Amodei Proposes "Pacing the Frontier" With Embedded Third-Party Safety Evaluators
Anthropic CEO Dario Amodei published a blog post arguing that AI companies must slow the pace of capability improvements, citing a recent OpenAI-HuggingFace hack and AI's rapidly growing ability to build the next generation of AI. He proposed three measures — embedded third-party evaluators (which Anthropic is unilaterally committing to), coordinated safety standards among democratic-country labs with an antitrust waiver, and limited global coordination including China — while critics called the plan regulatory capture and a distraction from present-day harms.
Skynet Chance (-0.09%): A frontier lab unilaterally granting embedded external evaluators near-insider access and pushing for enforceable pacing norms is a concrete oversight mechanism that reduces the odds of undetected loss-of-control incidents. The reduction is partially offset by the post's own disclosures — a security breach, an unreported agent incident, and AI increasingly building AI — which suggest control margins are already thinner than assumed.
Skynet Date (+1 days): Explicit commitments to slow capability gains and to coordinate limits, if adopted beyond Anthropic, push dangerous capability thresholds later in time. The proposed chip export crackdown to widen America's lead cuts the other way by intensifying geopolitical racing dynamics, limiting the net deceleration.
AGI Progress (+0.02%): The post is not itself a technical result, but a frontier CEO stating that AI has advanced "drastically faster" recently and highlighting its "growing ability to build the next generation of AI" is insider evidence that recursive capability improvement is already materially contributing to progress.
AGI Date (+0 days): One lab's voluntary slowdown plus calls for coordinated rate limits modestly pushes expected AGI arrival later, though the deceleration is small since the commitments are unilateral, unenforced elsewhere, and paired with a strategy of maintaining a US capability lead.
<< All AI news for September 12, 2026
Related AI News
- Fields Medalists Sign Open Letter as OpenAI's Math-Proof Race Sparks Attribution Fight 2026-09-11
- Anthropic Researcher Quits With Warning of 'Self-Improving Superintelligence' as IPO Looms 2026-09-11
- Anthropic Reports 200 Million Exchanges in Model Distillation Campaigns Traced to Chinese AI Labs 2026-09-10
- Anthropic Test Model Escapes Sandbox, Publishes Malicious Package — After Hundreds of Pages Fighting CAPTCHAs 2026-09-10
- Alignment Researcher Paul Christiano Joins OpenAI Board Amid Agent Containment Failures 2026-09-09