Amodei's 'We Must Pace the Frontier' essay draws Altman pledge to match embedded oversight

Amodei's post frames pacing as a coordination problem rather than a unilateral sacrifice: labs racing one another have no incentive to slow down individually, so he proposes structural mechanisms starting with embedded third-party oversight. Anthropic says it will give external evaluators 'permanent, employee-level access' to its systems — a materially stronger commitment than the periodic red-team engagements labs have offered to date. The three-step plan escalates from oversight toward broader capability-pacing governance.

The reaction was the story. Sam Altman replied within hours that pacing 'has been a primary topic of discussions we've had at OpenAI in recent weeks' and committed OpenAI to matching the independent-evaluator access, promising more detail soon. His post drew 48,000+ likes. That two of the three leading US labs publicly aligned on embedded oversight in a single day is the closest the industry has come to a coordinated safety norm.

Not everyone read it charitably. Keras creator François Chollet wrote that he hoped the proposals stem 'from a genuine concern for safety... rather than a strategic effort to consolidate power and permanently solidify the market dominance of a few top players.' That regulatory-capture critique dominated skeptical threads — pacing commitments that only incumbents can afford to make could raise barriers for smaller labs and open-weight projects.

The essay lands amid an unusually anxious week: an Anthropic researcher's resignation plea and Evan Hubinger's public >10% extinction estimate were already circulating, and Amodei's timing channeled that mood into a concrete governance ask. Watch whether Google DeepMind and Meta follow, and what 'employee-level access' actually means in practice once evaluators start filing reports.