Center for Human-Compatible AI (CHAI), UC Berkeley

Ce qu'il fait

Technical research lab founded by Stuart Russell, devoted to the theory, practice and policy of provably beneficial AI. Core research programme is assistance games / cooperative inverse reinforcement learning — designing AI that is uncertain about human objectives and defers to humans, rather than optimising a fixed specified objective. Interdisciplinary across CS, social science, law, philosophy, economics and governance. Runs an annual workshop and an internship programme.

Évaluation franche

Publishes a real administrative inbox, which most academic labs do not. The internship route is genuinely open to outsiders. Russell himself is extremely over-subscribed; do not expect a personal reply.

Préoccupations couvertes

Misspecified objectives and loss of control — Russell's thesis that 'optimisation alone won't get us to humane AI' and that the standard model of AI (maximise a given objective) is structurally unsafe.

Canal gouvernemental

Strong by proxy — Stuart Russell sits on the FLI AI Safety Index panel, advises the UN and numerous governments, and was a lead author on the International AI Safety Report.