x
Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers — LessWrong