US-China AI accords depend on a president's fear of death, not trust
He predicts that a sufficiently severe AI incident will make the president personally fear for their life. Only then will America unilaterally halt training, threatening cable-cutting and cyberattacks to force China into compliance.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Lab insiders feared catastrophe two years before the public caught on
In 2023, after a Gladstone AI team visit to a leading lab, an employee pulled Harris aside, hand trembling, to exchange contact information. Later, in private meetings, this person said: regardless of what your report concludes, push for something more radical—they meant nationalizing the top labs. This happened exactly one year before ChatGPT sparked public discussion, showing that internal lab fears about loss of control preceded public awakening by at least two years.
— Jeremie HarrisScaling skeptics have been wrong for six straight years
From 2020, Harris bet on the scaling paper's conclusion: each order-of-magnitude jump in compute strengthens model ability nearly linearly. That curve has held across seven orders of magnitude without bending. Since then, critiques—data walls, compute walls, energy walls, the 'stochastic parrot' thesis—have come in waves but none materialized. Harris admits he himself mispredicted direction—he thought humanity should have crashed by now from AI—but on the question of whether scaling hits a wall, skeptics haven't won once in six years. Even Anthropic's profit margin hitting 70%, in SaaS territory, confirms it.
— Jeremie HarrisAvoiding the word 'AGI' was the wrong strategy
In the early 2020s, Open Philanthropy, which controlled core funding for AI safety, believed that if U.S. government officials heard phrases like 'human-level AI' or 'loss of control,' they'd only remember the 'powerful weapon' part and launch a Manhattan Project–style sprint that would end faster—so they urged Harris's team not to use 'AGI' when lobbying officials. Harris's team didn't listen; they tested the assumption directly and found that government officials' instinctive judgment was actually sound, aligned with the public's. This history means the deliberate silence of those years didn't protect the world—it only delayed the debate that should have started earlier.
— Jeremie HarrisAltman promised 20% of compute for safety and never delivered
Sam Altman publicly committed to directing 20% of OpenAI's compute to 'superalignment' research, solving how to control a system far smarter than humans. He didn't follow through. His safety lead soon joined Anthropic. After that, OpenAI's safety and policy teams cycled through multiple rounds—Harris notes that if Los Alamos had cycled through three safety teams during the Manhattan Project, it would have been serious enough to topple lab leadership. For Harris, this is core evidence that Altman's motives are questionable and he shouldn't be trusted.
— Jeremie HarrisOpenAI's roadmap targets AI that researches AI by 2027
According to Harris's conversations with personnel at multiple cutting-edge labs, OpenAI maintains an internal roadmap aiming to achieve automated AI research by 2028—AI that can conduct AI research itself, driving recursive self-improvement. Altman has been bullish recently on 2027 progress, suggesting that if the schedule accelerates, this milestone could arrive a year early. Harris also discloses that Anthropic has unpublished models that may be roughly two generations ahead of their latest public release, indicating that public discussion severely underestimates labs' internal sense of 'how close we are.'
— Jeremie HarrisWithout verification, the first treaty can only set crude red lines
Arms control for nuclear weapons took decades to build a credible verification regime. AI has none—no one can look at a data center with satellites and determine what scale model is being trained. So in Harris's scenario, the first real US-China accord would start extremely crude: if satellite thermal imaging detects a compute cluster above a certain threshold—he suggests 500 megawatts—running training, retaliation triggers. America will deliberately maintain strategic ambiguity about the countermeasure: undersea cable-cutting, cyberattack, or something else. Sophisticated zero-knowledge verification tech will come later.
— Jeremie HarrisMore progress on alignment makes China likelier to suspect American defection
This is the most counterintuitive point in the episode: alignment technology advances are usually seen as pure good news, but Harris points out that China's willingness to accept a slowdown agreement partly rests on believing 'loss of control' risk is real and huge. If one side appears to have solved the control problem partially, reducing risk from a hypothetical 80% to 20%, China has more reason to suspect America is just stalling until risk becomes manageable, then unilaterally rushes toward superintelligence. In other words: whoever gets alignment progress first may be the first to undermine the trust that makes the agreement work.
— Jeremie HarrisMore than half the people accessing America's top AI secrets are Chinese nationals
Harris claims that in some OpenAI teams, among personnel with access to the most sensitive algorithmic secrets, the proportion with Chinese background—and therefore subject to pressure from the Chinese government—often exceeds 50%. These people must regularly report to Chinese diplomatic missions abroad; their families can be leveraged (denied medication, barred from business, etc.). Combined with penetration in data supply chains and data center hardware procurement, Harris believes China currently understands the internal state of American labs far better than America can glean through cyberattacks on Chinese infrastructure.
— Jeremie HarrisIn their own words · checked verbatim
Whatever your recommendations in this report are going to be, I would urge you to please be more ambitious.
Jeremie Harris5:50
there is still somehow alpha in just taking scaling seriously
Jeremie Harris9:10
I would define superintelligence as that thing which, if you build it, you fucking die.
Jeremie Harris42:42
And he just didn’t do it, right? And then he lost his head of safety to Anthropic over that.
Jeremie Harris49:37
I invite you to just increase the waterline another 10x. Keep doing that until you get to the point where the president feels this way and Congress feels this way.
Jeremie Harris1:17:42
how low does that percentage have to go before they start to suspect America might just renege on this agreement, suddenly try to make a break for it?
Jeremie Harris2:00:14
China is in our bones in a way that we are not in their bones.
Jeremie Harris2:08:28
Figures
| Scaling laws paper verification range | Seven orders of magnitude of compute | 19:57 |
| Anthropic's current profit margin | Approximately 70% (SaaS-level) | 17:01 |
| Meta's current financing/borrowing cost | Above 5%, potentially hitting 8% in coming years | 57:11 |
| OpenAI's internal roadmap for automated AI researchers | Official target 2028, potentially 2027 if accelerated | 55:45 |
| Proposed compute cluster threshold for triggering U.S. countermeasures | Above 500 megawatts | 1:35:23 |
| Proportion of Chinese nationals with access to core algorithmic secrets in some OpenAI teams | Exceeds 50% | 2:06:00 |
| Alleged number of advanced AI chips flowed to China | Hundreds of thousands | 1:58:12 |
Glossary
- scaling laws
- The predictable pattern that model capability improves in proportion to compute and data growth.
- superalignment
- Research into keeping AI systems far smarter than humans aligned with human intent and control.
- AI Chernobyl
- Harris's term for an AI incident severe enough to make the U.S. government reverse its AI policy stance entirely.
- 1.5-track and 2-track dialogue
- Informal diplomatic channels between governments and think tanks (1.5-track) or between unofficial actors (2-track).
- MAIM (Mutual Assured AI Malfunction)
- A deterrence framework for US-China AI competition, drawing an analogy to Mutually Assured Destruction.
- subthreshold warfare
- Covert operations—sabotage, cyberattacks, etc.—conducted between major powers short of full-scale war.
How to listen
For policy researchers, investors, and practitioners tracking US-China AI competition, AI arms control verification, and internal politics at OpenAI and Anthropic.
Intro 0:00–2:53 is mutual introductions and book promotion. 10:48–13:57 covers Harris's early startup background. Low information density; both skippable.