• Karmmah@lemmy.world
    link
    fedilink
    arrow-up
    72
    ·
    1 year ago

    I also noticed that chatGPT can’t actually correct itself. It just says “oh sorry, here’s something different” and gives you another crap answer. I noticed it with code specifically. If I remember correctly it was better when it was brand new.

    • squiblet@kbin.social
      link
      fedilink
      arrow-up
      25
      ·
      1 year ago

      The apology thing is sort of hilarious. I wonder what exactly they did to make it eternally apologetic. There was an article on HN recently about how it is basically impossible to get Chat GPT to stop apologizing, as in, if you ask it to stop, it will apologize for apologizing.

      • SokathHisEyesOpen@lemmy.ml
        link
        fedilink
        arrow-up
        12
        ·
        1 year ago

        I experienced exactly that! I told it to stop apologizing for everything and just respond with correct answers and it apologized for not being able to stop apologizing.

        • squiblet@kbin.social
          link
          fedilink
          arrow-up
          5
          ·
          1 year ago

          From HN,

          Me: Can you stop using the word apologize whenever you are talking to me?

          ChatGPT: I apologize (I will try to stop using the word “apologize” in my responses to you) if the use of the word “apologize” in my responses has been inappropriate or has caused any discomfort. As an AI language model, I am programmed to try to be as helpful and accurate as possible, and I apologize (I will try to use a different word) when I make mistakes or do not provide the information or assistance that is requested.

          I will make an effort to avoid using the word “apologize” in my responses to you in the future. Please do not hesitate to let me know if you have any further questions or need any additional information. I am here to help and will do my best to provide accurate and helpful responses to your questions.

          Also here is a recent HN discussion: https://news.ycombinator.com/item?id=36949931

      • relevants@feddit.de
        link
        fedilink
        arrow-up
        8
        ·
        1 year ago

        It’s because humans have rated potential responses and ChatGPT has been trained to generate the kind of responses that most consistently get preferred rating. You can imagine how an AI trained to say what people want to hear would become a people pleaser.

      • CeruleanRuin@lemmy.world
        link
        fedilink
        English
        arrow-up
        4
        ·
        1 year ago

        That’s what frustrates me the most whenever I try to use it. I tell it to be less verbose, stop over explaining and apologizing every time I correct it, and it just spits out another four paragraphs explaining why it’s sorry.

      • XEAL@lemm.ee
        link
        fedilink
        arrow-up
        1
        ·
        1 year ago

        The only solution I can think of is using it via API with Python and make a call with the final reply asking it to remove apologies from the text, but the token usage will increase.

        I do something similar when I need to tell the model to keep the language of a text before performing a task with that text. I send the model a chunk of text and ask it to respond with single word, indicating the language of the text and then I include that in the next prompt like “Your output must be in SPANISH”, or whatever.

  • BadRS@lemmy.world
    link
    fedilink
    arrow-up
    34
    arrow-down
    4
    ·
    1 year ago

    It cannot read. It doesn’t see words or letters. It works with Tokens which words are converted into. It cant count the number of letters in a word because it can’t see them. OpenAI has a Tokenizer you can plug a prompt into to see how its broken up, but youre asking a fish to fly.

    • ParkingPsychology@kbin.social
      link
      fedilink
      arrow-up
      5
      ·
      1 year ago

      And a single “S” is also a token. Which has vectors to all other words that start with an S.

      One thing to point out here is that the word sentences is severely mistyped as “sententences”. That’s not going to help.

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

      Is there a workaround to “trick” it into understanding letters? I’d love to use it to play with language and brainstorm some riddles or other wordplay, but if it literally can’t understand language on a human level, that’s a fools errand.

  • SimpleDev@infosec.pub
    link
    fedilink
    English
    arrow-up
    12
    arrow-down
    1
    ·
    1 year ago

    Idk what I’m doing wrong, thankfully it always seems to listen and work fine for me lmao

      • SimpleDev@infosec.pub
        link
        fedilink
        English
        arrow-up
        7
        ·
        edit-2
        1 year ago

        Now it’s broken, I guess I I don’t use it this way often enough. Interesting nonetheless!

        Edit - it’s very semantic, it matters if I include an uppercase “S” or not. That’s amusing.

        I wonder if the temperature settings adjustment would fix that or just make it even weirder.

          • SimpleDev@infosec.pub
            link
            fedilink
            English
            arrow-up
            3
            ·
            edit-2
            1 year ago

            But it’s able to correct unlike what’s shown in the OP messages.

            Extremely semantically it seems but it clearly listens. It’s neat to see how different each person experience is.

            Also different tuning parameters etc…could make outputs different. That might explain why mine is seemingly a bit better at listening.

      • SimpleDev@infosec.pub
        link
        fedilink
        English
        arrow-up
        0
        ·
        1 year ago

        Look at the first question in the my first screenshot. It gets that question correct for “mayonnaise” lol

          • SimpleDev@infosec.pub
            link
            fedilink
            English
            arrow-up
            1
            ·
            edit-2
            1 year ago

            I feel like ChatGPT itself probably has a fairly loose temp setting (just a hunch) and I tend to set my conversations up to be more on the strict side

            I imagine that’s why our results differ, it’s strange OpenAI doesn’t let ChatGPT site users or at least premium users adjust anything really yet.

  • CeruleanRuin@lemmy.world
    link
    fedilink
    English
    arrow-up
    6
    ·
    1 year ago

    ChatGPT doesn’t understand letters, or phonetics, or most other aspects of speech. I tried for an hour to train it to understand what a palindrome is, with the hopes of getting it to generate some new ones. Nothing stuck. It was like trying to teach a dog to write its name.

  • Dodecahedron December@sh.itjust.works
    link
    fedilink
    arrow-up
    8
    arrow-down
    2
    ·
    1 year ago

    Y’all seem to gloss over the word artificial when it comes to reading “artificial intelligence”. That or you’re leaning too hard on the first definition…

    1. made or produced by human beings rather than occurring naturally, especially as a copy of something natural. “her skin glowed in the artificial light”
    2. (of a person or their behavior) insincere or affected. “an artificial smile” 🤖
    • CeruleanRuin@lemmy.world
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      It’s just so counterintuitive for a layman to have this tool that can write long flowing passages of text and theoretically pass a rudimentary Turin test, but it can’t even begin to work with language on the level most toddlers can. We humans typically have to learn letters before we move up to words, sentences, paragraphs, and finally whole compositions. But this thing skipped right over the first several milestones and has no mechanism for reverse engineering that capability.

    • SokathHisEyesOpen@lemmy.ml
      link
      fedilink
      arrow-up
      2
      arrow-down
      2
      ·
      1 year ago

      It has not. ChatGPT has been a monumental achievement and has been capable of performing previously impossible and highly impressive tasks. This is new behavior for it.