I think the other thing this brings up is just the rise of China in the AI race. Where are we in that race?
The numbers in China in terms of investment are not what they are, or close to it, in the US. But we do see a very powerful fast-following strategy where China's not really trying to be at the frontier of capabilities or the most prominent of where, or the most elite of where US model providers are, but they want to stay a few months behind it. If they can be maybe three, six, or nine months behind, then perhaps that's not really going to make or break their competitiveness, especially if they continue to have access to US models through distillation.
Welcome to Prof G Markets. I'm Ed Elson. It is July 28th. Let's check in on yesterday's market vitals. The major indices were mixed as chip stocks sold off. Brent crude fell below $90 per barrel as the US and Iran paused their attacks. The yield on 10-year Treasuries declined. And finally, a Chinese chip maker, CXMT, soared more than 400% in its public debut, becoming China's new most valuable company. More on that later. Okay, what else is happening?
Chinese open models are gaining ground, and Washington and Silicon Valley are starting to notice. In recent weeks, Chinese startup ZAI and Moonshot AI have released models that are competitive with those from US frontier labs. The top five most popular models on OpenRouter, a marketplace that tracks usage across AI models, are indeed Chinese. And when a rogue, unreleased OpenAI model hacked Hugging Face during a security test last week, Hugging Face used a Chinese open-weight model to defend itself. Now, you might have thought that American AI executives would use this opportunity to encourage a crackdown on Chinese open-weight models. Instead, the opposite happened. In his first-ever post on X, Jensen Huang shared a letter titled 'Open Weights and American AI Leadership,' and within a couple of days, 50 companies signed the letter, including Microsoft, Meta, Palantir, and IBM. Together, they argue that US AI leadership depends on building not just frontier systems but a strong open ecosystem around it. And they warn Washington against, quote, 'premature restrictions that stifle competition and drive innovation overseas.' Which raises an interesting question, and that is: why would US AI leaders seek to protect the very thing that supposedly might destroy them?
Joining us to discuss this, we're speaking with Scott Singer, technology and international affairs fellow at the Carnegie Endowment for International Peace. Scott, thank you for joining us. I'd love to just start with the context here. This open versus closed-weight debate that has taken over the world of AI has a lot to do with US versus China, which is strange and interesting, but I would appreciate if you could just clarify why that is and how we got to this moment.
I think it's really a lot of happenstance more than it was any initial intentional strategy or ideology. If we look at the history of AI model development, a lot of the initial general-purpose frontier AI models that were coming out of places like OpenAI and then Anthropic were proprietary. The weights were not accessible to any sort of outside users. And so for China, there was a question of where and how they might be competitive on the global marketplace for advanced AI models. And so having models that, for the weights that are openly available and can be used by companies and startups and can be fine-tuned or adjusted so that they can fit whatever purpose they have, that was sort of the area where, you know, China couldn't necessarily compete on having the most advanced capabilities, but they could have leaner, cheaper models that would be useful across the startup ecosystem and for companies.
One of the things that has caused a lot of controversy is this idea of distillation. And this is what Anthropic and OpenAI have accused many Chinese AI companies of doing — distilling their models, which has a relationship between being an open-weight model as well. What is distillation, and why does it matter in this conversation?
Distillation is basically taking a much stronger model or a bigger model and using it to build a sort of smaller, cheaper model. And so that is the sort of crux of what's happening here. And you can, in an open model, you can take — in this case it's Chinese companies, but there's been allegations that American companies are doing this as well — taking what is American innovation, American products, and using that to power these Chinese products. And distillation is a super common practice in AI, and on its own is not sort of something to, you know, be too focused on or see as especially malicious. But the question is really tied to intellectual property, trade secrets, as well as fraud. So the way and scale that some of the Chinese companies have been accused of sort of leveraging American IP in this case is through, for example, building up a bunch of fraudulent accounts, getting the outputs from the American AI models, and then using that output to train their own models. And so there's this sort of thorny legal battle that gets at the heart of a sort of broader geopolitical question, which is basically, you know, what are the rules of the game in terms of this very technical process, and it is more representative of this sort of open versus closed battle, even distillation itself is sort of a widely accepted common practice across AI.
Right. I mean, just from the perspective of a Chinese company, Anthropic releases its model, and then you submit like thousands of prompts to fine-tune that model, and then eventually replicate that same model. And that is what many of these companies have accused companies like Kimi and their Kimi K3 model, which went viral and made a lot of headlines last week. That's what a lot of these Chinese companies are being accused of. Now, I mean, in the rules of, you know, Claude and Anthropic's privacy policy, you're not allowed to do that. I guess the question for Chinese companies is, do they care about that? The answer is probably no. So then it seemed like we were going to get some regulation because the administration seemed to have a view on this, and they were accusing these Chinese companies of distillation and saying that that was a problem, which is why I was quite surprised to see this open letter from Jensen Huang, co-signed by a lot of other American companies, saying, no, don't restrict this. Let this happen. Distilling is okay. Open source is okay. Open weight is okay. Why do you think they are down with this?
I think that there's maybe two things to point out. The first is that the American tech companies are not a monolith, and Nvidia itself benefits tremendously from, you know, exporting many chips to China. It can power Chinese AI development, and much of the American tech stack remains even in the era of sort of great firewall tech decoupling. So much of US tech is still integrated deeply with Chinese tech. And so I think in that sense it is beneficial to companies like Nvidia to make sure that open-weight access remains open.
Startups in Silicon Valley, the a16z's of the world, are running, you know, on a combination of American and Chinese models. And so American startups actually want to use a lot of these Chinese models because they're so malleable and adjustable. And so I would say that those are sort of the main factors that are really driving this sort of bifurcation between, on the one hand, you want, you know, the Anthropics of the world that really have not much to gain and are getting their profits trimmed up by the fact they have to face competition from the smaller Chinese companies, but then you have so much of the US that still benefits.
Yeah. It seems as though this might be kind of an attack in so many ways on OpenAI and Anthropic because, I mean, those are the two companies that really lose if the Chinese cheaper open-weight models continue to gain market share because it puts pricing pressure on them. Might that have a role to play in why someone like, you know, Mark Zuckerberg at Meta would be very excited about rolling back any restrictions on Chinese companies because essentially it means that you kind of slow the roll of Sam Altman and Dario Amodei? Is that kind of — are those the teams that are emerging in AI right now?
I think it's undoubtedly the case that for a company like Anthropic, which is exclusively building these proprietary closed-source models, that having really strong competitive open-source models is not really where you want to be in terms of, you know, you already have razor-thin expectations in terms of the data center buildout. You're investing a ton in compute, and if you have your profit potentially undercut, you definitely don't want that. I think that there is just sort of a question for the entire ecosystem of how far is too far when it comes to restrictions, and I think that in general there's going to need to be a complementary approach that includes a combination of closed and open models. We see from, you know, instances like the OpenAI Hugging Face incident last week that there are going to be some serious risks, including the possibility that there are serious cyber incidents that lead to loss of developer control of varying degrees. And we also have concerns around misuse, and a combination of open and closed-source models are probably going to be necessary to mediate the full range of concerns that we would have coming from those models.
I think the other thing this brings up is just the rise of China in the AI race. And I know this is something that you spend a lot of time focusing on. I mean, where are we in that race? Where are we in terms of Chinese investment in AI, Chinese development of AI, and then also Chinese adoption of AI just among the general populace?
The numbers in China in terms of investment are not what they are, or close to it, in the US. But we do see a very powerful fast-following strategy where China's not really trying to be at the frontier of capabilities or the most prominent of where, or the most elite of where US model providers are, but they want to stay a few months behind it. If they can be maybe three, six, or nine months behind, then perhaps that's not really going to make or break their competitiveness, especially if they continue to have access to US models through distillation. And so, what does this mean for China? It means that they have to operate and get the best they can within an environment where they don't have that many financial resources compared to the US companies, where they have far less access to hardware to train on. You know, they're making the best of a strategy that has been sort of forced on them.
I would bet on the US position overall. You know, the US is the model provider of choice across, you know, OpenAI, Google, Meta, Anthropic, xAI. Like, if you look at the global diffusion of these models, there's just not that many countries and companies that are really uptaking outside of the startups that are really excited about the Chinese ecosystem. America is still really dominating here. But I think the question is just how far can China compete outside of its own borders. And one really interesting indicator of this is, like, for example, at the launch earlier this month of the World AI Cooperation Organization in China. And if you look at the list of countries that signed on to the list for whatever this international cooperation organization is going to be, it's not exactly a really big, powerful group. You see absent on this list, you know, really none of the countries in the Middle East that have been pretty excited about Chinese investment. And so it is to say that amidst this moment where Kimi K3 and Moonshot are getting deserved attention for the legitimate, really strong capabilities that their models have, there is just like an overarching gap between how strong the US is, I would argue, and how strong the Chinese ecosystem is.
This seems to be one of the biggest concerns in the AI world right now: is China going to beat us? Do you think that investors are too concerned about China, more concerned than they should be?
