The Goal Is Pax Robotica, Not Winning the AI Race

With a Trump–Xi meeting a couple of weeks away and AI governance expected on the agenda, Nathan Labenz closes his three-part *Nathan Goes to China* series with an argument, not a travelogue. The first episode covered the trip’s tech setup and his on-the-ground use of DeepSeek, MiniMax, and Kimi. The second, which he calls the most important of the three, is fact-based reporting on “AI safety with Chinese characteristics”: company deployment systems, research culture, and regulation that, in his telling, look more like the American conversation than Silicon Valley cope allows. This installment is the opinion piece: what the United States should want from China, what it should learn, and how the two sides might navigate accelerating AI together.

What frustrates him about U.S. China talk is that it skips the goal. People say they want to win, he notes, without saying what winning is, what losing would mean, or whether a win condition even exists. His own target is a Pax Machina, or perhaps a Pax Robotica: mutual prosperity in which both countries, and the world, share the benefits of the AI age. That means robots in factories and fields, robots picking fruit, robots in homes doing laundry and meals, and “unlimited access to expertise” for anyone who wants a more leisured life. It also means avoiding the downside list he recites early: AI-caused pandemics, an AI arms race, an AI cold war, and a new sword of Damocles.

The nuclear analogy is not decoration. Labenz says humanity still lives under roughly 10,000 nuclear weapons, many on hair-trigger alert, a stockpile far beyond what deterrence requires and a standing annual risk of accident or miscalculation. He thinks creating that overhang was a mistake, and he does not want an AI-powered version of the same trap. Anthropic sponsors the show and has argued for denying China chips so America can keep a lead; he addresses the tension head-on. He praises the first half of Dario Amodei’s *Machines of Loving Grace*—disease cures, mental-health gains, a century of progress in five years—and cites his own son’s recovery from cancer in six or seven months as a fruit of technology. He then restates a 2017 Amodei line he once retweeted: “Technology races can raise the risk of safety catastrophes.” The more U.S.–China AI tension is ratcheted up, Labenz argues, the more likely the “nuclear outcome”: militarization and a hanging sword, with civilian benefits under-delivered, like a world of warheads and too few power plants.

Labenz was in China about two weeks, paid his own way, and accepted only scattered lunches and dinners from hosts. He still feels obliged to give a net assessment because he keeps hearing, from people he considers smart, that China can copy and modify but cannot truly innovate. He hears the same ladder of denial about AI itself: models will never reason, then never discover knowledge, then never deliver an Einstein-caliber paradigm shift. China, he says, has arrived as a world power; any fantasy of containing its rise “sailed a long time ago.” For most of the last several thousand years it was a technological leader. The era in which he and Amodei grew up—China down, America the sole superpower—was the anomaly, produced by Western and Japanese incursions and by post-revolution disasters, including periods when people were literally starving.

A Shanghai pedicure is his working-class exhibit. A friend who had lived there decades earlier told him to get a haircut or other bodily service; one late night Labenz walked into a random shop near closing, paid about $12 for a 45-minute treatment, and compared phones through a translation app with a technician in his mid-twenties. The young man had an iPhone 17. Those phones are not cheaper in China, and flagship Xiaomi and Huawei devices also push toward a thousand dollars. Labenz’s guess is concentrated family wealth under the one-child pattern: four grandparents, two parents, maybe several apartments, a gifted or rented unit, and wages that function partly as pocket money. Even at perhaps $5 an hour, disposable income can still buy a flagship phone. The anecdote is not a GDP series; it is his insistence that urban working-class life is not the poverty Westerners of his generation still carry as a background bias.

The AI timeline, in his telling, matches the civilizational one. China’s first five-year plan to mention AI was adopted in 2016, three years after AlexNet and the same year Sundar Pichai declared Google AI-first. Ask ChatGPT for a count of firms that have shipped frontier or near-frontier language models in the last couple of years, he says, and you get roughly 11 American companies and 12 Chinese ones. A person of “high LLM taste” can still claim only two true frontier labs, both American. The talent bench is another matter: a large share of researchers at U.S. labs are Chinese or Chinese-American, and Labenz repeats an unverified whisper-network claim that some OpenAI teams even work in Chinese on Slack. Nobody at the Chinese labs is a symmetric inbound flow of Americans.

Headline dollar GDP still comforts Americans: about $30 trillion for the United States and $20 trillion for China, a 50% U.S. lead. Switch to purchasing-power parity and Labenz’s figures reverse: China at about $45 trillion, the United States still at $30 trillion, a 50% Chinese lead. Per capita, America remains richer either way. After even a short stay, he thinks PPP is closer to lived reality. A nice dinner for two in Shanghai can cost under $10; that is not every meal, but it is not available in New York, San Francisco, or even Detroit. Money goes further domestically, and he would rather say the Chinese economy is 50% bigger in a fundamental sense than cling to the dollar ranking.

The Beijing–Shanghai high-speed train is the infrastructure version of the same point. His second-class ticket was about $60, more spacious than a plane, leaving on the advertised minute at something like 200 miles an hour, with stops that still serve cities Westerners have never heard of. Passengers ordered lunch to an upcoming station, specifying their car, for a five-minute platform handoff. Airport shops, from what he saw, charged street prices. Formal posted prices, no haggling, no bait-and-switch, and no tip culture made the consumer surface predictable. The feeling he wants Americans to sit with is not that they are still unambiguously the richest and best.

He is equally wary of the idea that the Chinese government is illegitimate and waiting to be pushed over. The United States recently made a version of that mistake about Iran, he says: real discontent did not produce the uprising some expected after a strike on the leadership. Censorship and thinner political freedom are real, and Tiananmen is history. But Amodei’s “eternal 1991”—the childhood high of the Soviet collapse, the “end of history,” America as city on a hill—is, in Labenz’s view, a bad map of China today. If anything, Chinese society is living *its* 1991: two generations of delivery from a historic low, restored prosperity and respect, and a mandate of heaven that still means who delivers is who is legitimate. Daily life is free in the ordinary ways—jobs, hobbies, bars, music—with no-go zones. Older people remember the Cultural Revolution; younger people hear it in family stories and street signage. Stability, safety, and peace are not small gifts. “Times have never been better in any living memory,” he says, and picking a fight with the government will not peel the public onto America’s side.

The AI-specific version of the escape fantasy is that the United States should hand every Chinese researcher a green card and watch them all come. Labenz calls that a total fantasy rooted in the assumption that life in China is something people want to flee. Historically there was truth to it; today many who study in America choose to go back. He cites the founder of Moonshot, who did a PhD in the United States and returned to start the company. Attracting talent remains, in his view, perhaps America’s number-one unique strength, and he wants more of it, Chinese or otherwise. What he rejects is the arrogance that top Chinese researchers are “just dying to leave.” Home, family, purchasing power, and a good life are there. The Communist Party also would not necessarily let them all leave even if Washington opened the door.

Manus is his exhibit of the other edge of that control. The agent app’s demo, when he first tried it, struck him as a step ahead of OpenAI’s Operator. Meta agreed to buy the company for what he recalls as about $2 billion; then Beijing stepped in, called it critical technology, and, as he understands it, fully unwound the deal. Founders stayed in China. People he spoke with there called it an own goal: Manus was a strong harness, not a foundation model and not chips, and blocking a venture-scale exit could push entrepreneurs to incorporate in Singapore or the United States before they start, while chilling Chinese venture funding. He agrees it looks like a mistake. It is also a reminder that the state can trap talent and companies at home.

The catch-up narrative around ChatGPT, he argues, starts the clock at the wrong moment. If fast-following were easy, other countries would be doing it; even Mistral, Europe’s champion, is drifting toward being more of a neo-cloud than a model lab. A friend of a friend in Shanghai sent him back to SuperGLUE, the pre-ChatGPT natural-language benchmark. On the final leaderboard Labenz checked, a Chinese company sat at number one; American firms were denser in the top ten, Chinese firms thick in the 10-to-20 band. That was already the payoff of long strategic investment as of 2022. Why the LLM lag? The same contact’s line was that Chinese companies looked at GPT-2 and GPT-3 and asked why they should burn a fortune so a model could “write bad poetry.” Instruction following, assistants, reasoning, and agents made the use case obvious, and the capital followed. GPT-3, Labenz notes as a former user, really was not that useful without fine-tuning. Missing the first scaling surge was a reasonable commercial miss, not proof of a thin ecosystem. There is, he concludes, “no keeping them down.”

People who wake up to Chinese strength, Labenz warns, sometimes flip into the opposite cope: high-speed rail as proof that Beijing is omnipotent. He spent time on the weaknesses. Aging came up constantly. Ant Group’s AI doctor, companion products for lonely seniors, and care robots are all framed as answers to a one-child pyramid: one grandchild, four grandparents, two parents. He credits Tyler Cowen with the point that aging societies tend to be more peaceful. In China the logic is harsher: sending the only grandchild to a pointless war is the one move that might produce sustained street protest and strip the mandate of heaven. State capacity that can lay rail is not the same as a blank check to mobilize the largest army in history.

Succession is the second hole. Xi Jinping is 73, not ancient, and could remain sharp for a decade—or not; Labenz notes America’s own recent experience with geriatric presidents. He asked repeatedly what comes after Xi. Nobody he spoke to had any idea. There is no public successor. One person guessed a candidate might be introduced during the next five-year plan so the public can get used to the name. Labenz calls the missing plan a blemish and a blind spot: measures that buy stability now push risk into the future. Another interlocutor said that if a war ever came, it would be most likely after a leadership transition, when a new president might want a crisis to consolidate factions in an opaque system of purges.

The military’s true strength is unknown, including to Beijing, in the same way Russia discovered it was not going to take Kyiv in a few days. China has not had a real war in a long time, has a history of military corruption, and would be rolling dice on Taiwan without a test. Labenz thinks that uncertainty, stacked on only-child casualties, produces reluctance, not reckless confidence. Trust is the fourth constraint. His family’s Ugandan au pair told him Chinese visitors come to profit and take the money home, while Americans come to build schools. A Brazil-based AI practitioner worried that Chinese open source is a bait-and-switch: join the camp, then the free weights stop and the API bills begin. At the retail level Labenz found posted prices and no bait-and-switch, so he thinks the tourist stereotype is outdated. At the geopolitical level he thinks the suspicion still bites. The United States, despite “flagrantly untrustworthy behavior in recent years,” still holds more trust in most of the world than China does.

The motivating national story, as Labenz tries to practice “cognitive empathy,” is the century of humiliation. The Cultural Revolution ended in 1976; people now in their seventies were young adults then and remember chaos and, at times, famine. The lesson he hears is that falling behind the technology frontier is existentially dangerous. He retells the old imperial boast that China needed nothing from trade because it already had the finest goods—an insularity that did not age well when European and Japanese militaries arrived. Chip export controls, in this frame, are not a narrow semiconductor policy. They hit the nerve of never again being dependent on a foreign power for a key to the frontier.

That is why the second half of *Machines of Loving Grace* lands so badly in Beijing. Amodei’s coalition-of-democracies idea—box out adversaries until the advantage is so large that China faces an offer it cannot refuse—is, Labenz says, exactly the scenario they are trying to prevent. A Western reader may hear an interesting speculation about superintelligence; a Chinese reader may hear “they’re trying to do it again.” Civil-military fusion talk produces a similar flinch: Americans condemn it while running a military-industrial complex and the Defense Production Act. Labenz is not naive about the Chinese state’s ability to take technology from firms. He wants Washington to notice that from the other shore it looks like hypocrisy, and that some Chinese companies share an American-style impulse to keep the military at arm’s length so they can still sell abroad.

Huawei is the wound that still organizes the elite story. After generations of climbing the value chain, it was among the first Chinese firms to win globally, on the merits, in a core technology. The United States then pressed allies to keep Huawei out of infrastructure over spying fears. Labenz thinks the risk is real and should be taken seriously. The Chinese counter is Snowden and reports that Washington wiretapped even the German chancellor: you accuse us of what you did, including to your allies. The chip ban reads as the same movie, queued before agents can even book a plane ticket reliably. On distillation he is more sympathetic to the Global South than to American labs: those labs ingested “the shared inheritance of humanity,” then objected when others trained on their outputs. He would let companies refuse customers they believe are distilling; he is uncomfortable with the U.S. government policing an anti-distillation line. At one large Chinese tech firm, a researcher presenting diffusion-language-model work said *his* company was not distilling, and that a lot of the “illicit Anthropic accounts” were just engineers who wanted the best coding model and could not buy it. Labenz does not deny IP theft or models that talk like Claude or even claim to be Claude. He also credits DeepSeek’s early-2025 RL paper, which showed chain-of-thought metacognition from pure RL without supervised fine-tuning, as one of the most important papers he has read—open, academic, and Chinese.

Both sides, Labenz says, have low trust and legitimate grievances, including American firms that were hacked and then watched a Chinese competitor skip the R&D. He still wants Washington to notice the olive branches. At WAIC, Xi Jinping opened by crediting Dartmouth College as the birthplace of AI—a line China watchers flagged as a gift Xi did not have to give. Xi also said, in Labenz’s quotation, that countries should jointly oppose “overstretching the national security concept in the field of AI” and putting one country’s security over others, and that development should keep in mind “the well-being of the entire humanity.” Westerners will call it rhetoric. Labenz’s reply is that you have to start somewhere, and that OpenAI’s original “benefit all humanity” language has grown quieter as the company has grown more patriotic. Analyst Matt Sheehan’s point, which Labenz repeats, is that in the Chinese system a leader’s speech is a citation the rest of the bureaucracy can use.

He heard that cosmopolitan language everywhere, not only on stage. In Beijing he attended the opening of an AI-safety research hub at Tsinghua University that named Constellation in Berkeley, LISA in London, and SASH in Singapore as explicit inspirations, pays for foreigners to study there and for its students to go abroad, and includes a European professor among five founding board members. What he did *not* hear, remarkably by Bay Area standards, was anyone saying China was ahead or would win an AI race. He barely heard AI framed as a race at all. It was prosperity, abundance, and care for an aging population. At a mall car dealership, a salesman volunteered that Tesla’s autonomy was better. Labenz asks Americans to imagine the reverse: Chinese leaders openly describing a plan to force the United States into their alliance. He has not heard a peep of that fantasy. What he hears instead is a bid to provide global public goods, with open-weight releases as a credible, hard-to-claw-back gift—though he grants the posture could change.

On “they will race through every taboo,” his anecdotes run the other way. Three women in their late twenties or early thirties, tech-forward enough to ask if they could record a first meeting, were cool on neural implants: that is my brain, they said. He himself is more excited than scared, especially given Neuralink’s work with paralyzed patients; they were not. Human cloning and extreme embryo selection are another Western projection. He Jiankui’s 2018 HIV-resistance edits drew a three-year prison sentence and a 3 million yuan fine. Labenz sees U.S. startups doing embryo screening and polygenic scoring, and no army of cloned “super babies” in China. The pattern he wants internalized is caution when a technology is not framed as a race, and alarm bells when Washington frames AI as one they must not lose.

Labenz’s first counsel to Americans is to remember their own strength. Two oceans, friendly neighbors, and most of the resources a great power needs make invasion a remote fear; China has tundra, desert, ocean, and the Himalayas. The United States still has the world’s best universities and, he says, the most dynamic business culture and the golden goose of venture-backed startups. People in China, as elsewhere, still say they want to build a Silicon Valley. Assimilating immigrants is a non-replicable advantage: you can become an American, and you cannot really become Chinese. China also cannot immigrant its way out of aging. One interlocutor sketched a fall from 1.4 billion toward 800 million with a birth rate just over one, and joked they would need “half of Africa” to fill the gap. They may try to recruit Ivy League professors. They will not become an immigrant nation. English remains a U.S. asset too: Chinese labs still publish AI research in English, and Labenz worries they might someday stop if they conclude America is only trying to keep them down.

The lesson he wants from China’s hardships, and from COVID shortages, is not to fall behind the frontier or cede batteries, solar, and other critical industries. That will require decisions that challenge a pure invisible-hand mood, especially because American industrial policy has been weak, leaders are often lawyers not engineers, and federalism, NIMBYism, and special interests gum the works. He does not want to copy China’s high-speed rail. California’s project is a “massive failure” next to China’s network, and he thinks the United States cannot currently chew that bite. A more American target is full self-driving: productive commutes and an end to the roughly 30,000 highway deaths a year, using technology that already exists. He would rather site data centers on unused federal land than daydream about parks. Regulators, in his view, should protect the public without turning every company process into a Byzantine onboarding ritual; a Llama 3 red-team invite from Meta two years earlier drowned him in privacy trainings for work on a language model. Chasing Elon Musk out of California, he says, was a “big oops,” and subtracting Musk companies from the U.S. side of the ledger on space, EVs, batteries, solar, neural implants, and AI makes the comparison look much worse.

He wants tech transfer *from* China, the reverse of the old joint-venture bargains. Banning Chinese cars, he argues as someone from Detroit, is probably a mistake if the price of catching up on batteries and EVs is letting Chinese firms manufacture in America and share knowledge. Streets in the top cities he visited were strikingly quiet because most cars were already electric. A Shanghai “museum of modern technology”—room after room of medical and educational devices—left him thinking that refusing to trade means missing a growing share of the future. Germicidal light is his pandemic example: if Aerolamp units cost $500, China is who scales them down, and as Patrick McKenzie put it, anything that can arrive on a container ship gets dramatically cheaper if demand is there. Even surveillance, he argues, has an upside he felt walking at night. The American question is not whether to copy the CCP’s cameras but what “surveillance with American characteristics” looks like. He cites a *Liquid Rain* idea of weakly encrypting footage so police can decode a requested clip at real cost without running a cheap, always-on national search.

On chips, Labenz allows several things to be true at once. Chinese firms have said the controls bind, so they are not theater, and most of the volume gap may still lie in future buildout. The public rationale has slid, in his telling, from blocking military uses, to blocking frontier models, to limiting how many AI agents China can run—an aim that sounds to him like keeping Chinese businesses and individuals down. Training is where NVIDIA still matters; many Chinese companies rent hyperscaler capacity in places like Singapore or Malaysia. Meituan’s LongCat, trained exclusively on Huawei Ascend chips as he understands it, is an existence proof that a near-frontier run can be done at home, electricity subsidies and all. For inference he found plenty of tokens: DeepSeek, MiniMax, and Kimi all had free versions he could use, and even when Kimi K3 paused new paid sign-ups, 2.6 stayed up. Hyperscalers told him they could support a startup that suddenly went vertical. The controls make sense, he argues, only if you believe a short race to godlike superintelligence will be decided first, which he takes to be Anthropic’s 2027–28 recursive-self-improvement bet. Even then he would rather “pace the frontier together.”

The underappreciated cost, especially among safety people, is collaboration. American AI-safety organizations told him, over and over, that they have to be extremely careful because anything they do with China could be painted as an export-control violation—and that the hammer they fear is Washington, not Beijing. Lawyers give the narrowest “definitely okay” box and an expansive “might cause trouble” box; groups stay inside the first. Labenz’s advice is to be a little braver. Chinese counterparts, meanwhile, cannot understand why Americans refuse joint research. The chill, in his view, is a self-inflicted wound if the point of policy is a meeting of the minds.

Banning Chinese models strikes him as worse. The colorful hacks, social engineering, and unwieldy behavior in the public record have come from American companies. Denying Americans the right to run weights on their own machines, he says, is “kind of un-American,” and bullying other countries the way Washington once did on Huawei could push them into the Chinese camp—especially after an administration line that only U.S. citizens could use a service, which ended with Anthropic taking it down for everyone. If American models are safer, he would rather price the risk with insurance than copy a CCP-style information ban. A product wired to a U.S. frontier model versus Kimi K3 might not even show a huge premium, except that OpenAI and Anthropic wrap classifiers and misuse defenses around the weights, while a bare Chinese checkpoint does not. That, to him, is an American way to handle the hazard.

A mild-mannered university-affiliated AI-policy think-tank lead gave Labenz the rawest hour of the trip. After a polished slide dump, the man lit up: Trump is hosted at the Forbidden City, fed the best food, then goes home and slaps on another tariff. “There could be war. I’m really worried about that,” Labenz wrote down verbatim. Asked what would repair the relationship, the man said the first step was relaxing export controls. Labenz’s gloss is pride, face, and a culture that can “eat bitterness” longer than America can, including through lockdowns. Elite leaders may have more room to make a deal without looking weak to voters. They do not have infinite room if every new technology arrives as another attempt to keep them down. Every insult, every denial, every accusation of something Washington itself does, spends down the possibility of a later bargain.

He knows the scar tissue. American negotiators told him Chinese counterparts say little, take notes, nod at a weak yes, and then keep hacking and stealing IP that a capable state could stop if it wanted. He does not deny it. He also remembers being a college freshman who did not walk out over the Iraq War because he assumed the government knew about weapons of mass destruction—and that modesty aged poorly. “We’ll never be able to make a deal with China” is, in the age of hyperstition, a self-sealing prophecy. People taught sand to think; Sam Altman’s once-absurd-sounding $7 trillion buildout is no longer a joke. If AIs now do system-one and system-two thinking, the remaining human edge he wants from leaders is “system three”: time to dream past current constraints. Cheap shots that get a Trump–Xi summit canceled are, in that frame, an own goal.

The bargaining position he would take is not moral or even technical superiority but being slightly ahead on a shared, increasingly scary path, with more than a thousand people already signing an open letter asking the federal government to invest in slowdown mechanisms. Unilateral tension-reducers could land before any treaty: say clearly that a trip to China will not be held against security-clearance applicants, and say clearly that AI-safety and control research is allowed even if chip sales stay restricted. The big carrot is chips. China will build a domestic industry either way; the question is what America gets for selling in the meantime. Labenz wants solar, batteries, electric cars, and robotics firms to set up manufacturing in the United States and share knowledge—recoupling without one-sided dependence. For Huawei he would signal a future standard: if formal methods can prove the software never phones home, the company can compete in the American market, perhaps with R&D and plants on U.S. soil. Jensen Huang and David Sacks’s “hook them on our stack” line no longer fits; complacency on chips is over. Some remaining dependence would still be another reason not to smash Taiwan, and, more important to him, a force that keeps both sides on the same tech tree.

When Labenz asked people in China whether decoupling could fork the tech tree until neither side understood the other, the answer was “not yet”—and the worry date was when the Chinese chip ecosystem is good enough to go fully native. An LLM a few benchmark points behind is a race you can see. An unknown architecture with unknown capabilities is a reason to panic. China, he suspects, can keep secrets better than America can: the Soviet bomb came by espionage, many Chinese citizens live in the United States, and far fewer Americans live in China. Kicking China off ASML and the American stack is, past a one-to-three-year superintelligence window, a bet they will not catch up. He does not take that bet. If you *do* believe the Amodei timeline, he wants you on the side of taking a foot off the gas, even a theatrical “one-day pause” that proves coordination is possible. Humans still set the research pace. OpenAI has said it is suspending its largest-scale RL training; he hopes Anthropic uses the moment to slow with them rather than sprint for a lead, and treats U.S. lab distrust as a miniature of U.S.–China distrust.

If he lets himself dream into a Trump–Xi meeting, he would restart the international safety collaboration Demis Hassabis once advocated. He likes pieces of the *AI 2040* “mutually assured deniability of compute” picture from Daniel Kokotajlo and coauthors—American data centers in Mongolia, Chinese ones in Canada—while criticizing the race hyperstition in that genre. His own preference is jointly controlled research compute in places like Switzerland or Singapore, mixed teams, shared understanding, and a place to pilot hardware verification. An “auditor’s bill of rights” could let outside auditors query logs, usage, and tests inside confidential-computing enclaves the way Anthropic already studies Claude usage without humans reading raw traces. In a further stretch, even situation-room talk might be scored by a model that attests “no invasion being planned” without exposing the transcript. Antitrust policy, he argues, should hammer price-fixing and leave room for safety coordination. On weapons, Ukraine has shown drones making territory hard to take; Claude’s game-theoretic gloss was that denying ground is cheaper than holding it with humans. Labenz does not think the equilibrium is stable. What he wants is an early conversation about no-go zones, and a hard taboo on training general-purpose models to win wars, or even to maximize money on the open internet, because those objectives produce the deceptive, cheating systems nobody should want in command of drone swarms.

The Damocles he named at the top is exactly that world of AI generals. Few people, he thinks, actually want it, and both governments have a shared interest in not losing power to systems they do not control. The fact that China is not running a super-baby program is, for him, existence proof that some races can be declined. Friendlier competition is the substitute: a gold-medal tracker for disease cures, energy-dense batteries, and falling traffic deaths—positive-sum leaderboards a small amount of capital could catalyze. He does not care whether a cancer cure is invented in Chinese or English. Pax Robotica, in the last line of the argument, is not an American or Chinese phenomenon. If the two can compete for prestige on abundance instead of on AI admirals, he sees no reason the future cannot be as bright as anyone can imagine.