On excluding dangerous information from training — LessWrong