Scaling Monosemanticity: Anthropic's One Step Towards Interpretable & Manipulable LLMs | Towards Data Science

From prompt engineering to activation engineering for more controllable and safer LLMs