I'd say that the critique points out that this "information from a previous model" itself needs tremendous amounts of data. Now, did we see any better generalization capabilities with all data counted?