Dario Amodei's 'Pace the Frontier' essay: Anthropic grants third-party evaluators permanent employee-level access
scottleibrand · x · 2026-09-13
Anthropic CEO Dario Amodei published a new essay, We Must Pace the Frontier, arguing the AI industry should deliberately slow frontier capability progress, with a three-part plan.
- First commitment already actioned: Anthropic is unilaterally committing to give third-party evaluators permanent, employee-level access to its systems, so they can verify safety-measure adherence, report incidents, and assess alignment during training.
- Core argument: The essay opens with AI's potential benefits before acknowledging risks — framed as the right way to present the technology to the public — and stresses that pacing must not cost democracies their AI lead over authoritarian regimes, chiefly the CCP. Key quote: "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."
Researcher deredleritt3r called it her favorite Dario essay since Machines of Loving Grace, praising the embedded third-party evaluator proposal as something every frontier lab should adopt, and noting it is the first pacing proposal explicitly tied to not ceding AI leadership to China.
Related event: Anthropic CEO Calls for Slowing Down Frontier AI Development(55 posts)→
More from AGI Musings
- "Pace the frontier" mocked as the new "two weeks to slow the spread" — StewartalsopIII · 2026-09-13
- Critics question METR's independence in Anthropic's 'Pace the Frontier' pledge — kristoph · 2026-09-13
- Scobleizer: Most in SF think OpenAI is winning — so rivals have big incentives to call for a slowdown — Scobleizer · 2026-09-13
- Nadella: the AI moat isn't the model, it's the learning loop only your company can run — rohanpaul_ai · 2026-09-13
- Models as perpetual teenagers: why capability growth keeps outpacing model maturity — repligate · 2026-09-13
- The 'Brooklyn Project' dubbed 'our generation's Manhattan Project' sparks discussion — MaliciousMussel · 2026-09-13