Superposition through Active Learning Lens — LessWrong