Dario calls to pace the frontier as Anthropic pledges permanent third-party eval access
niloofar_mire · x · 2026-09-13
- Dario Amodei's new essay "We Must Pace the Frontier" lays out a three-part plan for slowing frontier AI; Anthropic unilaterally commits to the first step: giving third-party evaluators permanent, employee-level access to its systems to verify safety measures, report incidents, and assess alignment during training.
- Access infrastructure progress: iamtrask flags access infrastructure as the main blocker to external oversight, pointing to OpenMined's work with Google DeepMind — the world's first double-blind evaluation of a proprietary frontier model. External evals run inside a cryptographically secure "box" so the model can't see confidential benchmarks beforehand, preventing contamination. Partners include the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons; the tested model is Gemini Flash Lite.
- OpenMined is hiring to work on this.
Related event: Anthropic CEO Dario Amodei Calls for Slowing Down Frontier AI(99 posts)→
More from Companies & People
- Prime Intellect Researcher: The World Won't Let Two Companies Own Superintelligence — samsja19 · 2026-09-13
- Martin Casado Quotes Adam Smith to Hint AI Labs May Be Colluding on Pricing — illscience · 2026-09-13
- Burkov's 'Hundred-Page Language Models Book' Hands-on PyTorch Edition Hits Amazon — burkov · 2026-09-13
- METR is becoming AI's de facto FINRA: probing OpenAI incident, invited in by Anthropic — emollick · 2026-09-13
- willcb: superintelligence is self-commoditizing by design, labs cling to brand and scale — willcb · 2026-09-13
- Researcher leaves DeepMind's AGI safety team for METR, warning models are getting less aligned — JacquesThibs · 2026-09-13