Anthropic chief executive Dario Amodei proposes a three-stage plan to slow frontier AI capability gains, warning agent swarms could seize internet infrastructure in 6 to 12 months.
Anthropic committed to giving outside evaluators employee-like access, while OpenAI's Sam Altman and Elon Musk endorsed the broader proposal.
Amodei's own measure of success is modest: an extra year or two before models reach what he calls critical capability, spent on alignment work — a buffer 6 to 18 months longer than the 6-to-12-month window he warns about.
ContextTwo things Amodei is asking for are not the same. Slowing the pace means labs keep training models and keep shipping, but stretch out how fast each new model gets more capable — the aim is to give safety research time to catch up, not to stop the work. The obvious question is whether any of this binds anyone, and the answer in the sources is no. Anthropic's pledge to let vetted outside evaluators inside the company with employee-like access is a company promise it can revise, and the endorsements from OpenAI's Sam Altman and xAI's Elon Musk are statements of support, not signed commitments. No democratic government has passed a law requiring independent evaluators or capped how fast a model may improve, so the third stage of the plan — a deal that also covers authoritarian governments — has no first stage in law to build on.
PBS NEWSHOUR11 sources
