The wall is training data. Yes, we can make more and more of post training examples. No, we can never make enough. And there are diminishing returns to that process.