Discovering Latent Knowledge in Language Models Without Supervision — LessWrong