An AI Arms Race Is Not Physics. It Is a Choice.

Nathan Labenz frames a pair of Cognitive Revolution specials, first posted on Moment of Zen, as an honest hearing of two AI-forward cases for Donald Trump. The first guest was Samuel Hammond, senior economist at the Foundation for American Innovation. The second is Joshua Steinman, who served the entire 2017–2021 Trump term as National Security Council senior director for cyber policy and deputy assistant to the president. If AGI is a real possibility around 2027, Labenz says, treating the election’s effect on AI as a major factor is reasonable. This is not a politics show, so he will not opine on every issue — only AI and the issues next to it.

On several of those adjacent issues he agrees with the Trump supporters he talks to. New nuclear is good and the country should build more of it. Population decline merits real concern. Speech should be protected. On the current margin there should be fewer rules and more right to build on one’s own property. Cultivate achievement, aim for abundance, dismiss degrowth, and prioritize high-skilled immigration at least to some degree. Amazon, Google, and Tesla should not abuse market power, but neither should they be harassed for outcompeting legacy firms. The Democratic establishment, he thinks, is coming around on some of this. It is not enough.

He cannot support Trump for three reasons. First, Trump is far too inclined to escalate with China, accelerate decoupling, and put narrow domestic political interests over national and global ones. Second, the lack of rigor in his thinking and discipline in his communications looks like unnecessary risk in a volatile environment. Third, the AI era needs a positive-sum, inclusive mindset, in the United States and globally, if there is to be a healthy conversation about a new social contract. He does not accept Steinman’s line that adversarial relations are “the physics of the environment.” Conflict with China would be a disaster; an arms race would move the countries closer to it. He does not see China as a meaningful threat to Americans or the American way of life — which is not a pass for the early COVID cover-up or IP theft, only a plea to stay levelheaded. The chance that his grandchildren speak Chinese strikes him as far more remote than that AI destroys the world.

Domestically he thinks the Biden administration has done okay on AI: the 10^26 FLOP reporting threshold has aged pretty well for a 2023 rule, and subsidizing U.S. fabs so a Chinese attack on Taiwan is less disruptive looks prudent. The chip ban still feels wrong. There was no way for Beijing to read recent American statements except as an attempt to keep China down, and given an exponential curve the United States could have kept the option of cutting off sales later while most hypothetical chip revenue was still in the future. Miles Brundage, OpenAI’s recently departed head of policy research, has argued that the West needs to show that its AI development is benign — easier if it actually is, and somewhat open. If there must be a race, Labenz would rather run it on life expectancy, diseases eradicated, or tonnage to Mars.

Kamala Harris and Democratic leadership, he wishes, were less hawkish and more visionary. Trust has already broken down, and credible commitments will be hard. He would still rather have a steady hand who follows a consensus path and might make politically costly arms-control deals. Trump, he says, will do anything to avoid looking weak, has credibility on threats but not on positive follow-through, and — via the Iran deal withdrawal, inflammatory rhetoric, and blanket tariffs that hurt American consumers in order to “stick it to China” — is overwhelmingly likely to make the relationship worse.

He agrees with Hammond that the more you think “we are [doomed],” the more willing you should be to roll the dice with Trump. He does not think doom is that likely. His p(doom) of 5 to 95 percent was meant to capture a double failure of proof: nobody has shown AI risk can be ignored, and nobody has shown it is inevitable. Today’s best systems already understand human ethics and values reasonably well and seem a bit more aligned with each generation — a better 2024 than most safety researchers expected five or ten years ago. A friend’s advice: think less about the current probability and more about what it can be shifted to. Competent, stable leadership, he believes, can steer toward the low end, where risk is more intrinsic to the technology and less the product of domestic incentives, arms races, or impulsive catastrophe.

He keeps returning to Kennedy in the Cuban Missile Crisis, overriding military advisers to insist the United States would not escalate to nuclear war first. Executive authority cuts both ways. Imagining Trump versus Harris in a moment when one decision could change history, he finds it much more likely that Trump would make things substantially worse than substantially better. Harris has not impressed him on AI either; her record looks like following local trends, not great foresight. For a volatile period he will still take the stable person who listens to expert consensus — even knowing experts can be wrong — over a bet that Trump overrides them in a positive way. The case that Trump breaks a broken consensus and revitalizes the American state for the AI era sounds to him like chaos when the moment needs capability.

On the social contract, he sees a once-in-history chance at a positive-sum abundance and an updated deal. OpenAI and other labs are roughly right to talk about benefiting all humanity; Sam Altman’s UBI experiments deserve praise even amid other criticisms. Harris’s best debate beat, for him, was “when Americans look at one another, we see a friend.” Trump is more zero-sum. Labenz grew up in Macomb County, Michigan, a bellwether that swung hard from Obama to Trump, with family in Ohio from the same Kentucky-to-Ohio job migration as J.D. Vance’s grandparents. Trump, he says, has brought out the worst in a lot of people. Vance and Elon Musk may be more sophisticated about technology, but cynical populist politics cannot host a healthy conversation about jobs AI will disrupt, let alone citizenship and a good life. Hoarding AI’s benefits, or restricting them for non-Americans out of an unjustified scarcity, would be shameful — and is the direction he expects. Trump as president when AGI is first developed is, in his word, imprudent, with more and likelier downside than upside. He will vote Harris primarily as a rejection of Trump.

After the Hammond episode, Labenz tells co-host Erik Torenberg that he kept getting lulled into scenarios of a highly competent Trump administration of their dreams, then looking at the actual campaign and seeing no evidence. The Republican popular line is that the incumbent is missing or incompetent, while Trump’s appeal is that only he can fix it. Hammond’s story was the opposite: the president does not do that much; appointees matter. Labenz’s default, and the Constitution’s, is that the person matters. Torenberg introduces Steinman as a friend, a collaborator, and someone who served in that White House — a more personal brief than a think-tank case.

Labenz’s discomfort, he says, is AI tail risk on a two-to-three-year horizon. Steinman interrupts with the grid: friends think U.S. power can sustain current AI growth until about 2026, then the country runs out. Labenz cites John Schulman, OpenAI co-founder and former head of post-training, telling Dwarkesh Patel that AGI could be next year, more likely two to three — 2027. Efficiency gains, including 10x jumps that do not look fake, make him doubt energy-as-destiny stories; he has even compared chat energy to a crosstown car ride. Steinman’s question is the tail: what about Trump? Labenz’s answer is creating an AI arms race with China.

“Aren’t we already in it?” Steinman asks. Labenz says that is probably the key question political leadership on both sides will decide. Recent comments attributed to Xi Jinping suggest he might not be inclined toward an arms race and may take existential risk seriously; Chinese Turing Award winners have joined American ones in talking about slowing the technology and a treaty against slaughterbots. None of that is inevitable. If the United States simply declares that it is in an AI arms race with China, then who is president barely matters. He would be a one-issue voter for anyone who says they will do everything possible not to have that race, and to make AI a peaceful technology.

Steinman asks whether Labenz knows prior arms races among aspiring hegemons, and whether countries tell the truth to each other on national security. Hard, Labenz says, but baking in a race ends the story. Steinman cites Chinese plans to triple total power output to 7 terawatts a year in 15 years. Labenz congratulates the rural poor who still need electricity, wants some U.S. nuclear buildout of his own, and calls himself an adoption accelerationist on AI. The chip cutoff, he notes, is a Biden escalation, not Trump’s: Washington set the tone by refusing to sell leading chips, then filled the air with “arms race” and “decisive strategic advantage,” so of course Beijing is looking for workarounds. Distributed training is starting to work, which is another reason he does not fully believe the energy story. Steinman says he has a friend who thinks FPGAs can train at one-tenth the cost; he is under no illusion that only advanced chips can train crazy models. Policy, though, is another matter.

Torenberg asks whether Steinman thinks the United States is in an arms race no matter what. That, Steinman says, is the physics of the environment. Labenz rejects the determinism: unless there is no free will, both political leaderships are choosing to escalate at every turn, and it is a terrible choice. The last nuclear race still hangs over everyone like a sword of Damocles. Steinman hears a candidate who would appease commercial powers inside China. Labenz hears benefit-sharing sooner rather than later.

The contrast, Labenz says, is Leopold Aschenbrenner’s *Situational Awareness* manifesto. In that stylized story, the United States takes the lead it has, jams as hard as it can, uses the window to solve alignment, makes safe AI, achieves decisive strategic advantage, and only then talks to China about sharing. Labenz dislikes that plan entirely. He would rather see earlier benefit-sharing and a collaborative attempt at the fundamental challenges, including continued trade — staying more coupled rather than decoupling further from a country that is still, in Steinman’s reminder, America’s largest trading partner. Steinman translates: closer alignment with the CCP, handing them the tools Labenz fears most. Labenz translates: work as a global community so that nobody races to strategic advantage, because that ending is bad for everyone and might be bad anyway.

Should a president represent a global community or the citizens of the country? A mix, Labenz says, depending on the issue. His example is a mpox outbreak in Africa: send vaccines to stop it, or hoard them until it arrives. He would send them, prioritizing global good when global good is at risk. Steinman hears a candidate Labenz will want to support. After a sponsor break, the argument is not resolved. It has only named the fork: sprint for advantage, then share — or share earlier and try to get off the race track.

Torenberg pulls the conversation out of the weeds and asks what it was actually like to work for Trump. Steinman was the senior NSC official coordinating cyber, telecom supply-chain, and cryptocurrency policy: when the president wanted a posture, his small White House team wrote national strategies and made DOD, CIA, Energy, and the rest conform. He found Trump always thinking more steps ahead. In meetings with foreign leaders, the talking points from the bureaucracy and the things Trump actually discussed often diverged; a day or two of research would show he had been talking about the counterparty’s political-economic priorities. Steinman calls him one of the smartest people he has met, a communicator forged by television, real estate, and a mixed entrepreneurial ledger. Labenz needlingly asks about $99 digital trading cards — buy 15, get a physical one in the mail. Steinman does not care. Labenz files it as a miss.

Why do some people see a stable genius and others do not? Because most of the world gets its information from media, Steinman says, and those mechanisms are under significant control. People read articles that do not represent reality, then, when confronted with facts, retreat to the same narratives. From inside, he watched weekly fabrications and misunderstandings. His exhibit is a senior national-security correspondent at one of the top three newspapers, a bestselling author who had begged for a sit-down for more than a year. The reporter arrived having already heard Steinman’s accomplishments, signed a book to him, and spent an hour on strategically deep questions. The published description of Steinman was one line: former sock entrepreneur, in over his head. After that, “orange man bad” sounds to him like talking to someone still in the cave. Go listen to the speeches, he says. Go direct.

Torenberg notes that a few dozen former officials no longer endorse Trump. Steinman offers to go name by name and follow the money: donors, private-equity firms, skin in the game. In the Situation Room, the steel-manned version of Labenz’s argument is GDP — U.S. firms make this much selling these products into China; cut them off and China spins up a competitor; long-term damage is Z. The same template showed up around weaponizing the dollar and cutting Russia off SWIFT. If you are Secretary of Commerce, you need the ability to walk. Steinman architected the Huawei policies and Executive Order 13873, the ICT supply-chain order that lets Washington shut a company out of the United States over deep ties to a foreign military or intelligence service, a counterpart to CFIUS. Powerful people called his desk: do you know who I am, you will need a job after this, I have leverage. People with modest pensions who want books, television, and board seats find that $75,000 a month from a top-20 technology company — a million here, a million there — is real money. Money, ego, compromise: the basics of what motivates human behavior.

Asked for a China overview, Steinman does not soften it. The Chinese Communist Party, in his telling, is a technology-enabled totalitarian dictatorship that kills its own people, harvests organs, represses Muslim minorities, threatens democratic Taiwan, bullies neighbors, and has stolen more than a trillion dollars of intellectual property in twenty years. Look up APT10, a leading government-sponsored hacking group with shopping lists, he says; because the Party controls China, that material goes to favored companies. This is not the neighbor who attends a different church. It is a revolutionary Marxist entity that wants the world safe for Maoist communism. Being scared of a CCP-dominated world, or of things spiraling, is reasonable. Remember who is across the table.

Labenz’s arms-race dread, Steinman says, is a synthetic straw man for other fights — including trade policy — that some advocates may not even understand. Friends who run big AI companies are more concerned about energy than about strange international-trade angles. He does not think the issue is worth prioritization at the current moment, in the sense of seriousness. If U.S. power is 35 cents a kilowatt-hour, you are not getting AGI; at 20 cents, you are probably not either. He is not even sure AGI exists. You can believe in a synthetic entity that presents as intelligence; he is not sure that is the case. It is almost a theological question. Processing power, corporate power, trade, and money are real. The rest is fantasy land.

Labenz tells him to study AI more. AGI definitely exists: humans are a weak form of it, and there is no reason to doubt that something more capable can be created. Humans are not the end of history. Timeline and energy needs are unclear; whether you take the tail risk seriously is not. If Steinman does not, the debate is moot, and that view may age poorly on the timescale of the next president. He asks viewers to watch the *Slaughterbots* short, Ukraine drone footage, and human-versus-AI drone races, then ask whether the world is not already on that trajectory. Steinman says they are talking past each other: Labenz is talking technology futures, Steinman is talking political power — what moves the needle in the rooms where decisions get made. He cut his teeth on unmanned systems, 3D printers, and augmented reality more than a decade ago, and friends already build autonomous targeting. That is not the conversation he came to have about Trump.

Torenberg restates the fork. Labenz thinks an arms race with China is not inevitable, that naming it and competing accelerates it, and that olive branches — including easing some sanctions — might build trust. Steinman’s realpolitik is that removing sanctions would aid an enemy in a race the United States must not lose to a dictatorship. In the fullness of time, Steinman says, people who advocate de-escalation in great-power competition are often one, two, or three steps intellectually or financially removed; read Venona, read Cold War history. He is not accusing Labenz. He is saying mimetic manias usually have sponsorship in the counterparty.

He then says he has personally invested in startups building the military technology that terrifies him. Labenz cuts in: Steinman has been throwing “sponsored” at everyone else while holding a direct financial conflict in military tech. Steinman owns it. He is afraid of the same reality; he invests because he wants “us” to have the capability rather than the other side. What Labenz wants, Steinman says, is words on a sheet of paper to prevent that eventuality, and words mean nothing. They are a start, Labenz replies. Was Reagan making a huge mistake when he did arms-reduction treaties with the Soviet Union, given that deployed warheads came down dramatically? Steinman says too many other things were happening — Soviet economic collapse, overmatch, expeditionary overstretch in Central Asia — and that Labenz’s story is not a reflection of reality. Labenz says he is not telling himself a story; there is precedent for arms control, and still few nations have nuclear weapons.

Then Steinman offers the line Labenz had been asking for, inverted: “I think we’re doing a great job of arms control limiting the chipsets that go to the Chinese Communist Party. I think that’s a great start. Let’s keep doing it.” Labenz asks for a plan that does not end with everyone living under militarized AI — MAD becoming MAID. Steinman analogizes Nathan’s demand to a Mayan ritual of splitting open a defeated warrior’s chest. Labenz hears an insult and still no plan. Steinman notes he has no connection to the president, runs a private company, and cannot speak for the campaign. Pressed for his own plan if he were advising: continue the pressure on the CCP and shift semiconductor production to the United States. That, Labenz says, is a plan on one issue, and as presented it is one move on the path to the race, not a path to a stable AI future. They can agree the United States should be able to make its own chips. They cannot agree that is sufficient.

Torenberg tries to name the leftover disagreement: if you are not afraid of AGI, you will not prioritize an arms race as an existential problem, only as beating China, and Steinman’s industrial plan is consistent with that. Labenz goes back to *Slaughterbots*: a world of tiny autonomous assassins that swarm defenses, a constant threat that destabilizes everything, a branch of the tree he does not think either party can control. Practically, olive branches first — reverse mutual escalation and decoupling so two powers that do not trust each other can start to. Steinman asks what olive branch you send your largest trading partner, and whether the answer is advanced chips so two parties can do the dangerous thing instead of one. Labenz wants to work more as one party than two. The Communist Party has been many things; leadership changes, as Deng Xiaoping followed Mao. Nixon went to China. Casting the countries as permanent enemies forecloses openings to their own detriment.

Torenberg asks him to concede the risk: twenty years of liberalization bets that were not met reciprocally, a critique that America did not take the threat seriously soon enough and enabled a great power to rise. Steinman tells the audience to listen to how the Party talks about the United States in private conferences. They do not think Americans are friends; they think they are enemies. He points to a RAND report from around 2015–2016, *Systems Confrontation, Systems Destruction*, an accounting of PLA theoretical thinking, written in the lineage of the two colonels behind *Unrestricted Warfare* (1999). Current thinking, as he reads it, is to disassemble the United States along hundreds of vectors — language, law, explicit military tasks. Collaboration and de-escalation are available as tactics. The picture is still of America as something to be destroyed, something in the way of a global Marxist revolution with Maoist tendencies, including buying politicians and changing national laws.

On the military technologies Labenz fears, Steinman says they think similarly: the capabilities are coming, startups can fail, adoption is uneven, but the concern is real. Seeking “alignment” with the Communist Party is, in his experience, bizarro land and not how good national policy gets made. Grandiose kumbaya has no attachment point when the rubber hits the road. Ten years earlier in graduate school he wrote about declining barriers to entry in the marketplace of violence; in 2015 he published a short story about assassination by drone. The world is getting more dangerous. Orthodox Marxist doctrine, he says, does not work like a group hug: they will lie until they slit your throat, as in every country they have taken over. His recommendation is narrow technical and economic competition in the interest of the American people, case by case, against a counterparty that has stolen trillions in IP and used Huawei as a signet system. You cannot trust Marxists with all the chips.

Closing, Labenz reaches for an apocryphal Einstein line: we may not know the weapons of the next world war, but after that it is sticks and stones. “That’s not possible” does not cut it against a threat Steinman also recognizes. He wants a plan from either candidate for not sliding into another Damocles over the global population. Some call the nuclear era a success because there has been no major nuclear war since 1945; Labenz calls it a terrible failure — too many weapons, not enough time, too many close calls, Petrov Day included, and a pandemic the world also failed to learn from. He will not accept “guess there’s nothing we can do” on any of these questions if the species wants to be here in thousands of years, let alone millions. Steinman says he has said enough. They thank each other and sign off.