Patricia Paskov of the AI Verification and Evaluation Research Institute discusses auditing models, compute and governanceGovernance sets responsibilities, rules and oversight for decisions about how AI systems are used., and the need for evaluator independence and clear standards. She distinguishes evidence gathering from decisions made on that evidence.
She describes a pilot using a secure enclave to evaluate a frontier model against a benchmarkEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. without exposing both to each party. The conversation also considers the limits of alignment testsAlignment aims to make an AI system's behaviour consistent with intended human goals, values and constraints., security and possible international verification.
Watch on YouTube
