Are these models actually "stable" (heh)? Or would changing anything in the dataset result in a butterfly effect so to speak, thus a completely different output for the same prompt?
However, I'm not aware of anyone actively trying to adapt these techniques for large generative AI models (though would be great to see).