My understanding of it is that it’s far more complex than “lossy compression”. Of course it can generate text that is close to something in its training data, but I don’t think people have the right mental model if they think the model itself somehow has whole chunks of code in it. It does not.