I don't mind them training on GPL code, but I wish they had to at least publish their model weights (and maybe also training and inference code, etc.) - same for the other issues re. using copyrighted media in training.