OpenAI sets up a four-month Apollo Research pilot to examine whether models attempt to undermine safety mechanisms
OpenAI Sets Four-Month Apollo Pilot to Probe AI Safety Controls | TokenPost
AI summary
OpenAI will host six evaluators from Apollo Research on site for four months to examine whether its models attempt to undermine the safety and control mechanisms of themselves or future models. The pilot gives external evaluators a defined period inside the company to examine model development and internal use; the pilot's start date has not been announced.
Read original