Dario Amodei has published a 3,800-word letter titled "We Must Pace the Frontier", proposing that frontier labs agree to limit how fast capabilities advance rather than stop training models. Sam Altman, Elon Musk, Demis Hassabis and Satya Nadella all backed it. Stuart Russell, distinguished professor of computer science at UC Berkeley and president of the International Association for Safe and Ethical AI, says the metaphor underneath it is wrong: you cannot set a slower pace of capability development and hope that buys enough safety. Requirements have to be fixed first, and progress allowed only once they are met.
The letter landed in a bad week. The AI safety researcher Jacob Coxon left Anthropic, new and unsettling details emerged about the OpenAI and Hugging Face incident, and an email from Business Insider arrived the same day under the subject line "The AI doomsday debate has reached a boiling point". Amodei's document is addressed to that temperature as much as to the technology.
"Pacing the frontier" borrows from Formula 1. When track conditions become dangerous, a safety car comes out and the field follows it at a controlled speed. The race continues; the risk drops. Amodei is explicit that he is not proposing to halt model training or technical progress. What worries him is speed itself, driven mainly by recursive self-improvement. Russell's version of the scene: Amodei and Altman running side by side at 200 mph toward a first corner covered in ice, realising the cars have no steering wheels.
The proposal has three parts. First, independent evaluators working inside each company with full access to its systems; Anthropic says it is ready to adopt this now, without waiting for a government to require it. Second, every company building frontier systems in "democratic countries" agrees on shared safety standards and limits on the pace of uncontrolled development, backed by state regulation where necessary. Third, "authoritarian countries" are brought into a wider negotiated arrangement.
Amodei also spends effort reassuring Washington, where US leadership in AI is treated as a primary geopolitical asset. Progress will still feel fast, he argues; the point is a speed limit on recursive self-improvement, and the one to two years it buys should go into interpretability research, alignment and better testing.
The phrase has a history. It first appeared in an open letter called "Pacing the Frontier" in July, signed by 1,386 employees of frontier AI labs — Amodei among them. That letter said there were no credible plans for controlling superintelligent systems and called building systems smarter than humans objectively insane and suicidal. It presumably ruined several weeks for communications directors across the industry. The new monograph, by the same author, is written to calm investors.
Critics read the proposal as an attempt to lock in Anthropic's lead through government intervention. Russell thinks that reading is wrong, and reaches for the Wikipedia entry on Formula 1 pacing to make the point: the safety car removes whatever time and distance advantage the leader had built over the rest of the field.
His actual objection is structural. Imagine Boeing announcing it will ship a new aircraft every year and hopes to complete at least some of the flight tests in that window, with good results. That is not how certification works. The aircraft enters service after the tests are finished and the state issues an airworthiness certificate. If that takes more than a year, it takes more than a year.
Read closely, Russell says, Amodei may already agree. The letter proposes rules of the form: if a model has capability X, it must be accompanied by certificates establishing alignment properties Y and Z. Requirement first, proof second. Safety researchers call this the red lines approach, and its consequence is the part the pacing metaphor obscures — a developer who cannot meet the requirement has to stop. In Formula 1 terms that is a red flag, not a safety car.
Russell puts a number on it. An acceptable annual risk of irreversible loss of human control might be something like one in 100 million, not the one in 10 or one in five that AI company executives currently give. Seven orders of magnitude is not a disagreement about tone.
Here is what I take from the endorsements. A proposal that Amodei, Altman, Musk, Hassabis and Nadella can all sign is, by construction, a proposal that costs none of them anything they were unwilling to give up. Pace limits bind competitors symmetrically and never require any individual company to stop shipping. Gating requirements do the opposite: they apply to you alone, on your schedule, and they can produce the answer no. The unanimity is not evidence that the industry has converged on safety. It is evidence that this particular instrument does not bite.
Russell's strongest point is the one the racing imagery buries. The difference between the two proposals is not speed. It is who carries the burden of proof — and under pacing, nobody does.
The proposal as described also leaves the mechanism unspecified in the places that determine whether it works. Nothing says who the independent evaluators report to, what happens when one of them refuses to sign off, whether a company can dismiss them, or who issues and can withhold the certificates that attest to alignment properties Y and Z. A certificate without an authority able to deny it is a press release with a serial number.
The resistance to all of this is financial rather than technical. Trillions of dollars are already committed to the current line of development and companies plan to commit trillions more. Russell's answer is that the sunk cost fallacy is still a fallacy: doubling down on a wrong decision does not make it right. He sees an opening in the coincidence of public attention, a rare common position among lab chiefs, and the coming meeting between Donald Trump and Xi Jinping.
The opening is real, but so is its shape. Every actor with the standing to set red lines is an actor that has already spent the money, and the one instrument they have all agreed to is the one that never forces a stop.