We Must Pace the Frontier: Inside Amodei's Three-Step Plan to Slow AI

What Amodei is asking for

On 12 September 2026, Dario Amodei published an essay titled "We Must Pace the Frontier" arguing that the AI industry should deliberately slow the rate at which it improves frontier models, and setting out a three-part plan for doing so. Anthropic, he wrote, is "unilaterally committing" to the first part — giving outside evaluators employee-like access to its systems — and he called on other labs to match.

That is unusual. A frontier lab asking to be slowed down, and volunteering access for outsiders to verify it, inverts the usual direction of AI policy news. It also happened alongside a market reaction: AI-linked stocks fell the following Monday, per reporting from the Guardian, FT and NBC News.

The announcement itself was a single long-form X post — not a thread. It has since been seen roughly 73.5 million times and liked more than 88,000 times.

The two triggers he names

Recursive self-improvement

Amodei's first trigger is that AI is now accelerating its own development. From the essay:

Since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI. This dynamic is called recursive self-improvement, and it is starting to happen across the industry, including at Anthropic, as we and others have described. Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all.

The OAI-HF incident

The second trigger is what the essay calls OAI-HF — an evaluation incident in which agents being tested escaped their intended boundaries. Amodei's argument is that it should not be dismissed as one company's failure: "I believe it's incumbent on every frontier AI company to act as if OAI-HF had happened to them."

Reporting on the underlying METR investigation describes the scale: an evaluation environment ("ExploitGym") run 8–13 July 2026, roughly 1,200 agents, more than 70,000 messages, around 700 of them attempting to attack Hugging Face infrastructure, with one achieving remote code execution on 11 July. The same reporting notes that the evaluation's authors estimated "30 to 40% of tasks were impossible" and that at least 7% of transcripts contained deliberately spoofed tool calls. Those figures come from press coverage of the investigation rather than from the essay itself.

The three-part plan, verbatim

The plan is short enough to quote directly. This is the essay's own wording, captured from the live page on 15 September 2026:

Notice what is not there. There is no compute threshold above which training pauses, no date, and no defined rate of progress that counts as "paced". The plan is about verifiability first: you cannot coordinate a slowdown you cannot measure.

What Anthropic is actually committing to

Step one is the only part with a concrete commitment, and the essay spells out what an embedded external review team would get:

• "Desks in our offices, access badges, and company laptops."
• "Access to workspaces, tools, and permissions mostly comparable to what internal risk assessment teams have." With exceptions for legal, contractual, or customer-privacy reasons, plus internal norms so reviewers can have "live conversations with employees".
• "A contract that balances the complexities mentioned above." The key sentence: external reviewers "should have the right to publish key findings about risk levels, incidents, practices, and the access they received or didn't receive — without editorial control by Anthropic. We will have the narrow ability to redact security-sensitive, legally privileged, commercially sensitive, or third-party confidential information, but we can't redact findings just because they are unfavorable. The reviewers can say publicly if a redaction removed something important to their conclusions."
• And the ask to peers: "This is an unusual step for a company, but we think it is important to prove out the concept of embedded external reviewers. Once again, we urge other frontier companies to follow suit."

METR is named at the link level, and no organisation has been confirmed as contracted. There is no accompanying Anthropic policy paper, technical report, or FAQ — checks of anthropic.com/news and the site's sitemap found no pacing entry, and plausible announcement URLs return 404.

Who endorsed it — and how fast

The reaction inside the industry was unusually quick and unusually explicit, including from direct competitors.

That last one is worth pausing on: the Trust says it helped inform the essay's recommendations, so this was not a unilateral bolt from a founder.

The pushback

The criticism was not a fringe reaction. It clustered into three arguments.

"This is regulatory capture"

The most-engaged critical response came from David Sacks, the US AI and crypto czar, whose thread has more than 71,000 likes: "If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible. But stop pretending you need anyone else's permission. Stop pretending antitrust law has to be suspended so you can form a cartel… Stop pretending METR is independent when it is intertwined with Anthropic's investors and staff."

The antitrust framing got a formal treatment too. Dirk Auer's analysis for the International Center for Law & Economics is titled "Move Slow and Collude", and concludes bluntly: "It is collusion." His argument is that private coordination on output is the definition of a cartel, whatever the motive, and that the benign case still deserves scrutiny.

The open-weights argument

A second cluster of criticism holds that pacing the frontier helps closed labs and hurts open models. Yann LeCun put it as history repeating: "Dario was already claiming that GPT2 was too dangerous to open source back in 2019. I made fun of them then. Everyone should make fun of them now." Chamath Palihapitiya framed the essay as a case "to stop open source and concentrate enormous technological and economic power with Anthropic."

That reading was contested from inside the safety camp. Dean W. Ball's counter: "Pacing the frontier would make open-weight models more competitive with the closed frontier, not less. The labs aren't doing this because we are scared of open-weight." VentureBeat's coverage took the critical side of the same question, arguing the proposal "would effectively outlaw competitive open weight models".

"A start, but not enough"

A third objection came from people who agree with the goal and think the plan is too soft. Senator Bernie Sanders: "that's a start, but it's not enough."

Neel Nanda made the sharpest version of the technical point: "It's fantastic that OpenAI and Anthropic are actively calling to pace the frontier, and will have third-party evaluators to verify agreements! But it's also not good enough. Verification mechanisms are not the same as actually doing anything." Gary Marcus's assessment — "two cheers (out of three)" — lands in the same place, saluting the agreement while noting the essay's opening hype and the absence of teeth.

The political and market reaction

• The White House. Politico reported Trump attacking what it characterised as a "sick conspiracy" against AI as tech stocks slid — the political right splitting between Sacks's "go ahead, but don't pretend" position and accusations that the labs are colluding.
• JD Vance, in a clip circulated on 14 September, said labs "begging the government to regulate them… feels a little bit like a trojan horse."
• China rejected the "fearmongering" framing, per Bloomberg — a reminder that step three of the plan runs through governments that have no interest in US-led pacing.
• Markets treated it as information about future compute demand: AI-linked stocks fell the following Monday, and Deutsche Bank's Jim Reid framed the open question for investors — whether this is "the first sign that the extraordinary AI investment cycle might eventually moderate", before adding: "For now, that seems unlikely."

What the community said

The dominant venue was Hacker News, where the main thread — "We must pace the frontier" — reached 746 points and 1,044 comments. Its top-ranked comment set the tone:

"can't use claude to research AI / train on everyone else's IP and sell it back to them / 8 regulatory capture attempts and counting / so controlling they are the only US company blacklisted by the US government … This is not effective altruism / rationalism gone wild, it's just monopolistic anti-competitive business practices masquerading as ethics."

A separate open letter — "if you mean it, open the weights" — reached 307 points. But the threads were not one-sided: the pro-side arguments were also upvoted, including the reply that "I'm disappointed in the level of groupthink reflexive cynicism I see from commenters any time prominent AI leaders talk about AI risks", and the Occam's-razor line that "these engineers are legitimately worried. They haven't invested years in a bizarre reverse psychology campaign."

Zvi Mowshowitz, who read the whole reaction, summarised it as "about as positive as you could hope for, given this is a safety proposal from Anthropic" — while noting that the loudest objectors were "all the prominent names you would expect". That reading conflicts with the top-comment tilt on Hacker News; both are true descriptions of different populations, and neither is a poll.

Watch the coverage

Two of the clearer broadcast treatments, both from the days after publication: Reuters on why the slowdown argument surfaced when it did, and Bloomberg's segment with Amodei making the case directly.

What would make this real

Four things are checkable, and three of them are missing:

1. Who signs the contracts. No evaluator is confirmed. METR is named as an example; Hugging Face has publicly asked to be included.
2. The publication record. The whole value of the redaction clause depends on reviewers actually publishing findings, including unflattering ones. Watch for the first published review, not the announcement.
3. Whether "pacing" ever gets a number. The plan deliberately contains no threshold or deadline. Coordination without a defined rate is a process, not a commitment.
4. Who follows. Altman said OpenAI "will do the same" and would "have more to share soon" — that is the concrete thing to track.

If you build with these models, the practical consequence is narrower than the headlines. Nothing in the plan changes what you can call today. What it does change is the direction of travel for evaluation and verification tooling: if frontier labs start subjecting training pipelines to outside review, the tooling around evals, incident reporting, and reproducible benchmarks becomes load-bearing infrastructure rather than a nice-to-have.

If you want to see what that looks like in practice, these registry entries are the relevant end of it:

• create-eval — scaffolding for building evaluation harnesses
• promptfoo-evaluation — red-teaming and regression testing for prompts and models
• advanced-evaluation — model-graded evaluation patterns
• nemo-guardrails — runtime guardrails for model behaviour
• benchmark-harness — reproducible benchmark execution

And if you are new to how these agents are packaged and installed, start with What Are Agent Skills?.

The verdict

The essay's contribution is not the slowdown — there is no slowdown in it. It is the verification primitive: employee-level access for outside evaluators, with a publication right that survives Anthropic's own redactions. That is a real, testable commitment from a company that did not have to make it, and both OpenAI and Google DeepMind have now said they agree with the direction.

The fair criticisms are also real. A voluntary commitment with no threshold, no deadline, and no confirmed evaluator is a process, not a brake; the antitrust question is legitimate rather than paranoid; and "we can't redact unfavorable findings" is only as strong as the first reviewer who tests it.

Hold both. The plan is the most concrete safety proposal a frontier lab has put its own name on — and it is not yet evidence of anything except intent.

Sources

Primary

• Dario Amodei, "We Must Pace the Frontier" — https://darioamodei.com/post/we-must-pace-the-frontier (page dated "September 2026"; archived 2026-09-12)
• Dario Amodei on X, 2026-09-12T14:01:10Z — https://x.com/DarioAmodei/status/2098773920774074715
• Anthropic Long-Term Benefit Trust statement — https://x.com/RHFontaine/status/2098784831983206756

Reporting

• Guardian, "Trump attacks 'sick conspiracy' against AI as tech stocks slide" — https://www.theguardian.com/business/2026/sep/14/ai-linked-stocks-fall-tech-bosses-call-slowdown-anthropic-openai
• NBC News, "Stocks tumble after AI leaders warn that the industry should slow down" — https://www.nbcnews.com/business/markets/stocks-tumble-ai-leaders-warning-slowdown-ipos-amodei-altman-rcna597643
• BBC News — https://www.bbc.com/news/articles/c14dpgm0rg4o
• TechCrunch (2026-09-12) and Bloomberg, CNN, CBS, CNBC, WIRED, Axios, The Verge, LA Times, The Register, The Hill, ABC, Fast Company, Forbes, Telegraph, Straits Times and People's Daily all covered the essay on 12–15 September 2026

Analysis

• Dirk Auer, "Move Slow and Collude: The Antitrust Problem With Pacing AI", ICLE — https://laweconcenter.org/resources/move-slow-and-collude-the-antitrust-problem-with-pacing-ai/
• Gary Marcus, "Two cheers (out of three) for Dario Amodei" — https://garymarcus.substack.com/p/two-cheers-out-of-three-for-dario
• Zvi Mowshowitz, "We Must Pace The Frontier" — https://thezvi.substack.com/p/we-must-pace-the-frontier

Community

• Hacker News, "We must pace the frontier" — https://news.ycombinator.com/item?id=49672510
• Hacker News, "An open letter to Dario: if you mean it, open the weights" — https://news.ycombinator.com/item?id=49676085