3 ms·

I think that argument falls down though, because a derivative work is an expressive work in its own right, and model weights aren't. It would seem more coheren

I think that argument falls down though, because a derivative work is an expressive work in its own right, and model weights aren't. It would seem more coherent to argue that a model output could be a derivative work, though it would need to include a significant portion of some given source. But even then, since the copyright office's position is that they're not copyrightable, I'm not sure they could qualify.

devit 2y ago
Model weights, if they can reproduce something like the original, are just a form of lossy compression (or even lossless for text), where the LLM answering the prompt is a more powerful version of asking software to retrieve a specific file from a Zip archive (or a webserver answering an HTTP query) of such lossy compressed data. So if model weights don't infringe, that would also imply that saving an image as a JPG or a video using AV-1 doesn't infringe, which would obviously effectively implies that copyright doesn't apply to images or videos on the web, which is not current law/policy, so I think that reasoning cannot possibly work.
- fenomas 2y ago
  That comparison would only make sense if compressed images were considered derivative works. They're not - copyright doesn't protect bytes on a disk, it protects creative expressions. Lossy compression doesn't affect the creative expression, so in copyright terms a compressed JPG is just a copy, and is covered exactly like the original image. In contrast a derivative work is one creative expression that contains elements of another - like when you take an image and add commentary, or draw your own addition onto it, etc. And I'm pointing out that a trained model is not that - it's not itself a copyrightable expressive work. (We could think of it as a kind of algorithm for generating works, but algorithms aren't copyrightable.)
  - devit 2y ago
    Well then the model weights would be a compilation of copies of the original works, which has the same effect as it being a derivative work unless the copyright holder chose to allow copies but not derivative works.

devit 2y ago

Model weights, if they can reproduce something like the original, are just a form of lossy compression (or even lossless for text), where the LLM answering the prompt is a more powerful version of asking software to retrieve a specific file from a Zip archive (or a webserver answering an HTTP query) of such lossy compressed data. So if model weights don't infringe, that would also imply that saving an image as a JPG or a video using AV-1 doesn't infringe, which would obviously effectively implies that copyright doesn't apply to images or videos on the web, which is not current law/policy, so I think that reasoning cannot possibly work.

fenomas 2y ago
That comparison would only make sense if compressed images were considered derivative works. They're not - copyright doesn't protect bytes on a disk, it protects creative expressions. Lossy compression doesn't affect the creative expression, so in copyright terms a compressed JPG is just a copy, and is covered exactly like the original image. In contrast a derivative work is one creative expression that contains elements of another - like when you take an image and add commentary, or draw your own addition onto it, etc. And I'm pointing out that a trained model is not that - it's not itself a copyrightable expressive work. (We could think of it as a kind of algorithm for generating works, but algorithms aren't copyrightable.)
- devit 2y ago
  Well then the model weights would be a compilation of copies of the original works, which has the same effect as it being a derivative work unless the copyright holder chose to allow copies but not derivative works.