Similar to what happened to NFT’s back in 2022

  • Korhaka@sopuli.xyz
    link
    fedilink
    English
    arrow-up
    1
    ·
    7 days ago

    Isn’t it just the training that costs a lot in comparison to running it? So if they stopped training new models and just sold their current ones it might be able to run at a profit from year to year, after a huge initial investment that may or may not ever be recovered.

    • Pirtatogna@lemmy.world
      link
      fedilink
      arrow-up
      1
      ·
      6 days ago

      Not really. The infrastructure to run language models is hardware wise demanding, consumes vast amounts of water (cooling) and electricity and requires frequent renewing and added capacity.

          • Korhaka@sopuli.xyz
            link
            fedilink
            English
            arrow-up
            1
            ·
            4 days ago

            This is the difference between running and training them. I can generate text output from an LLM faster than I can read it. But I am not training a new model any time soon.

            • Pirtatogna@lemmy.world
              link
              fedilink
              arrow-up
              1
              ·
              edit-2
              4 days ago

              Your PC is able to run comparatively small language models with limited number of parameters (a few billions). You can even do that on a modern mobile phone. Some of the more advanced models use hundreds of billions or even trillions of parameters. It’s not just the training. It’s more to do with model complexity.

              You can also do weather forecasting on a PC, but the result will not be comparable to a forecast made with a supercomputer.

              –Edit–

              Typo fixes, because while even the most whimsy LLMs can write, apparently I cannot.