Why don't we just, like, try and build safe AGI? — LessWrong