On presenting the case for AI risk — LessWrong