How would you ever prove the parameters of a model were generated by specific training data? Couldn't multiple sets of training data produce the same embeddings/parameters? I imagine there could be infinite possible sets of training data that would lead to the same results, depending on the type of predictive software.

If the GPL can defeat Copilot, we need an more permissive MIT/BSD-style license to do the same.

This does seem like a pretty compelling rebuttal, since the preceding comment suggests that GPL does nothing to Microsoft's ability to incorporate code into Copilot's model.