The operative word here is "just". The models are so large and the training is such that of course one of the things they are likely doing is memorizing the corpus; but they aren't "just" memorizing the corpus: there is some amount of regularization in place to allow the system to exhibit more generative outputs and behaviors as well.
The operative word here is "just". The models are so large and the training is such that of course one of the things they are likely doing is memorizing the corpus; but they aren't "just" memorizing the corpus: there is some amount of regularization in place to allow the system to exhibit more generative outputs and behaviors as well.