You know how Google’s new feature called AI Overviews is prone to spitting out wildly incorrect answers to search queries? In one instance, AI Overviews told a user to use glue on pizza to make sure the cheese won’t slide off (pssst…please don’t do this.)

Well, according to an interview at The Vergewith Google CEO Sundar Pichai published earlier this week, just before criticism of the outputs really took off, these “hallucinations” are an “inherent feature” of  AI large language models (LLM), which is what drives AI Overviews, and this feature “is still an unsolved problem.”

  • Excrubulent@slrpnk.net
    link
    fedilink
    English
    arrow-up
    167
    arrow-down
    13
    ·
    5 months ago

    No he’s right that it’s unsolved. Humans aren’t great at reliably knowing truth from fiction too. If you’ve ever been in a highly active comment section you’ll notice certain “hallucinations” developing, usually because someone came along and sounded confident and everyone just believed them.

    We don’t even know how to get full people to do this, so how does a fancy markov chain do it? It can’t. I don’t think you solve this problem without AGI, and that’s something AI evangelists don’t want to think about because then the conversation changes significantly. They’re in this for the hype bubble, not the ethical implications.

    • dustyData@lemmy.world
      link
      fedilink
      English
      arrow-up
      86
      arrow-down
      11
      ·
      5 months ago

      We do know. It’s called critical thinking education. This is why we send people to college. Of course there are highly educated morons, but we are edging bets. This is why the dismantling or coopting of education is the first thing every single authoritarian does. It makes it easier to manipulate masses.

      • Excrubulent@slrpnk.net
        link
        fedilink
        English
        arrow-up
        59
        arrow-down
        1
        ·
        5 months ago

        “Edging bets” sounds like a fun game, but I think you mean “hedging bets”, in which case you’re admitting we can’t actually do this reliably with people.

        And we certainly can’t do that with an LLM, which doesn’t actually think.

          • Excrubulent@slrpnk.net
            link
            fedilink
            English
            arrow-up
            11
            arrow-down
            1
            ·
            edit-2
            5 months ago

            A big problem with that is that I’ve noticed your username.

            I wouldn’t even do that with Reagan’s fresh corpse.

        • explore_broaden@midwest.social
          link
          fedilink
          English
          arrow-up
          5
          ·
          5 months ago

          I think that’s more a function of the fact that it’s difficult to verify that every one of the over 1M college graduates each year isn’t a “moron” (someone very bad about believing things other people made up). I think it would be possible to ensure a person has these critical thinking skills with a concerted effort.

          • Excrubulent@slrpnk.net
            link
            fedilink
            English
            arrow-up
            5
            arrow-down
            2
            ·
            5 months ago

            The people you’re calling “morons” are orders of magnitude more sophisticated in their thinking than even the most powerful modern AI. Almost every single one of them can easily spot what’s wrong with AI hallucinations, even if you consider them “morons”. And also, by saying you have to filter out the “morons”, you’re still admitting that a lot of whole real assed people are still not reliably able to sort fact from fiction regardless of your education method.

            • explore_broaden@midwest.social
              link
              fedilink
              English
              arrow-up
              3
              ·
              5 months ago

              No I still agree that we are far from LLMs being ‘thinking’ enough to be anywhere near this. But if we had a bunch of models similar to LLMs that could actually think, or if we really needed to select a person, I do think it would be possible to evaluate a bunch of the models/people to determine which ones are good at distinguishing fake information.

              All I’m saying is I don’t think the limitation is actually our ability to select for capability in distinguishing fake information, I think the only limitation is fundamental to how current LLMs work.

              • Excrubulent@slrpnk.net
                link
                fedilink
                English
                arrow-up
                3
                ·
                5 months ago

                Yes, my point wasn’t that it could never be achieved but that LLMs are in a completely different category, which we agree on I think. I was comparing them to humans who have trouble with critical thinking but can easily spot AI’s hallucinations to illustrate the vast gulf.

                In both cases I think there are almost certainly more barriers in the way than an education. The quest for a truthful AI will be as contentious as the quest for truth in humans, meaning all the same claim-counterclaim culture-war propaganda tug of war will happen, which I think is the main reason for people being miseducated against critical thinking. In a vacuum it might be a simple technical and educational challenge, but the reason this is a problem in the first place is that we don’t exist in a political vacuum.

        • dustyData@lemmy.world
          link
          fedilink
          English
          arrow-up
          1
          arrow-down
          9
          ·
          5 months ago

          Choose a lane, this comment directly contradicts you previous comment. I think you are just trolling and being an idiot with corrections to elicit reactions.

      • RidcullyTheBrown@lemmy.world
        link
        fedilink
        English
        arrow-up
        8
        arrow-down
        3
        ·
        5 months ago

        What does this have to do with AI and with what OP said? Their point was obviously about limitations of the software, not some lament about critical thinking

      • helenslunch@feddit.nl
        link
        fedilink
        English
        arrow-up
        1
        arrow-down
        1
        ·
        4 months ago

        It’s called critical thinking education.

        Yeah, I mean, we have that, and parents are constantly trying to dismantle it. No amount of “critical thinking education” can undo decades of brainwashing from parents and local culture.

    • scarabic@lemmy.world
      link
      fedilink
      English
      arrow-up
      2
      ·
      5 months ago

      Humans aren’t great at reliably knowing truth from fiction too

      You’re exactly right. There is a similar debate about automated cars. A lot of people want them off the roads until they are perfect, when the bar should be “until they are safer than humans,” and human drivers are fucking awful.

      Perhaps for AI the standard should be “more reliable than social media for finding answers” and we all know social media is fucking awful.

      • Excrubulent@slrpnk.net
        link
        fedilink
        English
        arrow-up
        1
        ·
        edit-2
        5 months ago

        The problem with these hallucinated answers that makes them such a sensational story is that they are obviously wrong to virtually anyone. Your uncle on facebook who thinks the earth is flat immediately knows not to put glue on pizza. It’s obvious. The same way It’s obvious when hands are wrong in an image or someone’s hair is also the background foliage. We know why that’s wrong; the machine can’t know anything.

        Similarly, as “bad” as human drivers are we don’t get flummoxed because you put a traffic cone on the hood, and we don’t just drive into tue sides of trucks because they have sky blue liveries. We don’t just plow through pedestrians because we decided the person that is clearly standing there just didn’t matter. Or at least, that’s a distinct aberration.

        Driving is a constant stream of judgement calls, and humans can make those calls because they understand that a human is more important than a traffic cone. An autonomous system cannot understand that distinction. This kind of problem crops up all the time, and it’s why there is currently no such thing as an unsupervised autonomous vehicle system. Even Waymo is just doing a trick with remote supervision.

        Despite the promises of “lower rates of crashes”, we haven’t actually seen that happen, and there’s no indication that they’re really getting better.

        Sorry but if your takeaway from the idea that even humans aren’t great at this task is that AI is getting close then I think you need to re-read some of the batshit insane things it’s saying. It is on an entirely different level of wrong.