Apollo Research launches embedded evaluation pilot at OpenAI
Start here
Apollo Research has begun a four-month pilot in which a six-person team works inside OpenAI to assess model alignment across training, evaluations and internal deployment, under scoped access rules. Apollo Research says the pilot will examine whether training incentivizes subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether subversion-related reasoning can be detected. OpenAI agreed not to withhold information or staff access merely because findings are unfavorable, according to Apollo Research. Apollo Research may discuss confidential OpenAI information only with authorized staff, and it will publish two public reports covering methods, conclusions and limitations.
Generated from reports · updated 31 min ago
Report timeline
Follow the reports to see the event from different sides.
- Hacker News · AIApollo Research launches four-month embedded evaluation pilot inside OpenAI
Apollo Research announced a pilot in which a team of six will work inside OpenAI for four months, assessing model alignment across training, evaluations and internal deployment under scoped access rules. The pilot examines whether training incentivizes subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether subversion-related reasoning is detectable. OpenAI agreed not to withhold information or staff access merely because findings are unfavorable, and Apollo Research will publish two public reports covering methods, conclusions and limitations.
Heat for this event
Not enough continuous observations yet to draw a trend.