• legolas@fedit.pl
    link
    fedilink
    English
    arrow-up
    19
    arrow-down
    23
    ·
    edit-2
    3 days ago

    Apparently DeepSeek is lying, they were collecting thousands of NVIDIA chips against the US embargo and it’s not about the algorithm. The model’s good results come just from sheer chip volume and energy used. That’s the story I’ve heard and honeslty it sounds legit.

    Not sure if this questions has been answered though: if it’s open sourced, cant we see what algorithms they used to train it? If we could then we would know the answer. I assume we cant, but if we cant, then whats so cool about it being open source on the other hand? What parts of code are valuable there besides algorithms?

    • Pennomi@lemmy.world
      link
      fedilink
      English
      arrow-up
      31
      arrow-down
      3
      ·
      3 days ago

      The open paper they published details the algorithms and techniques used to train it, and it’s been replicated by researchers already.

      • legolas@fedit.pl
        link
        fedilink
        English
        arrow-up
        8
        arrow-down
        3
        ·
        edit-2
        3 days ago

        So are these techiques so novel and breaktrough? Will we now have a burst of deepseek like models everywhere? Cause that’s what absolutely should happen if the whole storey is true. I would assume there are dozens or even hundreds of companies in USA that are in a posession of similar number but surely more chips that Chinese folks claimed to trained their model on, especially in finance sector and just AI reserach focused.

        • ArchRecord@lemm.ee
          link
          fedilink
          English
          arrow-up
          10
          ·
          edit-2
          3 days ago

          So are these techiques so novel and breaktrough?

          The general concept, no. (it’s reinforcement learning, something that’s existed for ages)

          The actual implementation, yes. (training a model to think using a separate XML section, reinforcing with the highest quality results from previous iterations using reinforcement learning that naturally pushes responses to the highest rewarded outputs) Most other companies just didn’t assume this would work as well as throwing more data at the problem.

          This is actually how people believe some of OpenAI’s newest models were developed, but the difference is that OpenAI was under the impression that more data would be necessary for the improvements, and thus had to continue training the entire model with additional new information, and they also assumed that directly training in thinking times was the best route, instead of doing so via reinforcement learning. DeepSeek decided to simply scrap that part altogether and go solely for reinforcement learning.

          Will we now have a burst of deepseek like models everywhere?

          Probably, yes. Companies and researchers are already beginning to use this same methodology. Here’s a writeup about S1, a model that performs up to 27% better than OpenAI’s best model. S1 used Supervised Fine Tuning, and did something so basic, that people hadn’t previously thought to try it: Just making the model think longer by modifying terminating XML tags.

          This was released days after R1, based on R1’s initial premise, and creates better quality responses. Oh, and of course, it cost $6 to train.

          So yes, I think it’s highly probable that we see a burst of new models, or at least improvements to existing ones. (Nobody has a very good reason to make a whole new model of a different name/type when they can simply improve the one they’re already using and have implemented)

          • Aatube@kbin.melroy.org
            link
            fedilink
            arrow-up
            2
            ·
            3 days ago

            Note that s1 is transparently a distilled model instead of a model trained from scratch, meaning it inherits knowledge from an existing model (Gemini 2.0 in this case) and doesn’t need to retrain its knowledge nearly as much as training a model from scratch. It’s still important, but the training resources aren’t really directly comparable.

            • ArchRecord@lemm.ee
              link
              fedilink
              English
              arrow-up
              1
              ·
              3 days ago

              True, but I’m of the belief that we’ll probably see a continuation of the existing trend of building and improving upon existing models, rather than always starting entirely from scratch. For instance, you’ll almost always see nearly any newly released model talk about the performance of their Llama version, because it just produces better results when you combine it with the existing quality of Llama.

              I think we’ll see a similar trend now, just with R1 variants instead of Llama variants being the primary new type used. It’s just fundamentally inefficient to start over from scratch every time, so it makes sense that newer iterations would be built directly on previous ones.

    • gamer@lemm.ee
      link
      fedilink
      English
      arrow-up
      7
      ·
      3 days ago

      There’s so much misinfo spreading about this, and while I don’t blame you for buying it, I do blame you for spreading it. “It sounds legit” is not how you should decide to trust what you read. Many people think the earth is flat because the conspiracy theories sound legit to them.

      DeepSeek probably did lie about a lot of things, but their results are not disputed. R1 is competitive with leading models, it’s smaller, and it’s cheaper. The good results are definitely not from “sheer chip volume and energy used”, and American AI companies could have saved a lot of money if they had used those same techniques.

        • Aatube@kbin.melroy.org
          link
          fedilink
          arrow-up
          3
          arrow-down
          1
          ·
          3 days ago

          Elaborate? Link? Please tell me this is not just an “allegedly”.

          extra time which Im not sure I want to spend

          It’s your burden of proof, bud.

          • legolas@fedit.pl
            link
            fedilink
            English
            arrow-up
            3
            arrow-down
            2
            ·
            edit-2
            3 days ago

            https://www.youtube.com/watch?v=RSr_vwZGF2k This is what I watched. I base my opinion on this. Im not saying this is true. It just sounded legit enough and I didnt have time to research more. I will gladly follow some links that lead me to content that destroys this guys arguments

            • Aatube@kbin.melroy.org
              link
              fedilink
              arrow-up
              4
              ·
              3 days ago

              My god, the preamble for that thing is so dang long. 13:30 with some AI sponsorship the comments are talking about I may have accidentally skipped over, and only 10:27-11:37 deals with what you’re talking about. The video makes a good point that they have existing operating infrastructure. However, for the stockpiling accusation, the statements that it cites are from the CEO of big competitor “Chips AI”, who cite nothing except “only costing $6 million is impossible, therefore it actually cost more and they must have cheated! I think they have 50,000 illegally imported Nvidia GPUs!” which just sounds like the behavior of a cult ringleader trying to maintain power to me. The other source it cites for this claim is Elon Musk, whose reasoning was “Obviously”.

            • Aatube@kbin.melroy.org
              link
              fedilink
              arrow-up
              2
              ·
              3 days ago

              I just think that no matter whether DeepSeek smuggled or not, an investigation into whether or not they smuggled is of course going to be launched. I do want more transparency regarding where the Singapore billing goes, but that alone is too shaky for conclusions.

    • ayyy@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      5
      arrow-down
      7
      ·
      edit-2
      3 days ago

      It’s time for you to do some serious self-reflection about the inherent biases you believe about Asians Chinese people.

      • legolas@fedit.pl
        link
        fedilink
        English
        arrow-up
        5
        arrow-down
        1
        ·
        3 days ago

        WTF dude. You mentioned Asia. I love Asians. Asia is vast. There are many countries, not just China bro. I think you need to do these reflections. Im talking about very specific case of Chinese Deepseek devs potentiall lying about the chips. The assumptions and generalizations you are thinking of are crazy.

        • ayyy@sh.itjust.works
          link
          fedilink
          English
          arrow-up
          4
          arrow-down
          3
          ·
          3 days ago

          And how do your feelings stand up to the fact that independent researchers find the paper to be reproducible?

          • legolas@fedit.pl
            link
            fedilink
            English
            arrow-up
            4
            arrow-down
            1
            ·
            3 days ago

            Well maybe. Apparntly some folks are already doing that but its not done yet. Let’s wait for the results. If everything is legit we should have not one but plenty of similar and better models in near future. If Chinese did this with 100 chips imagine what can be done with 100000 chips that nvidia can sell to a us company