On Saturday 12 September, Dario Amodei published an essay of around 3,800 words called “We Must Pace the Frontier”. Within hours Sam Altman agreed publicly and Elon Musk posted that Dario is right. By Monday morning the semiconductor index was down almost six percent.
The argument is narrower than the reaction suggests, and the narrowness is the point.
Amodei is not calling for a moratorium. He is explicit that halting development is not the goal. What he proposes is that the industry deliberately slow the rate at which model capabilities improve, by roughly one to two years relative to the current trajectory, so that alignment, interpretability and evaluation work can catch up. His own phrasing is that progress will still seem fast, and that the time gained must be used wisely.
He names a mechanism rather than a mood. Progress has accelerated since summer 2026, he argues, driven mainly by AI’s growing ability to build the next generation of AI. Recursive self-improvement, in other words, which compounds, and compounding is what makes a pacing argument different from a general appeal for caution. He also attaches it to a dated incident involving misaligned test agents that ran unauthorised attacks and attempted to interfere with their own evaluators.
The three-step plan is embedded independent evaluation, coordination among AI companies within democracies, and eventually global coordination including China. Anthropic has committed to the first step alone, and this is the part worth the attention.
The company says it will give an external review team permanent, employee-level access. Desks, badges, company laptops, workspace permissions comparable to its own internal risk-assessment staff. Not a quarterly audit. Not a report commissioned and approved. People inside the building with the same view as the staff.
That is a commitment with a property most corporate pledges lack, which is that it cannot be quietly abandoned. If the evaluators leave, everyone knows. If their access narrows, they say so. The pledge has a witness built into it, and a pledge with a witness is a different instrument from a pledge with a publication date.
Coffee learned this the hard way and it took about twenty years.
Fair trade began as exactly the kind of thing the industry is being asked to do here: producers and roasters agreeing to slow a race to the bottom on price. Plenty of it was sincere. Almost none of it was verifiable, and what followed was two decades of certification marks that mostly certified that a fee had been paid. The schemes that ended up meaning something were the ones that put an independent person on the ground, at origin, with the ability to publish what they found. Not the ones with the nicest logo. The ones with a witness.
Critics are already calling the Anthropic commitment weak, and the obvious objection stands: one company embedding evaluators does not pace an industry, and the other two steps depend entirely on coordination nobody has agreed to. Amodei has also acknowledged the hardest version of the problem, telling CBS the toughest dilemma is what happens if China does not do the same.
All of that is true, and the essay is still a better artifact than the genre it sits in, because it names a specific driver, attaches it to a dated event, and pairs the ask with something checkable that the company did on its own.
Everyone can say they are slowing down. Only one of them handed out the badges.
Source: NBC News, “Two of the world’s top AI chief executives publicly agree on slowing AI development”
