Dario pledges embedded evaluators as AEF-1 standard lands

The pacing debate returned this week after first emerging in July with Pacing the Frontier. Dario Amodei, lead author on the original, wrote a rare personal blog post spelling out how he sees pacing play out, and the AI Evaluator Forum — formed in December 2025 — happened to publish AEF-1, its expectations for what third-party evaluators should do, at the same time. The forum's members are presumably the leading third-party auditors that big labs would recruit for self-regulation.

The proposal has three parts: Embedded Evaluators, Democratic Coordination, and Global Coordination. Amodei's promise of desks, badges, laptops, and access to workspaces and tools mostly comparable to internal risk assessment teams goes further than standard domestic self-regulation playbooks. The AEF-1 baseline covers access, conflicts of interest, funding relationships, recusal, and transparency — the same issues raised by critics.

A sharp public split runs underneath the announcement. Bilal Chughtai left Google DeepMind and called for pacing and more transparency; Dan Selsam argued models may appear aligned under evaluation while hiding misalignment. On the other side, Shashank/Sayash Kapoor and Lennart Heim frame rogue agent incidents as a security and governance problem, and Aidan Gomez argued against a few Silicon Valley companies becoming AI gatekeepers. The test is likely whether embedded evaluators can be genuinely independent in practice, and whether Kevin Bass's allegations of financial entanglement between Anthropic and parts of the safety ecosystem hold up.