As an new AI image generators are not just replicating existing pictures but generating wonderful new ones, language models are not just replicating average of what they seen but are capable of generating novel ideas.
They're certainly capable of remixing things they've seen, and adding in randomness will add novelty. Whether that counts as "creativity" is something people can debate : - )
I think that the reason image ones have caught on better in some ways is because they don't need to be accurate. We're not asking them to understand anything, just produce images based on text prompts (which is amazing stuff all by itself).
Remixing objects and designs consistently like "van gogh" or "super mario" really implies some kind of internal model or "understanding" of the world.
Image generation didn't catch on because of lack of accuracy, but because of how GOOD the results are. It's made artwork immediately accessible to the masses without the huge learning curve. That's where these AI's are going to shine very very quickly.
Oh yes - you show them X and they make a model of X.
Show them enough pictures labelled "Van Gogh" and they get an idea of what "Van Gogh" looks like. They dp am awesome job of that.
The problem with the text ones is that people think that showing them words mean that they make a model of the thing the words are describing, rather than of how those words go together.
> The problem with the text ones is that people think that showing them words mean that they make a model of the thing the words are describing, rather than of how those words go together.
I believe a colossal truth is that the most efficient way to learn how the words go together is to at least make some approximate model of what the words are describing. And our optimization algorithms and model architectures are good enough to find theae solutions.
I guess you are on a different level of abstraction that I am.
It's clear the art AI's have a model of "van gogh's" style, and apply it to create very unique forms of art. The neural model weight's aren't storing compressed images of van gogh, but relationships and mathematical models about a concept.
I think that the reason image ones have caught on better in some ways is because they don't need to be accurate. We're not asking them to understand anything, just produce images based on text prompts (which is amazing stuff all by itself).