Into Asia
Into Asia
China's Grand Bazaar of Artificial Intelligence
0:00
-38:33

China's Grand Bazaar of Artificial Intelligence

Chang just returned from a extended trip to China and shares some of his reflections.

Chang just returned from a extended trip to China and shares some of his reflections. He visited frontier AI labs and stopped by the World AI Conference in Shanghai, where Xi Jinping spoke and announced a new AI governance body with 29 member countries. Chang and Ian discuss what the AI talks between the US and China might produce, what the Chinese worry about as they develop stronger and stronger models, and why Chinese generally feel much more optimistic about AI than Americans.

Topics discussed:

  • Differences between China’s post-pandemic malaise (2023-) and now (2026-)

  • Competition between Beijing and Shanghai and entrepreneurial dynamism

  • Who came to Shanghai’s World AI Conference?

  • Trump’s reverse course on AI safety

  • US-China AI talks in September and Track II dialogues

  • Kimi K3 and the transparency gap in Chinese safety evaluations

  • China’s image makeover as a responsible AI power—is it true?

  • Why Americans seem to be more unsettled by AI than the Chinese

Ian Buruma (0:00): Hello, and welcome to Into Asia, the podcast by Ian Buruma — myself — and Chang Che. Chang has just spent time in China. I’m in New York, and I’d like first of all to know what it was like. You normally live in Yokohama, where life is very different from life in China, especially under Xi Jinping. What is it like to be back in China after Japan?

Chang Che (0:27): To take a step back: I came to China in 2020, during the pandemic, and I witnessed a lot of tumultuous events, the most intense being the Shanghai lockdowns in 2022. In the years after that, I felt a kind of malaise — a sense that people were stuck in their careers and their lives, that they didn’t have the right to reach for more, to have ambition, because so much was out of their control.

Coming back in 2026, I feel like there are new pockets of energy. What’s structurally different is that in the 1990s and 2000s you could see this happening nationwide. You can’t really say that in the 2020s. There are little pockets — specifically the AI sector and the robotics sector, which is where I spent most of my time.

Ian Buruma (1:27): You spent time in Beijing as well as Shanghai. Are there significant differences between those two cities in terms of morale and entrepreneurial attitudes?

Chang Che (1:39): They’re both pretty dynamic. Beijing and Shanghai compete fiercely with each other to maximize industrial competitiveness in sectors like batteries, semiconductors, and artificial intelligence. What that means is that both cities invest heavily and try to recruit talent from all over the country and the world. You see tech zones where the government has given over entire districts to robotics facilities.

Take the three top frontier AI companies in China: DeepSeek, MiniMax, and Moonshot AI. [NB: there are others, Z.ai, ByteDance, Alibaba…etc.]. Each of them is in a different city. DeepSeek is in Hangzhou, MiniMax is based in Shanghai, and Moonshot is in Beijing. That gives you a sense of how distributed this is, and the reason is that it’s a competitive system. The cities are competing with one another.

Ian Buruma (2:42): So nothing in the south. The south is really, as it was thousands of years ago, outside the center of Chinese civilization.

Chang Che (2:50): Well, Hangzhou is sort of south. I think you mean Guangzhou and Shenzhen. They play a big role, but they’re not in the large language model business — they’re more in manufacturing. Shenzhen is the home of DJI, the big drone company. And the founder of Unitree, which is the hot robotics company in China right now, came out of DJI. So there’s a connection there.

Shenzhen is the crucible of the robotics industry. Even though you now see robotics companies all over China, a lot of the components and the talent came out of Shenzhen. It’s very, very important.

Ian Buruma (3:37): The last time I was in Shenzhen — which was quite a while ago — what was interesting was that it sat outside the mainstream. It attracted a lot of young people because the air felt a little freer there, socially and, I’d assume, intellectually. Do you think that’s still true, and that it has an effect on the development of AI?

Chang Che (4:02): At that time, the system was still trying to figure out the right balance of political control and economic freedom. In those economic zones, people were flocking there for all kinds of reasons — partly economic freedom, partly political freedom.

Now we’re in an era where the system has become very good at controlling political freedom, but it’s also become very good at unleashing entrepreneurial dynamism under extremely restrictive circumstances. So you see this strange dynamic in every city: very little civil society, but plenty of entrepreneurial energy. Entrepreneurs will go to Shanghai and Beijing for the entrepreneurship and won’t much mind the political controls. Something different is going on now in China.

Ian Buruma (4:51): You spent a lot of time on this trip at the grand bazaar of artificial intelligence in Shanghai — WAIC, the World Artificial Intelligence Conference. What were your impressions? What did you learn?

Chang Che (5:06): My first impression was that China is in some ways so closed and yet sometimes so weirdly open. China was basically a closed country from around 2020 until 2023. When it reopened, there wasn’t much people-to-people movement between the US and China. And then you have conferences like WAIC, where Western scientists and scholars from Europe and the United States come in and meet Chinese officials in this very choreographed spectacle. That was bizarre.

It also selects for people who want sustained engagement with China. There’s a self-selection effect: the Westerners I tend to meet there are people who have worked in China for a long time, who engage with the Chinese system, who have a certain optimism about China. I consider myself one of them, actually.

Ian Buruma (6:11): Are there more Americans who seriously engage with China, or is it coming from Europe as well, and other parts of Asia?

Chang Che (6:19): It was definitely heavy on the Europeans and Canadians. There’s vetting on both sides. If an American scholar wants to come to China, they probably have to notify their university or institution, and the institution probably has a say. And if a Chinese institution wants to invite somebody from America, they also have to go through checks. At any one of those points, something might go wrong, somebody might say no, and it doesn’t happen. So there were fewer Americans — but there were still some, which surprised me.

Ian Buruma (6:56): In a conversation we had earlier, you pointed out that those Americans did not include the so-called accelerationists — the people in Silicon Valley who have almost turned the development of AI into a religion and think the world will be saved by it. They weren’t well represented.

Chang Che (7:17): In the United States there’s a dichotomy between people, typically in Silicon Valley, who are focused on development, and the people they sometimes call the AI doomers, who focus on safety. What’s happened is that the accelerationists — the people who support unbridled development of AI — also tend to be quite hawkish on China. So naturally, people from the US labs are not coming to China.

What China is attracting is a lot of people who are not actually building AI capability but who work on governance and safety, who think about AI ethically. That’s the crowd WAIC draws. On the other hand, the Chinese labs bring their own people, so the Chinese accelerationists are available to talk to at WAIC.

Ian Buruma (8:20): My impression is that in the Trump world, the accelerationists have had more influence than the ethicists. Is that still true? Is it changing? And how does it compare to the Chinese scene?

Chang Che (8:37): It’s changing. It was definitely the case when Trump took office. He came in and scrapped the Biden-era executive order that was supposed to govern AI policy, which had a lot of safety components in it. He set a precedent. He also brought in people focused on stripping out the guardrails — David Sacks, major Silicon Valley figures. He brought them into government. So there was a period of maximalist AI development policy in the White House.

I think that’s changed with the arrival of Anthropic’s newest model, Mythos. Anthropic claimed it had the ability to discover bugs in decades-old software, and to hack into power grids and telecommunications systems. That freaked out the administration. I think that’s why, when Trump went to Beijing to talk to Xi, they discussed the prospect of AI dialogues. And it’s also why the administration export-controlled Anthropic’s release. Anthropic felt it had done enough safety testing to release a version of Mythos called Fable, and it was the White House that stepped in and said no — you cannot make this accessible to people outside the US.

Ian Buruma (10:08): How far have those talks gone? And where does Scott Bessent stand in all this? He has the president’s ear on AI issues, doesn’t he?

Chang Che (10:16): Bessent has been taking the lead on AI over the past year. It’s not entirely clear to me what’s going on behind the scenes between now and late September. My understanding is that there will be a round of AI dialogues before the summit, and that what’s happening now is the planning for that: what concrete proposals can the US and China agree on, so that we’re in some kind of managed competition rather than unbridled competition — the way we had nuclear treaties with the Soviet Union.

I have no idea what’s actually on the table. But there are things I hope are on the table, if that’s of interest.

Ian Buruma (11:03): It’s certainly of interest. How much were these ethical issues discussed at WAIC? Was that a topic of conversation, or was it all much more technical — what the breakthroughs were, and so on?

Chang Che (11:17): There were a lot of discussions about safety, and a lot about technical sharing, research and technology. There were hundreds of forums. Shanghai around the time of WAIC also becomes a convention center for Track II dialogues — unofficial talks among former government officials and experts, which are meant to feed into direct diplomacy later. So WAIC plays a role in shaping what happens between Trump and Xi in September.

I talked to some of the people involved. Cybersecurity seems to be one of the big topics this year, because of the release of Mythos and Anthropic’s claim that Mythos is a very good cyber attacker. That’s been top of mind for policymakers in both the US and China.

Ian Buruma (12:14): It’s interesting to think of the parallels with the early Cold War — containing the dangers of nuclear proliferation and the dangers of AI. One school of thought very early on, and I can’t remember all the names involved, possibly Oppenheimer, held that the best way to deal with the dangers of nuclear war was to share all the data with the Soviets. Others said absolutely not. And then of course it became an arms race, and later people tried to contain it. Is this notion of sharing technical knowledge on AI strong on either side?

Chang Che (12:58): In each of these cases there’s a competitive impulse and an impulse to do what’s best for humanity. In AI, that plays out around safety evaluations.

Right now, all the labs do their own safety evaluations. Every few months a company produces a stronger model, and companies in both the US and China have safety teams that test the finished model for risky behavior. Is this model going to accidentally or deliberately hack into things? Will it disobey its human interlocutor? Will it help somebody build a bioweapon?

The problem is that there’s no unified standard across the US and China about what counts as an agreed-upon risk — what the shared risks are that we, as a global community, should worry about when these systems come online. And in China especially, the labs are not very transparent about what they’re testing for. At least this year, they haven’t been transparent about their safety evaluations. When OpenAI or Anthropic release a model, they share some information about the level of safety and their confidence that the model isn’t going to hack into a power grid. That isn’t really happening on the China side, or it’s happening less.

This goes back to one of the proposals I hope can happen: a shared set of standards for testing AI models. That will require some kind of sharing. I don’t know exactly what needs to be shared, but there has to be some exchange of methodologies and experience between the companies about what’s working and what isn’t. And so far that’s obviously not happening.

Ian Buruma (15:08): The fact that there were so many Europeans in Shanghai shows the world is moving in a certain direction in the age of Donald Trump — something Xi Jinping is actively promoting: the idea that China is a more dependable power, less erratic, and that Europeans and Canadians therefore see it in their interest to work with China and become less dependent on a United States led by a government that is in many ways hostile to its allies. Does that mean Xi is right, that China is more dependable? Do the Chinese take ethics and safety more seriously than the Americans do today?

Chang Che (15:57): This is the realization I had on the trip. China really is setting up governance infrastructure for AI, which the Trump administration has historically not been keen on. China has filled the void on what everyone now calls AI governance. Xi gave a keynote at WAIC where he talked about this and announced a new organization based in Shanghai — the World Artificial Intelligence Cooperation Organization. Twenty-nine countries signed on as founding members. So China has made a real diplomatic push to present itself as a responsible actor.

But at the end of the day — the other thing that happened just before WAIC was that Moonshot released a model called Kimi K3, which, according to some evaluators, is about as capable as Anthropic’s Mythos. If we take Anthropic’s word for what Mythos can do, then Kimi K3 should be pretty dangerous. In the hands of bad actors, it should have the capacity to do damage to cyber systems.

So what does that mean? While China has made a big push to present itself as responsible, it still has a long way to go in thinking about what could really go wrong. There’s currently no transparent system of checks around model releases. Kimi K3 was released yesterday, open source, which means anyone can download it off the internet. You need computing power, but with significant computing power anyone can wield it. That has made me worried that China is dropping the ball on a real safety regime, a testing regime, for the systems that are coming out.

Ian Buruma (17:59): One thing is a safety regime for the AI systems themselves. The other concerns content. Here the Europeans and maybe the Canadians are between a rock and a hard place. The Chinese are very, very good at controlling content, for their own political reasons, because it’s a dictatorship — and you don’t want to go that route. On the other hand, the Americans, very much influenced by the accelerationists in Silicon Valley, don’t like government intervention in questions of free speech, especially when the speech suits them, and lean toward a free-for-all. The Europeans think that’s dangerous, because it opens the door to wild conspiracy theories and manipulation of public opinion — dangers Europeans do want to control and the present American administration does not.

Chang Che (19:00): There’s a funny anecdote here. I talked to one of the AI scientists in China who works on governance issues, and he told me he’s recently had requests from the Spanish government for advice on how to do censorship controls. And by censorship controls he doesn’t mean political censorship. He means these models have to be contained — contained from, for example, assisting somebody in taking their own life. There was a famous case involving OpenAI. These models can say things we don’t want them to say.

So the scholar was saying: look, the West used to make fun of us, or criticize us, for being political censors of our systems. But now there’s a general acceptance that yes, we need to contain our systems. And yes, we in China will do a lot more, and we’ll have political controls alongside these other controls. But people are now learning from us about how to censor AI systems, because the Chinese are extremely good at it. They’ve been doing it for years.

The rules around content control in China are very strict, and the Chinese labs are very good at complying with them. I remember in 2023, when these systems were first coming online, there was a general view that the Chinese would be late to AI because they’d have to worry about all the censorship controls — they’d never catch up to America, because Americans don’t care about that. That’s just not true. They dealt with the censorship controls very quickly, and they’re now knocking on the American labs’ door. They’re very good at it, and now the Europeans are trying to learn from them.

Ian Buruma (20:43): This is a tricky political question. It reminds me a little of Deng Xiaoping, who wanted to open China to Western knowledge — scientific, economic — in order to modernize the country, but didn’t want ideas that could promote democracy, free speech, and human rights to come in with it. He called that “spiritual pollution.” It’s very difficult to keep the two apart. If you import Western ideas to modernize, you’ll get ideas you don’t like.

On the American side, the Trump administration talks a lot about how Europeans want to crack down on free speech. But what they mean by free speech is ideas they personally like — anti-immigration rhetoric, blood-and-soil nationalism. So controlling content very quickly becomes a political question: what is it that you want to control? And that goes way beyond technology.

It also reveals differences between the United States, Europe, and much of the rest of the world. The Americans have the First Amendment, which allows anybody to say almost anything, even neo-Nazis, as long as there’s no imminent danger involved. The Europeans have a different attitude — there are laws in Europe banning hate speech. The Chinese have their own ideas, which are even more restrictive. So content control is a very tricky issue.

Chang Che (22:34): This dichotomy you’re describing, between economic innovation and so-called spiritual pollution — these Western values — China has been grappling with it for its entire modern existence.

A recent example is Jack Ma, the high-flying entrepreneur who founded Alibaba, the e-commerce giant. He effectively disappeared from public life in 2020 after a speech he gave in Shanghai at the Bund Summit, where he attacked financial regulators. Here was an entrepreneur who was a product of Deng’s opening getting a little too naughty with his words and speaking directly to politicians. The regime cracked down. And in the years that followed, Xi took a hammer to the private sector and emphasized state-owned companies for a time.

Then in 2023 you have the breakout of AI, and OpenAI’s ChatGPT shines brightly, and everyone in China is wondering why China couldn’t create something as great. And Xi reverses course a little. He starts inviting tech entrepreneurs back into the fold. Jack Ma eventually meets Xi in Beijing again, and realizes he had miscalculated. This push and pull is constant in China.

This time in China, I got the sense from the AI companies that they’re really afraid of a repeat of the Jack Ma episode, because they know what the regulators are capable of. They do not want a rogue AI agent from, say, Kimi K3 wreaking havoc on the economy — or saying something politically offensive about the leadership — and triggering a Jack Ma moment all over again. That’s top of mind for a lot of AI researchers in China.

Ian Buruma (24:41): Now you’re talking about chatbots and AI systems running out of human control — which leads us straight into the world of Frankenstein’s monster.

It seems to me that in the United States there’s another contradiction. On the one hand there’s this often self-serving obsession with free speech — J.D. Vance lecturing Europeans that they’re cracking down on free speech when they try to control noxious far-right propaganda, the First Amendment, the accelerationists in Silicon Valley. On the other hand, there seems to be much more fear among the general American public about the consequences of AI running out of control than there is in China. American public opinion is more skeptical about the possible consequences of all this than Chinese public opinion appears to be. And of course one has to be careful, because you can’t gauge Chinese public opinion very easily when there’s no free speech.

Chang Che (25:51): What you’re describing, in AI lingo, is the accelerationists and the doomers. The similarity between them is that they’re both millenarian ideologies. They both believe there’s a tipping point where some transformation of society is in store.

Structurally, that millenarian frame is less prevalent in China. I’m not saying nobody thinks that way — Liang Wenfeng, the founder of DeepSeek, is, I believe, a hardcore believer in artificial general intelligence. But generally in China you don’t see this belief that AI is headed toward an inflection point after which the world will be radically changed. I don’t really get that there.

I think that’s actually an explanation for a fascinating difference in the general population between China and the US. There’s a Stanford figure showing that a large majority of Chinese respondents see AI products and services as more beneficial than harmful, compared to well under half of Americans. That’s a stark difference toward effectively similar things.

So I wondered why. Part of it is science fiction — in the US we’re reared on tropes that talk about AI in a certain way. What I’ve noticed is that when Chinese people talk about AI, they tend to focus on much more immediate risks. When you talk to researchers, they talk about job loss, content controls, AI companions and chatbots that might get a minor hooked on an AI boyfriend or girlfriend rather than a real person. These are risks happening right now, and the Chinese government has rolled out regulations on them. In the US, you hear people talking much more about the future — there’s going to be a superintelligent machine that can do everything and solve all our problems. That’s the discourse. It’s a big difference.

Ian Buruma (28:12): There may be historical reasons for this too. I’m wondering whether there’s an equivalent in Chinese literary history of Frankenstein’s monster. Frankenstein was written by Mary Shelley at the beginning of the nineteenth century and was very much a product of Romanticism, which was intensely concerned with the autonomy of the human individual. The idea of creating a fake human being was alarming because it would usurp that autonomy.

I once heard a lecture by a British philosopher, a very devout Christian, who was extremely critical of AI because he saw it as human beings playing God. That’s not quite the same thing, but it comes from the same culture, I think — the fear that human beings will be taken over by, or become dependent on, something they created that then runs out of control.

I wonder whether similar fears play a part in Chinese culture, and whether there’s an equivalent of the Frankenstein myth. I doubt it myself, because even in Japan people seem much less uncomfortable with robots, with playing with machines, even with having AI boyfriends and girlfriends — which is a universal phenomenon now. It’s happening everywhere.

Chang Che (29:54): Your point about autonomy is interesting, because on the hierarchy of values in East Asia, autonomy isn’t one of the top ones. It’s much more about harmony and other collectivist values. If that’s right, it may explain why you see much less of the view in China that AI is analogous to God, or to aliens. That’s very common in the West. Many AI researchers I talk to in the West will use the words God or alien when they talk about AI. I don’t hear that in China.

So it may be cultural. There may be some cultural reason. Do you feel these differences can be explained by culture?

Ian Buruma (30:40): I don’t think anything can be completely explained by culture — just as you can’t explain anything totally from a political or technological point of view. These are all factors. But yes, I do think attitudes toward divinity and toward individual autonomy play a role.

Chang Che (31:03): Here’s another perspective, which isn’t cultural at all. Some people I’ve talked to say the reason the Chinese are more optimistic about AI is that they’re behind. Their AI hasn’t reached the wake-up moment — the “oh shit” moment — yet, whereas in the US we’re getting there.

That doesn’t explain a lot of the discourse, but it may explain why, when you talk to Chinese companies right now, you don’t get the same alarm you might get at a lab like Anthropic, where they’ve already been running these tests and are starting to realize how capable and how dangerous their systems are. There’s less of that in China right now — but what I’m hearing is that it’s starting to happen. There are companies and researchers in China now equally worried about AI’s cyber-offense capabilities. So it might just be the lag.

Ian Buruma (31:59): It might. But on the other hand, those statistics you were quoting —

Chang Che (32:04): Right, the optimism gap. That doesn’t explain that.

Ian Buruma (32:13): No, it doesn’t. What might explain it is not that China is lagging behind, but that public opinion in America, even under Donald Trump, still matters a great deal more than it does in China. It’s very difficult to gauge what most Chinese really think. If people are asked in official polls, they’ll say what they feel they ought to say and stay out of trouble.

Chang Che (32:40): I also detect a lot of FOMO — a kind of reluctant excitement. I have no choice but to be excited about this, because it’s come my way, like it or not. That feeling is very strong in China. Whereas in the US, people are more used to rebelling. If you don’t like something, you say it.

So the excitement the survey registers might be a kind of fatalism — excitement under fatalism. I’m going to lean into this because I have no other choice. Because if I don’t, everyone else is going to beat me and take my job.

Ian Buruma (33:19): I do think there’s a cultural difference, and it has to do with religion — hence the fact that someone deeply Christian can be very skeptical about AI. The notion that it’s deeply sinful for human beings to play God is important in monotheistic traditions, which is why Jews and Muslims traditionally were not allowed to make images of human beings, because that leads to idolatry.

The Frankenstein story is in some ways also about idolatry: creating something that is not human. So I think there’s a link between the Christian who is skeptical about AI because he feels it usurps the work of God, and Mary Shelley’s idea that it’s sinful and disturbing and dangerous for a human being to create a false human being.

What isn’t true is that the Chinese lack a humanistic tradition comparable to the Western one. In many ways theirs is longer. Think of Confucianism — cultivating your inner life, your virtue, becoming a better human being. It’s deeply ingrained. So they have their humanism too. But this idea of creating false gods, this fear of idolatry, is much less present in China. After all, the Chinese — and the Japanese, who were very much influenced by China — create gods all the time. They make gods out of human beings, out of animals, out of rocks, out of mountains. The divine is everywhere.

Chang Che (35:08): You’re saying the Chinese and Japanese may be more used to that — to the creation of gods.

Ian Buruma (35:14): Yes. They don’t have the same resistance.

Chang Che (35:17): They don’t have the same resistance. Interesting. Technology comes from the word techne — the idea of a craft, an art. And I think the concept is decoded differently in China. In the US, when you create something awe-inspiring and potentially intelligent, you may ask yourself: what have I done? Have I created something I wasn’t supposed to create? Whereas in China there may be more of: this is so useful. This is going to make our lives better. That’s deeply ingrained — we need to build things to make our lives better.

There’s another aspect that’s just as deeply ingrained: we’re going to catch up to the West with this. We built this technology and it’s going to catch us up. We have to accept it. Technology is very nationally coded in China. In the US that doesn’t exist — we think in more individualistic terms, or more religious terms. Those are the tropes that come to mind when we think about technology.

Ian Buruma (36:17): On the other hand, if you look at history, the opposite is sometimes true. Both the Chinese and the Japanese have resisted technical innovations for social and political reasons. When the Japanese first encountered Western firearms in the sixteenth century, they very quickly learned to build guns. But guns were then banned in Japan for a long time, because the samurai class had a monopoly on the use of force with their swords, and arming commoners would upset the social order. They kept guns out of Japan for a long time. Such attitudes existed in imperial China too: when Chinese rulers felt that Western technology or innovation might upset the social order, they resisted it.

Chang Che (37:15): That’s a good point. I really think the feeling of catching up dates to the post-Mao years.

Ian Buruma (37:21): I’d say post-1900.

Chang Che (37:24): 1900, yes. Of course there was a catch-up phase during the Mao years as well. But especially in the past decade, Xi Jinping has made it clear that he wants China to be the leading industrial powerhouse of the world. And he’s willing to do that without some of the basic civic freedoms the West is used to. That’s where the association between technology and the national project comes from. Over this past decade, that’s been the ambient message from the government.

Ian Buruma (38:09): I think that’s absolutely true. My only caveat is that I wouldn’t put it as his being willing to do without citizens’ rights and freedoms. He thinks democratic freedoms would hinder China in its pursuit of industrial and technological power.

Chang Che (38:28): That’s right.

Ian Buruma (38:29): Well, we’ve covered a lot of ground.

Chang Che (38:32): Thanks so much, Ian.

Discussion about this episode

User's avatar