Dario Amodei published “We Must Pace the Frontier” on his own site this month: a three-step plan for slowing the AI industry, of which Anthropic has unilaterally adopted the first. Step one is called Embedded Evaluators. Outside teams — METR is the example he names — get ongoing, employee-level access inside a frontier lab. Not a demo day, not a red-team weekend. Standing access to verify adherence to safety commitments, receive incident reports, and assess alignment during training rather than only on the finished model. The Hacker News thread ran to 656 points and 922 comments.
The reflexive conservative read arrived within the hour: regulatory capture in a lab coat. The man is behind, or ahead, or both, and the framework is a moat. It is a satisfying read, and it misreads the mechanism.
A Cartel Wants Licensing, Not Auditors
A cartelist does not invite a stranger to sit inside his own training runs. He asks the government to raise his rivals’ costs — licenses, registrations, a compliance department every competitor must hire. Licensure protects incumbents. Inspection exposes them. If you want the cynical reading of Amodei’s essay to hold, you need a cynical mechanism, and this is not one.
His own definition is worth reading twice. Pacing, he writes, “does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models, and for third party evaluators to confirm this.” The load-bearing word is confirm. The essay is not really a proposal to go slower. It is a claim about evidence — that the industry is running an experiment without an instrument, and that the instrument has to be built by someone with no stake in the result. That claim is either true or false, and no amount of moat-spotting answers it.
An Audit With No Reader Is a Diary
Here is where step one is thin, and where the amendment should be written. Employee-level access is a genuine concession. But the evaluator observes on behalf of the lab. Incident reports go into the lab. Nothing in the described commitment requires that the finding — or even the bare fact that an incident occurred — ever reaches a person outside the building.
Financial audit is the obvious comparison, and it cuts against Amodei. That edifice works because the auditor signs an opinion the market reads and a regulator can subpoena the workpapers. A self-reported incident log is a diary. So take the deal, then price it. Publish the incident logs on a schedule, in whatever redacted form protects genuine trade secrets. Give embedded evaluators a duty to escalate to a body that can act on what they find. And forbid — contractually, in the commitment itself — terminating an evaluator over the content of a finding. Anthropic gets to be first inside the tent; the public gets receipts. The company needs this to look credible, and it only looks credible from the outside.
The Only Bridge to Beijing Runs Through an Inspection Regime
Amodei is explicit that worldwide pacing would require cooperation with China — “the autocratic country with by far the most advanced AI capabilities” — and that there will be “stark limits on what can be achieved, especially at first.” He also warns against restraining American capability on the assumption that Beijing will do the same.
That is where the unilateral-disarmament objection lands, and it mistakes the deliverable. The scarce good in any eventual arrangement with China is not restraint. It is the ability to see inside a lab. Arms control is overwhelmingly inspection plumbing — definitions, thresholds, access rules, and a place to file the report. You cannot sign a treaty you cannot verify, and you cannot verify what you have never practiced verifying at home. Build the standard for what “verified” means and you write the terms of whatever arrangement eventually exists. Skip it and the reference standard arrives from somewhere else, in a document nobody in this debate helped draft.
Which is the case for the plan that requires believing nothing about Amodei’s sincerity. You build the inspection regime precisely because you don’t trust him.
Cheer or Jeer, Then Send the Invoice
The jeer has a price. If voluntary, private inspection gets laughed out of the room as theater, the vacuum does not stay empty. It gets filled by statutory compliance regimes drafted by legislators who cannot define a floating-point operation, and by discovery in the litigation that follows the first genuinely bad incident. That is the expensive version of the same outcome, arriving later and worse.
And the audience is sitting right there. Those 922 comments are not a mob; they are the only constituency on earth that will actually read an incident log if someone publishes one.
One evaluator at a competing lab put it plainly in a Slack DM the week the essay posted: “The access is real. The reporting line is the whole ballgame. If the only person who sees the incident report is the guy whose model had the incident, we’ve built a very expensive mirror.”
Amodei has offered a framework. The correct response is not applause and not contempt. It is a counteroffer with the reporting line drawn to the public — because the people who should be angling hardest for that are the ones who think the entire genre is theater.