AGI safety from first principles: Alignment — LessWrong