Running AI is so expensive that Amazon will probably charge you to use Alexa in future, says outgoing exec::In an interview with Bloomberg, Dave Limp said that he “absolutely” believes that Amazon will soon start charging a subscription fee for Alexa

  • LEX@lemm.ee
    link
    fedilink
    English
    arrow-up
    54
    arrow-down
    2
    ·
    edit-2
    1 year ago

    That’s already here. Anyone can run AI chatbots similar to, but not as intelligent as, Chatgpt or Bard.

    Llama.cpp and koboldcpp allow anyone to run models locally, even with only a CPU if there’s no dedicated graphics card available (although more slowly). And there are numerous open source models available that can be trained for just about any task.

    Hell, you can even run llama.cpp on Android phones.

    This has all taken place in just the last year or so. In five to ten years, imo, AI will be everywhere and may even replace the need for mobile Internet connections in terms of looking up information.

    • Zetta@mander.xyz
      link
      fedilink
      English
      arrow-up
      8
      ·
      edit-2
      1 year ago

      Yes, and you can run a language model like Pygmalion Al locally on koboldcpp and have a naughty AI chat as well. Or non sexual roleplay

      • LEX@lemm.ee
        link
        fedilink
        English
        arrow-up
        9
        ·
        1 year ago

        Absolutely and there are many, many models that have iterated on and surpassed Pygmalion as well as loads of uncensored models specifically tuned for erotic chat. Steamy role play is one of the driving forces behind the rapid development of the technology on lower powered, local machines.

          • LEX@lemm.ee
            link
            fedilink
            English
            arrow-up
            3
            ·
            edit-2
            1 year ago

            Huggingface is where the models live. Anything that’s uncensored (and preferably based on llama 2) should work.

            Some popular suggestions at the moment might be HermesLimaRPL2 7B and MythomaxL2 13B for general roleplay that can easily include nsfw.

            There are lots of talented people releasing models everyday tuned to assist with coding, translation, roleplay, general assistance (like chatgpt), writing, all kinds of things, really. Explore and try different models.

            General rule: if you don’t have a dedicated GPU, stick with 7B models. Otherwise, the bigger the better.

        • Zetta@mander.xyz
          link
          fedilink
          English
          arrow-up
          1
          ·
          1 year ago

          Which models do you think beat Pygmalion for erotic roleplay? Curious for research haha

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

            Hey, I replied below to a different post with the same question, check it out.

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

        Thanks for this, I haven’t tried GPT4All.

        Oobabooga is also very popular and relatively easy to run, but it’s not my first choice, personally.

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

        13B quantized models, generally the most popular for home computers with dedicated gpus, are between 6 and 10 gigs each. 7B models are between 3 and 6. So, no, not really?

        It is relative so, I guess if you’re comparing that to an atari 2600 cartridge then, yeah, it’s hella huge. But you can store multiple models for the same storage cost as a single modern video game install.

        • scarabic@lemmy.world
          link
          fedilink
          English
          arrow-up
          1
          ·
          1 year ago

          Yeah that’s not a lot. I mean… the average consumer probably has 10GB free on their boot volume.

          It is a lot to download. If we’re talking about ordinary consumers. Not unheard of though - some games on Steam are 50GB+

          So okay, storage is not prohibitive.

      • arthurpizza@lemmy.world
        link
        fedilink
        English
        arrow-up
        1
        ·
        1 year ago

        Storage is getting cheaper every day and the models are getting smaller with the same amount of data.

    • teuast@lemmy.ca
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      In five to ten years, imo, AI will be everywhere and may even replace the need for mobile Internet connections in terms of looking up information.

      You’re probably right, but I kinda hope you’re wrong.

        • teuast@lemmy.ca
          link
          fedilink
          English
          arrow-up
          3
          ·
          1 year ago

          Call it paranoia if you want. Mainly I don’t have faith in our economic system to deploy the technology in a way that doesn’t eviscerate the working class.

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

            Oh, you are 100% justified in that! It’s terrifying, actually.

            But what I am envisioning is using small, open source models installed on our phones that can answer questions or just keep us company. These would be completely private, controlled by the user only, and require no internet connection. We are already very close to this reality, local AI models can be run on Android phones, but the small AI “brains” that are best for phones are still pretty stupid (for now).

            Of course, living in our current Capitalist Hellscape, it’s hard not to imagine that going awry to the point where we’ll all ‘rent’ AI from some asshole who spies on everything we do, censors the AI for our own ‘protection’, or puts ads in there somehow. But I guess I’m a dreamer.