On top of that, what's the method to keep "errors" from compounding? It also seems like the capabilities of the trained model would approach an asymptote that is the limit of the training model, and never pass it.