It’s this and the fact that few/one shot learning seems to just emerge with with models of this size