Supervised learning and self-modeling: What's "superhuman?" — LessWrong