I was referring to the hypothetical image classifier, not the URL text classifier (the URL text classifier, as I mentioned, would perform very poorly).