How to get value learning and reference wrong — LessWrong