AI
OpenAI lets Apollo, METR, Transluce and Goodfire inspect models during training. What they get, and what OpenAI keeps back
OpenAI has signed its first "embedded assessments", which give four outside safety groups access to models while they are still being trained. Apollo Research will place six evaluators inside OpenAI to look for models that undermine their own safeguards. OpenAI still picks who they talk to and can redact the evidence behind their findings.