Extracting SAE task features for in-context learning — LessWrong