Do we know for sure how much national corpus of knowledge (like dutch) goes into these "global" models and how that affects "localized" model biases? What's wrong with specialized models?