Apollo Research says it has launched an embedded evaluation pilot with OpenAI; six evaluators will be stationed there for four months
AI summary
Apollo Research announced the launch of an embedded evaluation pilot with OpenAI; six evaluators will be stationed at OpenAI for four months to investigate whether models will attempt to undermine oversight and control of themselves. Author Marius Hobbhahn said the work focuses on how the training stack may incentivize scheming and subversive behavior in models, and said this is largely the direction his team's research over the past two years has been preparing for.
Full text · Original
Excited about this pilot towards embedded evals with OpenAI!
We're specifically focused on understanding how the training stack could incentivize scheming and subversion.
In many ways this is exactly what our research over the last 2 years has been preparing for!
We're launching a pilot towards embedded evaluations with @OpenAI.
Six evaluators from Apollo Research will work within OpenAI for four months, investigating whether models attempt to undermine their own oversight and control.View the quoted post on X