Aren't those models "skimming" off the frontier models? (forget the term, but basically training themselves by prompting Claude / OpenAI etc) If that's the case then once the data center / GPU money train runs out, the whole thing craps out