• SJ_Zero@lemmy.fbxl.net
    link
    fedilink
    English
    arrow-up
    29
    arrow-down
    1
    ·
    1 year ago

    I’ve been saying for a while now: AI demos really well, but when you actually need it to do a thing it often fails spectacularly.

    It’s a verisimilitude engine: It tries to make something that looks like it should be right, rather than actually trying to be right. Sometimes the easiest answer is the right answer so it gives that, but when you start asking it harder questions, it’ll just make something up that looks right but isn’t.

    • AzPsycho@lemmy.world
      link
      fedilink
      English
      arrow-up
      21
      ·
      1 year ago

      I experimented by asking it to write a SQL query for a platform that has its entire database map available online. The data I asked for was impossible to get without exporting some of the data in those tables into temp tables using sub queries and then running comparative omissions analysis.

      Instead of doing that it just made up fake tables and wrote a query that proclaimed the data was in these fake tables.

      • SJ_Zero@lemmy.fbxl.net
        link
        fedilink
        English
        arrow-up
        6
        arrow-down
        1
        ·
        1 year ago

        I asked it to write a review of beowulf in the style of beowulf. It wrote something rhyming which is not they style of beowulf. I said “rewrite this so it doesn’t rhyme” and it gave me something rhyming. I tried several times in several different ways including reasoning with it, and it just kept on kicking out a rhyming poem.

      • 50gp@kbin.social
        link
        fedilink
        arrow-up
        3
        ·
        1 year ago

        its good to remember that many of these chatbot AIs want to give an answer to the prompt instead of saying “sorry, thats not possible” and will then generate something completely garbage as result

      • fearout@kbin.social
        link
        fedilink
        arrow-up
        3
        arrow-down
        1
        ·
        1 year ago

        Out of curiosity, are you using 3.5 or 4? I found that gpt4 is pretty good at these tasks, while 3.5 is almost useless. A thing that often helps is to ask it “is your answer correct?”. That seems to make it find the errors and fix them.

    • discodoubloon@kbin.social
      link
      fedilink
      arrow-up
      6
      ·
      edit-2
      1 year ago

      I asked for help with a budget and it came up with $1000 extra out of thin air, cleverly spaced. I almost didn’t even notice!

      I’d say I trust it with language and literary ventures. I’ve heard it can be solid for slapping down code…

      But yeah the second day it came out I tried to get it to write spec sheets and it was always critically wrong somewhere. The gist was there but it couldn’t reference anything tangible.

      It’s a “first draft” engine. It’s fantastic for that, but anyone making claims or looking for anything else out of it is just looking in the mirror.

      • SJ_Zero@lemmy.fbxl.net
        link
        fedilink
        English
        arrow-up
        3
        ·
        1 year ago

        That’s a good way of putting it. I’ve made use of it to get a preliminary gist of things, but ultimately I end up needing to use my human brain to actually come up with final solutions to my problems.

    • Xanthobilly@lemmy.world
      link
      fedilink
      English
      arrow-up
      5
      ·
      1 year ago

      To be clear, you’re referring to generative AI and artificial general intelligence. There are many forms of AI that downright deliver. The field is evolving at such a speed, I think nay-saying will look naïve in hindsight.

      • donuts@kbin.social
        link
        fedilink
        arrow-up
        2
        arrow-down
        1
        ·
        1 year ago

        Too many forms of AI (for example AI “art”) rely so heavily on the idea that anything that can be seen on the internet is fair game to feed into a dataset regardless of license or intellectual property, that I don’t see any legitimate path forward.

        EITHER these companies start paying data/IP owners a license (which may be prohibitively expensive for the amount of data that’s needed to produce anything good) OR society is completely overhauled/evolved in a way so that copyright and intellectual property are no longer things that exist (which is unlikely considering the political resistance to anything even remotely socialist, and likely means the end of human-made commercial media as artists would no longer be able to make a living.)

        I don’t see any way in which the status quo, in which artists are just being flagrantly ripped off by AI companies, can legitimately continue.

      • SJ_Zero@lemmy.fbxl.net
        link
        fedilink
        English
        arrow-up
        1
        ·
        1 year ago

        I’ve been involved in a lot of failed AI projects (failed at the onset because the premise was wrong for the job). The field didn’t start yesterday.

        Not saying it can’t work at all ever, but it’s much more limited than people think it is based on a good demo in chatgpt.

    • Senicar@social.cyb3r.dog
      link
      fedilink
      English
      arrow-up
      3
      ·
      1 year ago

      The archive link to the AI article is hilarious. It’s neither chronological order nor canonical order. It’s not even a comprehensive list, a bunch of stuff is missing. It’s just bullet points of some star wars media vaguely ordered.

      Can’t wait for search results to become even more poisoned by this useless garbage!

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

        Agreed. The amount of effort required to validate the trash that LLMs will simply make up will be one of the biggest challenges for this type of AI to overcome.