Parsing Abram on Gradations of Inner Alignment Obstacles — LessWrong