I have not heard discussion of that, but trying to reconstruct what it would be my guess is: AI companies with closed, non-distributed models in the clear with the GPL (but not AGPL code). But Meta would not be able to distribute their weights without also distributing the GPL code with it that went into the training? Is the controversy that it "linked" all the other source code that the model was also trained with and therefore all thus source code must be distributed too if the model weights are distributed?

I'm really curious because my best guesses sound ridiculous to me when I type them out.