I think it can be argued that the set of weights of a neural network is a program, but not “the preferred form of the work for making modifications to it”.
That would mean the training set and training algorithm (and random number generator seed) would have to be distributed under the GPL, too.
They did the same thing though. Two differently trained neural networks will have different outcomes, in this case they'd pick different chess moves in some special cases.
I’m fairly sure there were (maybe still are) linkers that, given duplicate symbol names, would happily grab one of them without checking whether they had the same code.
Add in parallel linking, and program behavior can change.
That would mean the training set and training algorithm (and random number generator seed) would have to be distributed under the GPL, too.