Comment on OpenAI be like

dharmacurious@slrpnk.net ⁨6⁩ ⁨days⁩ ago

• dovah@lemmy.world ⁨6⁩ ⁨days⁩ ago

  Here’s a list of open source models: open-llms

  Models are only open source if the weights are freely available along with the code used to generate them.

  source

  • morrowind@lemmy.ml ⁨6⁩ ⁨days⁩ ago

    I would argue to be truly open source the training data needs to be as well.

    source
  • dharmacurious@slrpnk.net ⁨6⁩ ⁨days⁩ ago

    I really appreciate that! I was asking more for the information of it, I doubt I could do anything with the link. Lol. I don’t understand thing 1 about this stuff. I don’t even know wtf a weight is in this context lol

    source

    • edinbruh@feddit.it ⁨6⁩ ⁨days⁩ ago

      In this context “weight” is a mathematical term. Have you ever heard the term “weighted average”? Basically it means calculating an average where some elements are more “influent/important” than others, the number that indicates the importance of an element is called a weight.

      One oversimplification of how any neural network work could be this:

      • The NN receives some values in input
      • The NN calculates many weighted averages from those values. Each average uses a different list of weights.
      • The NN does a simple special operation on each average. It’s not that what the operation actually is, but it must be there. Without this, every NN would be a single layer.
      • The modified averages are the input values for the next layer.
      • Each layer has different lists of weights.
      • In reality this is all done using some mathematical and computational tricks, but the basic idea is the same.

      Training an AI means finding the weights that give the best results, and thus, for an AI to be open-source, we need both the weights and the training code that generated them.

      Personally, I feel that we should also have the original training data itself to call it open source, not just weights and code.

      source

      • MrMcGasion@lemmy.world ⁨6⁩ ⁨days⁩ ago

        Absolutely agree that to be called open source the training data should also be open. It would also pretty much mean that true open source models would be ethically trained.

        source
      • Vex_Detrause@lemmy.ca ⁨4⁩ ⁨days⁩ ago

        What does the open source training data include? I’ve read a few open source training data that is also tested for biased but I haven’t really looked at them.

        source
      • dovah@lemmy.world ⁨6⁩ ⁨days⁩ ago

        Yeah, good call. Training data should be available as well.

        source
      • dharmacurious@slrpnk.net ⁨6⁩ ⁨days⁩ ago

        Thank you!

        And yeah, it really does seem like the training data should be open. Like, not even just to be considered open source, just to be allowed to do this at all, ethically, the training data should be known, at least to some degree. Like, there’s so much shit out there, knowing what they trained on would help make some kind of ethical choice in using it

        source
• Radiant_sir_radiant@beehaw.org ⁨6⁩ ⁨days⁩ ago

  The closest one to true FOSS that I’m aware of is Apertus. Not sure whether it’s feasible to build anything meaningful from scratch without your own GPU farm though.

  source
• mindbleach@sh.itjust.works ⁨5⁩ ⁨days⁩ ago

  Do such models exist? Yes. Are they the big-boy models anyone’s really using? Ehhh not really.

  There are in-use models that are “here’s a thing do whatever good luck,” which is at least as open-source as any MIT project. (Permissive licenses being “here is the code, have a nice life.”) Very few models are properly reproducible, because even when their training data includes DVDs you probably own, it also includes a ton of random internet pages that maybe don’t exist anymore. The push for ever-larger models, trained on as much stuff as possible, makes the use of “open source” regrettable or even deceptive choice. But quite a few are unrestricted for whatever weird shit you want to get up to.

  source
• Jankatarch@lemmy.world ⁨6⁩ ⁨days⁩ ago

  I mean you could give the randomizer seed along with the code for training I guess that would count kinda?

  source