[banana_feather]

The idea is that if you violate the terms of the license to develop your own model, you lose your rights under the license and are creating an infringing derivative work. If I clone a GPL'd work and ship a derivative work under a commercial license, downstream users can't just integrate the derivative work into a product without abiding by the GPL terms and say "well we're downstream relative to the party who actually copied the GPL'd work, so the GPL terms don't apply to us".

[blendergeek]

If such a "derivative" model is a derivative work, then aren't all these LLMs just mass copyright infringement?

[dragonwriter]

If model weights aren’t copyrightable, derivative model weights are not a “work”, derivative or otherwise, for copyright purposes.

If they are, and the license allows creating finetuned models but not using the output to improve the model, then the derived model is not a violation, but it might be a derivative work.

[banana_feather]

At the end of the day it's not black and white, but there's a large and obvious difference in degree that would plausibly permit someone to find that one is and the other isn't. It's fairly easy to argue that using the outputs of LLM X to create a slightly more refined LLM Y creates a derivative work. The argument that a model is a derivative work relative to the training data is not so clear cut.

[dTal]

Exactly this. What's good for the goose is good for the gander!

[rodoxcasta]

If the weights are not copyrighteable, you don't need a licence do use them, they are just data. There's not a right to infringe if these numbers have no author. Of course, to use openAI API you must abide to their terms. But if you publish your generations and I download them, I have nothing to do to the contract you have with openAI since I'm no part of it. They can't impede me to use it to improve my models.

[stale2002]

No, because the premise of the hypothetical is that the weights aren't protected by copyright.

So, no matter what they TOS says, it's not an infringing work.

> Downstream users can't just integrate the derivative work into a product without abiding by the GPL terms

You absolutely could do this if the original work is not protected by copyright, or if you use it in a way that is transformative and fair use.

[mattl]

Something under the GPL is also copyrighted. The GPL is a copyright license.

[asadotzler]

The GPL depends on copyright but it not itself copyright. The GPL is a license that gets its legal standing from copyright but if you don't have a copyright on something, slapping the GPL on top of it doesn't make it copyrighted to you.

[mattl]

Absolutely.

[stale2002]

If the underlying work is not protected by copyright, it doesn't matter what license someone tries to put on it.

Similarly, if someone creates a fair use/transformative work then the license can also be ignored.

[dvdkon]

Thing is, the outputs of a computer program aren't copyrightable, so it doesn't matter if your improved model is a derivative work. What you say would apply if you derived something from the weights themselves (assuming they are copyrightable, of course).

[diffeomorphism]

Really?

Your customers bought that product under license A. Afterwards it turned out that you pirated some artwork from disney. Then your customer can sue you (not disney) to make things right. The specific license of the original work seems quite irrelevant here.

[pessimizer]

Not at all. The reason your customer can sue you is because Disney can sue your customer. Disney would be suing your customer under the specific license of the original work.

edit: you seem to see the customer as the primary victim here instead of Disney, but if Disney weren't a victim the customer wouldn't have a case.