Ben Pace comments on Why is pseudo-alignment “worse” than other ways ML can fail to generalize?