Tag
OpenAI accidentally allowed graders to see chains of thought during RL training; Redwood Research reviews their analysis and finds the evidence largely assuages concerns about dangerous effects, though minor risks remain.