• CineMaddie@lemmy.film
    link
    fedilink
    arrow-up
    34
    arrow-down
    1
    ·
    1 year ago

    Why are we relying on language models to answer questions. These things don’t really “know” anything right?

    • NightFantom@slrpnk.net
      link
      fedilink
      arrow-up
      14
      ·
      1 year ago

      They don’t, but they sound as convincing, (and are probably as correct) as a random blog you’d find googling your question

    • Catoblepas@lemmy.blahaj.zone
      link
      fedilink
      arrow-up
      4
      ·
      1 year ago

      They don’t, and people are way too blasé about how “oh it’s actually the same as googling because it’s just taking from sources online anyway,” when in reality it does nothing to “keep” the knowledge it gets from those sites and is just stringing together words that often go together. It’s like thinking your phone’s predictive text can answer your questions, if your phone also invented quotes and sources (this has already been an issue with journalists and lawyers using ChatGPT to “research”).

      • CineMaddie@lemmy.film
        link
        fedilink
        arrow-up
        2
        arrow-down
        1
        ·
        1 year ago

        Thank god someone who understands. I hated how towards the end reddit was so full of misinformation and people talking out of their ass with confidence. Hope Lemmy can steer away from those tendencies. It’s okay if we don’t have the answer sometimes.

    • Sethayy@sh.itjust.works
      link
      fedilink
      arrow-up
      4
      arrow-down
      5
      ·
      1 year ago

      No one knows anything get over yourself buddy - it gave a correct answer way more polite than I ever could so who’s gonna complain

    • jscummy@sh.itjust.works
      link
      fedilink
      arrow-up
      2
      arrow-down
      4
      ·
      1 year ago

      Don’t they pull from online sources? So it’s basically googling with extra steps and an unpredictable middleman

      • CineMaddie@lemmy.film
        link
        fedilink
        arrow-up
        3
        ·
        1 year ago

        That would be right if they understood/knew what they were talking about. It’s more akin to really advanced autocorrect that sounds/reads like something the ai was trained on. So it sounds correct but really has 0 basis on truth other than “the model predicts a human would say X next”. Truth is rarely the goal of any of these machine learning language models afaik.