America Helped China Catch Up in the AI Race ft. Kai-Fu Lee | Fortune Daily

0:00America tried to win the AI race by cutting China off from its best chips, but Kai-Fu Lee says that that plan has backfired. Lee is an AI pioneer, an author, and the former president of Google China, and he argues that the export limits imposed by the U.S. have pushed Chinese companies to build cheaper, open-source models that are now challenging America's most expensive ones. Those chips are, of course, now at the center of an eye-raising scandal at NVIDIA, we mentioned it yesterday, involving the company's CEO, Jensen Huang, who has allegedly become entangled in a Chinese espionage plot. The company's board is investigating after reports that Huang had an affair with a woman who U.S.

0:50officials suspect of spying for China. It's safe to say that the US-China battle is the leading story right now in the global AI race. And Li holds a very unique and important perspective, having held leadership roles inside of both countries' AI industries. In a first for this show, we invited Kai-Fu Li to join us for a roundtable discussion led by senior writer and editor Ali Garfinkel, who covers VC and investment for Fortune, and senior AI reporter Emily Fellini. Lee's work straddles both of their coverage areas, and they discussed what's really driving the US-China AI race. The interview you'll see right now is just a small part of the whole conversation, which we'll publish in its entirety on our YouTube channel later this week.

1:40So subscribe to make sure you don't miss that. I'm Ellie Austin in New York, and this is Fortune Daily.

1:54Dr. Lee, Kaifu, thank you so much for being here. Thank you. Here at Fortune, I focus on investors, Silicon Valley. My colleague, Emily Forlini, our senior AI reporter, is an AI expert. And there's always AI news, big AI news today. Emily, we were slacking about this this morning. Yes, this is an enormous story, global story, so we wanted to ask you about it, that NVIDIA is investigating its own CEO, Jensen Huang, for having an affair that he has owned up to. He did do this with one of his customers. So she is a data center executive in Asia, and they have had an affair, and she's a suspected spy for China.

2:33I thought maybe you could just give a little bit of context on the backdrop of some of the tension between the U.S. and China in terms of chips, which is really at the center of this. So it is public that the U.S. has restricted export of the most advanced NVIDIA chips to China. Why is that and how has that landed in China? Why is that such a controversial issue? Well, I think the whole LLM revolution began with the belief that whoever has more GPUs can train on more data, will build possibly AGI that will squash every other company from any country. So with that belief, then making sure that American companies have the most advanced chips and most quantity of chips seemed like a plausible strategy.

3:19But it ended up being completely ineffective because it actually did restrict the best chips from China. China did have to scramble. Chinese companies, I should say, did have to scramble. But in the scrambling, people put their engineering ingenuity to tweak everything to be 10 times faster than the American models. And I think that has led to a set of open source models made in China. That's maybe six to 12 months behind the American models, but it was made cheaply and it's now free for the world to use. So now we're in an iPhone and Android situation where the American companies have closed models, just like the iPhone.

4:04They make a lot of money. Chinese companies don't make nearly as much money, but they're going to get the footprint and the market share. And also the open models in China, if you look at websites like Open Router that have rankings of the most used models, the Chinese open models are consistently at the top. So why are people using open models so much worldwide? And when would they use an open model versus a closed model? Um, so first, I think the top American companies price their tokens extremely high with very high margins. It's somewhat understandable because they spend so much training, they need to recuperate the cost.

4:41But when the Chinese companies have figured out how to build nearly as good models for one 10th or even less the price for both training and inference, then that creates a natural alternative. So the types of people who like the open source models are the following. People who can't afford or don't want to pay the exorbitant prices. Secondly, they might have a less demanding task for which the open and closed models are equivalent. And third, it's a really big advantage, is that the open-weight models you can copy to your server and run it on-premise with no data leaking to large language model companies.

5:23And we have seen throughout the short history of LLMs how vertical application companies come out with a very cool product, but that's built on using closed models by sending lots of tokens to the closed models. And then the closed models gets all the data and train on top and eventually eliminating the business proposition of a startup. So with news like that really understood by enterprises, and also enterprises view their corporate data as sensitive and their competitive edge. So they just don't want to send their data out to a large language model company. They might be willing to send it to a hyperscaler.

6:13But then, you know, there are obviously people who are okay sending it to the LLM companies, closed models. Some will only send it to hyperscalers, and some really want to keep it on site. So the third group will move towards the open models. I am really curious. I feel like there's a lot of noise, there are a lot of headlines, but there's not a lot of context about the AI race between China and the U.S. What's fact? What's fiction? Where do you sort of feel like there are misunderstandings, either in the U.S. or in China? I think there are a lot of misunderstandings. A lot of them. Yeah. So first, there is a belief that the Chinese companies can't make money.

6:58But if you go back to the history of mobile internet, as I described in the AI superpowers, the Chinese companies are born to make money. That's the primary purpose. It's not so much to win AGI or win the race. So in the cases of Pinduoduo, Meituan, ByteDance, these were all companies that were questioned about how they'd make money, but they figure it out. So you may doubt, you may reasonably doubt China's ability to do a breakthrough, which is valid. The U.S. is way ahead on that, but never doubt Chinese ability to find a way to make money. So that's the first misunderstood area. The second one is there's some belief that China is an entity, that the Chinese government tells the model companies to do things like do open source.

7:53That's the way the national strategy wins. This is complete non-existent, made-up nonsense. The Chinese companies are venture-funded, just like the American companies. They have Chinese investors, American investors, they just want to serve their investors. And what really happened with the Chinese open source phenomenon was not by design, not by a company, not by the government, but rather when deep seek moment arrived that the best model in China was open source, that really raised the question for all the other LLM companies, can you be deep seek? If not, get out of the business. If you can, you must open source because there's no situation where a inferior closed source model will ever have any opportunity.

8:38So everybody follows suit and open source. And then the history began and the rest followed. Most are still staying in the race to be the best open model. I think that's the second misunderstood thing. The third is that export control is what limits the Chinese spending on GPUs. It is not. The Chinese companies are naturally parsimonious. Not just the startups. It's a piece of culture. Well, it's a piece of culture, yes. And it's also about prudence and survival instincts. That place is higher than the AGI dream. So a startup company is measuring its budget very carefully, raising money when there's chance using it prudently.

9:25And now they have money to last three or four years and we'll see what happens. And then the larger companies, Alibaba and Tencent, and they basically spend a fair amount on training their models, but nowhere like the American companies. And then if you look at... Do you think the American companies are spending too much? What does too much mean here? Yeah, so here's the American value proposition, right? Open-anthropic. To them, the most valued asset is talent and talent density. Very understandable. To attract the super smart PhDs, you got to give them GPUs. And you don't want to force them to tweak, you know, infrastructure, you know, firmware.

10:07That's not what they want to do. So to make these brilliant researchers happy, the cost of training is not just the cost of training. It's the cost to let a thousand people run a million experiments, which costs perhaps five or more times as training. But that's the American path. And the dream is that this team of genius will make an AGI that would become the most profitable company and the most valuable and remembered company in the world. That is the Silicon Valley model for open AI and entropic. And I respect that. And I think it's logical. The Chinese model is more like a study group. None of the companies have the talent density of Anthropic and OpenAI, but all of them have pretty good talent.

10:56Because they can't really compete alone, they do compete against each other, but they're basically open sourcing or at least open weight their model, and that allows each other to examine the model. they're actually publishing papers when the American companies are much less doing so. So the papers, the talent moving around in the companies, plus the open weight models that you can examine causes a rising tide to lift all boats. So there are actually quite a number, I would say four or five very decent Chinese open weight models. And that's how it came about, is deep seek success, forcing the hand of every company to have no choice but to also open weight or drop out.

11:46So this goes back, there are a couple of things here. I mean, the first is, it sounds like the deep seek moment was as big a deal in China as it was here. Bigger in China. It was a pride, a national pride for a lot of people. Was it all over the TV? Like, what did that look like? It was everywhere? Yeah, talking about it. It was cleverly done just before Chinese New Year So classic and then in Chinese New Year Everybody goes home and then the kids tell their parents and grandparents and just spread like wildfire and It was previous before DeepSeek. It was believed China was way behind maybe two years two and a half years behind the US And even longer before that but the moment DeepSeek came out It was like the one of the top three or four models in the world and I think that was fun to use.

12:34It was good in Chinese and then it's a national pride thing so people say okay we thought we were never going to catch up now we have a shot so it was a quite a special moment. A lot of the VCs who I cover at the Deep Seek moment were truly a gog I believe is the term over Deep Seek because of the low costs of training it and as we sort of think about the news, Anthropics S1 dropped a few weeks ago. The company is set to go public. I am really curious. Anthropics is a spending story for so many reasons, right? When you look at those financials, not only the revenue, but the massive net losses, those massive costs of compute, what do you see?

13:18Well, I've heard Dario speak that if you remove the cost of training the last model, the balance sheet actually is quite healthy. And that is a valid way to look at it. Right, because the - Still have to train the last model. Yeah. Well, the last model is - Training new models and new models, and they kind of replace themselves. So a lot of people are talking about, is that sustainable? Yeah. So his explanation is training the last model is going to bring revenue in the next two years. Then spending more money to train that model will make money in the two years after that. So they see a two-year horizon.

13:56The models are coming out every couple of months. Yeah, and that's the other thing that Anthropic should rightfully be proud about. It's because they have figured out some sort of RSI, recursive self-improvement. So AI is improving AI. So when he actually did the interview where I quoted, I think it was truly one and a half, two years per, one and a half years anyway, per release. Now it's going faster and they have better RSI than a lot of other companies. So these are they have RSI so RSI is here some they have some well There's still human in the loop, right? But when when the model is let's say 80 % trained by AI 20 % by the people It'll be trained, you know, three or four times faster.

14:42So the improvement will actually accelerate So these are the valid arguments for the anthropic valuation Another valid argument is that there are a lot of American enterprises who are willing to pay the exorbitant price prices, and that creates a very amazing growth of the revenue stream. If it weren't for the big spending and losses, it would be a beautiful spreadsheet, P &L. So these are the valid cases. That's very interesting that you think the first thing you said, that their valuation depends on them developing RSI, recursive self-improvement. It's partly. Partly, yeah. That's interesting.

15:21I have noticed they're putting out more papers to show that they are, and maybe that's perfect timing? So I think these are the valid reasons. So I can see reasonable people making an argument, they're buying. But on the flip side, there are also issues. Because one, their model is better than the open source models, but not by a huge margin, by six to 12 months. So people could say, I'll settle. I'll take a slightly worse model and save a bunch of token costs. and also their models are kept secret and then people would say, hey, I want on-premise. I hear that increasingly in the enterprise users.

16:05So the question is, will the enterprise users keep using more tokens or will some of the tokens move to the open-weight models? So will their growth really go at, they clearly won't grow at the same rate they've grown the last two years. Again, this was just a small portion of our conversation with Li. We also discussed Anthropik's path to an IPO, why businesses are switching to open source models, what Americans get wrong about China's tech companies, and Li's overarching thoughts on our AI future. That full conversation is dropping on our YouTube channel later this week, So subscribe to make sure you don't miss it.

16:50Up next, if AI can solve the problem, does it matter if humans don't even understand the answer? Well, OpenAI has published full or partial solutions to more than 370 outstanding math problems. But some of those answers have already been withdrawn. Jeremy Kahn now looks at why the release is sparking both excitement and backlash over the future of human discovery. A few weeks ago, OpenAI came out with a whole set of proofs to over 370 outstanding problems in mathematics. And this sort of stunned the mathematical community. It also has divided the mathematical community. A lot of them are very unhappy about the way OpenAI went about this.

17:32The concern is that by dumping all these solutions out there all at once, that essentially they've kind of undercut the entire field. that mathematicians are, a lot of them are very worried about, you know, what is their purpose now if not to try to solve these problems? There's a question about how inhuman interpretable those solutions are. And a lot of people are saying, you know, the field is, although gaining answers to these problems, potentially losing insight and losing the opportunity to perhaps develop new kinds of mathematics. Because in the past, when humans worked on these problems, they would often have to develop whole new areas of mathematics to try to solve them.

18:09And while there's some suggestion that the AI models for open AI have done this, it's very hard for humans to tell. But a lot of mathematicians are concerned that this isn't really the same process that you would have traditionally in mathematics, where a human would publish a proof, it would be peer-reviewed, they would then hold conferences and workshops about what they did, and the whole field would kind of gain this knowledge just by putting these out there, that none of that's going to happen. What becomes of the field? So I spoke to this guy, Dan Litt, who's a younger professor at the University of Toronto, and he is fairly hopeful.

18:44He thinks that there's lots to work through in these problems to, you know, that as human mathematicians go through the proofs, they're going to, you know, gain a lot of knowledge. He says it's really not that different from when humans would publish proof. That also often took mathematicians a lot of work to sort of understand And he also says that the whole field in general hasn't been very good at writing and explaining proofs and that this will become a new part of what mathematicians do. Is this the end of human driven math? Well, there are certainly lots of people on both sides of this debate who seem to agree that this does mark a kind of phase transition in mathematics.

19:21Terence Tao, who's often considered the greatest living mathematician at the moment, he's been pretty upset about what open AI has done, but he's sort of resigned to the idea now that the field has to move to this new phase. And he calls it sort of mathematics 2.0, which is much more about trying to explain proofs, trying to expand on the knowledge that AI generates. maybe a lot more about transmitting this knowledge to others outside of mathematics and looking for applications, that this will become really much more about what mathematical research is, rather than being so focused on problem solving.

19:59Now it may be much more about kind of exploring the kinds of math that the computer comes up with to solve the problem. That's it for an AI-packed episode of Fortune Daily. I'm Ellie Austin. We'll be back tomorrow. I've told them so many times to subscribe today I can't ask them again. Okay.