TechTakes

especiall when it goes into mangled, uncanny-valley territory

GPT-2 was the peak of LLMs, because it was interestingly broken shit

Just to be clear, I agree with you and am not debating or arguing with you in any capacity with anything I say from this point onward.

I always thought you could do interesting stuff with genAI, especiall when it goes into mangled, uncanny-valley territory. Though I can only think of examples for visual generators, like this album cover or the AI Pizza commercial.

The output of genAI can be interesting and thought-provoking, but ultimately, it is not art. When humans create art, they have a vision of what they are trying to make. That vision might be fairly concrete, like "I want to depict this apple," or abstract, like "I want to express sadness." Then, they craft in their medium until they have a work fulfilling their vision. LLMs don't do this. They don't have cognition, much less intent or understanding, so they can't have "vision". When they "create" something, they do it without understanding the artistic/creative language of the medium used. Whatever the output is, it is iteratively massaged noise that some algorithm evaluates to be statistically correlated with the input prompt.

<insert paragraph here that steelmans the idea of an "AI Artist", which I can't be bothered to do, but structurally would appear here in this comment>

I don't think someone who takes the output of an LLM and presents it as "art" is an artist, as I don't think the output of an LLM is art. If I did think that the LLM could produce art, then the person presenting the output still is not an artist; the LLM would be. But I don't think that. If someone were to take the output of an LLM and change it in some way, it might be art, much like how someone might create a collage, but generally you don't see that. You usually just see people take the output and flog it as art.