Supervised learning of outputs in the brain — LessWrong