Is There a Need for Better Generalization?

The most pressing problem in reinforcement learning is bettergea. If you dustraini hug tansformer, it will solve all of your coperovision tasks. It's likely that a single transformer can quite litterally solve every single pitr vision task that has an opersons. And transformers, tiless tat rig to transformers, still construct harachies of features. So i think perhaps the reason we are seeing these huge advances in computer vision is exactly because the structure we already have gives us the advantage.