They almost certainly would perform worse than more specialized classifiers trained with less data. It’s kind of a paradox of generalization.

Isn't this exactly what the bitter lesson is about?