• crabs@lemm.ee
    link
    fedilink
    English
    arrow-up
    9
    ·
    1 year ago

    When they first launched the bing AI powered by GPT I used it for everything, then it became pretty clear they nerfed it and I’ve been waiting for a competitor to catch up. Bard’s gotten a little better, but it hallucinates way worse still, making up answers.

    I’m secretly hoping for one of these open-source projects like Llama 2 or Orca to lead to a totally unrestricted chatbot even if it’s short-lived

    • Nate@programming.dev
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      Man I’d love to have the original bing ai back. Those hallucinations were something else. Probably a liability issue but I wish they had it available with a disclaimer.

  • fearout@kbin.social
    link
    fedilink
    arrow-up
    4
    ·
    1 year ago

    Yep, definitely. I have a plus subscription, and stuff that was easy for it just a few months ago now seems to take several back-and-forths to barely approach similar results.

    Science content is where I noticed the most degradation. It just stares at me using blank “it’s not in my training data” answers to questions that used to have comprehensive responses a while ago.

    I think they’re scaling down the models to make them cheaper to run?

    • NXTR@kbin.social
      link
      fedilink
      arrow-up
      6
      ·
      1 year ago

      They’re definitely reducing model performance to speed up responses. ChatGPT was at its best when it took forever to write out a response. Lately I’ve noticed that ChatGPT will quickly forget information you just told it, ignore requests, hallucinate randomly, and has a myriad of other problems I didn’t have when the GPT-4 model was released.

  • Naatan@lemmy.one
    link
    fedilink
    English
    arrow-up
    4
    ·
    1 year ago

    I wouldn’t be surprised if it is getting worse. It’s not “real” intelligence that “understands” your questions, and unlike more targeted solutions like GitHub copilot they don’t have a strong use-case focus that can guide their progress.

    But I think it’s also that people are coming to terms with what ChatGPT actually can and more importantly cannot do. It’s crazy sometimes to hear what the average person thinks the current iteration of AI’s is capable of.

  • Dax87@forum.stellarcastle.net
    link
    fedilink
    English
    arrow-up
    3
    ·
    1 year ago

    Yeah I just cancelled my plus sub because it’s not valuable to me anymore. It feels nearly as bad as 3.5 at times, and having to go back and forth with it on a 25 message per 3 hour budget is extremely stupid.

  • Llamajockey@lemmy.world
    link
    fedilink
    English
    arrow-up
    1
    ·
    1 year ago

    A month ago, as a man working in IT and graphic design I was getting the “ai is going to replace you” every day and I’m loving the ai decline headlines this week.

    • MHLoppy@fedia.io
      link
      fedilink
      arrow-up
      2
      ·
      1 year ago

      GPT-4 is included in the study and is not free though lol. It’s actually kind of expensive to use it for lots of queries.

    • dan1101@lemmy.world
      link
      fedilink
      English
      arrow-up
      2
      ·
      1 year ago

      It’s good at writing sentences. The content of the sentences may or may not be real/true.

      • Dr. Dabbles@lemmy.world
        link
        fedilink
        English
        arrow-up
        1
        ·
        1 year ago

        I’m just not impressed with guessing which words to use next. Especially when I have to verify what it produces.

    • cybersandwich@lemmy.world
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      It’s pretty great at writing short utility scripts and code. And it’s fantastic at explaining errors, warnings, and log file dumps.

      That’s what I use it for.

      • Dr. Dabbles@lemmy.world
        link
        fedilink
        English
        arrow-up
        2
        ·
        1 year ago

        Strongly disagree with explaining things, because you don’t know if it’s correct. And you have to validate code it creates, so 🤷‍♂️

        I’ve asked it to produce C for a specific product, and it effectively summarized and reproduced existing example code. Being able to so easily discover a source it used for training revealed the entire trick at once.