True answers from AI: "Miraculous" hypotheses — LessWrong