The AI in Mary's room — LessWrong