Safety
3d agoProposal Uses Admiration To Improve AI Alignment
Training models on transcripts of "admirable reasoning" previously reduced misalignment. This new proposal from LessWrong suggests using emotion-like representations to steer behavior. It draws on moral psychology to motivate AI emulation of virtuous patterns.
The Signal
Researchers can now test if admiration-like states provide a cheaper, more effective path toward AI alignment than current methods.