. [https://aiproem.substack.com/podcast] Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently. Season two will host a series of guests from early-stage investing, as well as builders, researchers, founders, and product managers. For more information on the podcast series, see here. [https://aiproem.substack.com/p/launch-of-differentiated-understanding] Chapters00:00 Insights from the China Trip11:51 Cultural Differences in AI Research18:15 The Role of DeepSeek in China’s AI Ecosystem25:26 Overview of Major Chinese AI Labs30:56 The Future of Open Source in AI37:50 Market Dynamics and Consolidation in AI42:28 Distillation and Model Convergence Controversies51:58 The Gap in AI Performance: US vs China61:09 Monetization Strategies in AI: A Comparative Analysis62:32 Government Influence and Misconceptions in AI Transcript (AI-generated for reference only) Grace Shao (00:00) Nathan, thank you so much for joining us today. Yeah, really, really excited to finally hear your thoughts on your big China trip, on what’s happening between the Chinese AI labs and the U.S. AI labs, what you think the potential compute constraints might mean for these labs and their performance in the future, and obviously the open-source ecosystem. So before we get into all of that, could you... Nathan Lambert (00:02) Yeah, thanks for having me. Grace Shao (00:23) Briefly tell us about how you ended up actually working on post-training and open language models. Just a bit about yourself. Nathan Lambert (00:29) Yeah. So I actually started my PhD at Berkeley in 2017, not working on AI things. I was an electrical engineer by training in undergrad, which is funny looking back, because that’s the same year that the Transformer paper came out. And I was like, I think I should do this AI thing, and tried to get the famous advisors to mentor me. And they’re like, we can’t take you. So I had my PhD as this wandering path to become an AI researcher. And then I ended up at Hugging Face after that, which was, realistically, the only industry research job that I had, but also a very hot startup and very fun to learn kind of at the intersection of these tools that people use a lot for AI and research, which is what I was doing. And then when ChatGPT hit, the kind of RLHF thing blew up as the hot word on the technical side of things. My PhD had ended up being in reinforcement learning, which is just the first half of reinforcement learning from human feedback. So it was kind of a natural pivot to be like, well, I might just do that. And Hugging Face was a good place for doing that, because the whole company is kind of all for that, which is like: figure out how to support the community on the hot thing and build platforms there. So they were very happy about that. And I helped build a team at Hugging Face. And then I was kind of burnt out on the remote-work time-zone thing and found out that the Allen Institute was doing such similar stuff. And I was like, wow, I have people that could be in-person friends and do similar things. I was like, quality of life — I need to do this. And a few years later, I ended up building a bunch of models. And I think being at a nonprofit opened me to this ecosystem vacuum of information, where there aren’t many people who can talk about what they’re doing. So then, with some luck and committing to write every week, I just feel like my influence filled the vacuum of nobody saying reasonable things. And it is this nice synergy between what I write about and what I work on in my day job, and it just kind of got bigger and bigger in a very fun way. I think that, generally, at the highest level, I’m motivated by wanting AI to go well on this trajectory. And I worry about a lot of near-term things, whether it’s social unrest in the U.S. and just kind of the massive hatred for AI — I think is a very big near-term problem — and then, medium term, concentration of power, because I think AI will be super powerful in ways that people don’t expect. So generally, open models are a nice way to curb both of them by being a bit more transparent to people, and it naturally is a hedge against concentration of power. There have been different reasons throughout that, but that’s kind of a recurring theme in my life in the last few years. Grace Shao (02:50) Definitely. I love your work because I think you help non-technical people like myself really understand what’s behind what’s happening in these labs a lot better. And then I actually just spoke to your former colleague, Tiejin Wang, and he was with APAC Hugging Face just last week. He was saying the same thing. Open source, in many ways, is kind of the best way to go forward as we know that this technology will not stop evolving, but it’s the best way to kind of put up guardrails and checks and balances for the monopolies. Okay, I don’t want to take up too much time on that side of things today because our focus really is about your China trip. Before we get into the weeds of all that, I want to hear about the trip itself. Most people who are writing about Chinese AI are getting their information secondhand. You really went there, you spent time with the researchers, you met with people who are building the models. Tell us about what you meant when you said you came back with great humility, right? Your eyes are a bit more open, whether it’s the good or the bad. Tell us about your trip. Nathan Lambert (03:50) I feel like I kind of went in — I mean, I had this horrible English phrase in my writing, which was like, “I knew I knew nothing about China,” which kind of tried to indicate that I knew going into the trip that I knew nothing. And it was still the fact in my current writing. This is a horribly written sentence that I had in there. And I only talk about it because somebody called me out on it. It’s like, what is this? And it’s like, leaving, which is knowing that it’s such a big country, there are just such vast amounts of talent working on these problems, and how unpredictable it is as a human to model people with very different worldviews and upbringings and training systems. Realistically, the way that people are trained in China is very different. And I just think that even being there, you can’t fully grasp: what are the pockets of three to six researchers doing that is actually a bit different than in the West, even if they’re working on the same goal? I think you could get down to that level of granularity and a sociological study and actually see differences in what they’re working on, and that’ll always change the output. I didn’t get to that level of granularity, but it’s just to start having real experiences and understanding how people explain how they work on these problems. And for me, realistically, a lot of it is coalition building, which is just like: I want there to not be vitriol at the level of the technical companies doing things in international bodies. So just meeting all the labs on both sides is really nice, because you need to do that for them to talk to you about more sensitive issues in the future. I got some criticism on the piece, which is like, this is how you shouldn’t visit China. And it’s like, well, what are you going to do if you’re going on an official visit to a bunch of companies? How do you expect to get in the door without being nice? You have to start somewhere, and I think it’s important to be respectful. Grace Shao (05:31) I think the piece was, frankly — I don’t think the criticism was fair, to be honest, because I think you were really transparent with the fact that you’re not a China person, right? It’s not like you’re going there and exoticizing everything. And if anything, a lot of people, even with China backgrounds, like to use certain dragons and tigers to describe things. I feel like you actually were really humble going and being like, I’m just a technical dude meeting with these labs, talking about their technical research, right? And then because you were physically there, you had observations of the culture and the people. So yeah, I actually thought your piece was quite good. And yeah, sorry. Nathan Lambert (06:05) I agree. I was willing to let that sail past, but I think it’s important for people who listen to realize how actively these companies are trying to court Western audiences, which is why we could get in the door. I mean, we had some prominent people on this trip, but that’s why we got all of them in the days that we wanted them, except for DeepSeek. So essentially some, like Catherine Rintel, who works with me at Interconnects, and some other creators... Grace Shao (06:23) How did you get everyone? Yeah, how did you get everyone? Nathan Lambert (06:29) He used to live in China and has connections in China. So he kind of orchestrated the mix of his connections and leveraging my connections to labs. We had some bigger names on the trip as well. Just stringing all of these together to get all the various labs in place is a few months of networking to make sure the trip lines up with people with established networks and contacts with the various labs. But these people want to look good to Western audiences, so they’re only going to say yes to the right researchers. And the researchers know that there are two to four comms/ops people in the room, hanging out, making sure that it goes well. Especially the bigger the company, the more comms people. You go to Alibaba and there are three to five various people, from the head of comms to some special offices. You’re not going to get these people in the office, or at all, without accepting the cost of these types of handlers. It’s the same thing in the U.S. You’re not going to just plop a senior executive into a chair. So it’s also good because now I have the WeChats of a bunch of researchers from China that I could just text about things. It’s like, hey, congrats on the new model release. It’s like Lay Lee works at Xiaomi, Xiaomi MiMo. It’s like, talk to this guy for an hour at a mall — I don’t remember the name of the tea store — but it’s like... Grace Shao (07:27) No, of course. No, of course. Nathan Lambert (07:49) Now we have these relationships, which is very useful, and that helps information spread across the ecosystem to these trusted parties, which doesn’t really exist. There are not that many, I think. And the opposite direction of the trip is very hard because Chinese researchers can’t really enter the U.S.; the visa purgatory is too complicated. A lot of us on the trip were either Canadian or entered on a transit-without-visa entry, which makes it very easy for American technical talent to go to China right now, which is why I think there are so many trips. I think there’ll be more of them. We’ve got a lot of inbound from VCs and open-source labs in the U.S. that want to establish collaborations with these various labs because they’re the best open-weight models, and they want to build a stack for companies in the U.S. building open-weight models. So I think there are going to be more prominent, but not gigantic, U.S. startups going to try to build these relationships, which I think is a really interesting technological development because we’ve never seen this type of professional work trip in China from U.S. tech companies. Most tech companies have a “bring a device to China, it auto-bricks itself, and you have to hand it into IT.” So to actually proactively send people in a professional capacity is a really big change. There are a lot of angles you could take this, and I think it’s cool to see how it unfolds. This isn’t even really about the trip. This is the follow-on that we’re hearing from people that are like, hey, how’d you do this? We want to do this trip. Grace Shao (09:06) Yeah, definitely. Actually, from my end, I hear about VCs or investors always being quite active going to China because previously American funds were very, very active during the internet era. People were kind of always trying to find a way to either get into these good deals or potentially keep their pulse on it. But I think it’s really, really positive for the whole AI ecosystem to have this kind of fair, transparent exchange in some capacity. But to your point, there’s no way that star researchers can come out and talk to you off the record without any compliance, because that doesn’t happen in the U.S. either. That’s just companies protecting themselves. I just think your trip was quite meaningful, and I want to bring it back to your observations. You talked a lot about the cultural aspects of it. You talked about how you felt like in China there was less of this star-researcher celebrity status around people. People were more humble, or there was more humility. It was very focused on execution. You argue that Chinese labs are particularly well suited to the current LM-building game because they’re very focused on meticulous stack-level work. And there’s less ego sometimes to work on the dirty work, or the non-sexy work. So kind of unpack that for us. Why do you think that is? You kind of touched on it — you said they were brought up differently, they were taught differently — but what’s so different? Nathan Lambert (10:27) So essentially, an interesting part that synergizes on this trip is that we stopped by some academic institutions. I think it was like AIR and Tsinghua and stuff. And you hear all of these academic leaders talk about how they’re pushing hard to try to change it. So yes, they know China is producing more papers than anyone else, but they still think that it’s not as transformative of research. And they think that they’re trying to cultivate the academic domestic ecosystem to change just the type of work it works on, and the distribution, and take more risk. And then you would talk to some industry leaders off the record behind closed doors, and you would hear things like, it’s never going to change because the education system is so structured. There are so many layers of the funnel that reward things like memorization and stuff that they’re just like, this research culture is not going to emerge. And then the follow-on with the AI labs is that these labs are doing fast-following. They kind of have a proof of concept, and they know what it needs to look like. Therefore, in that domain, you’re not trying to invent the new paradigm. You’re not trying to make the model that is o1 or o3, or the first model to work in Claude Code. You’re like, I see it, and I’m going to try to do that and make it the best thing. And I’m going to try to make it cheaper and just maximize that goal. A lot of companies don’t need to invent the new paradigm. OpenAI has done this so many times. That’s their bread and butter: never doubt OpenAI’s ability to release a blog post and a plot that changes how people think about AI. I still think it’s going to happen a few times in this massive boom over the next four years. OpenAI just kind of has that sense of what is the thing that you can push on a bit earlier and just transform things. But I don’t expect — and other people wouldn’t expect — the Chinese companies to do that as much, because it’s just such a culture of, I guess, building. I don’t know how to describe the positive version of this. Maybe it’s slightly more practical-minded, in terms of: it’s your job to build this thing. A lot of the researchers, maybe because they knew their managers — some of them had managers in the room — see their role in the company as being to make the models excellent. And especially for students, I work with students and that’s what they say. I work at the Allen Institute and we have students that will co-lead our language models. It’s not that surprising, because if you do an industry research job in the U.S., a lot of mentors will tell you that you’re kind of free of the burden of bureaucracy and politics. So the naivety of students, and the simplifying, is actually so good at just getting a lot of technical work done. There’s also the life-stage side. If you’re younger, you don’t have as much family, and you normally haven’t built up as many habits and other things you do with your life. Language models are so complex, and the amount of context that you need to absorb to understand what the bottleneck is — there’s so much information, and you have to be able to pick what the bottleneck is and break it. If you just don’t have the mental space to absorb all the context, you kind of end up doing things that are cute but don’t make breakthroughs on the model. So that’s kind of a difference that I’ve seen in people who were both very successful academically before language models. Some of them are able to pivot to this practical mind, which is: what is the state of the system? How do I improve it? And then some try to make kind of these abstract frames of what’s happening and approach it like an academic, and it normally doesn’t improve the model as much. So I just kind of see, if the academic system is a bit more practical-minded, a bit more structured, and the work you’re doing is structured in the language model — make this kernel implementation faster, make this idea work — then maybe it can be... I think it’s an oversimplification. I push on that a bit in the piece just to really contrast what you could think a U.S. lab would look like. And I have a few anecdotes. I’ve heard a U.S. lab paying off a researcher to be quiet about their thing not being in the model. All of these one-off things are more storytelling devices than anything, because most one-off things don’t matter at all. But also Llama 4 imploded, and that was because it was described as a Game-of-Thrones political-style environment, with all the VPs vying for influence and showing that their thing made the benchmarks go up. It kind of fell. Many, many people will tell you that. And we’ve had the Qwen turnover, but it doesn’t seem like it was quite the same type of thing as Llama 4 or xAI. xAI barely exists now. There have been some dramatic things in the U.S. with how these companies have kind of come and gone out of the fold. Grace Shao (14:55) Yeah, I kind of agree with you, but also I would push back on that. I think there’s obviously a more rigid and competitive academic system, which by default in East Asia results in a culture of students following the bureaucracy and authority a bit more. So I agree with you in the sense that they’re very pragmatic. They focus on the task that is given to them. However, I wonder if things will change with how AI will disrupt education. That’s number one. But also, a lot of the young researchers that you’re working with today seem quite different. At least a lot of the entrepreneurs I meet today are born in the ‘80s and ‘90s, some even younger and born in the 2000s. And I think there’s a kind of aura or confidence coming from them. If anything, you want to say they’re a bit more individualistic-minded. You went to Shanghai, right? They are dressed very, very uniquely. They have these outrageous outfits on the streets. People are seeking individual ways to showcase their personality. So I wonder if that will shift. But for sure, for the academic institutions like the Tsinghua and the Beida of the world, they are still very old-school. But I would say that is the same maybe in some academic institutions in the West still. Okay, I think on this topic we can go off on a tangent on academics, but let’s go back to China’s ecosystem. When DeepSeek V4 came out, we talked about it offline, the two of us, quickly about a piece I wrote saying how DeepSeek is starting to look a bit more like a base layer for China. And if anything, some of the labs kind of admitted to that. They’re like, we have very limited resources. And to your point earlier... Nathan Lambert (16:11) Yeah, you could take that in so many tangents. Grace Shao (16:34) Limited people — these labs are tiny. They’re run by 100 to 200 people max. Limited capital, obviously limited compute. They have constraints all around. And in that sense, in a way, the ecosystem’s looking less like a zero-sum game and more like different players optimizing their own strengths. So correct me if I’m wrong, but DeepSeek is providing a base layer where a lot of labs will quickly follow and basically adopt a lot of their engineering breakthroughs. And then Zhipu, Z.ai, will focus on the coding; MiniMax focusing on the multimodality, et cetera. There are a lot of these different players. ByteDance, obviously, very, very focused on their video models. And Qwen, like you mentioned, had the whole open-source saga break apart with Lin Junyang leaving. But in general, they’re still kind of the leader in hyperscalers on that front. So everyone’s doing their own thing almost, instead of really... Nathan Lambert (17:27) I agree with the people specializing, which I think is normal business evolution. You figure out a bit where you’re good at. And there’s so much opportunity that they are like, okay, I’ll follow this because they see that they’re good at it. I just am more skeptical of DeepSeek as a base because I have no idea what DeepSeek is doing. And some of the labs when we were there, because DeepSeek V4 had just come out, were like, yeah, we look at the things they’re doing, but they seem more intricate than needed. And if you read the paper, there’s just so much going on in this model. As a researcher, I’m like, some of it seems a little fake or a little dependent on their setup and not necessarily going to work in every model. Grace Shao (18:04) What does that mean? Break it down for me. Nathan Lambert (18:18) Essentially, I will say that building an LLM is dependent on where you have your GPUs, your pre-training dataset, your intended deployment setup, and stuff like this. So you make decisions based on your constraints, and you build the model. DeepSeek has these constraints and they end up with their model, but Moonshot and Zhipu have different constraints, maybe more flexibility, and they ended up building a different model. They will test the DeepSeek innovations. So they’ll say things like, X innovation doesn’t improve our model. These two organizations are on different development paths that have core similarities, like these large mixture-of-experts models and the general methods are similar, but a lot of the parts end up being a bit different. That’s why I’m like, I don’t know exactly. If DeepSeek was a base, you would see the Chinese labs just do post-training. We just take the base model that’s out there and we adapt it to our domain of specialty. And we have users that do that, which is something that I think about a lot. I’m thinking about starting a post-training lab and how to format post-training research better. So I think about this a lot. I think about what a shared base actually would be. They go through — some of these labs put an extreme cost on creating their base model. And if they didn’t need to do that, they wouldn’t. One of the labs told us how long their pre-training run was, and my jaw dropped. I was like, that’s way too long. Any U.S. advisor would be like, you’re taking way too much risk on this pre-training run. If they didn’t land that pre-training run from one of these past big MoEs at a Chinese lab, I don’t know if the company’s dead, but that’s a huge amount of time. Most U.S. companies now know that you don’t want your big pre-training run to be more than a few months because it’s just so much risk and time to put all your eggs in that basket. That’s a sign that, in that case, they don’t have as big of a peak-size cluster. Essentially, pre-training time can come down a lot when you have a bigger overall cluster; you can just get more throughput on it. But if your biggest cluster is smaller, it’s harder to get a certain amount of throughput, so you use that one for longer. That’s a compute constraint. To loop it back, I think the specialization is real, but I’m more like, I have no idea what DeepSeek is doing. I know they’re raising money now. I don’t know what the plan is there. They seem the most without a specialty in the Chinese ecosystem. Grace Shao (19:59) Dependency on. Yeah. Mm-hmm. No one knows, though. No one knows. They’re secretive. But that’s my point, right? I feel like they’ve been kind of nationalized, whether willingly or not, because they’re taking the Chinese government’s money. They’ve kind of gone secretive. And it’s not like there’s a secret that they prefer Chinese-educated researchers. They’re keeping a very domestic stack, from talent to capital to the whole stack. So to me, it seems like they’re being Huawei’d, in some ways, because they did well and they got their name globally, and then by default they’re becoming the next Huawei, willingly or not. Nathan Lambert (21:01) I don’t think nationalization makes you a base for the other companies, at least not at this stage. There could be something, but it’s hard to force. Grace Shao (21:06) But then you have some incentive, right? But then it is some incentive. You’re like, well, if you can propel one of the teams and propel the whole industry as a whole, it could be in your KPI or some kind of unspoken expectation. Nathan Lambert (21:17) The coordination problem is so hard. Essentially, both in the U.S. and China, even the open labs, what they do is they fork open-source code and match it to their internals, and every company does this. Therefore, all the improvements that could potentially be going to the open code and forming this base that is far more efficient — they’re not completing the feedback loop. I think China could be closer to it. If people really lean into DeepSeek as a standard architecture and DeepSeek shared their training code and all the specifics and how to do this, from a Chinese economic perspective, that would be a huge win because you’re just saving compute. But I think it’s too decentralized and too competitive to have that happen. It wouldn’t happen in the U.S. either. Grace Shao (22:04) It’s so cutthroat. Yeah. Nathan Lambert (22:08) Even though I think for open models to be closer to the frontier, it would be better. I talk about open models in the U.S. needing a consortium. But there’s definitely enough money to make a consortium in the U.S.; then you fail because the model won’t be good because you’re feeding too many asks into the model. That’s the only way to create a shared base. Grace Shao (22:25) Interesting. So it’s not really just commercial. Yeah. It’s not the commercial reason. Okay. So if you had to give a high-level commentary on each of the major labs, what would it be? If you look at ByteDance, Alibaba, Tencent Hunyuan, if they’re relevant, DeepSeek, Moonshot, Zhipu, MiniMax, Meituan, Xiaomi now being part of the ecosystem too. Nathan Lambert (22:46) You might have to prompt it or say more, but I could just kind of ramble through them, which is kind of fun. Alibaba: cloud-focused, understands that open models can enable more usage of platform. So I would say Alibaba is very, very cloud-focused. ByteDance: mostly characterized by everybody else being intimidated by them, and very user-focused, including multimodal. Kimi: vibes of the office were great. It would be one of the best startup vibes that you would visit among U.S. or China. Zhipu: very AGI-pilled, surprisingly cautiously excited about being entity-listed, even though they have no idea why they are, because they’re like, it stamps them as a big deal. And then there’s some... Grace Shao (23:27) I think they previously worked with SOEs. That’s the main reason. Or they still do, but that was one of their main sources of income. And unfortunately, because a lot of these labs spun out of Tsinghua, and Tsinghua is, for people’s context, in Beijing. It’s really close to the government, obviously. But the thing is, when it’s close to the government, it could mean there are three layers of agency underneath the actual government apparatus. But then people like to link it to the fact that it’s taking government money, so therefore they are suspicious. It’s very unfortunate, I think. A lot of companies get thrown into that category. Even companies like Lenovo and a few other Chinese companies have previously been called out by U.S. senators saying, they’re taking Chinese government money, but really it’s that their scientists or their research labs spun out of a certain government-affiliated or government-funded academic institution. That’s what it is. Anyway, yes, go on. Nathan Lambert (24:23) Yeah. Some more would be: Xiaomi — surprisingly great research vibes for a new team at a random company. They seem to be crushing it. Grace Shao (24:31) What do you think of Luo Fuli? The star researcher. Nathan Lambert (24:31) I didn’t get to meet her. I think she’s as close as they have to a star researcher right now. There’s the tier of star CEO, which there are obviously others — Dario and Sam, the analogies are there — but the star researchers, like the Sholtos of the world in the U.S., obviously you can come up with many more. She’s the closest you have to this. I need to watch more interviews. We’ll see. But she wasn’t in our meeting. But they just seem to be doing the right thing. They’re making general models. They don’t really have specialization yet. Florian, the person who helps me write about open models on Interconnects, and I took a detour to go see Meituan because we’re like, why is Meituan building these models? And they’re very practical about it. It was a less glamorous visit at a normal tech office. It wasn’t an official visit for them. They were like, yeah, we’re a major online platform. We obviously are going to use LLMs everywhere once we need to build our own LLM and specialize it to our products, which, surprise, is very practical-minded. I’m guessing there are many more companies in China like this. Grace Shao (25:39) That’s what Tencent’s saying too. It’s because they want to serve their existing consumers and optimize their LLMs for their own distribution and their own basic interface or activity loop. Nathan Lambert (25:52) Yeah. After I left, some people in the group went to Xiaohongshu, like RedNote, and they’re there. They’ve released some language models that are multimodal. They’re like multimodal data-processing things. So a lot of them are not that surprising. The startups just have different cultures. I have met some MiniMax people before, so I left the trip early before MiniMax on this one. But MiniMax was quirky. They have a ton of women in their company, which was very fun. And they have products. They’re maybe slightly more product-focused, but I feel like the quirkiness of the company kind of matches maybe Western confusion over what their products are doing and what they’re trying to do. But it kind of matches their language models that are a bit more efficient. Grace Shao (26:35) Well, they came out with a lot of very consumer-focused applications, right? They had Hailuo and Talkie, all these character companion-bot products before. Nathan Lambert (26:45) Yeah. And then the last one I went to was Ant Ling, which is also very corporate, but in a less intense way, because I think they see it as serving their own products, whereas Alibaba Cloud is like, this is the gold mine we have to win. It’s a much bigger deal for them than Ant Group. But a lot of these things, when you list them — I don’t know, eight to 10 companies — they’re all pretty reasonable with respect to the age of the company and what the company does best. There’s not as much confusion. Grace Shao (27:14) Yeah. And Ant is low-key best at medical chatbots right now, which I guess makes sense because everyone has access to Alipay. And then for seniors, apart from WeChat, it might be the only application they’re using on a regular basis. So it became the default medical consultation app, which is really random, but it’s their niche now. Yeah, I think you’re pretty spot-on. It’s pretty cool that you got those takeaways, even just meeting with them for a couple hours. Nathan Lambert (27:41) I have been reading about them for so long, so a lot of these priors are easy to confirm when they kind of fit with things you have seen. The Chinese showroom culture is so interesting, and also one of the most surprising things to have at software companies. It’s so funny. They’re definitely appealing to Western audiences. Z.ai had poorly translated merch. What was it? Something so — it would be borderline inappropriate translation in the U.S. It was like “ship big, go hard,” or something. Just some really weird translations. And they have live API statistics in their showroom. So Z.ai was like, we’re serving 5.5 trillion tokens a day. All the U.S. companies are so closely watched for when they announce token statistics. I know at least one of these numbers is wrong. It’s something like Fireworks does either 30 or 300 trillion tokens a day — or I meant Together for that one — and then one of Fireworks or Together, and the other one, are like 100 trillion tokens a day. Don’t take these as sourced; go look them up. There were some public announcements recently, but those were the first updates that anyone has on major infra companies in the U.S. Inference is a huge market. You don’t hear anything from Fireworks because they’re just struggling to demand and they’re making bank, because inference is a much better thing to sell than bare metal. Essentially, inference is selling the software implementation to serve tokens more efficiently, and you can just get more margin when you improve the stack for a fixed model. So a model comes out and you host it, and then you can make your stack more and more efficient on that model. You just get more margin and hopefully growing usage. That’s way different than GPUs, where the best case is that you lock in a huge commitment for a long term. Just being able to walk into an office and learn about their API is interesting because they also had geographic distribution, which was like: China was, I don’t know, two-thirds; U.S.A., 20%; and then the last percent was Singapore, Korea, Japan on the Z.ai API. So that’s cool. This is s**t that I always want to know about the companies, and I have no idea. One of the things I always want to know is: how are open models being used outside of the U.S. and China, and has this decades-long process of technological diffusion started to kick in in a way that any company can measure? I don’t think anyone has good data on it yet, but I think it’s obvious that at some point, open models that are cheap to run are going to have some interesting playbook across the globe for the long tail of countries. Maybe I’ll just walk into the front door of a Chinese open-weight company and get my answer. Grace Shao (30:30) But actually, I think the culture of these labs — a lot of them, because they’re run by really young, passionate people — you would feel like they’re a lot less commercialized or less corporate, or at least less sleek. They’re not sophisticated with, you can say, the capital-market side of things, but you can also say that they’re just really naive and open-minded and passionate about the product they’re working on, with less of a corporate guardrail built around them. Nathan Lambert (30:56) Yeah. Grace Shao (30:57) Okay, I want to talk about... Nathan Lambert (30:57) Yeah, go ahead. It’s like one of the people at Z.ai who’s known on X — I don’t know, 9,000 followers — it’s like Lu. She came up and was like, hi, I’m a student, I’m 20. I’m Lu from X. And I was like, that’s hilarious. There was a lot of s**t like that. It was like, oh, okay. I don’t want to call her a kid, but it’s like... Grace Shao (31:06) Yeah, yeah, yeah. And I think the one that runs Moonshot’s developer ecosystem or something is literally a girl fresh out of school, right? And she just posts hilarious memes all day long. There’s no filter on her social media. It’s funny. Okay, we go on these tangents, Nathan. We need to come back on track. Open source, open weight. Why? Why do you think Chinese labs are adopting it or embracing it, however you want to put it, especially after visiting them? Is it because they simply have to, because of what we talked about — they are leaning on each other because of all the constraints they have? Or do you think the philosophical drive is actually bigger in that ecosystem? Or is this a bigger strategic thinking for diffusion in the long run? Nathan Lambert (31:54) I actually don’t feel like it’s that special ideologically. I think it’s easy to say the ideological line when you are doing it. Now you can look at Zuckerberg: he said the ideological line when he was doing it, and then he stopped. I think it’s mostly just that, for one, distributing within the U.S. ecosystem, especially to enterprises, is the highest-value market, and they can’t sign many enterprise deals. And the closest best thing is things like Cursor adopting Kimi’s model. Even if Kimi doesn’t get paid for that, they’re happy. That’s the biggest sign of credibility for them, and they can figure it out in selling tokens or whatever in the future. Practically speaking, one, the only way to influence the U.S. market is by releasing these models. And two, it seems like they don’t feel like they’re losing as much if they release and share things. If the model was closed, they just think they would get less influence, they would be seen less, fewer people would use the model, their actual paid offerings would be adopted less. It just seems almost overwhelmingly obvious, because there are all these benefits and not as obvious of a drawback. There will always be better models, and just keep going. But I think every scientist loves... Grace Shao (33:08) Then why are so many U.S. labs against it, or not willing to? Nathan Lambert (33:12) Because they can make as much money without it. Anthropic and OpenAI make more money by not releasing them. They can just make so much money, so why bother thinking about an open model that doesn’t make money? There are different scales of influence. Same with Google. Google’s making so much money. I think Meta will make a lot of money by having good AI models in their products, if they get their act together. Even Google could release more models. They have so many surfaces other than Gemini that need AI to be commoditized and used, like the cloud and all of this. Meta could release the models. It’s just not worth the effort for some of them. They’re like, we need to do this high revenue target; it’s too much of a pain to go through legal and make it ready to release. Why bother? I don’t know, maybe it’s a little bit of a cynical take, but I think Microsoft and Meta could release their best models openly because they benefit if it’s a commodity layer. But I don’t expect them to, because it’s just kind of like the benefits of focus are so high, and they just kind of see it as something they don’t have to do. Grace Shao (33:56) And it’ll be good for them. But then eventually, we will see some consolidation in the market as well, assuming — because you can’t really have 10 labs in each dominant country right now all exist. Nathan Lambert (34:26) I do expect consolidation. I think this is potentially a subtle cultural point, which is that the U.S. labs are more likely to buy into “we’re special, we need to go fast, keep it closed,” and the Chinese labs are not. There could be something there. That’s also who the decisions funnel up to. I don’t know. I talked to the Alibaba people that make these decisions. I can’t say all the things that they say about them. Some of these were two-on-one and off the record, so I can’t say all these things. But at all the other labs, there is a person that makes the call, I’m guessing. I think those are senior leadership that we’re not talking to. So it’s kind of hard to know exactly what they really think. I definitely expect consolidation. My thing is that I expected it in China faster because the capital markets aren’t as strong as in the U.S., but I don’t have a model for that. I think you can model it, which is: what do you think the revenue growth would be? What do they need to do to raise to keep training bigger models? What is the compute cost? Then you look at the potential raises and think about which country would not be able to do that race first. But also, it’s this wild thing with OpenAI raising $120 billion. Are you kidding me? What is that? Grace Shao (35:47) Yeah, the valuations in the U.S. are not really understandable by anyone else right now. I think in China — so on your point on that, I’ve been writing about this and I think it would make sense for Tencent just to buy out one of the labs. They have the money, they need the capabilities, and frankly, they’ve really been struggling to compete with their LLMs, with all the labs talked about just now. So my... Nathan Lambert (36:05) Their licenses are so bad. They release all these models that have horrible licenses. They’re not that good, and the licenses are just horrible. Grace Shao (36:13) So I feel like it financially makes sense for a company like that to optimize and just buy out a lab. Then the labs can also lean on their distribution, because at the end of the day, how are they going to win consumer mindshare or distribution in China right now when it’s really just dominated by Alibaba, ByteDance, and Tencent? That’s my spiel. But when I spoke to some of the researchers... Nathan Lambert (36:33) I think big companies have a lot of inertia, and the senior leadership has the call, and they can have inertia. I still think Apple just ends up buying some lab for $25 to $50 billion. It’s not the worst thing. Just golden-handcuff the researchers. Some will still quit. Grace Shao (36:43) Yeah. But I think right now they don’t want to. The labs still have a dream. Some of the researchers still have a dream. So when I spoke to a lot of them, they’re like, no, we don’t want to do that. We want to commit to our own frontier research. If I wanted to join one of the big tech companies, I could have. So why would I want to sell? That’s what the researchers think. But to your point, we don’t know what actually the one person or two people at the very top think, especially if they continue to have hurdles with compute access and capital access, which brings me to the question. Nathan Lambert (37:14) It also depends on your view of inference. You can ask your next question. I don’t need to cut you off. It depends on your view of inference. If these agents are just so much inference, I do think it’s going to be an oligopoly-style market, not a monopoly-style market. And what’s the difference financially between two and four or five big companies with great models? Is that actually not sustainable if there’s so much demand? There are a lot of cases where we have two or three, like the cloud, but what’s stopping that from being four? Grace Shao (37:40) I think they will be the infrastructure providers. Yeah, yeah. And they would kind of lean into each of their existing ecosystems or distribution, whatever you want to call it, and serve certain specific models for specific uses. So enterprises can choose what matches their needs the best as well. I do want to bring the conversation to a more contentious topic, which is on distillation and model convergence. You raise the question of whether Chinese models are structurally different. Often we are hearing claims saying a lot of these labs are about three to six months or six to nine months behind U.S. labs. There’s obviously a lot of noise or allegations and accusations from certain U.S. labs saying Chinese labs are distilling them. How do you actually see that accusation or that kind of dynamic? Nathan Lambert (38:36) The biggest unknown that I don’t have an answer to, which actually has a lot of sway, is how much of the Chinese companies are actively trying to hack APIs versus just showing up as a customer and paying. If you’re trying to hack the APIs, normally you get reasoning traces out so that you can create a reasoning foundation that would be similar to the model that you’re trying to do this from. That’s very different than the API standard form, which is just the output of the model, which is a less direct process for learning from. I don’t know the magnitudes. If it’s more just like, I walk up to an Anthropic API and I use it as intended, but I’m making a competitive model, I’m not very sympathetic to Anthropic. They could ban it if they want to. And I think the impacts are kind of a standard practice. You can do it with many different models and so on. The evidence Anthropic provided is not large enough scale where I’m like, this is industry IP theft at mass scale going on 24/7/365. So there’s definitely some gray area to what is actually happening in distillation. That’s why, on the policy side, I try to push people to not call all of it the same thing. Essentially, using any API endpoint to make synthetic data to train your model is some form of distillation, but it’s very different if you’re trying to break this model so that it gives us a different behavior that is hyper-useful for training and not get caught. Those are pretty different actions, and they’re all looped into this common phrase of “distillation” right now. That’s my biggest problem, which is that academic researchers and small companies use distillation extensively as the core of their business and the core of research methods. So if the U.S. government nukes that as a thing that could be done in the AI ecosystem, it’s mostly bad for small players, bad for U.S.-China tensions, and bad for academics. That’s my primary concern. And then trying to get the labs to actually say more. There’s a distillation side and then performance is the other side, which on benchmarks, it does seem like the Chinese labs tend to be six to nine months behind. When it comes to general use, I’ve always found the closed models to be better in ways that are hard to measure. So I go very back and forth on whether the closed models are better. I think we will especially see Anthropic and OpenAI pull ahead on knowledge-work tasks like legal, healthcare, financial services, because I just don’t see the Chinese labs paying for that data. All that data is going to be people that charge hundreds of dollars an hour to annotate and create these environments. So it’s a whole new capital build-out that goes on there right now. It’s going to be billions of dollars if you’re going to buy a billion dollars of data and a billion dollars of compute and a billion dollars of talent to train your model. Grace Shao (41:30) They don’t have the money. Nathan Lambert (41:30) I don’t think they have that. Mercor has some of these evals, and I think there is a bigger gap there. So it’s very interesting. Florian, the guy that helps me, and I disagree on it. It’s this fine line between, yes, the evals — coding and lots of these things, and even random evals that surely the Chinese labs aren’t training on — the open models really are genuinely crazy impressive scores. So I think there’s also a tester’s bias, where I don’t use the open models as much. Maybe it’s hard to ground in my head what I was doing with AI six to nine months ago. I wasn’t even using Claude Code as extensively. I guess the question is, at the end of this year, can I use an open model in something like Claude Code and feel like it works at all? That’s the test on the performance gap, starting in June, June to August, and whether or not that hits. I don’t think the open models have hit that yet. I think it would be way more of a narrative if all the companies spending billions of dollars on Claude are like, oh, we can spend 1% and just use DeepSeek. These CIOs and all the big companies — some companies spend more on tokens for their employees than on headcount. These are normally startups. But they would happily reduce that token cost to 1% expenditure if it really was that similar, because then you could just use 10x the tokens. I don’t expect that to happen. And I expect things like the latest Claude and GPT-5.5. I expect more of these things through the year, and we’ll see if I end up being right. Both are right at the middle of us, as a world, getting more clarity on them. They’re like 18-month-long stories unfolding, and I feel like we’re just in the middle of performance gap and distillation and learning more. Grace Shao (43:25) Yeah, it’s interesting. You mentioned — it helped me recall a conversation I had with other people as well. The point on distillation is that I just had a conversation with your former colleague at Hugging Face, who leads APAC, called Tiejin Wang. He was just saying, look, the distillation accusations don’t really make sense because we’re all distilling off of each other as we speak. I’m learning from you; you learn from me. We’re distilling. It’s so vague of a terminology to just use that to accuse all these various behaviors. So to your point, I think people in the technical world who understand what’s happening actually want more clarity on what is the gray area, what is actually black and white, and what is not appropriate or unethical. That needs, I think, the industry to come together to really put guardrails and rules around. Now, number two on the compute side and the data side. Something anecdotally will be interesting to you is that when I spoke to one of the lab researchers in Beijing, I think in February around Chinese New Year, they were saying, look, they want to get better data, but they can’t because usually a lot of American labs would pay tens of millions, if not even more, like a hundred million dollars, for a set of very obscure or niche datasets, but they would have an exclusivity contract. What the Chinese labs will do is that they will literally wait out the exclusivity contract and then, say two or three months later, pay for it at one-tenth or one-twentieth of the price for that same dataset. So then once they start post-training on that dataset, that’s where the three to six months or six to nine months come in as well. Yeah. On that note, I want to... Nathan Lambert (45:00) Yeah. I think the data industry in the U.S. has two things. One, the lab asks the data vendor, we need this specific type of data. And the data vendor is a network that connects the people to the lab. The other thing is the data vendors know evals that are important, so they try to create good data for hill-climbing on specific evals. That data could be sold to multiple people, but is less expensive because they make it once and expect to eat margin or take margin on it. There could be a pipeline where once OpenAI is at the cutting edge, creates this new thing, they create deep research, then the data industry is like, let’s make things that are a little bit cheaper to sell. So there is time lag in these various things. But I heard the same thing on the ground, where they have a negative view of the data industry. It’s like, quality is bad, we don’t really have access, we do some in-house. That’s a very big difference from today, which is that you have the data companies in the U.S., which is insane. Grace Shao (45:53) Yeah, the American data companies are so mature. It’s its own sophisticated ecosystem. Before we get into data, I actually want to ask you this question. I think recently a lot of the narrative is now saying, look, Anthropic and OpenAI have kind of proven that pre-training scaling laws continue to hold, especially with the recent models. There’s an obvious compute constraint on the China side that we talked about. And then it will likely be even more amplified with the absence of Blackwells in the coming months. So as we move forward in this race, per se, if you have to put it in China versus U.S. in that sense, will we see a wider gap between the performance and benchmarks between the Chinese labs and U.S. labs? As in, will we see the gap going to 12 months, 24 months, as Chinese labs are very, very constrained on compute for pre-training breakthroughs? Nathan Lambert (46:40) I think it’s more of pre-training as a thing that you could actually finish. How big can you pre-train a model that you can finish and serve? The Chinese labs could train models that look like GPT-4.5, which is this giant model, but you can’t serve it. They end up training a model that is 2.5 trillion parameters and they release it, and no one can use it. They could barely serve it on their API because they don’t have Blackwell NVL72 racks or something — these racks that are definitely what are serving these large MoE models. They just don’t have the quantity of these. So there’s a difference between models that you can build and models that are actually useful. I think some of the Chinese labs are definitely like, we don’t need to release the gigantic models because nobody is going to use them in open weight. The biggest models end up getting served via API. So there might be some segmentation in that market. But I do think the inference and amount of economic resources that you have to serve your customers is becoming a thing that dictates what models are built. That’s why I think the gap will continue to rise. All signs point to GPT-5.5 being a bigger model, and I don’t expect that to stop. And then the economics of it is just the basics of: you need a certain volume to have the margin to support the research, because you can’t keep raising these ridiculous rounds forever. I think OpenAI, Anthropic, and Google are the only people with that AI usage volume to keep marching down the scaling laws to another 10x of training compute, which is mind-boggling amounts of investment in a model. That’s why, when the economic markets slow for fundraising, the model gap between these big three will just show a lot more. That’s the distilled way to say my prediction of when things will look different. It’s like these labs can’t fundraise, they go public, they can’t generate revenue more on their paid services, and then it’s just: look at how much training compute can be allocated or can’t be allocated. Grace Shao (48:41) Yeah. Basically, we’ll see a bigger gap, I think, in the coming months. Then what can make up for that? Domestic chips, or, like you said, better data. And why is it that sometimes people assume China has a very strong data ecosystem or data products, but actually the data vendor ecosystem is very weak in China? Nathan Lambert (48:41) So generally, I think I agree with what you said. I don’t know on the data side, but the way domestic chips could help is that if Huawei chips are fine for inference, and if they have sufficient volume to support the inference economics, which then trickles back into revenue, my read is that they just don’t have the volume of the chips, especially spread out across the amount of companies that they have. Essentially, the total FLOPs of Huawei, all the things produced, and it’s going to all these different places — it’s just not big enough. It could be something like ByteDance and Alibaba, with offshore data centers, can keep up a lot longer because they have access to Nvidia compute and have for a long time through this kind of offshoring. Maybe that stabilizes the ecosystem, and we’ll see what the AI startup, the younger startups like Kimi and Z.ai, end up doing. No one wants to do this, but if they pool resources, they last an extra year. You get another order of magnitude if they all pool together, but I don’t see them doing it. Grace Shao (50:00) But that’s the thing we were just talking about, right? MiniMax and Zhipu, how can they possibly compete with the hyperscalers at this point if you need offshore data centers? And the fact that Zhipu is on the Entity List doesn’t help, right? It’s not going to be easy for them to access these data centers either. Nathan Lambert (50:12) Yeah, I think they can’t. I think they won’t. Human nature will make it so they won’t collaborate. They’ll just do something smaller. They’ll just have successful businesses that are different. Grace Shao (50:22) They just have smaller ambitions, want a smaller piece of the pie. Yeah. Okay, so you wrote something like, nothing’s a secret, but everyone wants Nvidia chips. They want it, they don’t know how to get it, they’re fighting over it. Nathan Lambert (50:34) Yeah. They’re the only thing that works for training. All the models are trained on Nvidia. I don’t believe the DeepSeek propaganda that it’s trained on Huawei. The only models that are trained on Huawei are tiny. Inference on Huawei works. Every lab is like, inference on Huawei works. The labs that don’t have meaningful inference are like, we are told to get Huawei, so we buy them, but we don’t use them. Earlier research labs are like, we don’t have any inference and we don’t have a need for Huawei. Any company that has meaningful use of their models has figured out how to run them on Huawei for inference, which, to Jensen’s credit, is like — it’s happening when he said it was going to happen, but it’s not that surprising. Grace Shao (51:11) Yeah, the Dwarkesh interview. I don’t actually understand why he got so much hate for it because even without your political stance, what he said actually made sense logically by saying, if you don’t sell them the crappier versions of what we have, they will have an equally quite crappy version to serve themselves, or they would just want... Nathan Lambert (51:28) I think they would buy both. Buying both is actually true. The amount of Nvidia chips that you would have to sell to China for them to stop buying Huawei — because Huawei is almost surely way cheaper because Nvidia margins are insane — when would they actually stop buying both? Grace Shao (51:43) But then you have to go on CANN. You have to reroute everything back on CANN. The developer ecosystem is not there. That’s Jensen’s point, right? Or the habits are not there. So I think that’s what, when I talked to a research lab... Nathan Lambert (51:50) Yeah. But I’m saying they would also use Huawei. I think they are so supply-limited, they would use both. Anthropic uses everything. A lot of companies in the U.S. will use multi-platform. Meta is a huge buyer of AMD. Demand is so high that any chip that is potentially viable on the models within a few generations is very valuable. And the fact that you can run some reasonably large model on any Huawei chip is a big line crossed for Huawei. I don’t know if they can produce the volume of chips and scale that quickly, especially as they try to move to lower nodes. That’s the standard semi debate. But the question is: can Huawei scale production? That’s the only question. And if Huawei can manage to scale production, Jensen will just look really right. If Huawei can’t scale production, Jensen will look a little bit like a lunatic, but it will be outside of his hands. Grace Shao (52:43) And we don’t really know what happened during this trip. It seemed like nothing really substantial happened after this big Trump delegation. It was more like a high-profile tourism trip versus an actual deal trip. Okay, I want to ask you something you wrote about that’s a bit niche, not something you usually write about. It’s on the SaaS side of things. You said that there’s a common argument that China struggled to monetize AI because they’re unwilling to pay for enterprise software. We looked at how China tries to monetize on consumer AI, but clearly that’s not really been proven yet. In your piece, you push back on the claim and say that there’s a distinction between SaaS spend and cloud or inference spend. Tell us about what you think about that ecosystem and how Chinese AI labs are trying to make money maybe a bit differently from American AI labs. Nathan Lambert (53:32) I don’t know if it’s necessarily different, but I ask a lot of researchers about this. They say that everybody is trying the new AI tools when they come out. If they don’t like them, they stop using them. If they like them, they keep using them on the consumer side. So something like Claude Code would be an example: tons of people tried it. I’m guessing lots of them churn in China, just like in the U.S., but consumers are very quick to adopt and try new things, but won’t stick if it’s not actually serving them. And then the enterprise is like: there’s definitely cloud that exists. Digital services are gigantic. They essentially think that there’s more runway for making money on AI models that falls into that. And they all use coding agents; they all use Claude. It’s a hilarious thing. They’re all very Claude-pilled. There’s almost no mention of Codex, where in the Western media, Claude versus Codex is this whole thing. They all use Claude. And that is obviously a paid service. So I think there are cracks in the argument, and I expect AI models to be seen as a bit of cloud, but potentially it is the thing that changes some of the expectations, where it’s just so transformative because they’re so competitive, and it could be seen as a bit of a phase shift. Grace Shao (54:41) Yeah, and I think it’s a generational shift, a phase shift. Also, actually, recently Doubao raised their prices on Seedance usage and whatnot, and it’s a shift into trying to capture the prosumer market. You can say the average uncle and auntie on the streets still don’t want to pay for a consumer app, but I think there’s more prosumer market share that could be captured in China, maybe not fully enterprise either. I want to ask you about government roles and geopolitics. I know there is a common narrative that usually people assume Chinese AI labs are heavily subsidized. Actually, when I was in San Fran in March, I was at a dinner with a couple of investors, mostly public investors, and one guy asked me, “Hey, are all labs just basically subsidized by the government?” I was like, definitely not. The majority of them are not. If not, they frankly don’t want to take money from the government. It was really hard for him to understand that, because I think the misconception is all Chinese labs or Chinese tech are just funded by the government. Kind of to our point earlier, where any affiliation to any government agency, just by default, is assumed to be therefore backed. First of all, the government, I don’t even know if they have that much money to give out. Number two, I don’t think that’s how competition works, right? So what’s your thought on all of this? Nathan Lambert (55:55) It seemed more like a provincial government trying to help the companies do stuff, which is like get offices, get talent. I don’t know what the provincial government can do. In Beijing, there’s Beijing Academy for AI or whatever, which is a real research institute that’s just funded by a certain neighborhood in Beijing. It was like, okay, the U.S. could do that. But much less of the Ant Group-style thing, which is government takes major ownership stake in an investment round and goes on. Maybe Kimi’s latest round, there were mentions of government-backed VCs, and I don’t know how that kind of intermediary works. So I still think it’s very indirect. And because the government system is so competitive across the different layers, each of those layers are competing to help the companies, but they don’t have piles of cash sitting around to buy GPUs. Grace Shao (56:44) No, they don’t. And they frankly don’t know what they’re doing half the ti

19. Mai 2026 - 1 h 3 min

Matt Sheehan on China’s AI Policies: Employment, Anxiety, Safety, and State Priorities

Today, I’m joined by Matt Sheehan [https://substack.com/profile/222-matt-sheehan] who writes this insightful newsletter. [https://mattsheehan.substack.com/?utm_campaign=profile_chips] Matt is a senior fellow in the Asia Program at the Carnegie Endowment for International Peace. He researches China’s AI ecosystem, Chinese tech policy, and how technology shapes the country’s political economy. Matt lived and worked in China from 2010 to 2016 and later led China tech research at the Paulson Institute’s MacroPolo. He’s the author of [https://carnegieendowment.org/people/matt-sheehan]The Transpacific Experiment [https://carnegieendowment.org/people/matt-sheehan]. He speaks Mandarin, and he turns complex policy into plain English. [https://carnegieendowment.org/people/matt-sheehan] In this episode, he helps us understand China’s AI governance, about how Beijing is thinking through the social and political consequences of rapid AI adoption. We focus especially on a shift that became more visible in early 2025: rising concern inside China’s policy community about AI’s impact on jobs, worker anxiety, and social stability. Matt explains why China’s AI labor question is different from the Western debate. We also discuss how the Chinese government is trying to balance support for technological progress with the need to manage public anxiety, clarify labor rules, and avoid social instability as AI becomes more deeply embedded in the economy. He broke down the myths, explained the jargon, and the regulatory bodies in China. Our conversation started slow, but it became very, very heavy, what they call 干货满满 substance heavy. Also, a shoutout to Nathan Lambert [https://substack.com/profile/10472909-nathan-lambert]’s work in helping us better understand the open-source ecosystem and Rui Ma [https://substack.com/profile/25978-rui-ma]’s for helping us understand investing in China AI! Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently. Season two will host a series of guests from early-stage investing, as well as builders, founders, and product managers. For more information on the podcast series, see here. [https://aiproem.substack.com/p/launch-of-differentiated-understanding] To find the previous episodes of Differentiated Understanding, see here. [https://aiproem.substack.com/podcast] Chapters 00:00 Introduction to AI Policy in China 03:10 Matt Sheehan’s Journey into Chinese Tech Policy 05:55 Shifting Perspectives on AI and Labor 09:02 Public Concerns Over Job Security and Government Responses 15:09 Education and AI: Preparing for the Future 17:50 Regulatory Landscape of AI in China 34:00 Navigating China’s AI Regulatory Landscape 40:58 Misconceptions About Chinese AI and Government Funding 43:57 Understanding AI Safety and Security in China 52:03 Global AI Governance: Cooperation or Parallel Paths? AI-generated transcript Grace Shao (00:01) Hi Matt, thank you so much for joining us today. I’m so, so happy to finally have you on the pod for people who are listening. We’ve been trying to make this happen for like six months, but between us, there are like three little children running around with a bunch of viruses and have just not been able to make this happen. I’m really excited. ⁓ A few months ago, what really caught my attention about your work again is that you shared something on WeChat saying you were dissecting the new Chinese AI safety paper, like the big national one. ⁓like verbatim in Chinese. And I was like, wow, this is extremely impressive. It’s not an easy task. I commend you for doing that. So I really wanted you to help us understand the nuances of the AI policy world, especially how people are perceiving AI in China. I think there’s more more interest in how China’s governing AI ⁓ while we were hearing the backdrop of how the Chinese government is trying to push on AI diffusion, right? And then on top of all of this, like where areas where China’s AI governance seem to be leading, because in many ways it seems like China’s AI regulators are much faster to respond to how fast technology is evolving. But to start, we would love to hear about your personal story. Tell us about how you ended up studying China, studying Chinese tech policy. We met in Beijing years ago, maybe a decade ago. ⁓ Yeah, so tell us about that. Matt Sheehan (01:11) Sure. Yeah. Yeah, sure. Sort of stumbled into China stuff. I hadn’t taken Chinese or really knew anything about China until about halfway through college when I ended up getting a summer job in Beijing. I was just kind of like instantly fascinated and knew I wanted to move back there after I graduated. So took a little bit of Chinese my senior year, moved to Xi’an, taught English, kind of followed what at the time was a very like typical know, trajectory of like, go there, teach English and then go study Chinese at university and then get a job and get a slightly better job. And eventually I was able to kind of wiggle my way into journalism. And so I was a China correspondent for a publication called The World Post at the time. And that took me up. was there from 2010 to 2016. So kind of like the hinge period before and after she came to power. Pretty interesting thing to see. And When I moved back to California in 2016, I started working on a book about China-California ties. I’m from California and this was like the period of kind of explosion in cross-border investment and Chinese students come into California in the Silicon Valley-China relationship getting even more like twisted and complicated. China-Hollywood. So I wrote a book about that and as I was doing it, the kind of the tech section, the China-Silicon Valley, China-U.S. tech connections kept growing bigger and bigger and I ended up working a little bit with Kai-Fu Lee on his book, AI Superpowers, which was kind of my turn from like, it was like all things China, China, California, China, Silicon Valley, China AI. And since 2017, I’ve been working almost exclusively on AI issues in China. Maybe the first three years of that, like 2017 to 2020, was very focused on comparative capabilities. This was kind of a period right after the National AI Plan in China when there’s a big explosion in activity. And I think... This is kind of was like the first time America kind of got freaked out about Chinese AI capabilities. And so I spent a few years being like, okay, let’s try to like ground these assessments in some data. Let’s get like an actual grounded sense of where the countries are with each other. ⁓ And then starting in 2021, I sort of turned into focusing on Chinese AI governance, Chinese AI regulations. That’s when they first started rolling out their regulations or recommendation algorithms and sort of deep fakes. And I was kind of making a bet that I think If China continues to be at or near the frontier of AI, then how they choose to regulate it domestically is going to have huge implications for China’s own ecosystem. And then it’s going to really ripple out internationally on safety, security, growth, all this stuff. So I spent the last, now it’s like five years, ⁓ just deep in the weeds of Chinese AI policy and regulation. Grace Shao (04:04) Oh, great. I think I definitely want to double click on all the algorithm security and the kind of, you coined the answer versus what’s called security and versus what’s the other Chinese word? Yes, yes. Matt Sheehan (04:17) Anshun safety security. ⁓ Jeff Ding was talking about this very early on, but yeah, it’s a, it’s a constant thing that we have to negotiate for people who don’t know it’s the word Chinese, the Chinese word Anshun ⁓ means both safety and security. whenever you’re kind of translating documents on this front, you have to know, you talking about AI safety, which is kind of a different thing versus AI security, right. I think both you and Jeff definitely are some of the more nuanced scholars I follow. And I do want to kind of double click on that later on. But to start, I think I want to talk about something that’s top of mind for a lot of people. You just wrote a piece that you said was not super serious. It was just your scattered thinking put together on subset. I thought it was very well written about the growing anxiety around potential job losses. ⁓ your perspective is that you know there are more and more people voicing this kind of concern and I wanted to hear a perspective on that and I kind of wanted to share a bit of my different share my different perspective on this and what I’m hearing on the ground and kind of have a conversation around that as well. Yeah why don’t you start with sharing like what you found yeah Matt Sheehan (05:22) Yeah, sounds great. Yeah. So, I’m not an AI and labor person. That’s not been my focus for a long time, but I’ve been monitoring it for a long time and just lightly. And starting around, I guess it was early 2025, I just started to hear a lot more out of the Chinese policy community about worries about AI’s impact on labor and jobs. And this was kind of a surprise to me because ⁓ just a little bit prior to this, say early 2024, I had, I sometimes ⁓ in my job, I run these kind of like informal surveys or almost like a, what do call it, focus group of American and Chinese AI policy people and asking them like, you how would you rank these different risks? How concerned are you about ⁓ job risks versus privacy versus military AI? And we have both sides that like rank the risks and then talk about the results. And when I ran one of these in early 2024, it was very striking that the Chinese side I think it was at the time we had seven different risks and the Chinese side ranked labor impacts a second to last as six out of seven. And so my sort of baseline was like, OK, for a variety of reasons, this isn’t really too on the radar of China’s ⁓ policy community or wider policy community. And then starting around early 2025, some of those same people who I had been talking to about this before had really changed their thinking. They were saying that there was a big change in thinking within China, maybe especially within China’s of policy and government circles, but then I think also a little bit wider. And so that sort of sparked my curiosity. And for the past, now it’s like over a year, I’ve been just sort of tracking when does this AI and labor question show up in state media? When does it show up in kind of the online discourse? When does it show up in policy documents? And sort of the TLDR is like, I think this is really, really ramped up a lot. Over the past 18 months. It’s been maybe the single biggest change in how China perceives different sort of risks as it relates to AI. And I think I’m looking forward to sort of discussing how maybe like the policy world or the government’s perception of this differs from ordinary people or certain categories of people. ⁓ But I think it, from my perspective, it’s sort of it’s infused into both. think there’s been a fair amount of public concern. the policy community picks up on that and they want to both respond to like the actual problem, know, actual job losses, but they also really want to respond to people worrying about job losses. That’s kind of maybe the thing that actually made me write this piece now was discovering an interesting piece in state media. I think it was in science and technology daily. That was the headline was like ⁓ AI must be controllable, but people’s ⁓ anxiety about AI must also be controlled. And it was all about how sort of OpenClaw has triggered a lot of anxiety and a lot of people about, are they going to get replaced? You need to be building your own AI agent in order to not be left behind. And they’re sort of trying to tamp down those concerns in a few ways. So that’s what sparked the piece itself. Grace Shao (08:33) Yeah, I think definitely what you saw and you wrote about is like definitely kind of playing out in the China AI policy ecosystem that I see as well. And I think for sure, the open-claw frenzy have kind of opened the eyes to lot of the even average people what AI could potentially do. However, I guess my argument, not against it, but it’s just like, you know, we kind of cite each other’s work on subset. But my point was kind of saying, you know, this is a reflection of a relatively elite group of people end of the day, because the knowledge work economy in China end of the day is only like only 30 % of the workforce actually are the knowledge working economy. And end of the day, even though it’s 30%, because China has such a huge population, the mass, the sheer scale, it feels really large. However, I want to bring it back to the idea that like anyone who’s lived in China understands that the government’s like top top priority really is about social stability which leads to what they call social harmony, right? And I just think that, you know, the rising anxiety of job control a lot of times maybe is because there’s a fear of if there’s a lot of disruption to jobs then people will lead to social unrest which obviously gets a bit more sensitive but you know, a lot of what they do comes from that I guess thinking so I agree with you top down definitely have to understand what’s happening with technology and how advanced AI has become in 18 months. ⁓ have given them kind of, I guess even fear mongered a little bit internally, right? ⁓ But the nuance here is that I think ⁓ the rest of the 70 % of the Chinese workforce actually don’t work in anything structured that we know. I think even probably even 80%, you know, people in China, they most of them are actually like, you know, service providers ⁓ from, you know, rural areas and urban areas. A lot of people work in factories, even the entrepreneurs, right? They run like say hospitals, clinics, factories, bottle cleaning, like factories, whatever, right? ⁓ Car logistic rental businesses, these people aren’t actually ⁓ trained in the way that maybe the West by default think they are. They actually just run it from a grassroots way and they don’t have very, very streamlined processes. They don’t have documentation. They don’t run like what we think a corporate has run. So in that sense, I think it’s very hard for AI to replace any of their workflow. because it’s actually not a, we can’t really provide context and a lot of things, business is done is through one C, is through a wink, through a look, through a gesture, through, you know. So a lot of that, I think, in fact, will be harder to replace than even maybe some of the more mature businesses in the West where there are structured processes and everything. So that’s kind of my, I guess, a more nuance, I think, push on that. Yeah, wonder what you think of it. Matt Sheehan (11:20) Yeah. Lots of thoughts. And I think the sort of the fundamental distinction that you’re pointing at is very valid in that like, you know, companies in the West, in the United States, they’ve been like big companies have been running sort of digital databases for decades. They have like decades worth of data. They have pretty advanced like enterprise software. It’s just a much more I want to say something like a bias, but it’s a little more like put together sort of official structured ⁓ technological backend and not just technological, but like a process backend. Whereas in China, it’s just things have developed really quickly. A lot of it’s on the fly. lot of it is, know, enterprise software is just not, there’s not really a market for that in China in the same way. It’s a lot of stuff is pirated or they’re just not digitized in the same way. And it’s actually very interesting. This is in kind of the early days of the like China, US who’s ahead. ⁓ know, a lot of the debate focused on data. Matt Sheehan (12:18) And there was this idea of China has a billion people, so it must be this huge advantage in data. But my pushback on that was always kind of what you’re arguing, is like the US actually has very structured data, and it’s owned by corporations, it’s deployed by them, they’re already doing type of sort of lower end market intelligence type stuff. ⁓ So I think that that sort of backdrop is very real. think ⁓ maybe from there I’d like differentiate out to potential risks or debates. And it’s kind of what you were pointing out as well. There’s like the actual question of how many jobs are going to be impacted. How many people are, what is it going to do to people’s wages? What’s it going to do to aggregate employment? And then there’s the question of like, ⁓ how do people think about that? What are the fears due to sort of the Chinese social stability, even if the things haven’t manifested. So I think separating those out, definitely the government is... ⁓ their sort of initial response is related to public worry about this. So in the piece, I detailed the way that ⁓ sort of a robo taxi incident in Wuhan was in many ways the spark that really like ramped up the government thinking on this. And this is something that I heard from a couple of different Chinese policy people who both pointed to this incident, which I had totally missed at the time and wasn’t like major international news, but that had a big impact. And basically what it was is that Baidu was rolling out its sort of fully autonomous robot taxis throughout the city of Wuhan. There was this kind of like public letter, ⁓ open letter released by a taxi company that was kind of railing against, you know, both ride hailing platforms and autonomous vehicles as, you know, stealing the iron rice bowl or just smashing the rice bowl of taxi drivers and of companies. And even though it was a kind of like a small thing, a couple of days later, a Baidu taxi actually hit pedestrian, I don’t think they were seriously injured, but it kind of fed into this overall, a big kind of online reaction and discussion about like, what’s going on with AI? Is it going to take people’s jobs? And it, it’s one of those things, it’s funny to explain to people because it sounds like nothing, but it did lead to a pretty significant ⁓ imprint on the way that the Chinese government is thinking about it. So it was coming from, in many ways, public discussion of it. Like the discussion was happening online. This discussion might be happening among, you know, elites. chronically online people. ⁓ But it’s something that the government definitely picks up on. So I that’s one element. They’re worried about the worries and they want to, like among their sort of policy reactions in a ways, thus recently has been sort of directing platforms to say like, you kind of need to tamp down these articles or these viral videos that are telling everybody, like if you don’t adopt open claw, you’re going to be left behind. There’s been this kind of rash, both in China and here in the US of like, you know, kind of like fear mongering people into clicking and taking your course on building agents or just subscribing, whatever. And one thing the government is doing is telling you like, chill on that. Like, don’t be putting that narrative out there. ⁓ And so that’s part of this kind of like public opinion management thing. In terms of the actual impact on jobs and who will it hit? It’s a huge open question that I personally have gone back and forth on for years. I first kind of did a deep dive on this way back in 2017 when everything was still so speculative. At the time, was pretty not worried in part because of the reason you described it. I’m like, there’s just a lot of friction. There’s just so much friction in this economy. And just because an AI system can theoretically execute a test doesn’t mean it’s taking a person’s job. And for me personally, my thinking on this has changed a lot in the last year to 18 months, mostly because of how capable agents have proven to be. I expected agents to essentially be hitting a lot more roadblocks while they’re being deployed online. They haven’t been. They’ve been operating much smoother, or just they’re more relentless, and they can break through these bottlenecks. ⁓ It’s interesting that in China, the inciting incident was not about white-collar workers. It was about taxi drivers. ⁓ I don’t know this as a fact, but if I had to guess, I would guess that a much larger portion of the Chinese population’s job is driving a car or driving a scooter or something like that. And that maybe that’s a vulnerability that they might face depending on how self-driving vehicles roll out or delivery robots and stuff like that. The knowledge workers, yeah, it’s such a messy and unclear thing, but I think the government has at least started to take it seriously because it’s not, their policy responses are not just this like, public opinion management stuff. They’re also talking about, ⁓ like one of the more interesting pieces that I highlighted is, and maybe the most concrete thing they’ve done so far is, ⁓ according to Chinese labor law, there’s sort of reasons why you can and cannot fire a person. There’s like legitimate and illegal reasons to fire someone. And when someone is fired and they object, this gets taken to like a labor law mediation ⁓ body that’s under the Ministry of Human Resources and... forget what the second part social ⁓ security. ⁓ Yeah, Ren Li, Ziyuan, Shouhui Bao Zhang. Yeah, that’s what it is. ⁓ And in the last year, one of the things that really made a lot of made kind of a big splash is that those mediation bodies declared and it was echoed in like the biggest state media that saying that you replace someone with AI that AI can now do this person’s job is not a legitimate reason to fire someone and those people have to be reinstituted into their jobs. That’s a concrete policy thing that’s actually directed quite clearly at like actual impacts. ⁓ Is it going to work? I really don’t know. It might just be a little bit of friction and, you know, maybe China’s kind of doing what it always does, which is like, we’ll figure this out. We’ll kind of muddle through this. We’ll put some friction here. We’ll grow a little more here. But I think the concerns are real, whether they bear out, ⁓ whether they hit faster in the United States or China, which country is better positioned to sort of roll out a more redistributive welfare system. think these are all open questions, but I think the concerns at least are real. Grace Shao (18:24) Yeah, and I think you hit something that I feel like it’s being kind of missed in the headlines, which is the government actually cares more about the general mass, which are the people who are driving the scooters and the like the DDS cars more than the knowledge workers, which is kind of different from the Western kind of conversation right now, where a lot of the whether it’s fact, frankly, the power that can lobby and the power that the voice that have the voice are all really concentrated in. the white collar elite jobs that are very much concentrated in Silicon Valley and whatnot, right? And I think it really reminds me of the time when, you know, during the Hulianmang Shidae, like the internet era, you know, like the big tech only really got clamped down when the average consumers felt like they were really being pushed to R-Shrine Egypt 2-1. So that’s when they had to add the monopoly to probes. And then soon after, only maybe two years after the probes happened, there was a common prosperity rollout, which basically all the big tech in some capacity had to like showcase that they had a CSR aspect to them. I think this is something that we don’t really see in the US as much with all the big techs, because it’s kind of like they’re doing what they need to do. They have their profit driven interests. And then of course, everyone has a CSR, but it’s not really allegiance to the government CSR mission. It’s more like, we believe in ESG. We believe in climate. Amazon is going to have some like carbon footprint reduction plan, right? Whereas like the common prosperity thing rolled out. ⁓ you know, it kind of died on its own, like no one really talks about it anymore. However, during that phase, when it did get rolled out, it was like an understanding where, OK, if the government wants the, frankly, the poor or the middle lower class to feel protected, then you as a very large ⁓ moneymaker in the economy need to showcase that you are somehow ⁓ part of this kind of support. So I wonder how this will play out for the big tech in China when, the job protection policies really get rolled out in practice, like what you’ve mentioned. And obviously they can’t really say, you’re being replaced by AI. At least there’s that superficial guardrail there, I think. Matt Sheehan (20:24) Yeah, I think the sort of the political economy of these questions is going to be super interesting in both countries. know, essentially like business in many ways, it inverts the sort of technological impacts on employment that have been around for so long. Normally, like greater technologies integrated into the workforce, it hits sort of people working maybe low end manufacturing jobs or jobs that would be considered sort of repetitive and, you know, quote unquote, low skilled jobs, even if they’re not. ⁓ And, you know, in the United States, we’ve seen like three, four decades of this. And the people who are concerned about that basically didn’t get hurt because they are not, like you say, part of these influential classes. AI is going to, yeah, in the United States is going to be very different. This is going to be the first time that you have the way that my sort of mental model for it is like, if you’re a senator and you have kids or nieces and nephews or your friends, kids like what, what are their problems? And like how close do those feel to you? And you know, if you’re a 60 year old senator and you’ve got like a 23 year old niece and she just graduated from college and she got a degree in, you know, something that’s like is legitimately employable normally like in marketing or something like that, and those jobs just aren’t there, that’s just going to feel very close to home for people in power in the U.S. in the ways that it hasn’t felt in past waves of technology impacts. In China, I... partially agree with what you’re saying, but I think there is also going to be a significant element of this sort of the same dynamic as the United States. mean, yes, common, you know, Xi, common prosperity. He’s focused a lot on sort of eliminating extreme poverty and, you know, the CCP, it’s in its bones that like the rural, the working class are in many ways kind of the long term support base of them. But I mean, also if you look back at Chinese history, like a lot of the biggest and for the government most dangerous protest movements came out of elite schools, came out of students at elite schools who ⁓ either couldn’t find jobs or were facing inflation issues or had, for ⁓ ideological reasons. I think that stuff does hit close to home. think some of those same dynamics, if you’re a deputy director at the National Development and Reform Commission. your family, the people that are close to you, are going to be the type of knowledge workers that are going to be impacted by this. And I think that just can’t help but kind of like compress in on the thinking on this. ⁓ You know, how AIs can... Yeah, yeah, and like how... Grace Shao (22:58) Yeah. everything becomes personal in the end. Like in the end, it’s like politics is still personal. Yeah. Sorry. Matt Sheehan (23:08) Politics is personal and yeah, mean, like cities are where social instability is the most dangerous. Like cities are where people gather and you can have potentially dangerous incidents. These are the people who are very online and are sort of sparking or leading the conversation as much as that can be controlled and manipulated via censorship regimes or public opinion guidance. Like these people are gonna be vocal. yeah, I think if I was at the CSPI, I’d be concerned about Grace Shao (23:39) I want to kind of go into on education. Like you kind of touched on it, right? Like, you know, there’s been draft rules about children’s interaction with AI in China as well. There seems to be more guidance and obviously concerns around that ⁓ from at least from the top down ⁓ about their mental state, their dependency, or even what constitutes as an AI companion, how we should draw the line on that. We know like famously a couple years ago, China installed this rule where like, you know, kids under 16 cannot actually play online games on their own without the parents consent. However, that you know, there’s obviously loopholes in practice. But again, it goes back to there are, you know, rules and laws in place to try to protect minors. How do you view all of this? ⁓ Because in the with the backdrop of China trying really hard to diffuse AI into the real economy. And then there is this pushback like you just mentioned on concerns about AI taking jobs. I feel like there’s also almost like a ironic kind of contradiction happening where, you know, Tiger moms are like, okay, now we don’t need to learn math, we know how to learn AI. And Tiger moms are like saying, how do we optimize getting into, I don’t know, Harvard with AI’s help? And how do we get AI into the education, education, ASAP? I mean, honestly, we don’t even know what the education system might look like in like two decades. from our kids, but at this point, seems like there is like embrace. I don’t know. How do feel about that? Matt Sheehan (25:13) Yeah, a couple strands there. One just on the sort of regulatory side, like this is a long term strand in Chinese tech policy and tech regulation. They always put a pretty heavy emphasis on like how are kids using technology. They have, ⁓ they sort of mandated having like a minors mode on various ⁓ apps. ⁓ This is the regulation I think that you’re referring to as the newly passed. ⁓ I translate as anthropomorphic AI. That’s the word that’s the official translation. ⁓ So it basically means AI that, you know, behaves like a human. This could include AI companions that are, you know, literally like a character pretending to be your friend. They could also include, you know, the way that people interact with chat, GBT or Kimmy or whatever, you know, the phenomenon of having AI boyfriends and girlfriends and all this stuff. So there was a new regulation on this that was just finalized, I think, last week and It has some protections for everybody, specifically around ⁓ self-harm, addiction, and stuff like that. But it has really ramped up protections for minors and for elderly users. So there’s all these kind of specific add-on requirements. For ⁓ minors, it involves permission from parents. Parents can review at least some. They can set limitations on how the child uses the system. They can review. conversation that might have got toned down a little bit in the new version. ⁓ But I think, this is many ways it’s the same concern that surfaces in the US and elsewhere, like California just passed a just passed last year, passed a similar regulation on AI chat bots that I think also had believe it had extra protections in there for kids ⁓ on the education side of things. I guess there’s a couple of things. One, there’s just like the yeah, you say the tiger mom’s like this is ⁓ It’s a booming industry of like, I’m going to teach your, you know, four year old AI so that they can use it because this is going to be how they get a job and how they get into school and how they get a job. ⁓ you know, a lot of it is bogus. Maybe most of it is bogus, but it’s very attractive to parents who have grown up in a really, really cutthroat competitive education system where you’re looking for every single edge that you can find. And so that’s, that’s a piece of it on the, from the policy side, they have both sort of AI and education policies, AI plus education policies that they’re pushing in a bunch of ways. I have some friends who are working with teachers over there who are described to me pretty like sophisticated and interesting ways. The teacher that are using AI to lesson plan, to create like really interesting games that keep the kids engaged and learning stuff like that. So you have those, and then on the labor. the labor side of things, they’re also viewing AI, they’re also viewing education as something of an antidote ⁓ to AI fuel job disruption. This is in the, I in the five year version, it is in the five year plan, it’s in the AI plus plan, it’s in a few other places where they say, we’re really gonna prioritize lifelong education. So maybe you used to be an accountant, you lost that job and ⁓ you’re gonna retrain as something else. ⁓ which I think is a good, you know, it’s a good attitude to have. If you’re a person, you should always, I’m always trying to learn, you know, books. Um, think they’re great, but I don’t know if at a totally like a, you know, macro population, I know if you’re going to get 500 million people to be constantly staying one step ahead of AI in terms of what jobs it can do now. I mean, a lot of the things that we would have told you go back like two, three years and say, what jobs are going to be disrupted by AI? A lot of the recommendations would have been totally backwards. People would have thought that, ⁓ coding jobs are great. jobs involving creativity, ⁓ illustration, ⁓ stuff like that. AI can’t do those things. It can’t be creative in that way. And it’s like, that’s actually kind of what it’s best at now in some ways. I mean, you can argue about the level of creativity, but like generation of content, generation of images, videos, language. So I’d say it’s a piece of the Chinese sort of response on labor concerns. a fad, but maybe like a ⁓ useful fad within like the sort of education industry. But I’m a little bit skeptical of this as like ⁓ an actual antidote to the disruption that I at least imagine is coming. Grace Shao (29:49) Yeah, it’ll be really hard to be like upskilling, like re-skilling like hundreds of millions of people. Like, it’s just, you don’t even have the capacity to do so if there’s actually mass disruption. alright. ⁓ Matt Sheehan (30:00) mean, this was always the response, in the United States on coal miners. We’re going to teach them to code. everybody, all these manufacturing workers, we’ll offer like a job retraining program. I’m like, ⁓ maybe, yeah. Grace Shao (30:05) Yeah. I’ll take generations. I’ll take generations for things to shift, know, resources to shift, people’s mentality shift, you know, for a while, like, when many, when remember the first wave of like, a basic rural kids no longer wanted to work in factories and wanted to go to urban cities, there was a surplus essentially service providers and then like everyone eventually became a DD driver or a food delivery man and then Now we’re seeing a reshuffle in that population again, where people want to move back to their rural, cities. so I think based on how society is evolving, opportunities will arise without even us realizing anything, hopefully in the best case scenario, where people will find opportunities to reskill. ⁓ But I want to talk about something that’s a bit Grace Shao (30:57) I guess not heavy, but actually not many people understand, even including myself. So you really look at the government and the policy structure of ⁓ China’s regulators in the cybersecurity space and whatnot. ⁓ There’s so many players. There’s the CAC, then there’s the NDRC, there’s the MMIT. I can keep on naming acronyms, but can you give us a really, really quick high level understanding of who’s regulating whom? ⁓ how do they actually work with each other and are their KPIs aligned before we get into more about how you know, how China’s policy is shaping the technology and AI ecosystem. Matt Sheehan (31:36) Yeah, maybe I’ll do it. ⁓ I’ll introduce a couple of the players and I’ll do it somewhat chronologically in terms of like when have they become important or rise and fall in importance. So AI policy, like the first really big policy document was the 2017 National AI Plan. It was released by the State Council, effectively sort of China’s cabinet, sort of the highest level of government. But ⁓ people who are in the know say that that was largely sort of drafted and pushed by the Ministry of Science and Technology. So this is really like the policy wave of like 2017 to 2020 more or less. And it’s the Ministry of Science and Technology and it’s also the Ministry of Industry and Information Technology, MIIT. So these are really the organizations whose job it is to promote science, promote innovation, and MIIT is more of like the industrial applications of the technology. So they were kind of in the driver’s seat in that period of time. They were the most relevant actors. They were the ones who were driving real activity. starting in 2020, 2021, you had the CAC, the cyberspace administration of China, really like come to the center and become the most important actor in AI policy. The CAC, it’s basically the internet regulator. It was created in 2014. It was largely created to kind of like get the Chinese internet under control from a sort of political content ideology perspective. They’re connected to the Ministry of Prop, or the propaganda department. publicity, as they say now. ⁓ So from 2021 through 2023, the CAC was the one rolling out these binding regulations on recommendation algorithms, on deepfakes, on generative AI. And these are the regulations that actually force companies to do things. They actually force companies to register their models, to do pre-deployment testing, at this point to label AI-generated images in different contexts. There’s for that 2017 to 2020. It’s kind like the go-go period. Let’s just like push this industry forward. You have the Ministry of Science Technology, MIIT. And then from 2021 to 2023, it’s really the CAC. This corresponds roughly with the tech crackdown of 2020 through the end of 2022. That was a period when the CAC, the CAC is kind of at least historically, it’s kind like the bad cop of tech policy. They’re the ones who are like telling companies like come in and drink tea and we’ll tell you what you’re doing wrong or, finding companies in different ways. Cyberspace Administration of China, yeah, CAC. ⁓ Some people call it CAC. ⁓ And then sort of one of the more significant changes from 2023 to now is the rise of the NDRC, the National Development and Reform Commission, Chinese Fagawei. ⁓ And they are a macroeconomic regulator. They are like what grew out of the sort of state planning apparatus. And they’re Grace Shao (34:01) This is a cyberspace administration of China, right? Matt Sheehan (34:30) really powerful, they’re kind of a super, super ministry within the bureaucracy, but they’re not sort of directly, there aren’t that many direct connections to AI. They deal with, they deal a lot with money. They have money to give out for projects that funnels into compute projects and stuff like that. ⁓ But they wouldn’t be like who you would think of as the go-to AI regulator. What I was told and what I feel pretty confident ⁓ is what happened is that in some time in, I think, 2023, maybe mid to late 2023 and then into 2024, the top leadership in China essentially said, hey, we need a little more balance in our AI policy. The last three years it’s been led by the CAC. They’re kind of a bad cop. They’re really focused on controlling the technology, controlling the sort of output, the content, the ideology from it. And that is important. That’s kind of their first priority. But we need to rebalance this a little bit. We need to move out of our total tech crackdown era. And now we realize like our economy isn’t doing great. We realized we’re behind the US after CHAT GPT came out, and we need to balance this out. And so they empowered the NDRC to be a of a coordinator across AI policy, someone who is intended to take the input from the various ministries, from Ministry of Science and Technology, MIIT, CAC, and to try to make it little more coherent and balanced. And so that’s kind of the role that they have played for the past few years. The details of how that works out are shrouded in secrecy, you know, you hear little tidbits here and there. But there have been like visible manifestations of it. They had not, when they released these regulations, usually there’s a sort of a lead regulator on it or a lead policy document person on it. And then various other ministries, they co-sign it and they’re like listed below. NDRC hadn’t been on any regulations prior to 2023. And then starting in 2023, they were listed second as like the second sort of most important organ. policy body on these things. essentially we have this kind of like 2017 to 2020 is this like go-go period. Let’s diffuse. Let’s push the technology. Let’s push innovation. 2020 to 2023 is this more constrictive crackdown. Let’s build the regulatory infrastructure for things. And then 2023 to today is just like, let’s balance this out. Let’s not be purely focused on the content and ideology concerns. Let’s also be thinking about development. Let’s be thinking about employment. The NDRC is actually allegedly one of the groups that is very concerned about the employment impacts. you know, tons more details that I will love to go in on, but maybe that’s a starting point. Grace Shao (37:06) No, I think it’s super, super helpful. I just understand the nuances of like what their actual KPIs even are and like, you know, who does what, how they work together. I think that’s really helpful for lot of listeners and even investors who are trying to follow the space and just confused by acronyms. But help me understand now, like you say that 2023 to now essentially is in the same kind of era. However, I feel like at least from the capital market perspective, you know, the last year might have seen a bit of a shift again, you know, it was a bit of a let’s go AI, big tech AI, all the labs, let’s go, let’s go IPO. Then obviously the deals, some of the deals didn’t come through, some of the IPOs didn’t come through. ⁓ There seems like you even said people are being told to tamper down their excitement a little bit. Is that aligned with what’s happening with the policy side of things? Or is that actually more a reflection of just, frankly, you know, the AI space not being that exciting right now, you know, since the Gentic ⁓ kind of breakthrough. We’ve not seen more consumer and breakthrough. Also, there’s a lot of talk about, you know, there’s no obvious proof ROI on all the spending from all the big tech right now. Help me understand all that, I guess. Matt Sheehan (38:15) Sure. Yeah. When I was breaking down those errors, is largely its policy, but it’s already kind of like government attitude towards it. It’s like which, you know, they’re always in some ways swinging back and forth, going back and forth on the seesaw between, you know, control development, control development. And that 2023 to now being one era is sort of in that sense. It’s the period of rebalancing more towards development. There’s tons of sort of wiggles in that process and things they’re pushing more and retreating on. But from a positive perspective, that’s the overlay. ⁓ In terms of like the capital markets, investments, I mean, I think this is kind of at least for people in the United States, it’s kind of like the one of the most misunderstood things about the Chinese AI ecosystem is that it is really like cash constrained, that it is not like the United States where, you know, open AI is just like sucking in. the tens of billions of dollars from a huge variety of investors are just spending huge capital outlays, which people talk about, is it a bubble? Is this going to come back to bite them? That’s an open question. But in China, you don’t have the concern about that bubble because there just is not the same level of infusion of cash. when a couple of the companies did IPO recently, Z.ai, formerly Jerpool and Minimax IPO in Hong Kong, and I think I’m not Grace Shao (39:28) 100%. Matt Sheehan (39:39) really an IPO guy. think the IPOs were like modestly successful, but the valuations are just, yeah, the valuations are, yeah, not even close. And ⁓ it reflects a lot of things, but it largely reflects like a funding environment, a business environment, a macro economic environment, and the general sort of attitude towards risk investment. think I was just reading something that ⁓ Ray Ma from ⁓ TechBuzz. Grace Shao (39:42) So that’s six to eight billion dollars. The valuation is tiny compared to American peers. Matt Sheehan (40:06) China was writing on this. She’s always very good on these topics. yeah, it’s just people kind of assume that there’s like infinite money in China. They’re like, yeah, the government, whenever they want to, they just like turn on the taps and then, you know, it’s like, no, that’s not how it works. And like the VC ecosystem is much smaller, much more new and immature. And so it’s a different story. Grace Shao (40:28) on that note you know I was just in SF like last month and I met with quite a lot of investors and people’s kind of I guess misunderstanding was often twofold. One is exactly your point, people are just like oh China’s so rich the government just gives money all these AI companies are backed by the Chinese government I was like 100 % no first of all like there’s some other issues happening in the background but like the government doesn’t even are you know it’s kind of cash constraint and not even that much right now second all these companies are definitely not being backed by the government in any sense in fact Most of don’t want to take municipal governments or provincial government money because you get kind of tied into, you know, what we’re seeing is, you you get forced into working with the government and it constrains your profitability and commercial goals. On the other hand, another really big misconception was, I thought quite funny was that people often ask, was open-claw frenzy because the Chinese, average Chinese consumer or user were really, really concerned about privacy issues. So they wanted everything on edge. I was like, hmm, like again, it’s kind of like just not, a major conversation people have. Like I think I hate to generalize, but I think because of how the internet ecosystem is in China, people by default have kind of ceded to not thinking too much about privacy or personal data issues as much. So that definitely isn’t. So I kind of want to bring the conversation that this, you know, like What are some biggest misconceptions you think people have and how do we help them understand and bridge that gap a little bit better? Matt Sheehan (41:59) Yeah, I think yeah some of the stuff that you point out is correct like If you’re if you’re a if you’re a startup if you’re like a small medium company You I’ve talked to these people they’re like actually like we do not want to take government money if we can avoid it not just because we get kind of in mesh but like Entrepreneurs are legitimately afraid that if they take government money and then their company doesn’t work out and they lose the government’s money like they could end up like on the hook like in jail this legitimate fear that it was stated to me by someone. like, you know, is that happening to entrepreneurs everywhere? No, but it’s like you don’t. ⁓ The government. It does a certain amount of sort of VC-esque investing, but there’s not really that VC mentality of like high risk, high reward. Like we know that most of this is going to go under. It’s kind of local governments at least have been trained on like real estate investment, which is like 10 percent, 10 percent, 10 percent every year. And this idea that most of these companies that you invest in are going to fail is not really ⁓ deeply embedded there. I do think some of the companies do rely on government funding in different ways. ⁓ mean, Z.ai, Drupal, one of their biggest, maybe their biggest single revenue stream is from ⁓ building custom models, but custom applications for ⁓ state-owned enterprises, local government, stuff like that. Matt Sheehan (43:28) When you listen to them in interviews, they’re like, it’s not that big. Maybe it’s 40 % or something like that. But it’s a significant revenue. It’s part of their business model. So there’s that type of a connection to government. With DeepSeek, that’s a company that’s kind of quite mysterious. And we don’t know exactly where all their money comes from. Is it all earned? I think the government got more hands on with them in the sort of aftermath of the DeepSeek moment. You had reports about the government taking passports away from people who worked there to make sure like you guys stay local ⁓ or the government was like vetting investors was another this is reporting the information. ⁓ But the idea that like these companies are just they just have the kind of like the hose of government money just flowing in at all times and therefore they don’t have to think about anything else is just not it’s just not real. ⁓ They’re they’re much more constrained cash constrained. ⁓ terms of like trying to misconception on Chinese AI regulation, AI policy, this is like my, you know, much of my job is like first getting across like, the trend does actually like seriously regulate the technology. And then, you know, the next layer being like, it’s not all people think, you know, it’s an authoritarian system. She didn’t think he must just kind of like sit down and just like write the regulation. So like nothing matters except what he thinks. And we don’t know what he thinks. It’s like, no, like he doesn’t. There’s this actually very complicated and sophisticated policy ecosystem of, legal scholars and the companies are doing their lobbying and their thought leadership and, you know, they’re responding to public outcry over things. And I think that’s a, know, you can get this across to people, but it’s certainly not the people’s default mental model of how China works on policy is ⁓ just does not reflect the kind of sophistication in this zone. And it’s somewhat understandable. Like there are policy areas where Xi Jinping just like makes a decision and that’s. that’s where things are going. Like I think we saw a lot of this in the kind of 2020 to 2022 era. But as an AI policy, COVID, AI policy, it’s not that way. certainly things are not, people are not gonna like, you know, actively push things that are totally against the will of the top leaders, but they are within the constraints of like, Matt Sheehan (45:52) the direction of travel, what the CCP is good with, what she wants to do within that kind of very wide lens. It’s really individual people, scholars, bureaucrats, companies that are filling in all the details on this. And it’s a very sophisticated system because they’ve just had a lot of ⁓ had a lot of bites at the apple. They have like passed, I don’t know, eight, nine different A.I. already. The regulators at the CAC have been getting documentation from AI companies for three, four years. They’ve been building evaluations. been like, and they kind of, got their reps in with AI policy and it leads to a more sophisticated ecosystem. Grace Shao (46:33) ⁓ yeah, so the last kind of section I want to focus on is just getting into the nitty gritty about, you know, the policy and the security and safety kind of side of things we touched on in the beginning of our conversation. you are one of the few in the West, I think, and talk about the nuance of the word, which you just explained, it’s security, but also safety. ⁓ help us understand. how to interpret that when we read about that. It actually even helps us understand a little bit of what’s happening in the West. Like, I feel like there’s the governance people, the security people, the safety people, but from someone who might not be in that ecosystem, people are conflating it a little bit. And I just want to understand, you know, how do we understand each of their objectives, again, KPIs, or even their goals? Matt Sheehan (47:20) Yeah, yeah, basically, it’s really complicated. It’s very context dependent and it’s always changing. ⁓ I think maybe the first key thing to understand here is like the very particular meaning of AI safety in the West like that. The West AI safety ⁓ largely refers to kind of a specific camp ⁓ of AI development and policy people that are, you know, believe that AI is going to achieve human and superhuman capabilities. And this could pose like serious, maybe catastrophic risks to people. like, that’s a somewhat coherent community in the United States that has a certain amount of power. Their power kind of ebbs and flows depending on things. But like when you say AI safety in Washington, D.C., it means one quite specific thing. ⁓ In China, that community, it has started to emerge, but it’s much newer. It’s much more recent. It doesn’t have the deep roots that it has in the West. And ⁓ the way that the word is used in policy documents is both confusing and has changed over time. So a lot of times when ⁓ just to kind of put a little color on the terminology, Anquan, when ⁓ when you’re talking about cybersecurity in Chinese, you say Wang Luo Anquan. So it’s like network Anquan, network security and cybersecurity means something very different from AI safety from super powerful AI systems posing risks. And so there’s one sort of category of mistakes, which is to be very naive and to read all the Chinese policy documents. And every time they say a word that’s translated as safety to believe, wow, they’re talking about AI safety, they really, really care about this. That is a very naive and incorrect reading of things. ⁓ But in the past, I would say, 18 months, two years, you have seen a pretty significant uptick in the way that people sort of in and around the system and to a certain extent in and around the companies, their level of attention to what we would call in the West, AI safety to these more kind of large scale, potentially catastrophic risks from powerful AI. I’d say this is, there’s like sort of levels and degrees of this. There’s people talking about this. There’s it showing up in government documents in one way or another. And then there’s actually implementing this either through like binding regulations or through sort of companies doing their own testing and evaluation to try to their own sort of safety research and their own safety testing. I’d say what we’ve seen so far is a large increase in rhetoric, a large increase in sort of awareness within the policy community about these safety issues. We’ve seen it to start to show up in more significant documents. I think the one you’re referring to early on that I was working working on analyzing is called They call in Chinese the AI safety and governance framework 2.0, which is in many ways put out by some organizations underneath the CAC, the internet regulator. And it’s kind of ⁓ their attempt to diagram and do an initial discussion of how they see different risks from AI ⁓ and how are they going to mitigate these risks. Oftentimes they’re focused on technical standards as a mitigation. And there was a AI safety governance framework 1.0 in 2010. fall of 2024 and there was a 2.0 in fall of 2025. And just between those two documents, you can see real increase in the frequency and to a certain extent, the sophistication of the discussion around these risks in China. I’d say it’s pretty significantly below the sort of the AI safety discussion in the United States, but it’s on the radar. will counter, they’re like, okay, that’s great that they’re talking about it, but are what, you Are they just trying to trick us? Are they trying to make us believe they believe in safety? Are they saying it but not doing it? And ⁓ they’re like, we have not seen much in the way of like, we certainly have not seen like concrete binding regulations that sort of implement safeguards on this front. And in terms of what the companies are doing, it’s quite opaque, but ⁓ we don’t think that they’re doing the, I’d say. pretty confident they’re not doing nearly the level of sophistication or intensity of safety testing as you see at places like OpenAI and Anthropic. To me, this seems somewhat normal. This is kind of a process. Chinese companies have been behind. The government has perceived itself as being behind. When you’re behind the frontier, you’re not as worried about frontier risks as you’re like other people are going to get to those first and we need to catch up. ⁓ So I see this as kind of like a long-term process. And I think that the sort of increase in discussion about this is you encouraging if you’re concerned about these issues. But it’s a really don’t want to be just kind of reading the documents and say every time we see Anquan being like, wow, China cares about AI safety. Look at all this stuff. It’s much more ⁓ nuanced and evolving, ⁓ evolving quickly, I would say. Grace Shao (52:21) I know like when we spoke a couple months ago, just catching up, you were saying a big part of your job is also trying to help, you know, bring the two sides together. Obviously, you know, ⁓ it’s been challenging given the geopolitical backdrop, but how do you think the global AI governance space can work together? ⁓ Are we going to see, you know, kind of the two world superpowers and two super AI powers, ⁓ you know, guide? in different directions or do you think there are certain issues where they need to come together and they will come together and are coming together? ⁓ For example, to your point on safety issues around protecting humanity, protecting children, are these things that you are seeing collaboration? Matt Sheehan (53:08) Yeah, I’m kind of ⁓ both an optimist and a pessimist on this front in that, like I said, I have very, very low expectations for the United States and China to work together on anything. I have very, very low expectations of any type of a binding agreement or some type of detente where we both shake hands and kumbaya and we’re both going to be very safe with AI and we agree and it’s great. I just don’t expect that. ⁓ So in that way, I’m pessimistic. I think attempts to try to sort of preemptively create these global governance structures that are going to bind both of the countries in advance so we never reach these dangerous thresholds. ⁓ That’s just not where I’m putting my bets. I think it’s good. We need to make all kinds of bets on this front, and it’s good that people are working on this, but that’s not where I’m putting my bets. Where I’m putting my bets is on a much more limited kind of narrow bore, but I think potentially highly effective form of ⁓ engagement. wouldn’t even say cooperation. I wouldn’t even necessarily say coordination. ⁓ my sort of the term, my mental model for it is something I call AI safety in parallel, which is that like the two ecosystems are going to be moving somewhat in parallel. They’re both going to be pushing the technology forward. They’re both going to be working through safety issues from a technical perspective, from a policy perspective. And as we kind of move forward in parallel, we’re not going to be telling each other what to do. And we’re not going to be like, okay, I’ll do the safety thing because you are. You told me you’re going to do it, so I’m going to do it. We’re not like sort of moving in lockstep on this, moving in parallel. And we need to have these touch points. We need to have touch points where the two sides develop some form of mutual understanding of what the other side is doing. They understand how other side is thinking about the issues. They understand how they’re perceiving these risks. That’s one of the reasons I do the risk ranking. stuff and in some cases trying to share best practices, explain kind of explain what we’re doing and why we’re doing it and have the Chinese side explain what they’re doing and why they’re doing it and then where possible share good ideas that we think are sort of uniformly good in the U.S. and China. If we think that we have a policy intervention maybe it’s around ⁓ certain types of pre-deployment testing. ⁓ It’s good to communicate that. to the Chinese side. And it’s good to have the Chinese side communicate some of ⁓ the reasons and the sort of the specifics of what they’re doing on these fronts. We’re not here to just like trust each other. I think a lot of people are very worried that the Chinese side is going to, is they’re going to trick us. They’re going to say they’re doing it and they’re not, which is like legitimate concern. know, that’s, this is high stakes like geopolitics and powerful technology. So you don’t take anybody’s word for it. But when you have these conversations in talking with people, you can get a, a sense of their level of sophistication when talking about the issue. If someone is talking about AI safety and they’re like, yes, humanity first, protect the humans, control the machines, that’s our policy. And it’s like, okay, is there anything more to that? They don’t have more than you kind of know that they’re actually not really thinking about it. But if you can get into a more ⁓ deeply engaged discussion, you can see like, actually, yeah, they’re working through these problems themselves. You can see it in the way that they’re. discussing it. You can see it when they talk about their specific regulatory mechanisms. You can see sort of the connection between sort of action and outcome or thinking and action. And so my model for this is like, we’re not going to agree on things. We’re not going to sort of trust each other. But there are ways that we can both be moving forward at the same time and comparing notes, checking in, getting a sense of what the other side is thinking and doing that I think could contribute to safety in a meaningful way. Grace Shao (57:02) I think that’s fair and I think the word you kept on using trust is quite interesting because I feel like whenever I speak to people in the industry, ⁓ there’s just such a lack of trust even within whether you want to say countries or communities or beliefs and value systems and a very, very optimistic, naive way, I really hope that there can be a bit more consensus on certain things like that need to be protected in practice, like such as children, right? And how we go ahead with that. ⁓ But okay, I don’t want to end on a super somber note or anything, but... ⁓ The takeaway is trust nobody. That was... Matt Sheehan (57:34) It’s optimistic in a way. think this can’t... When you do see... Well, trust nobody, but talk and see if you can share some good ideas along the way. think there is real... ⁓ I’ve seen some real sort of traction from these type of things and I think it’s limited. We’re not going to get some kind of hard guarantee that China is going to be perfectly safe or we’re both going to... Matt Sheehan (58:04) do the right thing. But within with those low expectations, with those kind of pessimistic expectations, there are there’s progress that can be made. Grace Shao (58:13) ⁓ I do want to ask one question that’s kind of been happening around right now. That’s been kind of happening like the whole idea of Chinese open models seem to take a little bit of a sidetrack and starting to kind of only release their most frontier related models and close weights. ⁓ Obviously from a very like, you know, capital perspective, where I study it is I feel like it’s a lot of it is because they need to see our eye. They cannot keep doing this because they’re not making money. API sales not enough to, you know, sustain the kind of long-term ⁓ business as well as research costs. Are you seeing anything from the policy side? Like do you think there’s been a policy shift? That’s also kind of why I asked earlier if this year somehow, know, last year there was a public embrace by the government saying we should open source our technology. Has there been a shift? Matt Sheehan (59:04) So ⁓ I’ve seen sort of little tidbits around ⁓ sort of public policy concern about open weight models, but not enough that I would call it a shift. ⁓ In that document, the AI safety governance framework 2.0, it was interesting because it was the first time that there was a fair amount of ink spent on potential risks from open source models. The risks they were primarily talking about were essentially if there are vulnerabilities in these models in some way, either maliciously inserted or just a vulnerability mistake, those could proliferate throughout the ecosystem because you have all these downstream models and that could lead to impacts. There’s a little bit of a mention of like, maybe open models will be used by criminals and stuff like that. I certainly don’t see this shift in specifically Alibaba strategy as Matt Sheehan (1:00:04) in reaction to a significant policy shift. mean, it kind of makes, yeah, corporate decision. It makes if you’re going to be spending tens of billions of dollars building models or at least hundreds of millions, billions, bi

27. Apr. 2026 - 1 h 1 min

Differentiated Understanding

2 Monate für 1 €

Mehr Differentiated Understanding

Alle Folgen

Nur bei Podimo

Beliebte Hörbücher