Directions and desiderata for AI alignment — LessWrong