Running llama-2-7b-chat at 8 bit quantization, and completions are essentially at GPT-3.5 levels on a single 4090 using 15gb VRAM. I don’t think most people realize just how small and efficient these models are going to become.

[cut out many, many paragraphs of LLM-generated output which prove… something?]

my chatbot is so small and efficient it only fully utilizes one $2000 graphics card per user! that’s only 450W for as long as it takes the thing to generate whatever bullshit it’s outputting, drawn by a graphics card that’s priced so high not even gamers are buying them!

you’d think my industry would have learned anything at all from being tricked into running loud, hot, incredibly power-hungry crypto mining rigs under their desks for no profit at all, but nah

not a single thought spared for how this can’t possibly be any more cost-effective for OpenAI either; just the assumption that their APIs will somehow always be cheaper than the hardware and energy required to run the model

  • zoe@lemm.ee
    link
    fedilink
    English
    arrow-up
    0
    arrow-down
    1
    ·
    1 year ago

    well running ai on consumer gpus isn’t supposed to be efficient: i assume when node sizes get smaller cores will be more efficient and consolidating vram (and gpu cores) on one big circuit board would be cost effective: just cores running fp16 or whatever ai specific. gpus like the a6000 exist for a reason. tbh pessimistic (or misleading) take on op’s part. the thing could replace lawyering jobs, save on grafic design costs, no more language teachers, youtube videos can be transribed in text format and used as learning material, why should this be bad tech ?

    • self@awful.systemsOP
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      why should this be bad tech ?

      because it is godawful at:

      • replace lawyering jobs
      • save on grafic design costs
      • no more language teachers
      • youtube videos can be transribed in text format and used as learning material
      • froztbyte@awful.systems
        link
        fedilink
        arrow-up
        4
        ·
        edit-2
        1 year ago

        one of the things that I like using as an example here is: just make it do something it isn’t currently trained on

        e.g. try to make it render content in zulu or isixhosa or [insert list of thousands of things that the developers barely/never touch] - it’s near guaranteed to have been trained on a very, very narrow set of that subject (if anything at all)

        “then just train it on more data” comes the refrain

        you: “okay, find me sufficient data of that”

        them: “it’s just a curation problem”

        you: “then who will create that?”

        the absolute very minimum of thinking beyond the second order just so entirely evades so many of these promptfans it’s astounding

        edit: TIL lemmy doesn’t do single newlines well

        • self@awful.systemsOP
          link
          fedilink
          English
          arrow-up
          4
          ·
          1 year ago

          I keep flashing back to eliezer being smug on Twitter about how good ChatGPT is at chess, and it turns out once you get past book openings and extremely well-documented games, it completely shits the bed and stops acting like it knows the rules of chess or even basic chess notation. and this is a very obvious outcome if you know how LLMs work, but most promptfans don’t

        • froztbyte@awful.systems
          link
          fedilink
          English
          arrow-up
          1
          ·
          1 year ago

          sidethought: I just thought up “promptfans” on the spot, but it doesn’t look like it exists anywhere else? so I guess that’s a word now

        • froztbyte@awful.systems
          link
          fedilink
          English
          arrow-up
          1
          ·
          1 year ago

          also, as an interesting semi-segue on this thought: there is an active group of researchers between a number of entities in africa working on creating better online corpuses of african languages (because what exists online is so scant)

          they’ve been at it for over 2 years now, as far as I know. when I last looked, fairly little of their work had gotten wider recognition

          “too small, too niche” to “address properly” is how each of these large outfits treat things like this. if they ever do give it some attention at all, it would likely be as part of some wider (batched) brush-stroke push to “improve our support for non-english languages” (or “$x art” or or or), and each will be given their respective 5% of attention for 3 hours then never again

    • Fanny Matrice@piaille.fr
      link
      fedilink
      arrow-up
      0
      arrow-down
      1
      ·
      1 year ago

      @zoe @Instrument_Data This dependence on progress in semiconductor processes makes me wonder…
      Up until now, this industry has always managed to surpass itself, but one suspects that we’ll eventually reach a physical wall.

      This RTX 4090 uses one of the world’s 3 thinnest processes currently in production: TSMC’s 4N. This makes transistor gates as long as 35 silicon atoms.
      How much lower can we hope to go? 20 atoms? 10 ? 5 ?

      • zoe@lemm.ee
        link
        fedilink
        English
        arrow-up
        1
        arrow-down
        2
        ·
        1 year ago

        idk, vram is also inefficient since it wastes heat too (since its a variation of dram which implies that it combines a transistor and a capacitor, and a transistor dissipates heat).

        alot of stuff need to witness a significant upgrade to cut down on Joule’s effect.

        now process nodes require 2 years to go down 0.5 nm in size, and probably 4 years when smaller