What about the training data? Is everyone 100% confident that models are not a derived work of the training inputs now, even if they can reproduce input exactly?