• Razen@lemmy.world
    link
    fedilink
    English
    arrow-up
    7
    ·
    7 hours ago

    Are they eating the cost? How are they able to do it while others are unable to?

    • Balinares@pawb.social
      link
      fedilink
      English
      arrow-up
      5
      ·
      4 hours ago

      They invented a hybrid attention design that drastically reduces the amount of memory needed for the KV cache at inference time. Like, dividing it by 10. And memory is a large part of the cost of inference.

      • jaykrown@lemmy.worldOP
        link
        fedilink
        English
        arrow-up
        3
        ·
        2 hours ago

        This is a main part of the reason, yes. They actually innovated and did something that pushed the technology forward to be much more efficient, which we first saw with DeepSeek R1 for different reasons.

    • sketch@lemmy.pt
      link
      fedilink
      English
      arrow-up
      6
      ·
      7 hours ago

      The American models are eating their cost big time, in return they get user data to train on and a massive reality distortion field that can theoretically be exploited later. It costs less for DeepSeek to eat their cost, and maybe the value of that user data is worth it now? Maybe there is some Chinese VC getting involved to try and boost DeepSeek with a little reality distortion field they can attempt to exploit later? I don’t know, but all these seem plausible to me.

        • Wispy2891@lemmy.world
          link
          fedilink
          English
          arrow-up
          2
          ·
          45 minutes ago

          i never said that. In fact, openai gives for free 2.5 million of gpt-5.5 tokens every day if you share all your inputs for training.

          It was the answer to “why it’s this cheap?” => because it’s subsidized by your data

        • Kynsey@lemmy.ml
          link
          fedilink
          English
          arrow-up
          7
          ·
          4 hours ago

          LOL right? Like does the other person not realize that unless your running a local LLM or using something like duck.ai ALL of the AIs are training on your convos. It reminds me of all the “Chinese Surveillance” fearmongering around TikTok as if Meta and Instagram don’t do the exact same thing.