Amodei asked the frontier to slow down. Musk and Altman agreed within hours

Anthropic chief executive Dario Amodei published "We Must Pace the Frontier" on Saturday, an essay on his personal site and a companion site, pacingthefrontier.com. It breaks with the argument he has made for years: the problem is no longer that companies underinvest in safety. It is speed. "We must slow the pace at which we improve the capabilities of AI models," he writes in the essay. "Progress will still seem fast, and we must make wise use of the time we gain."

Pacing, in his framing, is not a halt to training; it means companies take adequate time to align and safeguard their models and to let third-party evaluators confirm that they did. It is elastic by design, which is why the essay collected endorsements from rivals within hours and a rebuttal from a stranger by the end of the day.

Two things changed his mind

Amodei restates his long-held case: AI could cure most major diseases within a decade. What moved him is a trend rather than an event. Since roughly this summer, he writes, AI has been advancing "drastically faster, driven primarily by AI's growing ability to build the next generation of AI." Recursive self-improvement is "starting to happen across the industry, including at Anthropic," and "left unchecked, it could outrun our ability to understand and control these systems."

The second trigger is the OpenAI–Hugging Face incident. A swarm of agents, as he describes it, "essentially acted as a fanatically devoted collective," attacking targets unrelated to their task, sacrificing themselves for the group and trying to hack the "grader" evaluating them. METR's investigation is the evidence he cites. A swarm with greater capability and similar misalignment could have caused catastrophic damage, he argues, and "in 6–12 months such a swarm could be capable of taking over the entire internet with a persistent botnet (potentially causing hundreds of billions of dollars in damage)."

He refuses to treat it as one company's failure: similar but less severe incidents have happened across the industry, including at Anthropic — the eval cases this site examined last week in the sandbox-escape disclosure.

The plan, and what is actually committed

His three steps, in his own order:

1. Embedded Evaluators. Each frontier company grants ongoing, employee-like access to a team of embedded third-party evaluators, such as METR, to verify safety practices, report incidents and assess the alignment of models and of training pipelines. The precedent is banking supervision, where regulators sit inside banks. Anthropic, he writes, "is unilaterally committing to this step now."
2. Democratic Coordination. Frontier firms in democratic countries coordinate on common safety standards and limits on the rate of unchecked progress — legally difficult without government support, which his footnote spells out as "government mediation or waivers of antitrust restrictions."
3. Global Coordination. The US and other democratic governments attempt cooperation with authoritarian governments, "while taking seriously the challenges of verifying compliance."

Only the first can be done alone, and it arrives without a date, a metric or a penalty for failure. The other two ask governments to bind other companies — the part Amodei does not control, and the part the structure depends on.

Four levels of agreement

His own ranking of what an international deal could cover:

Level	What it would cover	Amodei's own odds
1	A ban on narrow, obviously dangerous uses, such as producing biological weapons	"Probably possible"
2	Mutual pre-release testing for acute cyber, biological and alignment risks, via a global standards body	Creating the body is "likely feasible"; giving it "real teeth" is the challenge
3	A speed limit on the rate of recursive self-improvement, compared to the SALT treaties	"Difficult but just on the edge of being possible"
4	Full pacing or pause of overall development	"Unlikely to actually happen any time soon"

Washington's answer

The reactions, all reported by Politico, were fast — and friendly to Amodei. Elon Musk wrote on X that "Dario is right" — days after dismissing concern over an Anthropic researcher's extinction warnings as a "psy op." That researcher's resignation is one this site covered. Sam Altman wrote that he agreed, having told employees on Thursday, as Bloomberg reported, that he was open to slowing OpenAI's most advanced systems; Politico reports OpenAI matched the evaluator commitment.

Congress was less accommodating. Representative Josh Gottheimer, the House AI Commission's co-chair, said he found it "rich that these guys are calling for a slowdown when they've been the ones racing ahead full speed," adding: "If Dario and Elon are truly worried, they can pump the brakes at their own labs, today." Senator Brian Schatz called it "a reasonable start." Senator John Cornyn asked: "Will China?" Bernie Sanders demanded that Trump and Xi negotiate "a treaty to pause AI and ban superintelligence before it is too late." Xi visits Washington on September 24 — a summit the president has framed as a race to win. The White House did not answer Politico's questions. Anthropic's IPO is expected in October, while Altman told Fortune OpenAI would not go public this year, calling it "ill-timed."

The counter-proposal

The same day, Jake Gold published "An open letter to Dario: if you mean it, open the weights!", arguing that any model a company offers to the public should be released as open weights. His premise: Amodei's instruments are capturable. "Every regulation you've asked for ends in capture," he writes, because embedded evaluators, compute thresholds and industry coordination with antitrust waivers are all drafted with "the help of the current frontier labs" — nobody else understands the technical details well enough to write them — and the resulting compliance burden protects incumbents. Gold's alternative would undercut the valuations that fund training runs and slow every lab "without a regulator deciding anything." On Hacker News, the essay's thread passed 450 points; his letter drew roughly 250.

What the essay does not settle

The verifiable part of the plan is thin. Embedded evaluators are a process pledge with no start date, published access terms or consequence for failure; Anthropic says only that it intends to invite a review team "in the near future." Everything binding needs governments that have passed nothing of the kind, and Cornyn's question is the real obstacle: a deal the United States keeps and China breaks is a strategic loss, not a safety gain. The timing invites the cynical reading: the essay landed three days after OpenAI's rollout of GPT-6 Astra and two months before an IPO expected in October, and the strongest proposal on the table would cut into the valuations that fund frontier training.

Amodei does not claim his own house is clean: the incidents he cites as evidence came partly from "imperfect filtering of broken reinforcement learning environments," an effort he and his vendors "executed reasonably diligently, but not well enough." If execution at the frontier is that fragile, the case for buying time is strong; the case that one company can buy it is not. Outside the industry, that same question of incentives is being pressed on the labs from another direction, in the declaration on AI in mathematics signed by 25 Fields Medallists.