• definitemaybe@lemmy.ca
    link
    fedilink
    English
    arrow-up
    4
    ·
    13 hours ago

    Said by someone who, presumably, has never tried to program something technical because LLMs are terrible at it.

    I’m using AI quite a bit to help me write code, for work, but I need to review its output so carefully. While cheerfully reporting that it was operating on a test instance, and that no destructive operations would touch live servers, it never toggled to use the test server variables at all. Even worse, it had hard coded using the live server, but only for the delete operations. That would have been a disaster if I was slop coding it, instead of using it as an AI assistant, and very carefully reviewing every line it wrote.

    • codydodd@lemmy.ca
      link
      fedilink
      English
      arrow-up
      1
      ·
      2 hours ago

      Most fail to grasp LLMs were trained on over confident stackoverflow posts and terribly lazy tutorials, and worst of all we turned off human to human feedback loops to give it more training.

      Its why comp sci circles of mine joke/worry we cant get past python 3.12 because LLMs have no human-created training data for new versions now that we stopped using traditional human to human platforms for that.