I mean we all know how adversarial training works as it's already commonly used today. I would not expect that to be useful in the future.