@richiekhoo @catsalad Ironically, they actually still need artists to create things and post them; because they can't feed ouput of models back in as input for too long without it destroying the quality of the model in ways that are very difficult to fix (at least for now)

And if these models were impossible without the original training data, the owners of the model shouldn't be OpenAI or some group of billionaires, but the artists who it could not exist without

@capriciousday @catsalad Yes, OpenAI is upset that they took their training data but due to the nature of how it can obscure the origin, they can't prove it; but just like with OpenAI stealing everyone's copyright'd data, its obvious because it wouldn't work otherwise, just by the nature of how these models work

The advancement here is building something that obscures the origins of many copyright violations at once

@meow @catsalad I don't like to be the one saying this, but there are legitimate applications; just those ones don't lead to the absurd amount of investment, and potential limitless profit margins

And the ability to propagandize kids in their classrooms

@Talon1024 @catsalad But they can crawl for all the images on the internet?

Seems like a matter of will, also, just because something is inconvenient doesn't alleviate them of the liability

That argument is very lazy (obvio in more than 1 way)

@ASprinkleofSage @catsalad We don't even understand how the human brain works; we have not made sentience or anything that can actually be described as AI unless you accept a very recent change to the definition

So the hubris of thinking we were able simulate human intelligence, let alone god is very silly to me

@Rin3d @catsalad Its not transformative, its just obscured by doing much more copyright violations at once. By definition, LLMs and Stable Diffusion do not create novel things. And so more copyright violations they do the higher the quality gets; and just obscures the origins more. Without the training data it won't make images.

The argument amounts to, if you steal enough training data to successfully obscure origins, it becomes legal?

Any image is actually reproduction of many stolen images

@Rin3d @catsalad And most importantly, the same Copyright Act that protects movies and music would apply to them, and once they go over 2,500 USD dollars in sales it becomes a federal felony.

This isn't even just about potential civil suits.

If the law was applied equally, they would have to face charges and anyone else who profited from it, like building tools using the data, would all be criminally liable too

Reason being the music+movie industry have been expanding the scope for a decade+

@Rin3d @catsalad To be clear, I have 100's of examples of these models generating quite literally Mario, or Sonic; and they were not in the prompt, it was just related to their probabilistic nature and the amount of content related to those topics in the training data

That wouldn't be possible if it were actually transformative