• ZkhqrD5o@lemmy.world
    link
    fedilink
    arrow-up
    5
    ·
    3 days ago

    I just want to rent them for my life. I love being a modern serf.

    Seriously though, go local or go home. Never paid a single cent for renting a LLM. Any PC that can play video games is fully sufficient to generate text with LLMs.

    • aksdb@lemmy.world
      link
      fedilink
      arrow-up
      6
      arrow-down
      2
      ·
      3 days ago

      Any PC that can play video games is fully sufficient to generate text with LLMs.

      I am all for local inference, but that is a super ignorant and uninformed take. I hope you are not running around complaining about the hallucinations and bad quality of LLM generated stuff, if your baseline are models that fit in 8 GB of VRAM or something. And no, 16 or even 24 GB wouldn’t bring you close to what the big models can do.

  • Blurntout@lemmy.ca
    link
    fedilink
    arrow-up
    8
    ·
    3 days ago

    Honestly whoever designs how to plumb all the local/edge inference together for a proof of work “token” to exchange for use of that inference or sale is gonna be the crypto that’s actually useful lmfao

    I understand why big ai wants to control the compute but that’s gotta be on the foss radar right?

      • Mika@piefed.ca
        link
        fedilink
        English
        arrow-up
        2
        ·
        3 days ago

        Kinda the opposite. Openrouter exist and you can host any model, but even within the model there are bunch of params that affect how good is the output you get.

        And while most hosts list these upfront, & you just gotta be smart enough to request with specific requirements, I’ve heard there are still good reasons to just stick to a list of trusted hosts sometimes, I don’t really remember what was the issue, quantinizastion level that they aren’t honest upfront maybe.

    • surewhynotlem@lemmy.world
      link
      fedilink
      arrow-up
      5
      ·
      3 days ago

      Mesh LLM? SETI but for tokens. I can half see it.

      Most of the value in openai now isn’t the model, it’s the prompts and process. That’s where FOSS could shine.