Four reasons I find AI safety emotionally compelling — LessWrong