“is there some kind of you know mode collapse risk where you know we kind of we overfit to data by training it on data that came out of AI itself”