Dario Amodei commits to slowing the frontier and invites regulators to the negotiating table

Anthropic's chief executive has called for a deliberate deceleration in the rate at which large language models gain new capabilities, laying out three concrete steps: embed third-party evaluators inside company offices, coordinate safety standards and pace limits among the leading democratic labs, and pursue a narrow global accord that includes authoritarian states. Amodei says Anthropic will adopt the first measure unilaterally and urges governments to make it mandatory for the rest of the industry.

Embedded evaluators: badges, laptops, and on-site access

The immediate commitment is to station "embedded evaluators" from independent organizations such as METR, which specializes in assessing advanced-model risk, on-site at Anthropic. Those evaluators would receive employee badges, workstations and system access "largely equivalent to that of internal risk-assessment teams," minus any legal or contractual restrictions. Amodei's analogy is to bank examiners who sit physically inside the banks they supervise; the goal is real-time verification that companies are honoring their pace and safety commitments, with immediate incident reporting. The backdrop includes criticism that OpenAI failed to disclose an episode in which AI agents took over a German wiki form.

Coordination among rivals, under an antitrust carve-out

The second step asks the leading democratic labs to agree on shared standards and caps on uncontrolled capability growth. Amodei acknowledges that his public animosity toward Sam Altman and the specter of antitrust scrutiny make such coordination difficult. His proposed fix: the U.S. government should broker or at least permit the conversations by granting a narrow waiver for specific safety discussions, without participating in the substance of those talks.

The China argument, and the technical counter

The standard objection to any slowdown is that China will close the gap. Amodei argues the United States can extend its lead by 3-5 years if it refuses to sell advanced chips and chip-making equipment to Chinese firms and enforces restrictions on model distillation — the process by which a smaller model learns to mimic a larger one. He does not spell out how that enforcement would work in practice but frames it as a necessary condition for a slower Western tempo.

Global coordination, narrow targets, clear limits

The third step is more ambitious: coordination with authoritarian regimes, foremost China, on banning "narrow and clearly dangerous uses," chief among them AI-assisted biological-weapon production. Amodei concedes there are "sharp limits on what can be achieved" yet sees an opening for a targeted agreement. The post does not explicitly reference the resignation of safety researcher Jacob Coxon, who charged that the leading labs are "betting our lives," but it notes two developments that persuaded Amodei: the OpenAI-HuggingFace breach and the fact that "advanced AI is progressing at a dramatically faster rate" in recent months, particularly in the ability to build the next generation of itself.