• gravitas_deficiency@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      10
      ·
      edit-2
      4 hours ago

      Or just realize that nobody fucking likes LLMs as much as the Captains of Industry want us to believe, and that the true power of this technical domain lies in more targeted and bespoke ML model generation and usage.

      ML is good and enables - and has enabled - some genuine generational leaps in science and technology. But LLMs are such a fucking waste of the technology’s potential. Not to mention, I’m extremely irritated that (largely due to Nvidia cornering the market) everyone is super gung-ho about a digital approach which amounts to brute-forcing neural nets digitally with shitloads of memory and highly-parallel compute, when it’s obvious to anyone with more than a passing familiarity with electrical engineering that an analog approach is going to be FAR more efficient in terms of resource and energy usage.

      • eleitl@lemmy.zip
        link
        fedilink
        English
        arrow-up
        3
        ·
        6 hours ago

        Yeah, for 4 and 8 bit quantization at least analog charge buckets or memristor-likes and analog multipliers would dramatically reduce substrate size, reduce power burn and speed up inference. Even better killer drones, here they come. Yay.

        • gravitas_deficiency@sh.itjust.works
          link
          fedilink
          English
          arrow-up
          2
          ·
          4 hours ago

          I mean… yeah, DARPA will probably be one of the first adopters of that stuff, it’s true. But DARPA is pretty much always a first adopter of any new tech, because they’re basically the research wing of the US military, and they have effectively infinite resources at their disposal (note: I am not debating whether or not that is a good thing here; simply stating that it is a thing). But just because they’ll likely do something military-ish with it first doesn’t mean that it’s a “bad” technology. The internet itself was, after all, initially a project of DARPA’s predecessor, ARPA, and was initially named “ARPAnet”.

    • datendefekt@feddit.org
      link
      fedilink
      English
      arrow-up
      11
      arrow-down
      2
      ·
      13 hours ago

      Don’t worry, Chinese LLM vendors without access to the newest hardware will take care of that.

      • Mwa@thelemmy.club
        link
        fedilink
        English
        arrow-up
        2
        arrow-down
        4
        ·
        edit-2
        12 hours ago

        based,hopefully they use this to optimize local LLMS so i have higher tokens per second.