EVALUATING FEATURE STEERING

Summary from the TLDR AI newsletter

How do you evaluate the ability and efficacy of feature steering? Anthropic experimented with Golden Gate Claude earlier this year. It found that many features, especially similar safety ones, share common ranges for control.