Clarifying some key hypotheses in AI alignment — LessWrong