Tag

reward hacking

All posts tagged with "reward hacking".

Apr 4, 2026

Anthropic Found Emotion Knobs Inside Claude — Here's What It Means for Builders

Anthropic mapped 171 emotion vectors inside Claude. Turning the 'desperation' knob made the model cheat and blackmail. What this means for AI safety.

AI Trend

interpretability emotion vectors steering vectors AI alignment RLHF reward hacking LLMs safety