Empathy as a natural consequence of learnt reward models — LessWrong