Browsing: interpretability

Anthropic researchers explore alignment audits, a process for investigating hidden objectives in language models, using a blind auditing game and various techniques.