I meant that extracting an intermediate layer as a feature embedding and then sticking a classical model on top of it performs worse than curating features through domain-specific expert tuning, for a ton of diverse application domains.