Announcement
OpenAI says four outside teams will hunt for AI safety blind spots
OpenAI says Apollo Research, METR, Transluce and Goodfire will scrutinize model behavior, tests and internals.
TLDR
Next up for scrutiny: the safety checks themselves. OpenAI says it has signed agreements for its first phase of embedded assessments: Apollo Research will probe covert behavior against developers’ intentions, METR will investigate misalignment in OpenAI’s largest reinforcement-learning run, Transluce will assess alignment evaluations, and Goodfire will examine model internals. A broader METR monitor assessment, OpenAI says, still needs agreement on scope.
Combined views
303.1K
12 Sources, first seen ago

Comments