Amodei: Pace the Frontier With Embedded Evaluators; Altman and Musk Say They Agree
On September 12, 2026, Anthropic CEO Dario Amodei published “We Must Pace the Frontier,” committing Anthropic unilaterally to METR-class evaluators with employee-like access. Sam Altman said OpenAI will do the same; Elon Musk said Dario is right. Pacing is not a halt.
TLDR
Dario Amodei on Saturday, September 12, 2026 published “We Must Pace the Frontier” (~3,800 words) on darioamodei.com—not the Anthropic newsroom. Drivers: summer RSI (AI building the next AI) and the OpenAI–Hugging Face swarm. Warning: in 6–12 months a more capable misaligned swarm could take over the internet via a persistent botnet, hundreds of billions in damage. Pacing ≠ halt. Three steps: (1) embedded evaluators (METR-class) with employee-like access (desks, badges, laptops) and publish-without-editorial-control—Anthropic unilaterally commits; (2) verifiable pacing of capability gains; (3) keep US/China lead via chips + anti-distillation. Sam Altman on X: independent evaluators with employee-like access is “a great idea” and “we will do the same.” Elon Musk: “Dario is right.” Not the September 9 METR incident SOW—this is industry-wide permanent embedding.
What he is (and is not) proposing
| Item | Essay Sep 12 |
|---|---|
| Halt training? | No |
| Anthropic commit | Unilateral embedded evals |
| OpenAI | Altman: will do the same |
| xAI | Musk agrees (no process detail) |
Product-line de-dupe: not Coxon resignation (Sep 9), not METR incident assessment (Sep 9), not Aug 27 cyber letter.
Why this story matters
The three loudest frontier CEOs just agreed in public to inspectors in the building. That is cheaper than a halt and more than a blog. Watch: badges and laptops, not adjectives—and whether China policy in step 3 guts the coalition.