deep comments on When is AI safety research harmful?