• SkaveRat@discuss.tchncs.de
    link
    fedilink
    English
    arrow-up
    7
    ·
    1 day ago

    that’s the annoying part.

    LLM code can range to “doesn’t even compile” to “it actually works as requested”.

    The problem is, depending on what exactly was done, the model will move mountains to actually get it running as requested. And will absolutely trash anything in its way, From “let’s abstract this with 5 new layers” to “I’m going to refactor that whole class of objects to get this simple method in there”.

    The requested feature might actually work. 100%.

    It’s just very possible that it either broke other stuff, or made the codebase less maintainable.

    That’s why it’s important that people actually know the codebase and know what they/the model are doing. Just going “works for me, glhf” is not a good way to keep a maintainable codebase

    • turboSnail@piefed.europe.pub
      link
      fedilink
      English
      arrow-up
      6
      ·
      22 hours ago

      LOL. So true.
      On top of that, an LLM can also take you on a wild goose chase. When it gives you trash, you tell it to find a way to fix it. It introduces new layers of complication and installs new libraries without ever really approaching a solution. It’s up to the programmer to notice a wild goose chase like that and pull the plug early on.

      That’s a fun little mini-game that comes with vibe coding.