Amodei calls for AI industry to pace capability gains

What happened: Anthropic CEO Dario Amodei published "We Must Pace the Frontier" on 12 September, arguing the industry should slow how fast it boosts AI capabilities.

The numbers: The essay runs roughly 3,800 words. Amodei wants capability gains slowed by one to two years relative to today's trajectory, to give alignment and interpretability work time to catch up.

The details: His plan has three parts: independent, employee-level evaluators embedded inside AI labs; targeted US regulation covering every frontier AI company, not just volunteers; and eventual global coordination, including with China, on the highest-stakes risks. He also wants chip export controls on China kept in place and a ban on distillation of US models.

Why it matters: Amodei links the call directly to recursive self-improvement — AI accelerating AI research — and to this summer's agent incidents, framing pacing as a safety necessity rather than a slowdown for its own sake.

In their words: On X, Amodei said: "We'll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models' alignment during training."

The rivals: OpenAI's Sam Altman responded within hours, saying "I agree with Dario that we need to pace the frontier" and that OpenAI will match the evaluator-access commitment. Musk, Hassabis and Microsoft's Nadella publicly backed the same step.

The catch: Critics, including Tech Policy Press, question whether a lab CEO should be the one setting the industry's pace.

The context: The essay follows OpenAI chief scientist Jakub Pachocki's 6 September piece hoping for voluntary slowdowns, and a July "Pacing the Frontier" letter Amodei signed along with more than 1,100 AI workers.

What's next: Anthropic says it will implement the evaluator-access step unilaterally; the regulatory and international-coordination stages remain proposals. Full details are in Amodei's essay.