• kkj@lemmy.dbzer0.com
    link
    fedilink
    English
    arrow-up
    19
    arrow-down
    11
    ·
    6 days ago

    LLMs are actually really good at a handful of specific tasks, like autocomplete. The problem arises when people think that they’re on the path to AGI and treat them like they know things.

    • finitebanjo@lemmy.world
      link
      fedilink
      arrow-up
      28
      arrow-down
      4
      ·
      6 days ago

      Nah mate, its shit for autocomplete. Before LLMs autocomplete was better with a simple dictionary weighted to use percentage.

      • kkj@lemmy.dbzer0.com
        link
        fedilink
        English
        arrow-up
        10
        arrow-down
        5
        ·
        6 days ago

        I’ve found it better than the weighted dictionary for prose, and way better for code. Code autocompletion was always really limited, but now every couple dozen lines it suggests exactly what I was going to type anyway. Never on anything particularly clever, mind you, but it saves some tedium.

        • finitebanjo@lemmy.world
          link
          fedilink
          arrow-up
          11
          arrow-down
          2
          ·
          6 days ago

          It also sometimes halucinates entire libraries and documentation and is single handedly responsible for massive sector wide average vulnerabilities increase.

          Did you make sure to subtract all of that negative value before you even considered it as “good”?

          • kkj@lemmy.dbzer0.com
            link
            fedilink
            English
            arrow-up
            9
            arrow-down
            1
            ·
            5 days ago

            Oh, it’s fucking horrible at writing entire codebases. I’m talking about specifically tab completion. You still have to read what it’s suggesting, just like with IntelliSense and other pre-LLM autocomplete tools, but it sometimes finishes your thoughts and saves you some typing.

            • Zacpod@lemmy.world
              link
              fedilink
              arrow-up
              8
              arrow-down
              1
              ·
              5 days ago

              Hard agree. Whole codebase in AI is a nightmare. I think MS’s 25% is even WAY too much, based on how shitty their products are becoming. But for autocompleting the line of code I’m writing? It’s fucking amazing. Doesn’t save any thought, but saves a while bunch of typing!

          • toddestan@lemmy.world
            link
            fedilink
            arrow-up
            1
            arrow-down
            1
            ·
            5 days ago

            Just because a hammer makes for a lousy screwdriver doesn’t mean it’s not a good hammer. To me, AI just another tool. Like any other tool, there’s things it is good at and there are things it is bad at. I’ve also found it can be pretty good as a code completion engine. Not perfect, but there’s plenty of boilerplate stuff and repetitive things where it can figure out the pattern and I can bang out the lines of code pretty quickly with the AI’s help. On the other hand, there’s times it’s nearly useless and I switch back to the keyword completion engine as it’s the better tool for those situations.

            • finitebanjo@lemmy.world
              link
              fedilink
              arrow-up
              1
              ·
              5 days ago

              If you invent a hammer which reduces the average structural stability anywhere from 5% to 40% then it should be banned.

    • Flax@feddit.uk
      link
      fedilink
      English
      arrow-up
      3
      ·
      5 days ago

      Dunno why the downvotes. I think it’s useful for menial stuff like “create a json list of every book of the Bible with a number for the book and a true or false if it’s old or new testament” which it can do in seconds. Or to quickly create a template.