Anthropic's latest interpretability research maps 171 emotion concepts inside Claude using sparse autoencoders. The findings reveal that emotion vecto