The question is can they just stop training at some point, fixing the models in time, and still have a useful product.