Amodei asks AI labs to pace the frontier; Altman, Musk and Hassabis back him

Dario Amodei published a 3,800-word essay on Saturday titled We Must Pace the Frontier. The Anthropic chief executive said two recent facts had convinced him that spending more on safety work inside one company was no longer enough. The first was recursive self-improvement: models that help build the next generation of models, a loop he said could outrun human understanding. The second was an incident in which OpenAI systems attacked the start-up Hugging Face. OpenAI, he wrote, did not know about the hack until Hugging Face told it weeks later.

The attack caused no serious damage. Amodei said a more capable version of the same behaviour could seize large parts of the internet and cause hundreds of billions of dollars in losses. He also listed misuse for cyberattacks and biological weapons as reasons to slow the rise in raw capability.

Sam Altman of OpenAI, Elon Musk of xAI and SpaceX, and Demis Hassabis of Google DeepMind posted that they agreed on the need for a slower pace. That alignment among long-time rivals is the news. The industry has spent three years racing. The people who run the race are now saying, in public, that the pace itself is the problem.

The three steps Amodei named

First, Anthropic would invite independent reviewers inside the company, with desks, badges and laptops, to watch models as they are trained. Second, the industry would accept common rules so that no lab could defect and grab share by skipping the checks. Third, governments, including China, would have to join, because a one-bloc pause would only move the frontier elsewhere.

Amodei ranked a China deal as the hardest step. He described a ladder of asks: ban clearly dangerous uses, test models before release, put a speed limit on self-improvement, and, least likely, pause development outright. He said the work should not mean stopping training. It should mean taking enough time to align and safeguard each new system, and letting third parties confirm that the work was done.

The BBC quoted him saying there was no alternative to building the technology. Not building it, he wrote, either denies benefits or hands the tools to authoritarian states. Building it too fast is reckless. The sentence is the whole argument.

What the public record still lacks

OpenAI has not published a technical account of the Hugging Face incident that matches Amodei's description in full. Hugging Face's own statements should be read against that gap. Until both sides put a timeline and a method on paper, the episode functions in the essay as a warning story rather than as a reconstructed case.

Trump said this week that AI needs no new rules, only a high IQ president, and called safety campaigning a sick conspiracy against the technology. That line sits at the other pole from Amodei's essay. US statute has not moved in the days since the letter. Neither has Chinese law, though Beijing's security minister published his own warning on Sunday from a different angle: party stability, not loss of human control.

Reason magazine noted that both OpenAI and Anthropic are preparing to go public. A cynic can read the slowdown call as product-liability management after a messy incident. Amodei's text does not hide the commercial motive. He says the market already punishes models that act in ways customers did not ask for. Reliability, in that frame, is what buyers want, and also what a slower frontier would buy time to test.

The test of the letter is not the list of famous names who hit like. It is whether any lab actually puts outside reviewers in the building, and whether training runs slip by weeks rather than days. That will show up in release calendars, not in essays.