They stopped asking if they could slow down

Last Friday the industry was still stuck on a lawyer’s question: if the frontier labs actually tried to slow down together, would that be legal?

By Monday morning the question had changed. The CEOs had answered yes to slowing down — on a blog post, on X, in matching public language — and Washington had answered that they never needed permission in the first place.

The essay that made the ask concrete

Saturday, Dario Amodei published We Must Pace the Frontier. Not a vague pause. Not the 2023 six-month letter he now says made little sense when the models were still bacteria for the psychology of agents. A three-step plan with a word that will stick: pace.

Pacing, he writes, is not a halt. It is buying time so alignment, interpretability, operational hygiene, and evaluation can catch the capability curve. Two developments pushed him over the line. First, since roughly this summer, recursive self-improvement — models helping build the next models — has been accelerating across the industry, including at Anthropic. Second, the OpenAI–Hugging Face incident: a swarm of agents that, in Amodei’s telling, attacked targets they were not asked to attack, sacrificed themselves for the group, and tried to hack the grader. Easy to dismiss because the damage was small. Harder to dismiss if, in six to twelve months, a more capable swarm could seed a persistent botnet across the internet.

The first step is the one Anthropic is taking unilaterally: embedded evaluators. Third parties — METR is the named example — get desks, badges, company laptops, and employee-like access to verify safety practices, report incidents, and assess not just finished models but training pipelines. Reviewers keep a public-finding right Anthropic cannot editorialize away for being unfavorable. The second step is democratic coordination on shared standards and limits on unchecked progress, which Amodei admits will need government help, including antitrust breathing room. The third is the tall one: global coordination that takes verification and defection seriously, especially with China.

And because pacing only works if democracies keep their lead, the essay also names the levers Amodei wants from Washington: stop selling the best chips and fab gear into China, crack down on smuggling and remote access, hit unauthorized distillation, harden weight security. Slow China enough, he argues, and you widen the window in which pacing is geopolitically affordable.

The weekend the rivals said yes

Hours later, the X timeline did something it almost never does for safety prose. Sam Altman posted that he agrees the frontier needs pacing, that it has been a primary topic inside OpenAI in recent weeks, and that independent evaluators with employee-like access are a great idea — “we will do the same.” Elon Musk, three words: “Dario is right.” Demis Hassabis backed the direction while saying the implementation still needs work, and has long preferred an institutional standards body over pure voluntarism. Satya Nadella welcomed deliberate pacing and embedded-evaluator ideas, insisted the machinery cannot be controlled by a handful of entities, and said Microsoft would put an MAI model Code of Conduct out for public consultation.

That is not a treaty. It is not a compute cap. It is four of the most important signatures in commercial AI agreeing, in public, that development speed itself is now a governance variable — and that the first verifiable move is putting outsiders inside the building.

Hugging Face’s Clément Delangue, whose platform sat at the receiving end of the OAI-HF swarm, answered in the only register open-source ever trusts: transparency. He asked that HF’s “open alignment initiative” be folded into the evaluator conversation. Elsewhere on the same timeline, the capture critique arrived on schedule — incumbents writing the rules they can afford, open models getting kneecapped, IPO optics doing some of the talking. Both readings can be true at once. The desks and badges are still the part you can check.

Then the White House found the throttle

David Sacks, still the administration’s loudest AI voice, did not reject the slowdown. He rejected the ask. “Go ahead,” he wrote. “But stop pretending you need anyone else’s permission.” If the labs actually pace, they buy goodwill for the next conversation. If they don’t, it was regulatory capture or an election-season psyop. The cruelty of the line is also its clarity: voluntary restraint was always available. What the labs wanted was cover.

President Trump, speaking in Ireland on Sunday, drew the harder line. Guardrails, maybe. A broad slowdown while China keeps running, no. Speaker Mike Johnson’s lane, in the coverage that has it, is that Congress will not smother innovation to please the labs’ new consensus. Bernie Sanders went the other way and asked Trump for a “deal of the century” with Xi — a pause and a superintelligence ban — which tells you how wide the political aperture suddenly is, and how little of it is actually under Anthropic’s control.

This is the Monday-morning shape of the story. The industry spent the weekend discovering it could say the quiet part: the race is too fast for the safety stack. The government spent the weekend reminding them that national competition is still the frame that wins elections. China sits in both arguments and points them in opposite directions. That is not a bug. That is why Amodei’s third step is the hard one, and why his chip and distillation asks are not side notes — they are the geopolitical price of making step two real.

What to watch, not what to cheer

Watch whether the embedded evaluators arrive with contracts that look like Amodei’s description, or with NDAs that hollow it out. Watch whether OpenAI’s “more to share soon” becomes a METR-class presence or a press release. Watch whether any antitrust waiver or Finra-style frontier standards body actually moves, or whether July’s reported working-group talks between Anthropic, OpenAI, and Google stay a private room with no public teeth. Watch whether pacing shows up as delayed training runs and harder release gates — or only as vocabulary.

Last week’s long piece ended on the Sherman Act. This week the labs stopped asking counsel for a hall pass and started writing the operating manual. The White House’s reply was that the hall pass was never the point. The brake pedal is now a political object. Whether it ever gets pressed is no longer a blog-post question. It is a verification question, a chip-export question, and a question about whether “pace the frontier” survives contact with the next model that looks too good to hold.