With AI closing in on general intelligence, using just binary numbers and some calculations, I’m going to place my bet that quantum entanglement is not necessarily for general intelligence.
That’s what the hype men say, but increasingly I don’t believe it. I read yesterday that most major AI systems cannot correctly answer “which is larger, 7.11 or 7.9?”.
Often I read of these strange examples, and struggle to remember each one or provide accurate examples when people challenge my claims that these systems still have many flaws. And fair enough. I’ve thought about that recently, and I think this is partly due to the fact that intelligence is a complex highly dimensional phenomenon, and failures in artificial systems manifest as all these weird little quirks and strange behaviors and failures and they are honestly hard for me to remember as an interested lay person.
That said, these systems have loads of shortcomings, and in particular they’re not good at learning from small datasets which means they can’t easily generalize to humanoids (for example) doing the millions of little real world tasks humans do. And they can’t serve as general algorithms for all software problems. They work well with text and pretty well with images but I think we will find for a long time to come significant shortcomings that prevent them from really manifesting as a truly general intelligence.
And I don’t know about this quantum stuff but until we actually solve AGI I don’t think we can really say with confidence what it takes to make intelligence function.
>I read yesterday that most major AI systems cannot correctly answer “which is larger, 7.11 or 7.9?”.
It's largely irrelevant whether an AI system (say, GPT for example) can or can't correctly answer this number question. Many otherwise normal, socially functional humans would be stumped by it too though they without a doubt have self directed cognition and more or less the same sense of self awareness that you or I feel.
The real distinction with GPT or any known modern AI system is that it lacks that sensor of distinct self and its essential cognition, regardless of how well they do with number puzzles.
Something like GPT being largely able to pass the Turing test in the eyes of a human interacting with it doesn't demonstrate that the AI system has cognition, it instead evidently demonstrates that the Turing test was always a poor metric of advanced AI performance..
I agree with this! I simply find it hard to explain the shortcomings in a way that people accept, and clear examples like this are the best I can do. But you’ve said it nicely.
It’s ambiguous with software releases but if someone says “which number is bigger, 7.11 or 7.9?” With no other context, I doubt any human with basic mathematical understanding would provide an incorrect answer. That is not an ambiguous question unless the concept of software releases was specifically introduced in to the conversation.
Worst part about these LLMs is they don’t ask clarifying questions, they just provide false answers.
I don't believe that we are closing in on general intelligence, and I have some ideas as to why.
The universe is a chaotic system, the only thing in it that might be deterministic is the universe in totality, but even then we have relativity and everything in it is happening in it's own universe with respect to time. It's like a giant real time distributed system with a speed of light latency. Human made computers are designed to isolate as well as possible determinism, built using the chaotic substrate, the material of the universe which is the universe itself. It's not perfect because you can't entirely isolate any system, so you get bit flips that are caused by the chaos of the universe, but they do pretty good. The brain is evolved, it doesn't care if it's deterministic or not, it only emerged to work properly, whatever that means. Digital computers are designed with a goal in mind, to isolate from the chaos and deliver determinism for practical use. Nothing else in the universe works this way, not brains, not gas giants, nothing is constructed to isolate a deterministic space.
I believe that this chaos is where intelligence and consciousness come from, or at the very least that it is a core component of what they are. With deep learning and tensor models and the like, we are attempting to emulate the chaotic nature of a brain on aachine isolated from it's chaotic substrate. Even if you can get something that kind of resembles a mind, you've got massive unnecessary overhead due to this emulation. It's legacy architecture all the way down. And I believe this isolation has it's own impact, an emulated chaotic mind on a digital computer, prevents it from interacting with the world around it fully on the bare metal of the universe, low latency via the API that is direct interaction. Until we decide to scrap the emulation and build a physical architecture specifically for the purpose of generating thought I don't think we will make much headway on general intelligence. Such an architecture in my view can have no abstraction or emulation in it's design, it will have to be entirely analog, less than analog even, like a car cigarette lighter.
That’s what the hype men say, but increasingly I don’t believe it. I read yesterday that most major AI systems cannot correctly answer “which is larger, 7.11 or 7.9?”.
Often I read of these strange examples, and struggle to remember each one or provide accurate examples when people challenge my claims that these systems still have many flaws. And fair enough. I’ve thought about that recently, and I think this is partly due to the fact that intelligence is a complex highly dimensional phenomenon, and failures in artificial systems manifest as all these weird little quirks and strange behaviors and failures and they are honestly hard for me to remember as an interested lay person.
That said, these systems have loads of shortcomings, and in particular they’re not good at learning from small datasets which means they can’t easily generalize to humanoids (for example) doing the millions of little real world tasks humans do. And they can’t serve as general algorithms for all software problems. They work well with text and pretty well with images but I think we will find for a long time to come significant shortcomings that prevent them from really manifesting as a truly general intelligence.
And I don’t know about this quantum stuff but until we actually solve AGI I don’t think we can really say with confidence what it takes to make intelligence function.