AI safety mindset — LessWrong