• leftzero@lemmynsfw.com
      link
      fedilink
      English
      arrow-up
      2
      arrow-down
      1
      ·
      edit-2
      6 months ago

      LLMs are incapable of “recognising” any patterns they haven’t been trained on.

      And they don’t really even recognise those, they’re just fancy auto complete engines, simply outputting the highest scored token from their training base based on their input.

      They’re pattern matching machines; there’s no recognition, inner modelling of new knowledge, self referencing, or understanding of any kind, merely blind statistics.

      They’re just bigger and fancier Eliza’s, and just as distant as Eliza was from any practical form of intelligence, artificial or natural.

      While I personally do believe that achieving AGI¹, on a Turing machine is possible, LLMs and how they work are an excellent example in support of John Searle’s arguments against it with his Chinese room though experiment.

      1— Or at least something equivalent to human intelligence, or better, in the measures by which we consider ourselves to be intelligent, though it’s arguable whether we can really be considered intelligent at all, or we’re just better, more complex, Chinese rooms.

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

        But since we don’t understand how cognition works in living beings almost at all – who’s to say that’s not how ‘actual thinking’ works other than 'I know it when I see it!"

        • leftzero@lemmynsfw.com
          link
          fedilink
          English
          arrow-up
          2
          ·
          6 months ago

          Because there are many aspects of what we understand as “actual thinking” (understanding concepts, learning, or solving puzzles, for instance) that LLMs are fundamentally incapable of achieving no matter how larger or more complex we make them or how much we optimise them.

          They do one single thing (which, granted, they do relatively well): they take an input, they apply it to every token in their training data, generating a score for each of them, and they output the one with the highest score. And that’s all they do.

          And that’s why, for instance, you’ll never be able to make a LLM that’s any good at playing chess, because there simply wouldn’t be enough atoms in the universe for it to store all possible states of the game, which it would need to have in its training model in order to auto complete its next move (and that’s not even accounting for the actual score computation, both in space and time).

          They’re a cool fancy gimmick, possibly useful in certain cases as long as you can account for their hallucinations, but they’re not any closer to actual intelligence than Eliza ever was.

          • lanolinoil@lemmy.world
            link
            fedilink
            English
            arrow-up
            3
            ·
            6 months ago

            you’ll never be able to make a LLM that’s any good at playing chess,

            They said that about machines and then we all laughed at the mechanical turk hoax. Now machines can almost beat you in Go.

            I’ll say it again – It is hubris and you will obviously be wrong to try to predict the future or what will have value.

            like come on – superpositioning exists and we’ve no clue how consciousness works (Bostrom thinks its just maths) but you have this crystal ball full of certainty. It smells…

            • leftzero@lemmynsfw.com
              link
              fedilink
              English
              arrow-up
              1
              ·
              6 months ago

              I’m not talking about “machines” or any other generic term.

              I’m talking specifically about LLMs. And their limitations are evident. For instance, maths is one of the many things they can’t do (and will never be able to do in any efficient way).

              We have indeed, developed programs that play chess better than people (though sadly, until the LLM bubble pops we probably won’t get any further). But they’re not LLMs, or anything resembling an LLM. Because one of the other many things an LLM can’t do is play games of skill. Or reason. Or solve puzzles. Or even have a concept of strategy.

              LLMs, again, can only do one single thing. And that’s to pick up the one card from their deck that’s been picked up most often after the sequence of cards on the table according to their training model.

              That’s all they do. That’s all they’ll ever be able to do. Because that’s how they work. And, sure, with that you can make it look like they’re holding a conversation (until you ask them something that isn’t in their model), but that’s it.

              They’ll put words after another according to statistics (not, keep that in mind, according to meaning, or strategy, or anything like that; they don’t, and can’t know or care what the words mean, or whether the sentence they’ve put together makes any sense, or whether what it’s stating is true or false), and that’s that.

              They won’t play chess, they won’t write good innovative code, they won’t write original stories, and they won’t drive your car.

              We don’t need to know how what we call consciousness works to know that. We just need to know how LLMs work. And that we most definitely do.

              • Petter1@lemm.ee
                link
                fedilink
                English
                arrow-up
                1
                ·
                6 months ago

                Most LLM distributors analyze the output (or outputs) using another AI…

              • lanolinoil@lemmy.world
                link
                fedilink
                English
                arrow-up
                1
                ·
                6 months ago

                Sure steam engines may not fit every use but from them we learned to make other kinds of engines right? But yeah I’m sure ‘LLM’ will either change scope/definition or we’ll make new stuff to fit other use cases kind of like diffusion models for images vs llm for text generation.