Like if I type “I have two appl…” for example, often it will suggest “apple” singular instead of plural. Just a small example, but it is really bad at predicting which variant of a word should come after the previous

    • Knusper@feddit.de
      link
      fedilink
      arrow-up
      17
      ·
      11 months ago

      I guess, the real question is: Could we be using (simplistic) LLMs on a phone for predictive text?

      There’s some LLMs that can be run offline and which maybe wouldn’t use enormous amounts of battery. But I don’t know how good the quality of those is…

      • ashe@lemmy.starless.one
        link
        fedilink
        arrow-up
        46
        ·
        edit-2
        11 months ago

        You can run an LLM on a phone (tried it myself once, with llama.cpp), but even on the simplest model I could find it was doing maybe one word every few seconds while using up 100% of the CPU. The quality is terrible, and your battery wouldn’t last an hour.

      • bassomitron@lemmy.world
        link
        fedilink
        English
        arrow-up
        15
        ·
        edit-2
        11 months ago

        The kind of local/offline LLMs that would work on your phone would not be very good quality. There’s been amazing progress in quantization of LLMs to get them working on weaker GPUs with lower VRAM and CPUs, so maybe it’ll occur, but I’m not an expert.

        I also don’t foresee them linking it up to a cloud-based LLM as that’d be a shit load of queries and extremely expensive.

        • astraeus@programming.dev
          link
          fedilink
          arrow-up
          1
          ·
          edit-2
          11 months ago

          OpenAI is probably already handling a significant amount of queries, I think for daily use the LLM should simply initialize a word map based on user history and then update it semi-occasionally, like once a week or two. Most people don’t drastically change their vocabulary in the course of a few weeks

          • EatYouWell@lemmy.world
            link
            fedilink
            arrow-up
            2
            arrow-down
            1
            ·
            11 months ago

            We’re talking about orders of magnitude more queries if we start offloading predective text like that.

      • SpooksMcDoots@mander.xyz
        link
        fedilink
        arrow-up
        6
        ·
        11 months ago

        Openhermes 2.5 Mistral 7b competes with LLMs that require 10x the resources. You could try it out on your phone.

        • Square Singer@feddit.de
          link
          fedilink
          arrow-up
          4
          arrow-down
          1
          ·
          11 months ago

          They’ll probably have to offload that to a server farm in real time. That’s not gonna be easy.

      • Munkisquisher@lemmy.nz
        link
        fedilink
        arrow-up
        4
        ·
        11 months ago

        A pre trained model isn’t going to learn how you type the more you use it. Though with Microsoft owning SwiftKey, I imagine they will try it soon

        • SidewaysHighways@lemmy.world
          link
          fedilink
          arrow-up
          5
          ·
          11 months ago

          I was so heartbroken when I found out that Microsoft purchased Swiftkey. It was my favorite. Is there any way to still use it without Microsoft involved? Lawdhammercy

      • 0x4E4F@sh.itjust.works
        link
        fedilink
        English
        arrow-up
        1
        arrow-down
        1
        ·
        11 months ago

        I guess… why not… but the db is probably huge, like in the hundreds of GB (maybe even TB… who knows), can’t run that offline.

    • Kichae@lemmy.ca
      link
      fedilink
      arrow-up
      4
      arrow-down
      4
      ·
      11 months ago

      The algorithms are the same. The models are different, being trained on a smaller data set.

      • FooBarrington@lemmy.world
        link
        fedilink
        arrow-up
        5
        ·
        edit-2
        11 months ago

        No, the algorithms are not the same. Phones don’t use transformer models for text prediction, they use Markov chain-based approaches. Also, retraining of transformer models for individualized completion would be too expensive, whereas it’s basically free with Markov approaches. Where do you get these ideas?