Self-explaining SAE features — LessWrong