The AI model distillation paradox - MR Online
mronline.org
When U.S. labs train models on the world’s writing, Washington calls it fair use. When Chinese labs train on U.S. outputs, it’s a “malicious attack.” It can’t have it both ways.

Last month, Washington insisted that training models on other people’s work without permission is so obviously fair that licensing requirements would be ruinous to national innovation. It also insisted that training models on other people’s work without permission is an act of industrial predation posing a grave threat to national security.

Both propositions cannot hold simultaneously. Either the outputs of these systems are protectable under the law, in which case the writers, musicians, artists, and programmers whose work American models were built on can hold a claim that the Justice Department has denied. Or they are not, in which case the six Chinese laboratories accused of “malicious and targeted” distillation attacks have not done anything that Elon Musk has not freely admitted to under oath, without consequence.

  • P03 Locke@lemmy.dbzer0.comEnglish
    2·
    1 day ago

    It would be like KFC saying you can’t try to figure out how we made this food by eating it and taking the food apart and tasting it and testing it. And, again in my opinion, contract law (i.e., the terms of use for using the models) also does not provide the companies a viable claim because of a weird section of copyright law that basically doesn’t allow state law claims, such as contract claims, to be brought when they are in effect basically copyright claims.

    This is what most people don’t understand. Looking at a picture or looking at a web page is not stealing. Also, you can’t compress the entirety of “the internet” into a 20-200GB model. The claim that the data they store are even close to exact representations is bunk.

    Oh, you’re using this argument to claim that models shouldn’t be considered confidential and “trade secrets”. Well, yeah, but also that argument can be turned the other way. Fair use still stands.

    And I know you’re an IP attorney and all, but I would be okay if LLMs end up destroying the entire framework of copyright. It is a system created by the rich, to be only abused by the rich. We’re already seeing it slowly get dismantled first-hand with all of these decompiling experiments, and I’m sure that will be followed up with rich megacorps doing what they do best: playing the hypocrite by claiming one thing is infringing in their copyright, while what they do is totally legal.

    • zod000@lemmy.dbzer0.comEnglish
      2·
      24 hours ago

      Also, you can’t compress the entirety of “the internet” into a 20-200GB model.

      You clearly can, LLMs are just an extremely lossy method of compression. It does make one wonder at what point does a highly compressed version of an original work no longer qualify. If we compressed a licensed image enough that it was a blur that vaguely resembled the original, does that still count?