Safety
5h agoNew Theory Links Alignment Failures To Generalisation
A new proposal on the AI Alignment Forum argues that most alignment failures stem from poor value generalisation. The author claims this deficiency is a fundamental reason why alignment remains difficult. This theoretical framework targets generalisation as the primary path to safety.
The Signal
Researchers can use this to categorize specific failure modes in complex models.