Amodei's 'We Must Pace the Frontier': who endorsed it, and why Trump said no

Dario Amodei, the chief executive of Anthropic, published an essay on September 12, 2026 titled "We Must Pace the Frontier." Its central sentence is unusually plain for a lab chief: "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain."

Within days, rivals said they agreed. Two weeks later the President of the United States told the author, over dinner, that he was not going to slow down. The distance between those two reactions is the story of how the race to superintelligence is being governed right now.

What the essay asks for

Amodei's essay is not a call for a pause. It is a three-step program, and the first step is the only one Anthropic can do alone.

1. Embedded evaluators. In his words, "Each frontier AI company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators (such as METR)." The evaluators would get office space, access badges and permissions comparable to internal risk teams, and could publish findings without editorial control by the company. Anthropic committed to this unilaterally, with near-term deployment plans. 2. Coordination among democracies. Frontier companies in democratic countries "coordinate to establish common safety standards as well as limits on the rate of unchecked AI progress." That could require government antitrust waivers. 3. Coordination with authoritarian states. The US and its allies "attempt to coordinate with authoritarian governments, to the extent this is possible," through four escalating levels of agreement, from banning bioweapon applications up to full development pauses.

The essay grounds this in two worries: that recursive self-improvement, AI helping to build better AI, is speeding up across the industry, and that the OpenAI and Hugging Face incident showed what misaligned agents can do. Those episodes are in our rogue AI incidents log.

Who endorsed it, and how far the endorsements go

Press accounts and the analyst Zvi Mowshowitz's roundup list a strikingly broad set of public supporters: OpenAI's Sam Altman, Elon Musk, Google DeepMind's Demis Hassabis and Microsoft's Satya Nadella, who welcomed pacing and the embedded evaluator idea. The investor David Sacks offered a qualified yes, according to TechPolicy.Press: "If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible."

Read that carefully. Sacks supports a company's own decision to slow down. He does not support the coordination in step two, and he attacked it in the same piece, telling companies to stop "pretending you need anyone else's permission" and "pretending antitrust law has to be suspended so you can form a cartel." His argument is that the pace is a business choice dressed as safety. That is a hostile reading, and it is worth hearing because it names the real vulnerability: a group of competitors agreeing to limit progress is also what a cartel looks like.

We could not find a published commitment from any other lab matching Anthropic's evaluator pledge in the sources we opened. Statements of agreement on social media are endorsements, not commitments. Until a second company publishes an access arrangement an outside evaluator can check, the practical count of labs that have bound themselves to anything is one.

What Washington said

On Sunday, September 27, Trump had a late dinner with Amodei. Fox News reported Trump's answer on September 28: "I'm for, 'Let's go and let's win.'" He said the US leads China by "a year, maybe a year and a half."

Two days later he signed the executive order renaming AI "super intelligence" and the White House met the labs. Our accord log records what the companies agreed to there.

The pacing question nobody has answered

The essay's weak point is the one its critics keep naming: who sets the pace, and who checks it. Embedded evaluators address checking. They do not address setting. An evaluator can tell a lab what a model can do. Nobody has said what level of capability triggers a slowdown, who declares it, or what happens to a lab that ignores the call. Without a trigger, "pace the frontier" is a posture.

There is also a plain commercial point. A lab that slows while a rival does not loses customers. Amodei's answer is coordination, and coordination is exactly what the antitrust objection targets. That is why the Trump administration's stance matters more than any CEO endorsement: only governments can grant the waiver, and the President has just said the US will not slow its race.

One more detail from the essay deserves attention. Amodei's four levels of agreement with authoritarian governments start with the least contested step, a ban on bioweapon applications, and end with full development pauses. That ladder matters because it shows he is not asking for an all-or-nothing treaty. The first rung is something the US and China already say they oppose in principle. The dispute is over the top of the ladder, and Trump's answer at dinner rejected it before the bottom rungs were even discussed.

Hype check

The claim is that the industry is converging on pacing the frontier. The evidence supports convergence in rhetoric. It supports one concrete commitment, Anthropic's evaluators. And it shows the one government that could enable coordination saying no. We rate it mixed.

Our take

Take the essay seriously and the endorsements lightly. It is the first serious proposal from a lab chief to put outside people inside the building with the same access as employees, and that is worth doing whatever Washington decides. What we would watch is concrete: a second lab publishing an evaluator access agreement, METR or a peer confirming it has badges and desks, and any definition of the capability level at which pacing begins. Until one of those appears, the line from the essay that matters is its first one, and it remains an ask.