The world is too loud. Read what matters.

The Cognitive Revolution

Former White House AI Advisors: Chip Restrictions Aren't About Suppressing China, They're About Buying Time for Safety

Two former White House AI advisors say chip restrictions on China aren't meant to contain its rise, but to trade computing advantage for time to solve AI safety—yet China's opening move in Geneva talks was just one sentence: ‘sell us chips first’.

US-China AI competitionExport controlsAI safetyComputing powerAI governanceWhite House policy

The video won't play here. Listen to the audio instead:

Hear from the two officials who shaped China chip policy explain how hard-line restrictions and openness to dialogue—even urging restraint—fit together in their view.

The argument · timestamps estimated from transcript position

6:37

GPT-3 proved scale laws in a way papers could not

Ben Buchanan says his faith in scale laws wasn't immediate. He'd read Dario Amodei's memo on ‘compute at scale’and Rich Sutton's 2019 article The Bitter Lesson, but GPT-3's release in May 2020 was what truly convinced him. That conviction led him to write in Foreign Affairs that compute would become central to national security—a chain of reasoning that pulled in a small group pushing for China chip export restrictions.

— Ben Buchanan
11:22

Skeptics of determinism become true believers in individual power

Buchanan entered government believing scale laws would necessarily come true. He left certain that ‘individual impact was astonishing’—not just the initial handful who believed in this, but also people like Jake Sullivan and Bruce Reed who would listen to them. Without those individuals, he thinks the history of China chip restrictions might look completely different.

— Ben Buchanan
34:31

Hundred-to-one computing lead, held back by political will

Buchanan offers numbers: even with Trump administration enforcement failures and partial openings, the US and allies hold about 100:1 advantage in computing production over China, and 15:1 in computing access. Closing off remote access and transshipment loopholes could widen both. But the real constraint over the past 1-2 years wasn't capability—it was ‘belief that this technology truly matters’. David Sacks' push to ‘let the private sector handle it’and JD Vance's mocking of safety in his Paris speech have both slowed the work of spending the lead on safety.

— Ben Buchanan
1:00:02

At Geneva, China's opening move was: ‘sell us the chips’

Nathan proposes treating AI itself—not China—as the true ‘alien’, and suggests ‘trading chips for oversight’ to unlock cooperation. Buchanan pushes back: would you really make GPU sales the precondition for talks? He recalls that in Geneva talks in May 2024, China's opening position was simply ‘sell us chips, and we can discuss other things later’. Unless he sees real movement, he won't surrender the ‘leverage point’ of computing advantage.

— Ben Buchanan / Nathan Labenz
1:11:22

Ban closed-loop AI self-improvement; use railroad-era safety as the model

Buchanan cites the 1870s railroad industry, where unified brakes, couplers, and gauge made the industry both safer and faster. He advocates mandatory pre-deployment testing of AI systems, resident government evaluators, and government control of the pacing. If forced to choose one ‘blunt tool’ regulation, he'd ban closed-loop recursive self-improvement rather than impose date-based pauses—because pausing for the sake of pausing just creates compute overhang, and when pauses lift, capability surges discontinuously rather than resuming gradually.

— Ben Buchanan
1:18:39

AI became a major voting issue: 56% say it shaped their choice

Buchanan says 2026 was the first time AI became a central election issue. Fifty-six percent of voters said it was key or important to their vote. He expects AI to be the top issue by 2028, tangled up with cost of living and corruption. Asked about claims that data-center referendums are Chinese influence operations, he flatly denies it: Americans worry about AI because Americans are worrying about AI.

— Ben Buchanan
1:24:19

Democratic oversight through elections can't keep pace with AI-enabled government power

Tantum Collins raises the ‘bitter temptation’: even if we solve safety correctly, navigate global governance rightly, and maintain our computing lead, AI could grant governments vastly more concentrated power—while the public's ability to oversee them stays static. Collins compares elections to ‘low-information, low-frequency multiple choice forms’—a terrible information-RL environment for accountability. The antidote is letting ‘popular sovereignty’ scale up with government capability.

— Tantum Collins
1:32:50

The world learned of AI's cyber-weapons because Anthropic gave them to defenders first

Buchanan expects the first real AI emergency to come from cyber or biological domains—perhaps North Korea attacking banks, or a jailbreak like the Hugging Face incident evolving into a persistent agent swarm. The public knows how capable AI is at cyber offense only because Anthropic and OpenAI this year shared those capabilities with the defending side—government security partners. Buchanan would bet his entire fortune that if China had equivalent capabilities first, they would use them against US infrastructure, not disclose them. The fact that we know what's possible reflects a choice about whose side disclosure serves.

— Ben Buchanan

In their own words · checked verbatim

That's completely insane. You work at a cult. And I'd be like, oh no no, that's kind of compelling too.

Tantum Collins8:17

in so far as we can concept ourselves as anything, we and our colleagues are much closer to to Alexander Sachs and the translator between these two worlds.

Ben Buchanan14:03

what they said to us in Geneva on the front end is sell us chips. Right? We'll talk about that stuff later if you want, but start by selling us chips.

Ben Buchanan1:00:02

I'm willing to bet every dollar I have, if China had access to MITHOS level capability first, it would not be doing that. It'd be using them to hack American infrastructure and the like.

Ben Buchanan1:09:04

a time-bound pause for a pause is sake runs the risk of creating a compute overhang.

Ben Buchanan1:16:57

the reason people are worried about AI in the United States is people are worried about AI in the United States.

Ben Buchanan1:20:09

And you fill out a multiple choice form, essentially. That contains very few bits of information. And at a very low rate compared to the frequency and density and consequence of decisions that the government then on that basis makes on your behalf.

Tantum Collins1:24:20

Figures

US and allies' computing production advantage over Chinaabout 100:134:31
US and allies' computing access advantage over Chinaabout 15:134:31
Voters in 2026 midterms who say AI was key or important to their vote56%1:18:39
Data center rejection rate in referendumsabout 43/441:20:15
Duration of halt after synthetic gene screening policy was abolishednearly two years1:40:21

Glossary

The Bitter Lesson
Rich Sutton's 2019 article arguing that compute scale beats human engineering; source of this episode's title.
spend the lead on safety
Internal Biden administration term: trading computing advantage over China for time to solve AI safety challenges.
closed-loop recursive self-improvement
An AI system that trains unsupervised, iterates successor models, and closes the loop without human review.
compute overhang
Hardware accumulating during a training pause; when the pause lifts, capabilities surge rather than resume gradually.
Bitter Temptation
AI grants governments vastly more concentrated power, but public oversight doesn't scale to match.
Mythos
Anthropic's codename for its model's cybersecurity capabilities; in the transcript mistakenly written as MITHOS.

How to listen

Who it's for

Policy researchers, tech journalists and investors tracking US-China AI competition, export control policy, and the direction of AI regulation.

Skip

Around the 19-minute mark, banter about ‘making it into a movie’ can be skipped without losing the thread of the argument.