> Current generation of AI models can't think of anything truly new.
How could you possibly know this?
Is this falsifiable? Is there anything we could ask it to draw where you wouldn't just claim it must be copying some image in its training data?
How could you possibly know this?
Is this falsifiable? Is there anything we could ask it to draw where you wouldn't just claim it must be copying some image in its training data?
We got brass bands with brass instruments, synth music from synths.
We know therefore, necessarily, that they can be nothing novel from an LLM -- it has no live access to novel developments in the broader environment. If synths were invented after its training, it could never produce synth music (and so on).
The claim here is trivially falsifiable, and so obviously so that credulous fans of this technology bake it in to their misunderstanding of novelty itself: have an LLM produce content on developments which had yet to take place at the time of its training. It obviously cannot do this.
Yet an artist which paints with a new kind of black pigment can, trivially so.
> Everything is simply a blend of prior work.
I generally consider these two to be the same thing. If novelty is based on something else, then it's highly derivative and its novelty is very questionable.
A quantum random number generator is far more novel than the average human artist.
> have an LLM produce content on developments which had yet to take place at the time of its training. It obviously cannot do this.
Put someone in jail for the last 15 years, and ask them to make a smartphone. They obviously cannot do it either.
> I generally consider these two to be the same thing.
Sure words themselves bend and break under the weight of hype. Novelty is randomness. Everything is a work of art. For a work of art to be non-novel it can only incorporate randomness.
The fallacies of ambiguity abound to the point where speaking coherently disappears completely.
An artist who finds a cave half-collapsed for the first time has an opportunity to render that novel physical state of the universe into art. Every moment which passes has a near infinite amount of such novel circumstances.
Since an LLM cannot do that, we must wreck and ruin our ability to describe this plain and trivial situation. Poke our eyes and skewer our brains.
> Since an LLM cannot do that, we must wreck and ruin our ability to describe this plain and trivial situation. Poke our eyes and skewer our brains.
Change the temperature of a model, and you will have your randomness.
There is no explanation of how a biological brain can have randomness but a silicon process cannot.
Or, as the kids might say, AI couldn't feel the vibe shift occurring in the world at the time.
The claim was that current LLMs (though I assume they meant generative AI in general, since we're talking about image generation rather than text) are unable to produce anything novel. Meaning they either don't think Picasso's work is novel, or they don't think a human could have prompted an AI to make a new type of abstract art before having trained it on it. Whether the AI would want to do this is irrelevant. AIs don't want anything. They do what a human prompted. And while WWI may have shaped Picasso, learning data from WWI isn't necessary in order to make a cubist painting when prompted to stitch multiple perspectives into a painting. It's blending perspectives that are available from old data. And blending things in a new way is novelty. Most novel art falls into that category.
How could you possibly know that? Could you prove that an image wasn't copying from images in its training data?
The commenter flessner asserted it couldn't, despite having no way to demonstrate it. They are passing off faith as fact.
Assuming someone did want to show AI models can make new stuff...
> How could you possibly know that? Could you prove that an image wasn't copying from images in its training data?
This isn't even that hard. You just need to know what images are your training data when you train your model. A researcher with a small grant could do this.