Artificial intelligence is worse than humans in every way at summarising documents and might actually create additional work for people, a government trial of the technology has found.

Amazon conducted the test earlier this year for Australia’s corporate regulator the Securities and Investments Commission (ASIC) using submissions made to an inquiry. The outcome of the trial was revealed in an answer to a questions on notice at the Senate select committee on adopting artificial intelligence.

The test involved testing generative AI models before selecting one to ingest five submissions from a parliamentary inquiry into audit and consultancy firms. The most promising model, Meta’s open source model Llama2-70B, was prompted to summarise the submissions with a focus on ASIC mentions, recommendations, references to more regulation, and to include the page references and context.

Ten ASIC staff, of varying levels of seniority, were also given the same task with similar prompts. Then, a group of reviewers blindly assessed the summaries produced by both humans and AI for coherency, length, ASIC references, regulation references and for identifying recommendations. They were unaware that this exercise involved AI at all.

These reviewers overwhelmingly found that the human summaries beat out their AI competitors on every criteria and on every submission, scoring an 81% on an internal rubric compared with the machine’s 47%.

  • ArbitraryValue@sh.itjust.works
    link
    fedilink
    English
    arrow-up
    26
    arrow-down
    7
    ·
    edit-2
    3 months ago

    The important thing here isn’t that the AI is worse than humans. It’s than the AI is worth comparing to humans. Humans stay the same while software can quickly improve by orders of magnitude.

    • WalnutLum@lemmy.ml
      link
      fedilink
      English
      arrow-up
      4
      ·
      3 months ago

      LLMs as they stand are already approaching the improvement flatline portion of the sigma curve due to marginal data requirements increasing exponentially.

      It’s a known problem in the actual AI research field that nobody in private industry likes to talk about.

      If it scores 40% this year it’ll marginally increase by 10% next year then 5% 3 years later and so on.

      AI doesn’t follow Moore’s law.

      • ArbitraryValue@sh.itjust.works
        link
        fedilink
        English
        arrow-up
        4
        arrow-down
        1
        ·
        edit-2
        3 months ago

        So far “more data” has been the solution to most problems, but I don’t think we’re close to the limit of how much useful information can be learned from the data even if we’re close to the limit of how much data is available. Look at the AIs that can’t draw hands. There are already many pictures of hands from every angle in their training data. Maybe just having ten times as many pictures of hands would solve the problem, but I’m confident that if that was not possible then doing more with the existing pictures would also work.* Algorithm design just needs some time to catch up.

        *I know that the data that is running out is text data. This is just an analogy.

    • krashmo@lemmy.world
      link
      fedilink
      English
      arrow-up
      4
      arrow-down
      3
      ·
      3 months ago

      Theoretically that’s true. Can you tell techbros and the media to shut up about AI until it happens though?

    • ContrarianTrail@lemm.ee
      link
      fedilink
      English
      arrow-up
      10
      arrow-down
      12
      ·
      3 months ago

      The AI we have today is the worst it’ll ever be. I can only think of two possible scenarios where AI doesn’t eventually surpass human on every single cognitive task:

      1. There’s something fundamentally different about computer made of meat (our brains) that cannot be replicated in silica. I personally don’t see this as very likely since both are made of matter and matter obeys the laws of physics.

      2. We destroy ourselves before we reach AGI.

      Otherwise we’ll keep improving our technology and inching forward. It may take 5 years or 50 but it wont stop unless either of the scenarios stated above is true.

      • Ultraviolet@lemmy.world
        link
        fedilink
        English
        arrow-up
        10
        arrow-down
        3
        ·
        3 months ago

        LLMs are fundamentally a dead end though. If we ever create AGI, it will be a qualitatively different thing from an LLM.

        • ContrarianTrail@lemm.ee
          link
          fedilink
          English
          arrow-up
          10
          arrow-down
          3
          ·
          3 months ago

          It’s not obvious to me as to why this is for 100% certainty going to be the case. Even if it’s likely true, there’s still a chance it might not be.

          • Wooki@lemmy.world
            link
            fedilink
            English
            arrow-up
            4
            arrow-down
            4
            ·
            edit-2
            3 months ago

            Zero chance IBMs most likely word predictor will become anything more than what it is programmed to be. It is not magic, witches dont exist.

            • rottingleaf@lemmy.world
              link
              fedilink
              English
              arrow-up
              3
              arrow-down
              1
              ·
              3 months ago

              People were being shown deus ex machina in supposedly sci-fi movies and series for many years.

              Only there it was always meant as 1 in a billion event, as a miracle.

              Here a lot of people want to streamline miracles, while even one hasn’t been produced yet.

              It’s the difference between Tolkien’s dwarves and Disney’s gnomes.

      • ArbitraryValue@sh.itjust.works
        link
        fedilink
        English
        arrow-up
        6
        arrow-down
        2
        ·
        edit-2
        3 months ago

        It would be odd if AI somehow got worse. I mean, wouldn’t they just revert to a backup?

        Anyway, I think (1) is extremely unlikely but I would add (3) the existing algorithms are fundamentally insufficient for AGI no matter how much they’re scaled up. A breakthrough is necessary which may not happen for a long time.

        I think (3) is true but I also thought that the existing algorithms were fundamentally insufficient for getting to where we are now, and I was wrong. It turns out that they did just need to be scaled up…

        • Wooki@lemmy.world
          link
          fedilink
          English
          arrow-up
          6
          arrow-down
          2
          ·
          edit-2
          3 months ago

          It would be odd if AI somehow got worse.

          No its not odd at all, its the opposite, it is happening and multiple studies are showing its decay is being caused by feedback entropy which is a real problem to remove

          • ArbitraryValue@sh.itjust.works
            link
            fedilink
            English
            arrow-up
            3
            arrow-down
            2
            ·
            edit-2
            3 months ago

            Multiple studies are showing that training on data contaminated with LLM output makes LLMs worse, but there’s no inherent reason why LLMs must be trained on this data. As you say, people are aware of it and they’re going to be avoiding it. At the very least, they will compare the newly trained LLM to their best existing one and if the new one is worse, they won’t switch over. The era of being able to download the entire internet (so to speak) is over but this means that AI will be getting better more slowly, not that it will be getting worse.

        • ContrarianTrail@lemm.ee
          link
          fedilink
          English
          arrow-up
          3
          arrow-down
          2
          ·
          3 months ago

          It’s possible that the way of generative AI and LLMs is a dead end but that wouldn’t be a stop, only a speed bump. It would only mean it takes longer for us to get there, not that we wouldn’t get there.

          • ArbitraryValue@sh.itjust.works
            link
            fedilink
            English
            arrow-up
            5
            arrow-down
            1
            ·
            3 months ago

            I don’t disagree, but before the recent breakthroughs I would have said that AI is like fusion power in the sense that it has been 50 years away for 50 years. If the current approach doesn’t get us there, who knows how long it will take to discover one that does?

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

              Right and all the dogs in the race are now focused on neural networks and llms, which means for now, all the effort could be focused on a dead end. Because of the way capitalism is driving AI research, other avenues of AI research have almost effectively halted, so it will take the current AI bubble to pop before alternative research ramps up again

              • Jesus_666@lemmy.world
                link
                fedilink
                English
                arrow-up
                4
                ·
                3 months ago

                Like every time there’s an AI bubble. And like every time changes are that in a few years public interest will wane and current generative AI will fade into the background as a technology that everyone uses but nobody cares about, just like machine translation, speech recognition, fuzzy logic, expert systems…

                Even when these technologies get better with time (and machine translation certainly got a lot better since the sixties) they fail to recapture their previous levels of excitement and funding.

                We currently overcome what popped the last AI bubbles by throwing an absurd amount of resources at the problem. But at some point we’ll have to admit that doubling the USA’s energy consumption for a year to train the next generation of LLMs in hopes of actually turning a profit this time isn’t sustainable.

                • ArbitraryValue@sh.itjust.works
                  link
                  fedilink
                  English
                  arrow-up
                  2
                  ·
                  3 months ago

                  The issue I have with referring to the current situation as a bubble is that this isn’t just hype. The technology really is amazing, and far better than what people had been expecting. I do think that most current attempts to commercialize it are premature, but there’s such a big first-mover advantage that it makes sense to keep losing money on attempts that are too early in order to succeed as soon as it is possible to do so.

              • rottingleaf@lemmy.world
                link
                fedilink
                English
                arrow-up
                2
                ·
                edit-2
                3 months ago

                I think that’s intentional. Nation states and other powers that be have working propaganda mechanisms.

                A real AGI is a change most important in the sense of power, not in the sense of economy (because we know how to make new humans and educate them, it wouldn’t be a qualitative change there).

                All this AI gaslighting is intended to stall real advancements there.

                The Web in some sense was produced in the context of AI research. In general semantic and hypertext systems were. And look what it has done to the world. They may just not want another such cataclysm.

                EDIT: Also notice the shift from the hypertext paradigm to the application platform paradigm in the Web.

            • ContrarianTrail@lemm.ee
              link
              fedilink
              English
              arrow-up
              2
              ·
              3 months ago

              The timeline doesn’t really matter to me personally. As long as we accept the fact that we’ll get there sooner or later it should motivate us to start thinking about the implications that comes with. Otherwise it’s like knowing there’s an asteroid hurling towards the earth but we’ll just dismiss it by saying: “Eh, it’s still 100 years away, there’s no rush here”