Casper: 'more evals' is a weak pacing lever, implementations range from safety-washing to real oversight

StephenLCasper · x · 2026-09-23

In his thread on embedded evaluations, Stephen Casper argues: (1) calls by Dario and others for embedded evaluation have not been paired with stricter pacing measures like hardware controls, making 'more evals' a remarkably weak mechanism for slowing AI progress; (2) embedded evaluation spans a huge range — from net-harmful safety washing to rigorous oversight where auditors can access records, data, models and meetings, and whistleblow.

Related event: Researcher criticizes embedded evaluations as weak pacing substitute(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →