Empericially, CNNs generalize better on image recognition tasks than hand built features. This comment doesn't make much sense and is needlessly obtuse in the face of progress, tbh.