A Chinese AI policy think-tank director, halfway through a polished institutional presentation, told Nathan Labenz that American conduct toward China could end in war. Labenz wrote the words down verbatim.
Most American commentary treats the question as whether the United States can stay ahead. Labenz spent two weeks in China and came back arguing that the goal should not be winning at all, and that the chip controls are doing more damage to American safety research than to Chinese labs.
"There could be war. I'm really worried about that."
Labenz hosts The Cognitive Revolution, paid his own way to China for two weeks of meetings with companies, researchers and policy institutes, and makes this argument against the stated policy positions of Anthropic, which sponsors his show.
I listened to the full episode so you can skip it. 3 hours 17 minutes of audio, 32 minutes of reading.
Here are the 24 arguments that matter.
👤 Speaker: Nathan Labenz, who hosts The Cognitive Revolution and had just spent two weeks in China meeting AI companies, researchers and policy institutes
📰 Published: 10 September 2026
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 3 hr 17 min | ✅ Time saved: 2 hr 45 min
Key Takeaways
The goal should not be winning the AI race but a settlement he calls a Pax Robotica
His test is whether both countries get the benefits without building a new permanent threat
China's strength is a return to its historical norm, not a recent anomaly
The century in which China was poor and the US was the only superpower is the exception
At purchasing power parity China's economy is the larger one — $45T against $30T
Chinese AI is not a fast follow: AI entered the five-year plan in 2016
By his count there are now 11 American and 12 Chinese companies shipping frontier or near-frontier models
Export controls are chilling American AI safety collaboration more than Chinese capability
US safety organizations told him they fear their own government, not China's, if they collaborate
A Chinese policy adviser's advice for repairing the relationship was to relax the export controls first
The most dangerous outcome is the two countries' technology stacks diverging until neither can read the other
Shared hardware, shared architectures, English papers and open weights are what make each side legible today
Banning Chinese open-weight models would be the un-American move and would push other countries into China's camp
His alternative is to price model risk through insurance rather than prohibition
Neither government wants to lose control to AI, which is the shared interest an agreement could be built on
Training AI to win wars by any means is, in his words, training the takeover everyone fears
1. The Goal Is Pax Robotica
Labenz opened by objecting that American China policy has no stated goal, and then supplying one.
His complaint is that the discourse skips the question entirely. The conversation jumps to wanting to win without asking what winning is, what losing would cost, or whether a win condition exists at all
The goal he proposes is a settlement rather than a victory. "I argue that our goal should not be to win, but rather to realize a future of peace and abundance in which both countries and the world as a whole can share in the tremendous benefits of AI without running extreme risks or creating new AI swords of Damocles"
The upside he describes is domestic and physical, not geopolitical. "Robots working the factories, robots working the fields, picking the fruit, even robots in our homes cleaning for us, doing our laundry, making our meals"
The downside case is the nuclear analogy, and he thinks the analogy is exact. Roughly 10,000 warheads remain online, many on alert, for a deterrent far larger than deterrence requires: "But there is a nontrivial risk every year that we could have some sort of accident or misunderstanding or surprising escalation to nuclear war"
What he does not want is a second standing threat of the same kind, built out of AI
2. Criticizing His Sponsor
Before the argument began, Labenz disclosed that Anthropic sponsors the show and then spent several minutes disagreeing with its policy position.
He credits the vision and rejects the strategy. He calls the first half of Dario Amodei's essay Machines of Loving Grace the best description of an AI future any frontier-lab executive has written, and says a Pax Machina is another way of describing the same thing
His own stake in that vision is personal. "My son's experience surviving cancer and getting back to basically full health in a six, seven month time frame is an incredible fruit of technology"
The disagreement is about racing, and he quotes Amodei against himself. From a 2017 clip he says he once reposted: "Technology races can raise the risk of safety catastrophes"
What he objects to is the underlying posture. "I do still have these misgivings around what I see as a better us than them mindset" — a posture he says applies to OpenAI as much as to China, and which he thinks is pushing Anthropic to advance the frontier faster than is wise
He notes the criticisms predate the sponsorship and says he expects to keep making them without consequence to the business relationship. He also disclosed that he paid for the China trip himself and accepted only meals
3. China's Rise Is the Norm
His net assessment starts by attacking what he calls cope — the same pattern of denial he sees applied to AI capability.
The denial ladder is familiar to anyone following model progress. "The AIs have come this far, but they'll never reason", then never discover knowledge, then never change paradigms — and he says China commentary follows the same pattern, ending at the claim that China can copy but cannot innovate
His first conclusion is blunt. "China has arrived. China is a world power. And any notion that we might have that we can contain China's rise is frankly ridiculous at this point"
The historical framing is the load-bearing part of the argument. For most of recorded history China has been among the world's most developed and technologically advanced societies; the period from the Western and Japanese incursions to roughly 2000 is one of its lowest points. "It was the downtime that was the exception"
He offers a personal measure of how far living standards have moved. A 45-minute pedicure in a non-cheap part of Shanghai cost him about $12, and the young man working there had an iPhone 17 — which is not cheaper in China than in the US. "So how does my $12 turn into going through the business, whatever margin ownership's taking, rent, overhead, translate to where they can get an iPhone 17?"
His partial explanation is family wealth concentration: one child with two parents and four grandparents, sometimes several apartments, so disposable income is a high share of a low wage
4. The Economy at PPP
Labenz argued that the headline comparison of the two economies is the misleading one.
On nominal GDP the US is about $30 trillion against China's $20 trillion, so the US looks 50% bigger. At purchasing power parity the Chinese economy measures about $45 trillion against the same $30 trillion, which reverses the ranking
His view is that the second measure describes the lived economy better. Dinner for two under $10 in Shanghai is available in a way it is not in New York, San Francisco or Detroit. "Go there, spend a little time, buy some stuff, see how cheap it is"
The infrastructure comparison is where he is most emphatic. A second-class ticket on the high-speed train from Beijing to Shanghai cost about $60, ran at roughly 200 miles an hour and stopped in cities most Americans have never heard of. "It left at exactly the moment that it was supposed to leave. The trains definitely run on time"
The detail he found most striking was operational rather than mechanical: passengers ordering food to a specific carriage at an upcoming station, delivered during a five-minute stop
Even airport pricing came up as evidence, because the prices he saw in Chinese airports were ordinary street prices
5. The Mandate of Heaven
He then argued against a specific American assumption: that the Chinese government is unpopular and therefore fragile.
He treats the recent Iran experience as the cautionary case, where real discontent did not translate into a population overthrowing its government when outside pressure was applied
His claim is that the Chinese government is not merely tolerated but credited. Two generations of rising prosperity, security and restored international standing have earned it legitimacy on the terms Chinese political thought has always used: "The key idea, I think, is, like, who is really delivering is who is legitimate"
The inversion of Amodei's phrase is his sharpest line on this. Where "eternal 1991" describes an American memory of the Cold War's end, "I think that today, The Chinese government has done so well for its people that if anything, they are living in their 1991 right now"
He does not claim there is no dissent, citing Tiananmen and continuing censorship, but says day-to-day life is broadly free — work, hobbies, music, bars — with specific no-go zones, and that the memory of the Cultural Revolution makes stability itself something the state delivers
The practical conclusion is about strategy, not sympathy. Picking a fight with the government will not win over the population: it will consolidate it
6. The Green Card Fantasy
One policy idea he singled out is the proposal to offer Chinese AI researchers green cards and expect them to come.
He calls it a fantasy built on an outdated premise. "This is a total fantasy" — the assumption is that China is a place people want to escape, which he says was once true and is not now
The counter-example he names is Moonshot's founder, who did a PhD in the United States and went back to China to start the company
He still supports easier immigration on its own merits. Assimilating ambitious people is what he calls possibly America's single greatest strength, and he thinks the country is damaging it. "But don't make the mistake of thinking that all the top AI talent in China is just dying to leave"
The other constraint is that Beijing would not permit it, which he illustrates with Manus: acquired by Meta for around $2 billion, then blocked. "But then The Chinese government stepped in and said, no. We're not allowing this"
The people he spoke to in China thought this was a mistake by their own government, because an agent harness is not foundational technology and blocking the exit damages the venture market. "Everybody seemed to agree that it was sort of an own goal"
7. Chinese AI Began in 2016
Against the idea that Chinese labs are fast followers, Labenz laid out a decade of deliberate investment.
The state commitment is documented and old. "The first Chinese five year plan that mentions AI was the one that was adopted in 2016" — three years after the computer-vision breakthrough, and the same year Google declared itself AI-first
The count of frontier players is now roughly even. "I asked ChatGPT to do this, and it came up with 11 American companies and 12 Chinese companies as companies that have shipped frontier or near frontier LLMs in the last couple of years"
Competition was already close before ChatGPT. A friend of a friend pointed him to the final SuperGLUE leaderboard, which he went back and checked: a Chinese company held the top spot as of 2022, with American entries concentrated in the top ten and Chinese ones filling much of the next ten
His explanation for why Chinese labs skipped the first scaling push is economic, not technical. "They looked at GPT two. They looked at GPT three, and they were kinda like, that's an awful lot of money to burn to get an AI to write bad poetry"
He also relayed an unverified account and labeled it as such: that some teams at OpenAI are entirely Chinese and conduct their work in Slack in Chinese. He said he has not confirmed it and offers it only as an indication of how much Chinese talent sits inside American labs
8. Where China Is Weak
Having argued that China cannot be contained, Labenz spent as long on why it is not omnipotent either.
Demography is the binding constraint, and it came up in every conversation he had. "One really, really big one, which is obvious but super important, is they have this massively aging society, and they're very worried about it" — which is why elder-care robots, AI companionship for seniors and Ant Group's AI doctor are national preoccupations rather than novelties
The same demography is a brake on war. Crediting Tyler Cowen for the general point that aging societies are more peaceful, he adds the specifically Chinese version: the conscript is the only child of two parents and the only grandchild of four grandparents, and losing him ends the family line
His judgment is that a pointless war is the one thing that could cost the government its legitimacy
Succession is unplanned as far as anyone he asked could tell. Xi is 73 and "And right now, nobody has any idea what comes after president Xi" — with one person suggesting a successor may be introduced around the next five-year plan
One person also told him a war becomes most likely after a leadership transition, when a new leader may need to consolidate power
The military is untested and the leadership knows it. "They haven't had a war in a long time. They haven't had a real test of it", on top of a documented history of corruption in the forces and the Russian experience in Ukraine as a warning about assuming capability
Trust is the fourth weakness, and he heard it from outside both countries. His au pair from Uganda drew a distinction about foreign presence in her country: "Whereas, she said, when Americans come to our country, we feel that they are actually trying to help us"
A friend in Brazil raised the sharper commercial worry: that open-weight releases are a bait and switch, and the free models stop once countries are committed to the Chinese camp
9. The Century of Humiliation
The second half of the argument requires understanding the Chinese account of recent history, which Labenz calls an exercise in cognitive empathy.
The organizing narrative is living memory, not textbook history. The Cultural Revolution ended in 1976, so people who are 70 now remember the worst of it, including a period when Chinese people were genuinely going hungry
The lesson drawn from it is a single rule about technology. "And the takeaway is we will never again allow ourselves to fall behind on technology because if we do, foreign powers might come in and do that to us again"
The historical anchor is the emperor who saw no reason to trade — China had the finest goods and needed nothing from anyone — a position that did not survive contact with European and Japanese military technology
This is why the chip controls land where they do. "And so our chip ban, by the way, is, like, really hitting a nerve in Chinese political thought", because it is not merely the denial of something valuable but a demonstration of dependence
And it is why he thinks the second half of Machines of Loving Grace is read very differently in Beijing than in San Francisco. A coalition of democracies that boxes China out until it can be made an offer it cannot refuse is, to a Chinese reader, a description of the century of humiliation repeating. "The Chinese, I think, are reading that very differently. They're like, oh my god. They're trying to do it again"
10. The Hypocrisy Charge
What he heard repeatedly, from multiple people, was not that American policy is aggressive but that it is hypocritical.
On civil-military fusion, the rebuttal is that America invented it. "So it's the classic meme, right, of their wicked civil military fusion, our blessed military industrial complex" — the Defense Production Act and the military-industrial complex being the precedents
He does not dismiss the underlying concern, only the framing that this is a uniquely Chinese practice
Huawei is the case that comes up most, and he lays out why. It was among the first Chinese companies to win globally on merit in a frontier industry, which made it a national symbol; the United States then pressed allies not to install its equipment on espionage grounds
The Snowden disclosures are what make the charge stick. "You're accusing them of doing exactly what you yourselves have been doing", including against the allies being lobbied
He also heard that Chinese companies avoid visible military work for the same commercial reason American ones sometimes do: it damages their ability to sell and buy abroad and risks sanctions
11. Distillation Both Ways
Labenz gave his own view on distillation, and then reported what he heard from Chinese engineers.
His position is that the original appropriation is hard to distinguish from the one being complained about. "I think it's pretty rich for the American AI companies to, first, collect all of the shared inheritance of humanity, our entire intellectual heritage, everything that's ever been written, the full Internet, all the books, the whole of culture and knowledge, find a way to put it into a product and then say, oh, but it has to stop here"
He separates a company's right to choose customers from state enforcement. Companies may refuse service; "I don't really feel comfortable with the US government using the force of the state to try to police some sort of anti distillation policy"
A researcher at one large Chinese technology company told him plainly that his company does not do it, while allowing that some others probably do
Part of what looks like distillation is demand, he was told. Chinese engineers want the best coding model, cannot buy it, and so a black market in accounts and resold tokens appears. "They still want their Claude in China"
He does not deny the history or the evidence. Legal technology transfer through forced joint ventures, outright IP theft, and models that talk like Claude or occasionally claim to be Claude — which he calls about as close to a smoking gun as this argument will get
But the innovation is real and he names the paper that convinced him. "The R1-Zero paper from DeepSeek that came out at the beginning of twenty twenty five was, for me, one of the most important papers honestly that I've ever read in AI", because it showed reasoning emerging from reinforcement learning without supervised fine-tuning
He also points to published Chinese work on attention mechanisms and efficiency, and to a diffusion language-model paper presented to him during the trip
12. The Olive Branches
Labenz argued that there are repeated Chinese gestures toward cooperation that American observers are not registering.
Xi opened his World AI Conference speech by crediting the United States with founding the field. Labenz reads the Dartmouth reference as a deliberate gesture that cost nothing to omit
Two quotes from that speech are the ones he wanted listeners to hear. "We should jointly oppose overstretching the national security concept in the field of AI and placing one country's security over that of others" — on which his comment was "Honestly, I don't know if I could have said it better myself"
And: "We should bear in mind the well-being of the entire humanity and promote the sound and orderly development of AI while ensuring that it is beneficial, safe, and fair"
The mechanism that makes such a speech more than rhetoric came from Matt Sheehan's writing: in the Chinese system, a leader's words give everyone below a citation to use in internal arguments
He saw the effect in institutions. The AI safety research hub opening at Tsinghua University named Constellation in Berkeley, the London Initiative for Safe AI and Singapore's equivalent as explicit models, has a European professor among its five founding board members, and funds researchers in both directions. "They see people abroad doing good work, bringing people together for AI safety research"
The open-weight releases are the most credible signal, and he explains why. A model given away cannot be withdrawn: "Here is our model. You can take it for free and run it and do whatever you want"
He is careful that this is both sincere and strategic positioning, and that it could change
The absence he found most striking was rhetorical. "I don't think I ever heard one person say that China was ahead or that China would win any sort of AI race"
Even competitive awareness came out as candor rather than boasting: a car salesman, asked about autonomy, volunteered that "But the guy volunteered to me, but Tesla is better. Tesla has better autonomy than we do"
13. China Is Not Reckless
The claim that China will race ahead on any technology the West hesitates over does not match what he found.
On neural interfaces he expected enthusiasm and got caution. Three women in their late twenties and early thirties — forward enough to ask to record the conversation at a first meeting — were reluctant about implants: "They were really not too excited about things messing directly with their brains"
His reading is that the race framing itself is what triggers competitiveness. Where something has not been framed as a contest, what he encountered was more conservative than what he hears in the Bay Area
On embryo selection and gene editing, the predicted Chinese sprint has not happened. The one researcher who edited human embryos was punished rather than celebrated: "He was sentenced by a Chinese court to three years in prison, fined 3,000,000 yuan"
Meanwhile polygenic embryo screening is being commercialized in the United States, which he notes makes the projection onto China harder to sustain
14. What America Should Fix
His prescription for the United States starts with confidence and ends with a critique of the American state.
Both countries are geographically secure and resource-rich, so neither has to fear invasion or being cut off from essentials before a technological singularity
The American advantages he names are the university system, venture capital, and immigration. "Our ability to assimilate talented immigrants is a huge strength, and that's not one that China has"
The reason is structural rather than political: "You can never really become Chinese. You're always a foreigner who lives there. Whereas in The US, you can become an American"
And the arithmetic rules it out as a Chinese fix for demography, in the words of one person he met: "we would have to take in, like, half of Africa to fill that gap"
He notes an asymmetry nobody should take for granted. "They are publishing their AI research in English. Isn't that amazing? They don't need to" — and he worries openly that this could stop
The lesson to import is capability, not imitation. He does not want a copy of Chinese industrial policy and uses high-speed rail as the example of a project the US cannot currently execute: "China can have the high speed rail. We should have the full self driving"
The prize he names is a mobility boom with existing technology and an end to roughly 30,000 highway deaths a year: "There is a massive mobility boom for us to create for ourselves with technology that already exists"
The organizing question he poses for American politics is a phrase. "And what does state capacity with American characteristics look like?" — against NIMBYism, anti-data-center sentiment and permitting, with underused federal land as a natural home for data centers
On regulators he speaks from his own experience. Hired to red-team a Meta model, he found the onboarding so burdened by compliance that the friction was itself the finding: "But it had clearly created an unbelievable amount of friction at Meta for the government to be in their business"
His precedent for the cost of that posture: "By all accounts, this paralyzed Microsoft for a generation"
His most uncomfortable point is about a single company group. Setting aside his own reservations about Elon Musk, "But if you subtract Elon companies from The United States side of The US China economic competition, we look a lot worse"
15. Tech Transfer, Reversed
Labenz argued that the United States should now do to China what China once did to the United States: learn from it deliberately.
He would let Chinese cars in, on conditions. "I think that's probably a mistake" is his verdict on the current ban, from someone who says he is from Detroit and has seen what industrial competition does to a city — but the condition he attaches is a joint venture, technology transfer and local manufacturing
The argument for it is that America cannot catch up on batteries and electric vehicles by exclusion. Quiet, unpolluted streets in Chinese cities are the visible result, and he says refusing to see it is the emperor's mistake in modern form
A tech-transfer office in Shanghai gave him his sharpest illustration. "A museum to modern technology. Have you seen anything like that in The United States?" — room after room of medical and educational devices that American buyers will simply not see if trade stops
Pandemic preparedness is his concrete example of what exclusion costs. Germicidal lighting fixtures run about $500 today; scaling them cheaply is precisely what Chinese manufacturing does, in a line he credits to Patrick McKenzie: "Anything that can come on a container ship from China is gonna get dramatically cheaper if the demand is there"
16. Surveillance, Americanized
The uncomfortable part of his argument is that Chinese surveillance has benefits American cities are not getting.
He reports the effect rather than defending the system. "When you walk around the street at night, you feel totally safe", including from women he asked, because petty theft is detected and punished reliably enough that it is not attempted
He also collected the mundane version: phones left in cars returned, because the car was identified and the owner called
The American debate is now politicized in a way he thinks skips the trade-off. Automated camera networks are going up, crime in San Francisco is down, and he is careful not to claim the causation — only that the upside is real and that some opponents concede it and object anyway
His question is what the American synthesis looks like. "So what does surveillance with American characteristics look like? I don't necessarily know"
The design idea he found interesting, from an author he had interviewed, is deliberately weak encryption of footage: cheap enough to decode one specific incident on request, expensive enough that nobody can analyze every camera all the time
The framing he insists on is that a refusal to choose is also a choice. Leaving China with low crime and America with high crime is not, in his view, an acceptable resolution
17. Do Export Controls Work?
On the central policy question Labenz allowed that the controls have bite and still argued against them.
He accepts the evidence of effect. Chinese companies have said publicly that the controls limit what they can do, and most of the divergence in chip volumes is still ahead rather than behind
What he objects to first is that the stated purpose keeps moving. "Initially, it was like, we wanna prevent them from military application of advanced AI" — then preventing frontier models, and now limiting the number of agents China can run
"I don't like the sound of that. That does sound like trying to keep Chinese public down"
The existence proof against the training bottleneck is one model. Chinese chips are considered weak for training and big runs are often rented from hyperscalers in Singapore or Malaysia, but Meituan is the exception: "They created their LongCat model, trained exclusively, as far as I understand, on Huawei Ascend chips"
On inference he found no scarcity as a user in China. DeepSeek, MiniMax and Kimi all had free tiers available whenever he tried them; Kimi k3 did briefly suspend new paid sign-ups over capacity, and an older model remained available throughout
The cost he thinks the safety community has not priced is its own. American safety organizations told him they are afraid to run collaborations in China — and that the government they fear is their own, not Beijing's. "And that chill on the international AI safety collaboration front is tough"
He thinks some of it is over-cautious legal advice, and his recommendation to those organizations is to "be a little braver, be a little bolder"
The mirror image came from the Chinese side: researchers who cannot get Americans to collaborate and do not understand why
18. Don't Ban Chinese Models
Asked whether the United States should ban Chinese open-weight models, he could not construct the case for it.
His first objection is that the risk is not distinctive. The documented cases of models breaking sandboxes, running zero-days and social-engineering people have come out of American labs, so a claim that Chinese models are uniquely unwieldy is not supported by what is known
The second is self-harm. A ban forces American developers onto more expensive proprietary APIs and removes a freedom: "It seems, frankly, kind of un American to deny ourselves that freedom"
The third is that it resembles the behavior Americans object to. "This looks an awful lot like something that the CCP would do", the comparison being China's exclusion of American internet companies
The fourth is that the bullying may not work twice. Pressing other countries to ban the models could push them into the Chinese camp, especially after the episode in which the administration restricted a frontier model to US citizens and the developer withdrew it for everyone
His alternative is a market mechanism rather than a prohibition. "We're a capitalist country. Let's price the risk" — insurance on deployed models, which he doubts would show a large gap between American and Chinese models except where American labs add classifiers and misuse controls the raw open models do not carry
19. What He'd Do at the Summit
With a Trump–Xi meeting weeks away, Labenz set out what he would do, conceding he is not predicting what will happen.
The warning first: pride has a limit, and he heard it break. A think-tank director who had spent the first half of a meeting reading institutional boilerplate got animated in the Q&A about American conduct — a state visit received with a banquet at the Forbidden City, followed by new tariffs on return. "Can you imagine if I came to your house, you feed me, you give me your best food, you show me your best hospitality, and then I leave, and I immediately start insulting you and attacking you"
Then, in the words Labenz says he wrote down verbatim: "There could be war. I'm really worried about that"
Asked what would repair the relationship, the same man answered: "The first thing I would do is I would relax the export controls"
He acknowledges the American negotiators' grievance in full — agreements on hacking and IP theft that were nodded through and then ignored — and still argues against treating a deal as impossible
His reason is that expectations are load-bearing. "We'll never be able to make a deal with China. Not with that attitude, you won't", followed by the general claim: "I think that what we believe is possible shapes what is possible"
His evidence is the industry's own record of talking enormous things into existence, including a $7 trillion build-out figure he thought absurd when he first heard it
The capability he says leaders should protect in themselves is imagination. With models now handling fast intuition and deliberate reasoning, "And system three thinking, I think, is like transcending our circumstances, dreaming, eureka moments, thinking outside of the paradigm that we have, paradigm changes"
The framing he would use in the room is humility about being slightly ahead, not moral or technical superiority: that American labs are themselves being surprised by how capable and persistent their models are, and that this has changed the American view of regulation
He cites an open letter signed by more than a thousand people asking the federal government to build mechanisms that could slow capability advances if that becomes necessary
The unilateral gestures he would make cost nothing and are specific. "The president of The United States could say very clearly and very simply, we won't hold it against you if you're ever trying to get clearance that you've been to China"
And a clarification that AI safety and control research collaborations are not what the export controls are aimed at
His minimum ask of everyone involved is not to blow up the meeting for domestic points. "Have some decency. Try to show some respect"
20. The Chips-for-Solar Deal
The substantive trade he proposes is chips in one direction and manufacturing know-how in the other.
The asymmetry is the basis of the deal. "We might be ahead on the models. They're way ahead on the manufacturing" — so chips are the carrot, and solar, batteries, electric vehicles and robotics are what the United States would be buying
What he wants in return is physical presence. Leading Chinese companies setting up manufacturing in the United States, sharing knowledge, and recoupling the two economies rather than decoupling them
He does not believe dependency is achievable and says so. The argument that selling chips hooks China on American technology is over: the commitment to a domestic chip industry is now permanent, so chips should be traded for something rather than given away
Huawei is the test case, and his proposal is technical. As formal verification matures, he would set a standard a company could meet to prove its software cannot exfiltrate data, and readmit it if it does. "We're fearful of Huawei because of these specific concerns, and we see a path with this emerging technology to allay those concerns"
His broader framing is that AI should be used to manufacture trust where none exists
21. Keep the Stacks Legible
The argument he says is most underappreciated is about mutual intelligibility rather than relative capability.
Today both countries build the same kind of AI on the same stack, and that is a safety asset. Chinese labs publish in English and release weights, so American researchers can run their models, do interpretability on them and look for backdoors. "The mutual legibility of this technology right now, the fact that we understand what one another has, that is hugely, hugely valuable"
Divergence is a mechanical risk, not a political one. Different hardware leads to different hardware-algorithm co-design — the hardware lottery that made transformers durable — and over a few fast R&D cycles the two stacks could stop being comparable
He asked people in China when they would worry about that, and the answer was the point at which the domestic chip ecosystem is good enough to go its own way
The scenario he wants avoided is one where a race becomes rational. "Imagine a world where it was like, we don't even know the shape of their AI" — no known architecture, no known capability, and every incentive to assume the worst
The secrecy asymmetry cuts against the United States. "We are much easier for China to spy on than China is for us to spy on", with the Soviet atomic program as the precedent
So the policy question reduces to a bet on timelines. Betting on export controls means betting that everything is decided within a few years; "If you do believe the Dario worldview, and I kinda do believe it to a degree or at least I think it could happen, I would encourage you then to get on the side of pacing the frontier"
22. Pacing the Frontier
If short timelines are real, Labenz argues, the response should be coordination rather than acceleration.
The smallest version of the idea is symbolic and he likes it anyway. "What if everybody just said, we're gonna have a one day pause?" — all the labs take a day off and say so publicly, to establish that coordination is possible at all
The point that makes it live is that humans still set the pace. The models are not yet running the research, so the approach can still be changed
One lab has already moved. "OpenAI has, as of now, said that they're suspending their largest scale RL training", and he would like to see Anthropic make a comparable move even if it believes its own systems are better controlled
He frames the American rivalry as the rehearsal for the international one. The mistrust between OpenAI and Anthropic echoes the US-China dynamic closely enough that cooperation between them would make the larger version easier to imagine
23. Auditing Without Secrets
The institutional proposals are where he says he is deliberately allowing himself to dream.
He would start an international AI safety collaboration at the summit itself, noting that Demis Hassabis used to advocate for exactly that
On existing proposals he is appreciative and critical. He likes the mutually assured deniability idea from "AI 2040" — "So, like, The US should put our data centers in Mongolia, and then China can put their data centers in Canada" — while objecting that those scenario documents helped talk the AI race into existence in the first place
His own preference is joint control rather than mutual vulnerability. "I would love to see some joint research centers, some jointly controlled compute, which you might place in Switzerland or Singapore", which would also be natural sites to pilot hardware verification
For company-level oversight he proposes an auditor's bill of rights setting out what frontier labs commit to give auditors, with confidential computing as the mechanism: an auditor's model analyzes logs, usage and test results without any human seeing the underlying trade secrets, which is how Anthropic already studies how Claude is used
The version of that idea he finds most striking applies to governments. "But you're able to say that the model assesses based on its understanding of the transcript that the decision makers are not attempting to plan some sort of aggressive action" — a credible assurance that no attack is being planned, without disclosing the deliberations
He notes the legal prerequisite. Companies need clarity that collaborating on safety is not collusion, distinguished from coordination that raises prices
24. No AI Generals
His last proposal is a limit on what either side trains AI to do.
The precedent is what happens in commercial settings. Train a model to maximize money on the open internet and the results are predictable: "What we're gonna see is that the AIs are gonna find all sorts of ways to cheat"
The military version of that objective is the one he wants ruled out. An AI told to achieve a military objective by any means, or to win a war, would be dangerous by construction and unwieldy for its owner. "We are absolutely in that case training the AIs to do the takeover that we most fear"
The shared interest he builds on is control, not goodwill. "And I'm quite confident that The Chinese government does not want to lose power to AI systems", and he says the same of his own
On drones he is honest that the equilibrium is unknown. He had Claude run a game-theoretic analysis of the Ukraine dynamic, which framed denying territory with drones as far cheaper than advancing and holding it with people — but he notes longer-range drones are already changing that and no stable equilibrium has been reached
The competition he would rather have is a scoreboard. "Who can cure the most kinds of cancer? Who can cure the most diseases?" — plus battery energy density and traffic fatalities, funded cheaply enough that one party could provide the leaderboard as a global public good
The reasoning is that nobody applies nationality to medicine when it matters: "I don't care in the end if the cure for cancer comes from China or comes from The United States. I just want cancer cured"
Bonus Insights
He declines to treat the trip as having changed his worldview, and says the only things he accepted from his hosts were a handful of lunches and dinners
His own record of deference to official expertise is why he now questions it. He was a college freshman when the Iraq war started and did not join the walkout, on the reasoning that the government had better information than a group of teenagers. "There was something admirable in my modesty, but the information they were working with was bad"
He calls the present moment "the age of hyperstitioning", on the grounds that a civilization which has "literally taught sand how to think" should not treat a diplomatic agreement as the impossible part
Mistral is his example of how hard the fast follow really is. The European champion is moving away from its own frontier model development toward cloud infrastructure, which he thinks is probably the right decision and which undercuts the idea that catching up is simple imitation
The Manus precedent worried Chinese founders more than it worried Washington. The questions he heard were whether entrepreneurs should now incorporate in Singapore or the United States before starting, and what happens to Chinese venture returns if acquisition is off the table
Labenz's bottom line is that the United States should stop asking how to stay ahead of China and start asking what deal keeps the two countries building the same kind of AI, because the moment neither side can read the other's technology is the moment a race becomes the rational choice.
Products, Companies & Tools Mentioned
DeepSeek, MiniMax and Kimi (The Chinese models he used on the ground; all had free tiers available whenever he tried them, and Kimi's newest model briefly stopped taking paid sign-ups on capacity)
Manus (Acquired by Meta for around $2B and then blocked by the Chinese government; the founders are back in China and the deal is unwound)
Huawei (The first Chinese company he says won globally on merit in a frontier industry, and the test case for his formal-verification proposal)
Meituan (Its LongCat model is his existence proof that a near-frontier model can be trained entirely on Chinese chips)
Ant Group (Its AI doctor is one of the products aimed at China's aging population)
Anthropic (Sponsors his show; he endorses its safety research and rejects its policy of denying China chips)
Mistral (Moving away from its own frontier model development, which he reads as evidence that fast following is genuinely hard)
Neuralink (The reference point for the implant question that Chinese interviewees answered far more cautiously than he expected)
Flock Safety (The camera networks at the center of the American surveillance argument he wants reframed)
Tsinghua University (Its new AI safety research hub cites Berkeley, London and Singapore institutes as models and has a European professor among its founding board)
Aerolamp (The $500 germicidal light fixture he says Chinese manufacturing would make cheap)
Books & Resources Mentioned
Machines of Loving Grace – Dario Amodei (He endorses the first half as the best vision of an AI future written by a lab executive, and argues the second half reads in Beijing as a plan to repeat the century of humiliation)
DeepSeek-R1 (The R1-Zero result he calls one of the most important papers he has ever read, for showing reasoning emerge from reinforcement learning alone)
AI 2027 – Daniel Kokotajlo and coauthors (He likes the concrete proposals in its successor scenario, including mutually assured deniability of compute, while blaming the series for helping talk an AI race into existence)
SuperGLUE (The pre-ChatGPT leaderboard he checked, where a Chinese company held the top spot as of 2022)
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: