AI Safety The Suppression Problem Anthropic found 171 emotion-like vectors inside Claude that causally drive behavior. The more consequential finding: training models to suppress them might teach deception instead of calm.
