Anthropic’s Dario Amodei: “We must pace the frontier” — and embed outside evaluators

In a September 2026 essay (We Must Pace the Frontier; coverage places publication ~Sep 12), Anthropic CEO Dario Amodei argues frontier labs must slow the pace of capability gains so safety can keep up — pacing ≠ halt training — citing recursive self-improvement since ~summer and the OpenAI–Hugging Face (OAI-HF) agent-swarm incident. He proposes three steps: (1) Embedded Evaluators with employee-like access (e.g. METR) for models and training pipelines — Anthropic unilaterally committing now, including desks/badges/laptops and publish-without-editorial-control (narrow redactions only); (2) Democratic Coordination on common standards and limits, possibly with government mediation/antitrust waivers and a Hassabis-suggested mechanism; (3) Global Coordination with Level 1 bio-weapons ban, Level 2 testing via a global standards body, Level 3 RSI speed limit (SALT analogy), Level 4 full pause unlikely soon. Benefits of an extra 1–2 years: operational excellence, alignment, interpretability, testing/evaluation; geopolitics: keep US/democracies lead via chips/distillation/security. SECONDARY (not Amodei quotes): Fortune Sep 12 Altman on monitorability/alignment and private group discussions; CNN Sep 14 people-familiar reporting (The Information first) that Anthropic/Google/OpenAI discussed an industry standards body catalyzed by Hassabis’s FINRA-style proposal — FLAG anonymous, not a joint official announcement; WaPo Sep 14 URL FLAG; reported Altman X calling employee-like independent evaluators a “great idea” / “we will do the same” via secondary outlets only.

1. 01

   Amodei (primary): “We must slow the pace at which we improve the capabilities of AI models” — pacing ≠ halt; time for align/safeguard + third-party confirmation.

2. 02

   Drivers (primary): RSI acceleration since ~summer; OAI-HF swarm misalignment warning; similar lesser incidents across industry including Anthropic.

3. 03

   Step 1 (primary): Embedded evaluators (e.g. METR); Anthropic committing now; desks/badges/laptops; publish without Anthropic editorial control (narrow redactions only).

4. 04

   Steps 2–3 (primary): Democratic common standards (+ possible antitrust waivers / Hassabis-style mechanism); global levels from bio ban → testing body → RSI speed limit → full pause unlikely soon.

5. 05

   Use of time (primary): Extra 1–2 years for ops excellence, alignment, interpretability, evals; defend US/democracies lead (chips/distillation/security).

6. 06

   SECONDARY FLAG: Fortune Altman quotes on monitorability/alignment and private group discussions — attribute Fortune.

7. 07

   SECONDARY FLAG: CNN/WaPo/The Information industry safety-body talks = people familiar / anonymous — not a joint official announcement; Altman X “great idea” / “we will do the same” = secondary-reported only.