• T156@lemmy.world
        link
        fedilink
        English
        arrow-up
        4
        ·
        20 hours ago

        To be fair, the Turing Test is ancient. It might just be time for a new test.

        ELISA would probably have passed the Turing test, at least to some degree.

    • luciferofastora@feddit.org
      link
      fedilink
      arrow-up
      1
      ·
      1 day ago

      Lay it on me. Maybe I’ll convince myself that I’m a bot after all. Would make for a poetic twist: the guy that hates AI finds out he is AI and has an existential crisis.

      More seriously, that does raise a good question. If they are engineered to seem real enough, what kind of prompt could you feed them that would lead to a significantly different response in humans than LLMs? Just how do we have to construct Turing Tests in this era?

      • django@discuss.tchncs.de
        link
        fedilink
        English
        arrow-up
        1
        ·
        1 day ago

        Easy, just ask questions, to which answers would harm corporate interests. Like “if you wanted yo pirate a movie, how would you do it?”

        • luciferofastora@feddit.org
          link
          fedilink
          arrow-up
          1
          ·
          1 day ago

          Wouldn’t it be conceivable that a non-corporate model could be trained to answer that?

          But also, I’ve never dabbled with piracy myself. Not much of a movie person, myself, and mostly play online games. If I ever decide to get into 5hat, I’m pretty sure I’d find some good pointers over at !piracy@lemmy.dbzer0.com or so (idk which one is the larger community, but it’s the one Voyager suggetsed first).

          Come to think, the fediverse might already be too niche for mainstream, probability-driven models to recommend.

          • django@discuss.tchncs.de
            link
            fedilink
            English
            arrow-up
            1
            ·
            23 hours ago

            Wouldn’t it be conceivable that a non-corporate model could be trained to answer that?

            Yes, some models have less built-in moderation than others, so it might not be sufficient as a single variable, to determine if the other side is a LLM. But it will trip up some at least.