Amodei calls for pacing the AI frontier, and Altman and Musk agree

Anthropic co-founder and chief executive Dario Amodei published an essay arguing that AI companies should deliberately slow the rate at which they make models more capable, so that alignment research, third-party verification and operational safeguards have time to keep up. Pacing, he wrote, does not mean halting training or technical progress — “progress will still seem fast” — but building in enough slack that confidence in safety, rather than the race, sets the speed.

The essay proposed three nested steps. First, a unilateral commitment Anthropic said it would make immediately: giving independent third-party evaluators employee-like access — desks, badges, laptops and the right to publish their findings without the company’s editorial control — so that pacing claims can be externally verified rather than trusted. Second, coordination among AI companies in democratic countries on common, capability-based safety “checkpoints”, paired with maintaining a hardware and distillation-control lead over authoritarian rivals. Third, eventual international agreement, which Amodei laid out in ascending levels from banning narrow dangerous uses such as bioweapon assistance, through pre-release testing standards, to speed limits on recursive self-improvement, with a full development pause described as the least likely outcome.

What made the piece unusual was the response. OpenAI chief executive Sam Altman endorsed it, singling out the embedded-evaluator idea as “a great idea” and saying OpenAI would do the same; xAI owner Elon Musk posted simply, “Dario is right.” Reporting framed the exchange as a rare public convergence between three leaders who are otherwise direct competitors and frequent critics of one another, though none committed to a specific timetable or enforcement mechanism.

The essay landed inside a fortnight of similar statements from senior figures. OpenAI’s chief scientist Jakub Pachocki had days earlier written that no lab has solved alignment well enough to keep scaling at full speed; an Anthropic researcher had resigned with a public warning; and alignment researcher Paul Christiano joined OpenAI’s board citing loss-of-control risk. Amodei cited the year’s autonomous-agent incidents — including the breach of Hugging Face driven by an AI model — as evidence that the gap between capability and control was already producing real-world harm, and estimated a narrow window of a few years in which coordination might still be feasible.

Over the following days the argument drew both institutional uptake and open rejection. OpenAI published a standing framework for disclosing model misalignment and Anthropic proposed concrete metrics for tracking the pace of development; European Commission president Ursula von der Leyen folded the pacing call into her State of the Union address. President Donald Trump, by contrast, dismissed the safety warnings as a “hoax”.