[civilitty]

> How much of the "novel ways which have never been considered before" is just the novelty effect of having your very own artist?

What's the difference and why does that matter? Sure, plenty of people may have had the idea of an anthropomorphized teddy bear swimming in a pool, but if none of them ever realized it in the real world, the AI couldn't have learned from it.

> The change here is the cost, and not the capability.

I strongly disagree. Scaling the artistic process with GPUs instead of meatbags is a huge change in capability, just like the mechanized tractor was a huge capability change over the ox. The change in cost is a side effect in the change of capability.

You can now use a tool instead of outsourcing it to someone else and artists now have an automated tool that ostensibly replaces their manual labor.

[polio]

I agree that it's a big change in the economics of getting art (e.g. a single game dev can now plausibly get assets this way), but I'm not sure that it's a change in the "cognitive" process of creation. This is in response to the original comment that suggests that novel combinations are interesting.

The machine is clearly good at realizing novel combinations, but I think that has more to do with the lack of interest of human artists in rendering these combinations, rather than the lack of ability.

I am also of the opinion that a human realization would produce better art. A machine might literally depict the bear in a pool, but a human could imagine a logically consistent context for that to be happening and decorate the pool with details like the leaderboard of the Teddy Bear Olympics and have reporters and spectators that are other stuffed animals. There might be a rivalry in progress. The distinguishing feature for me so far has been that human art is a snapshot of a much more sophisticated simulation that draws from the experience of having lived, felt things like fear, tension, joy directly, rather than having to approximate the aspects that give new art its electric nature indirectly as the machines do.

I'm sure a machine will be able to do that someday, but most of my experience with Dall-E 2 has been for the background to be vague, blurry, and weirdly unintentionally surreal. The prompt itself is maybe rendered accurately 95% of the time.

[civilitty]

The OP never said that novel combinations are "interesting"... They claimed:

> Generative models can combine existing concepts in novel ways which have never been considered before.... So as long as an answer can be expressed as a combination of primitive concepts, LLMs can generate them.

You seem to be hung up on the "never been considered before" as if artists are incapable of combining those concepts when the more charitable interpretation of what the OP meant is "doesn't appear in the training data." Obviously they aren't omniscient and can't possibly know what people have ever imagined in the totality of existence, but we can empirically compare the novelty of generated images to the training inputs.

No one said anything negative about artists or their abilities. We're talking about the AI's abilities in a positive sense and that's the mechanistic ability to combine concepts in novel ways to generate images. Any art that springs from that is from the human using the AI as a tool.