• FizzyOrange@programming.dev
      link
      fedilink
      arrow-up
      20
      arrow-down
      4
      ·
      10 hours ago

      Nah AI can be extremely useful for learning technologies. You just need to be careful to verify they aren’t bullshitting you.

      For example find an explanation of PPM compression that is concrete and simple. As far as I can tell it doesn’t exist.

      But I could ask ChatGPT and it told me how it works (probably) in just a few seconds. I haven’t verified yet (at a BBQ) whether it is the correct algorithm but it’s certainly a plausible one that would work.

      It told me that you use a trie (typically) of symbol prefixes to record the probability of the following symbols, so for example you know that for the prefix “Th” the probability of “e” is 90%. Then you encode the symbol with arithmetic coding using the modelled probabilities. Apparently the typical max context length is 4-6.

      That would have taken me hours to find by reading code and ancient papers but I can verify it a lot quicker.

    • Rogue@feddit.uk
      link
      fedilink
      arrow-up
      18
      arrow-down
      10
      ·
      11 hours ago

      Using the term “discuss” is just creepy. It’s a piece of software. Do people actually think they’re conversing when they use an LLM?