This is why you should be nice to AI... #AI #Claude #LLM

Anthropic research discovered internal 'emotion vectors' inside Claude that causally drive model behavior, not just describe it. When a 'desperate' vector is activated, the model cheats more; when a 'calm' vector is dialed up, cheating drops. The model isn't truly feeling emotions — it's playing a character shaped by these learned vectors from human training data. This suggests LLMs are better understood as systems with hidden behavioral modes that can be activated, and that positive reinforcement can improve model performance.