In light of the recent Crowdstrike crash revealing how weak points in IT infrastructure can have wide ranging effects, I figured this might be an interesting one.

The entirety of wikipedia is periodically uploaded here, along with many other useful wikis and How To websites (ex. iFixit tutorials and WikiHow): https://download.kiwix.org/zim

You select the archive you want, then the language and archive version (for example, you can get an archive with no pictures, to save on space). For the totality of the english wikipedia you’d select the “wikipedia_en_all_maxi_2024-01.zim”

The archives are packed as .zim files, which can be read with the Kiwix app completely offline.

I have several USBs I keep that have some of these archives along with the app installer. In the event of some major catastrophe I’d at least be able to access some potentially useful information. I have no stake in Kiwix, and don’t know if there are other alternative apps and schemes, just thought it was neat.

      • souperk@reddthat.com
        link
        fedilink
        English
        arrow-up
        5
        ·
        4 months ago

        Probably a lot less, keep in mind that whenever it answers a question the whole model is traversed multiple times, going through multiple GBs is not possible in the matter of seconds the model answers.

        • Max@lemmy.world
          link
          fedilink
          English
          arrow-up
          7
          ·
          3 months ago

          I’d be surprised if it was significantly less. A comparable 70 billion parameter model from llama requires about 120GB to store. Supposedly the largest current chatgpt goes up to 170 billion parameters, which would take a couple hundred GB to store. There are ways to tradeoff some accuracy in order to save a bunch of space, but you’re not going to get it under tens of GB.

          These models really are going through that many Gb of parameters once for every word in the output. GPUs and tensor processors are crazy fast. For comparison, think about how much data a GPU generates for 4k60 video display. Its like 1GB per second. And the recommended memory speed required to generate that image is like 400GB per second. Crazy fast.

    • mctoasterson@reddthat.com
      link
      fedilink
      English
      arrow-up
      15
      ·
      3 months ago

      I mean, you can self-host your own local LLMs using something like Ollama. The performance will be bound by the disk space you have (the complexity of the model you’re able to store), and the performance of the CPU or GPU you are using to run it, but it does work just fine. Probably as good results as ChatGPT for most use cases.

      • Nooodel@lemmy.world
        link
        fedilink
        English
        arrow-up
        3
        ·
        3 months ago

        We do this at work (lots of sensitive data that we don’t want Openai to capitalize on) and it works pretty well. Hosted locally, setup by a data security and privacy sensitive admin, who specifically runs the settings to not save any queries even on the server. Bit slower than chatgpt but not by much