Comment on Proton’s Lumo AI chatbot: not end-to-end encrypted, not open source

• wewbull@feddit.uk ⁨1⁩ ⁨day⁩ ago

  I think it’s different. The fundamental operation of all these models is multiplying big matrices of numbers together. GPUs are already optimised for this. Crypto was trying to make the algorithm fit the GPU rather than it being a natural fit.

  With FPGAs you take a 10x loss in clock speed but can have precisely the algorithm you want. ASICs then give you the clock speed back.

  GPUs are already ASICS that implement the ideal operation for ML/AI, so FPGAs would be a backwards step.

  source

  • astanix@lemmy.world ⁨1⁩ ⁨day⁩ ago

    Thank you for the explanation!

    source
• steelrat@lemmy.world ⁨8⁩ ⁨hours⁩ ago

  Wildly different, similar in that ASIC was tuned to specific crypto issues, everyones making custom silicon for neural nets and such.

  I wouldn’t plan on it going away.

  source
• cley_faye@lemmy.world ⁨1⁩ ⁨day⁩ ago

  It’s probably different. The crypto bubble couldn’t actually do much in the field of useful things.

  Now, I’m saying that with a HUGE grain of salt, but there are decent application with LLM (let’s not call that AI). Unfortunately, these usages are not really in the sight of any business putting tons of money into their “AI” offers.

  I kinda hope we’ll get better LLM hardware to operate privately, using ethically sourced models, because some stuff is really neat. But that’s not the push they’re going for for now. Fortunately, we can already sort of do that, although the source of many publicly available models is currently… not that great.

  source

  • KumaSudosa@feddit.dk ⁨1⁩ ⁨day⁩ ago

    LLMs are absolutely amazing for a lot of things. I use it at work all the time to check code blocks or remembering syntax. It is NOT and should NOT be your main source of general information and we collectively have to realise how problematic and energy consuming they are.

    source
  • Zos_Kia@lemmynsfw.com ⁨1⁩ ⁨day⁩ ago

    There’s absolutely a push for specialized hardware, look up that company called Groq !

    source

    • KingRandomGuy@lemmy.world ⁨1⁩ ⁨day⁩ ago

      Yes, but at this point most specialized hardware only really work for inference. Most players are training on NVIDIA GPUs, with the primary exception of Google who has their own TPUs, but even these have limitations compared to GPUs (certain kinds of memory accesses are intractably slow, making them unable to work well for methods like instant NGP).

      GPUs are already quite good, especially with things like tensor cores.

      source
• brucethemoose@lemmy.world ⁨1⁩ ⁨day⁩ ago

  If bitnet or some other technical innovation pans out? Straight to ASICs, yeah.

  source