The models lose the ability to inject subtle and nuance stuff as they scale up, is what I’ve observed.