Anthropic found a 'panic' word inside Claude
anthropic used a new technique to peek inside claude and found millions of hidden features influencing outputs, revealing that models can monitor their own internal states.
anthropic released a study on 'mechanistic interpretability', mapping millions of internal features within its claude models to decode their decision-making processes. researchers identified a 'j-space' where concepts—like 'panic' or 'protein'—appear internally to guide reasoning without surfacing in the final text. this demonstrates that large language models maintain complex internal states and self-monitoring capabilities, suggesting that future alignment work could target these hidden layers to ensure safety rather than just observing external behavior.