What I learned from conversations with China's star AI scientist

On more structured approaches to AGI, the Chinese approach to AI, and the disillusioned emigre scientist

This spring, I had several lengthy conversations with Zhu Song-Chun—one of the world’s most cited AI scientists and the director of the Beijing Institute for General Artificial Intelligence—about China’s AI landscape and his perspectives on America.

Zhu made quite a splash in 2020, when he decided to leave his tenured job at UCLA and pursue his research on AGI in China. His decision set off seismic political shocks. Washington lawmakers sent furious letters to US institutions that offer Zhu money and Chinese outlets thanked Trump’s China Initiative for Zhu’s repatriation. Chinese netizens labeled him the modern-day Qian Xueshen, a famous MIT-trained rocket scientist who worked on the Manhattan project, was spurned by the McCarthy-era US government, and later helped lead China’s nuclear and satellite programs.

The Guardian Long Reads recently published my profile of him, where I detail why he left the United States, the geopolitical pall hanging over his decision, and some of the unresolved conflicts in AI research that brought Zhu to China. China has a more eclectic AI development ecosystem than the US, which has bet the farm on the belief that large language models will lead to AGI. (There is an evangelical tone to the AI debates in the US that I rarely detect when speaking to Chinese researchers.) Finally, the polarization of American politics from the late 2010s repulsed Zhu and probably many other bright immigrant scientists, whose American Dream doesn’t fit neatly into the MAGA coalition or the progressive left.

Here, I want to share some of these wider observations about the Chinese AI sector, problems with the American project, and where I see AI progressing. It draws from my chats with Zhu as well as AI engineers and researchers in the Bay Area, Beijing, Shanghai, in technology companies like ByteDance.

1. LLMs vs. More Reasoned, Structured Approaches

Nearly all American tech founders are proponents of the “scaling hypothesis,” the belief that large language models are on a path toward some kind of runaway inflection point, or AGI. It drives most of Silicon Valley’s investment decisions, the White House’s AI policy, as well as the Biden administration’s sanctions on high-end computer chips to China. This year, as improvements in LLMs have slowed and disappointed, it has bolstered the position of skeptics who argue that improvements in AI from scaling will soon plateau, or fail to achieve some important piece of AGI. Zhu is among this chorus, which also includes academics like

Gary Marcus

at NYU, Subbarao Kambhampati at Arizona State University, Joshua Tenenbaum at MIT, and, most notably, Yann Lecun, a neural network pioneer.

Broadly speaking, the skeptics see frontier models like GPT-5 as lacking something critical for AGI. They point to certain common sense intuitions—for example, LLMs are still notoriously bad at playing simple games—as indicators that something is missing. They emphasize in-built structures, with explicit tools for reasoning and planning, as opposed to neural networks synthesized on one internet’s worth of data.

To some AI researchers, there was always something unsettling about neural networks. Before they became dominant, the AI field was more fragmented, with diverse approaches that included the statistical modeling Zhu’s team practiced at Harvard and Brown, and the “cognitive architecture” he practiced later at UCLA. What neural networks did was show how most of the tasks researchers wanted AI to do, which they thought would require some creative, ingenious human design, could actually be tackled by brute force: relying on neural nets, data, and computing power, which, as Moore’s law dictates, keeps getting better.

AI it turned out wasn’t like physics, with one lone genius coming up with a very creative theory; it involved absorbing the internet and buying Nvidia chips. It felt like a short-cut: amazing and transformative yes, but a shortcut nonetheless. Still, there was no way to prove it because LLMs kept trouncing alternative methods. This beguiling dynamic was called the “bitter lesson.”

But we’re not at AGI yet, and its still not guaranteed neural nets will get us there. As of late 2025, it feels to me like LLMs will plateau before AGI, which says nothing about how earth-shattering it will be for us as a species. In the meantime, the older in-built methods weren’t wrong, per se, they just take way too long. It’s much easier to scale, and much harder to design the in-built “structures” that will allow machines to reason and mimic the human capacity for learning, as Zhu wants to do. The former feels like discovering a big oil reserve, the latter feels to me like an Einstein-level paper away. I think the skeptics might be the proverbial tortoise to the LLMs’ hare, probably on the right track to AGI, but really far away. I worry about the “group-think” in the US. If the Trump administration cares about AI, it should also be investing in alternative paths to AGI. Ironically, Trump is gutting the research funding of the universities that are equipped to do that kind of basic research.

Some of Zhu’s collaborators and researchers told me of one inspired possibility: There was a universe where, had Zhu stayed in the US and was a bit more openminded, he could have teamed up with the “Godfathers of AI.” In a way, Zhu’s work on systems that can plan and reason with minimal data input might not be incompatible with neural networks. There is a field called neurosymbolic AI, where researchers are working on hybrid systems of this kind. Zhu might have been the antidote to the current LLM stagnation, using his methods to improve LLMs weaknesses. That door — where in-built reasoning-based systems and data-driven models might have converged under a collaboration between their leading proponents—is likely closed, as geopolitics has hardened the AI community into rival camps.

2. Different Conceptions of AI in China and the US

In the much-discussed AI 2027 report, two superintelligent systems are set to either dominate or destroy humanity. The report is meant to highlight the apocalyptic stakes of AI competition, but it also revealed what Americans imagine to be the battle terrain: all-powerful AI, designed by manic founders and coders in a Bay Area or Hangzhou office, powered by vast computing clusters in Texas and Inner Mongolia. Two years ago, I was at an NYTimes conference in the Bay Area, when Sam Altman was describing what he imagined AGI to look like. I don’t recall exactly what he said, but he made a gesture with his finger: he pointed to the sky.

This is not the way Chinese talk about AI because over the past few years, “AI” has already been transforming their physical surroundings. In the Beijing district of Yizhuan, driverless cars glide past on congested roads. In Hangzhou, mall guests order KFC via drones at designated drop-off points. At the “dark” factory of the electric vehicle company Zeekr, in Ningbo, most of the production has been fully automated so the lights can be shut off. A Shenzhen nursing home is trialing humanoid robot companions, which will soon be rolled out for other tasks like garbage disposal and police patrols. Americans know that Chinese companies are racing toward the god machine, but few see that the rest of China is also racing—just not on the track Americans are watching. Last month, I went back to the United States to visit family and friends. As I landed in Newark Airport, took the AirTrain and the New York subway, I could hardly see the “gentle singularity” Sam Altman talked about. While Americans brace for the AI “take off”, a physical takeoff is happening inside Chinese homes, intersections, and factories.

There is something culturally peculiar about how Americans think about AI. It is suffused with philosophical concepts from the effective altruism movement or scifi cultural references like the StarTrek holodeck. There are religious elements, too, and now a suffocating dose of geopolitics: the Trump administration’s AI plan keeps saying the US needs to “dominate” in AI.

This strategic language does exist in China, but Chinese seem most excited about what AI—referring not just to LLMs, but robots, and smart devices—can do for the economy. This is because their old economic model of manufacturing exports and high investments in infrastructure is dead, and Chinese planners have been looking for a defibrillator to shock its economy out of its torpor, ideally to go back to the economy vitality that the real estate sector once held up in the 1990s and 2000s. The immediate value of AI for China is as an economic life raft. To use the Party vernacular, its the “new productive force” that (hopefully) can break China out of stagnation, out of “garbage time,” cure its “lying flat” youths, and, of course, while we’re at it, beat the Americans and show the world China is the future.

There are definitely “boosters” in China, especially at the big tech firms like ByteDance, but they don’t seem to take up much space in the public debate—when there is one. Part of this might be self-selection: Those who join OpenAI may have been inspired by sci-fi movies growing up or have breathed the Silicon Valley fumes that see technology as a tool of salvation; those who enter the AI field in China may have more prosaic drives. The ones I meet are PhDs in machine learning from top US universities who just want a high paying salary and fulfilling work in this lackluster economy. What’s weirder is there doesn’t seem to be “doomers” in China either, people who think AI will displace jobs and take over the world. AI doesn’t seem to elicit these polarizing views. Most seem cautiously optimistic or resigned to it. I’m not sure why.

3. The “Silent” Emigre Scientist

One of the definitive studies of the reverse brain drain phenomenon is the Princeton sociologist Yu Xie’s survey research on Chinese scientists after the Trump’s Justice Department launched the China Initiative. The study is often cited for the link it draws between the national security crusade and the surge in Chinese return migration, but the most unsettling parts of his survey were the individual comments he got from scientists:

If it were not because the COVID pandemic cuts off international traveling and I am a U.S. citizen, my family would have left the U.S. permanently without any intent to come back in the future. What I have experienced at my former institution was not only disgusting, but a system[ic] corruption that I believe [is] illegal. I had never thought of somewhere in this count[r]y to be dark and corrupted like this. If I had, I would not have become a naturalized U.S. citizen, which I regret now. What I ha[ve] experienced not [only] ruined my academic career, but also destroyed my American dream.

Zhu’s case was less severe and more gradual than the Chinese scientist above, but they share a common theme: the disillusionment with the American Dream. Given that we’re in a period of supposed American restoration, going back to the old moonshots and megaprojects of yore, it’s worth asking what Zhu and this scientist imagined America to be like when they first immigrated and what went wrong?

Another bright Chinese mind arrived in the United States four years earlier than Zhu in 1992: Wang Huning, the fourth highest ranking official in China. If you read his book America against America, the picture Wang paints is quite similar to the one Zhu described to me at his office in Beijing. Wang spends a lot of time on the Apollo program and American space exploration and the university system—MIT, Stanford, and Harvard—as bastions for the “dissemination of knowledge and wisdom.” Because America was the pre-eminent power for so long, it inevitable absorbed the projections of many contradicting views of what it stood for. For Chinese talents of the post-Cold War generation, though, America was not primarily a progressive march toward racial justice, nor the libertarian dream of self-sufficient enterprising individuals; it was a cosmopolitan laboratory of free scientific inquiry, where humanity’s best and brightest came together to push the limits of human possibility.

There’s important context for this. In the 1980s, Deng Xiaoping had practically turned scientific education into the new religion, and the state press inundated the public with coverage of scientific feats from new American chess programs to NASA’s Voyager missions. The message was clear: we need to learn from America to modernize China. (Though I don’t know for sure, I think its likely that this vision of America resonates among emigre scientists in other less developed countries who apply for university in the US. ) Over the next two decades, most aspiring Chinese scientists only followed the first half of that prescription—they went abroad and never came back. The image of America as an Eden of scientific innovation is not unlike the mythical portrait in Silicon Valley, among figures like Peter Thiel and Elon Musk. Several sources of mine confirmed that Thiel was the first to commission an English translation of Wang’s book into English and its easy to see why: Wang shared Thiel’s view of America as the cathedral of scientific advancement and also shared the same conservative critiques of its social ills (mostly because he read the same conservative intellectuals as those that influenced Thiel.)

This context helps explain how restorative campaigns in American politics from domestic political movements—Occupy Wall Street, the Tea Party, MAGA, the New Right—fail to speak to, and alienate, some of America’s brightest scientific minds. (None of the so-called “Godfathers of AI” were born in America: Yoshua Bengio is Canadian-French, Geoffrey Hinton is British-Canadian, and Yann LeCun hails from France.) If Americans care about winning the competition for technology supremacy with China, then keeping the country appealing to the world’s most talented emigre scientists should matter. The Chinese produce quadruple the number of engineers as America, so if all of that talent stays in China, to say nothing of other countries, that’s not a sustainable future for the United States.

Zhu did not see the same pathologies in America as Wang, who was heavily influenced by neo-conservatives like Allan Bloom. But his American Dream drains quickly. (Wang never had it to begin with.) During a one-year stint as a lecturer at Stanford in the heyday of the dot-com years, Zhu effectively told me that he realized many Americans care more about making money than doing something truly innovative. The “commercial ethos” in the Bay Area, Zhu told me, was very “thick.” Disappointed, he moved to Ohio State University, where he was even more disillusioned. Colleagues didn’t seem as obsessed with research excellence and adapting to a globally competitive era as he was. “In Ohio, the mentality was: don’t move up, just stay where you are,” Zhu told me. When the university proposed a fellowship to attract international students, Zhu recalled that some local media argued against the proposal because these students would eventually move to the coasts like New York or the Bay Area. The Ohioains “felt like they were subsidizing California,” Zhu told me. He later linked that culture to the insular, nativist elements of the MAGA coalition that contributed to his decision to leave the US.

I think there is some inevitability to these perceived short-comings of America. The country is much too vast, multi-faceted, and contradictory to survive such a one-dimensional prism painted in part by the Chinese government itself. This rosy image of America among a cohort of Chinese intellectuals and scientists has been labeled, critically, by one scholar, as “beaconism.” Most of the beaconists threw their lot with Trump’s restorative campaign in 2016 because they believed, at the time, that the chief threat to their image of what America stood for was “woke” progressive ideology. At the time, of course, that seemed to be the most destabilizing for the institutions they hold dear, such as free speech and the American university system. Zhu was among them; he had no patience for “political correctness.” He referred to disciplinary investigations of his fellow professors based on perceived racist remarks in lectures, though he declined to name names. He also expressed frustration with affirmative action hiring policies in his field, as well as when deciding awards and selecting speakers at conferences.

But this alliance between disaffected Chinese beaconites and Trump is weak because the former still hold a high degree of reverence for America’s university system, the immigration system, and freedom of speech. They don’t want to see these pillars of their American Dream crushed by the government, especially because many of their kids and family friends still participate in them. Though they don’t want their kids reading Derek Bell and Ibram X. Kendi, they also don’t entertain the monarchical obsessions of Curtis Yarvin. There is no space for them in American politics right now.

The math of US electoral politics dictates that this “genius emigre scientist” vote will never be taken seriously, but it’s under-theorized. After all, this is the cohort that moves the needle on the technologies of the future: only a handful of scientists have the ability to push the frontier in any given scientific area including semiconductors, artificial intelligence, batteries, critical mineral processing, nuclear energy, and biotech. If they vote with their feet, as they seem increasingly likely to do, America is doomed.