“The promise of generative AI is that it can learn the underlying meta-structures of the training data, so that it can recompose and extrapolate them in ways that are novel or not represented well in the training data. I say that’s the ‘promise’ because in reality you'll find that generative AI actually does a pretty terrible job of producing truly novel outputs. …
This isn't to say that these models can’t interpolate between things represented in their training data; that they’re quite impressive at- this is basically ‘style transfer’ and they do an excellent job at it. But genAI advocates really want us to believe that they are capable of extrapolation, not just interpolation, and I have seen no evidence that they are.” — Qwerasd
https://typedrawers.com/discussion/comment/71757/#Comment_71757