• Zephyr@sh.itjust.works
    link
    fedilink
    English
    arrow-up
    5
    arrow-down
    1
    ·
    4 days ago

    Sidebar it’s kinda crazy how the expectations of AI grew both by users and most definitely product / service providers. You remember that first Will Smith spaghetti video? That was just 3 years ago. A lot has changed but also depending on the model someone is using nothing has. Even frontier models are not yet strong AI aka they seriously need human supervision to check what’s coming out. We’re in this awkward stage where they really do work half of the time but also don’t the other half. The applications are far exceeding their consistent capabilities. In short the deployment and messaging about capabilities needs to be scaled back drastically.

    • zarkanian@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      5
      ·
      3 days ago

      The problem is they have to over-hype it because they aren’t making their money back.

      Even frontier models are not yet strong AI aka they seriously need human supervision to check what’s coming out.

      What do you consider “strong AI”?

        • zarkanian@sh.itjust.works
          link
          fedilink
          English
          arrow-up
          5
          ·
          3 days ago

          That’s actual AI (what people are now calling AGI), and LLMs aren’t capable of that. They’d have to develop a new technology.

          • Zephyr@sh.itjust.works
            link
            fedilink
            English
            arrow-up
            1
            ·
            3 days ago

            I’m not saying this current push strictly is. That said it’s far more capable than one would have assumed for an algorithm just doing next word prediction. There’s something in there that can be distilled.

    • antonim@lemmy.worldOP
      link
      fedilink
      English
      arrow-up
      3
      ·
      3 days ago

      Even frontier models are not yet strong AI

      I believe that would require a whole new paradigm of how AI works, the sort of shift that won’t be achieved by mere refinement of the current models. So I’m not expecting surprises on that front right now. On the contrary, what has surprised me the last time I tried to use ChatGPT is that it produced ungrammatical sentences with incomplete words. That’s a sort of “chickenpox” issue that should’ve been fixed long ago by now…