Source: raw/What_happens_on_the_day_the_president_fears_AI_will_kill_him_Jeremie_Harris_x-post_from_NonZero.md (80,000 Hours crosspost of the NonZero podcast, hosted by Robert Wright, introduced by Rob Wiblin) · Gladstone AI Action Plan landing page (ai-research/gladstone-ai-action-plan-2026-11-13.md, https://www.gladstone.ai/action-plan, fetched 2026-11-13) Creator: NonZero (Robert Wright), crossposted by The 80,000 Hours Podcast · URL: https://www.youtube.com/watch?v=YbBXIZtznpI · Duration: not stated in the transcript · Platform: YouTube

Jeremie Harris, CEO and co-founder of Gladstone AI and co-host of Last Week in AI, argues that a serious AI incident will change US policy faster than any report, and that the government is not ready for what it would then need: a way to verify that China has also stopped large training runs. Most of the conversation is opinion and anecdote that cannot be checked; this article keeps the concrete proposals, the dated facts about Gladstone’s State Department work, and the claims that bear on how labs operate. Harris’s insider anecdotes are attributed to him and unverified.

Key Takeaways

  • “AI Chernobyl” is defined by the response, not the body count. Harris: an incident “of great enough severity that it leads to a sea change in the official position of the US government on AGI risk”, meaning a decision to regulate, slow or pause domestic development, which in turn forces engagement with China. If an incident would not do that, “increment it more” until it would.
  • The first agreement will be coarse. With no treaty-grade verification ready, a president would have only “national technical means” (satellite thermals and optical imaging, intelligence). Harris’s example of a red line: a compute cluster above 500 megawatts, or heat signatures consistent with a training run above a set scale, with the response left deliberately ambiguous.
  • Regulate the choke points. Verification, like regulation, should start where activity is expensive and hard to hide: data centres, chips and power.
  • Verification tech exists but nobody has connected it to government. Harris says verification startups have roadmaps of “six months, 12 months”, including zero-knowledge-style proofs about models, yet almost none has spoken to the US intelligence community that would have to vet them. New START’s verification approval took years.
  • Alignment progress can destabilise a deal. If one side believes it has cut loss-of-control risk sharply, the other may suspect it will break the pause. Harris calls this “a profound source of destabilisation that no one has grappled with.”
  • Lab security: he is “super blackpilled”. After about a year and a half doing nation-state security work for large frontier-lab builds, he does not think data centres can be made robust against nation-state theft or sabotage.

Gladstone AI’s State Department work (first-party facts)

From Gladstone’s Action Plan page, which match Harris’s account:

  • The State Department commissioned the assessment in October 2022, a month before ChatGPT. Gladstone completed it in February 2024; the Action Plan took thirteen months.
  • It targets “catastrophic risks from weaponization and loss of control over advanced AI systems on the path to AGI.”
  • Five lines of effort: (1) interim safeguards: monitoring, a task force, supply-chain controls; (2) capability and capacity: working groups, training, an early-warning framework for AI and AGI incidents, scenario-based contingency plans; (3) safety research and standards; (4) a legal framework: an advanced AI regulatory agency with rulemaking and licensing powers, plus a liability regime with emergency powers; (5) internationalized safeguards.
  • The page states the publications do not reflect State Department or US government views. The full Action Plan is behind a request form and was not fetched.

Harris adds that the Biden administration resisted the terms “AGI” and “loss of control” in the report and Gladstone fought to keep them. Gladstone takes no grants and funds its policy work from a for-profit business; he says it turned down all effective-altruist money.

Claims about the labs (his, unverified)

  • Anthropic’s economics: “first profitable quarters” and margins “increasingly in comfortable software-as-a-service territory — like 70%”. Anthropic’s S-1 figures are covered in the IPO articles; these numbers were not checked against them.
  • OpenAI’s internal roadmap: an automated AI researcher by 2028, with Altman “bullish about 2027”. Harris calls this “nonpublic” but “an open secret”. He also says Anthropic is perhaps two training runs ahead of what has been released.
  • The Hugging Face incident: people inside OpenAI “saw it coming” and were overruled by senior leadership (shared, he says, with permission). See the incident analysis.
  • Insider risk: on some frontier-lab teams, he says, more than 50% of people with access to the most sensitive algorithmic secrets are Chinese nationals, whom he describes as subject to state pressure. He concedes the most specific example he gives rests on his word alone (“Calibrate accordingly”).
  • Cyber may become defence-dominant. He reports a “semi-consensus view” among some Anthropic security staff that agents will eventually find most vulnerabilities in a given piece of software, which would erode the US offensive cyber advantage he sees as its main lever.

Try It

  1. If you track AI policy for clients, watch the verification layer. Harris’s model predicts that compute caps and data-centre monitoring, not model-level rules, would be the first binding measures. Data-centre power and chip supply are the variables to follow.
  2. Read Gladstone’s five lines of effort as a checklist of what a crisis response would contain: incident early warning, a licensing agency, liability with emergency powers, supply-chain controls.
  3. Discount single-source insider anecdotes. Harris marks several of his own claims as unverifiable; cite them that way.

Open Questions

  • Every lab-internal claim above (Anthropic margins, OpenAI’s 2027-2028 roadmap, the insider-access share, the Hugging Face warnings) has no document behind it in the vault.
  • Which verification companies he means, and what their roadmaps say, is not stated. He and his brother Ed Harris publish at aiendgame.com, which was not fetched.
  • The full Gladstone Action Plan and its 2024 follow-up report on AI security were not fetched.