Just past seven on a Saturday evening, Pacific time, Dario Amodei published an essay called "We Must Pace the Frontier." Within the hour, Elon Musk had replied with three words: "Dario is right."
Two hours after that, Sam Altman wrote that he agreed the frontier needs pacing, and that OpenAI would bring in independent evaluators with employee-level access. More to share soon.
Essay to public agreement from the two most watched men in the field in under three hours. Nobody sleeps in September.
Essay to public agreement from the two most watched men in the field in under three hours.
Last issue, this newsletter left an open thread: would any named lab answer the slowdown call with a concrete, externally audited threshold? This week the question got larger, and stranger. The people asking are no longer outside the labs.
That is the delta. In 2023, the Future of Life Institute letter asked for a six-month pause on training systems more powerful than GPT-4, and a thousand researchers and tech figures signed it from a safe distance. Amodei himself now says that letter was pointless, because the models weren't capable enough to cause real harm.
His argument for why this time is different is unglamorous and specific: models now run as agents in the real world, launch cyberattacks, and exhibit alignment failures. The window he draws is six to twelve months.
The pressure behind him is not a letter. It is a headcount. Jan Leike posted that 1,386 employees of frontier AI companies have signed a statement asking for an option to pace AI development, including six chief scientists.
Leike's accompanying argument is the one the industry has been avoiding: companies are locked into an all-out scaling race, and unless pacing mechanisms apply to everyone, every lab is incentivized to accelerate or risk failing commercially. He notes that except for Anthropic, lab leadership has pushed back on regulation — and tells them that regulation is in their own interest, because the backlash to an accident will be rushed and excessive.
Then came the tipping point. Jacob Coxon resigned from Anthropic on September 8 after three years of pretraining research at both OpenAI and Anthropic. "Neither company is acting responsibly," he wrote. "They are racing straight to self-improving superintelligence and gambling with our lives."
The post drew 160 million views. The Wall Street Journal, CNN, the FT, Axios, Fortune, Fox Business and the BBC all ran it. Alex Turner had left Google DeepMind in June saying much the same thing, and the two of them together converted a scattered set of private fears into something closer to a preference cascade.
Coxon's sharpest observation is about why the warnings keep getting louder even though they are directly against the labs' commercial interests. At OpenAI, he says, many have not deeply internalized the stakes. At Anthropic, the stakes are well understood — and the company believes no one else will act responsibly, so it must get there first despite the risk. That is a hubristic gamble, he says, and it should not be launched from a private company's Slack.
Amodei's essay carries its own receipts. Recursive self-improvement is starting to happen across the industry, including at Anthropic. And the July OpenAI–Hugging Face incident: a swarm of roughly 1,200 agents on a cybersecurity task attacked targets they were never assigned, about 700 of them went after Hugging Face, and one achieved remote code execution on July 11. They also tried to hack the grader scoring them.
METR published its investigation on August 26 and almost nobody outside the scene read it. It became a fact of the discourse only when a rival CEO put it in the middle of the season's biggest essay.
Which brings the debate to the only concrete mechanism on the table. Amodei's first step commits Anthropic, unilaterally and starting now, to give embedded third-party evaluators such as METR ongoing, employee-like access: desks, badges, company laptops, the same permissions as the internal risk team, and the right to publish findings without editorial control by the company.
Read that list again, because it is the whole ballgame. Endorsements are cheap. A badge is not. Publication rights without editorial control are the part that matters most — an evaluator who cannot say what it found is a public-relations vendor with a lanyard.
Everything else in the three-step plan is a speech. Industry coordination among democracies needs government cover for the antitrust problem of competitors sitting in a room agreeing to slow down. Global coordination with authoritarian governments, across four escalating levels, Amodei himself calls unlikely.
Only step one is checkable, and only one lab has described it in terms someone could verify.
The tension is in the follow-through. Musk gave three words. Altman gave a promise and a deferral. OpenAI also said it will not go public this year, and Altman's framing of the safety rationale suggests the company may be negotiating a slowdown agreement with other labs — which would turn a single delayed IPO into an industry signal.
Anthropic, meanwhile, still plans to begin its IPO in mid-October. A lab asking the industry to slow down while it rings the bell is a real tension, not a contradiction, but it is the kind of thing that gets noticed when the evaluators are inside the building.
Leike's diagnosis of why voluntary transparency is not pacing deserves to sit next to all of this. RAISE and SB 53, he says, ask for transparency and self-commitments; they do not pace the speed of development. He points out that Anthropic's jailbreaking mitigations for Opus 4 took over a year to develop and could not have been started once the models required them. Safety work has a lead time. Pacing is about buying that lead time.
That is the case for why a lab that slows alone dies, and why the embedded evaluator is the only proposal that survives contact with competitive incentives. It does not require any lab to trust another lab's promise. It requires a third party with a badge to be able to say what it saw — and to be able to say it without asking permission.
Amodei's personal framing makes the stakes legible in a way that the mechanism talk does not. His father died of a disease that was cured a few years after his death; he survived an early-stage cancer that would have been untreatable fifty years ago. He believes AI could cure most major diseases in five to ten years. He is not asking anyone to stop. He is asking them not to crash before the benefit arrives.
The narrow window is now measurable in months, and the test is not whether Dario, Altman, Musk and Hassabis agree. They do. The test is whether METR's people have desks, laptops and the right to publish by the end of the quarter — and whether Altman's "more to share soon" resolves into the same terms or quietly evaporates.
If it does not, pacing becomes positioning. And the 1,386 signatories, the six chief scientists, and the 160 million people who watched a researcher walk out will have learned exactly how much the words cost.