IQ is defined as having a normal distribution with a standard deviation of 15 points, so there’s no need to assume anything. The raw scores are transformed as necessary to give the right distribution.
About 117 billion people have ever lived, so one could argue that a couple of them were that far out. But most of those were ancient humans and it probably doesn’t make sense to include them (if you did, a lot of modern humans would necessarily get a boost to their IQ).
The smartest person currently on the planet would be ~6.3 standard deviations from the mean, which would be an IQ of ~195. Close to but not quite 200.
IQ tests are scored in a way to fit a Gaussian distribution with a standard deviation of 15, but that is a data-fitted model that is neither statistically rigorous nor particularly meaningful from a modern neuroscience perspective of ‘intelligence’ as a valid metric of the wide array of cognitive abilities, even those that can be measured in terms of intellectual (problem-solving) tasks. IQ should be regarded as a general and somewhat imperfect assessment of specific types of abstract problem-solving ability, particularly in regard to complex and novel algorithmic tasks, but there are plenty of people with high IQs that can write only functionally with a coarse command of language, can’t give a coherent lecture to save their lives, and fail in many basic social norms, while there are people of mean IQ who are socially adept, emotionally attuned to others, and capable of highly complex psychophysical skills that are not measured by a sit-down test. And rRegardless of what IQ tests actually measure, no widely accepted intelligence test purports to assess IQ beyond 4 standard deviations which makes any quantification or estimate of distribution of the tale meaningless.
But setting aside this hijack, IQ is not really a good way to measure the intellectual potential of ‘artificial intelligence’ or any kind of actual machine cognition system (which, as yet, does not exist and a path to true cognition in silicio isn’t even known); the kind of ‘intelligence’ we see in the operations of an LLM or generative AI are not comparable to how a person learns or solves problems, so it is difficult to say what limitations or restrictions machine cognition might experience in terms of developing ‘super-intelligence’, but the most obvious limitations are those associated with its ability to perceive and interact with the physical world in the way that humans (and other animals) do.
I agree that IQ isn’t a great measure at these levels. Even for humans it breaks down well before you get to these extremes (as an aside: if eventually the population reaches levels where a >200 IQ is possible, then you should also expect someone with a negative IQ).
Thinking out loud here, it might make sense to distinguish between two types of superintelligences, which I’ll call Type I and Type II.
A Type I is as if you packed every smart person on the planet in a box and increased the clock speed by a factor of a million. Any problem that could be solved by someone in some years would be solvable by the box in minutes. Such a box would be very useful but it wouldn’t exceed civilization as a whole except in how fast it operates. It does seem to be plausible outcome for AI development, though.
A Type II is a fundamental step past this. It solves problems that no human ever could, no matter how many of them you put together or how much time you give them. It’s not clear that a Type II is possible or even a coherent concept. But it seems like it should be–after all, you can pack a billion very smart dogs into a box and they will never come up with General Relativity, whereas a smart person can. So why shouldn’t there be problems that humans can never solve but a sufficiently advanced AI can?
It’s possible it doesn’t matter if a Type II is possible. A Type I is advanced enough already to completely change civilization, for good or ill.
I would argue that the Type II ‘intelligence’ (using your categorization) already exists in the form of computational tools that can process enormous volumes of data held simultaneously in memory and produce results or tease out patterns that no individual or team of humans could just by dint of being able to perform so many algorithmic operations without error, and this has become so ubiquitous in areas of complex signal processing, finite element analysis, computational fluid dynamics, protein dynamics, and other complicated and multidisciplinary system modeling that we don’t even think of it. Of course, these systems don’t learn by independently experiencing the woirld and don’t have any volition or way of setting up up their own problems to solve, but they can provide insights (albeit with limitations) that we could not get by any normal analytical methods with pen and paper.
I suspect you are actually describing a system capable of literally inhuman insights, like the ability to ‘visualize’ fifth or higher dimensional geometry or physically impossible topographies and develop laws of physics beyond the ‘knowable’, but often these kind of discoveries, like general relativity, are a consequence of reframing the problem in a paradigm of that makes it intelligible. And as complicated as general relativity is in any practical implementation, it is fundamentally quite simple in concept once you comprehend the math.
I remain doubtful that LLMs and the current backpropagation approach to ‘AI’—which is in essence trying to ‘brute force’ the heuristics of building a world model by feeding every scrap of textual and image data that exists in electronic form just to write coherent responses to prompts even though adult humans do fine with a tiny fraction of that amount of ‘learning’ plus the practical experience of walking around in the real world for a couple of decades—is actually scaleable to the point of being able to formulate a ‘Type I’ workable human-like artificial general intelligence (much less a super-intelligence) that can actually do the things that a typical human can do in terms of overall cognition. Chatbots seem impressive because we don’t expect a computer to be capable of forming coherent and seemingly worldly responses to arbitrary questions and prompts, but if you look at them with a critical eye they are also shockingly gullible and unreliable in their ability to distinguish fact from fabulation and are mostly good at ‘faking it’ when they mess up. They’re purpose-designed to be ‘deception machines’, sounding like a real person but without any of the common sense or genuine empathy of a real person, but sufficient in facsimile to make someone who wants to believe that there is a ‘spark of consciousness’ find the evidence they seek to validate their wishes.
I’m reminded of a line from an old sci-fi short story, don’t recall the title but I think it was by Norman Spinrad; “The side that gives the machines the most freedom always wins”.
In the setting AI had largely displaced human in all areas of importance due to the simple fact that which ever side in a competition or conflict that relied on the machines more consistently came out ahead, because the machines were genuinely just more capable than humans. What was interesting was that there wasn’t a hint of “AI rebellion” or even the machines being interested in it; the AIs displaced humanity because of human behavior.
And it’s a scenario I find plausible. If superhuman AGI exists somebody is going to try to use it to advance their interests or agenda, and anyone who wants to be competitive with them will need their own AGI. The AGI will at that point eventually end up running the world without any need for them to even try as people push more and more responsibility on them just to keep up.
“Intelligence Explosion”?! The more sophisticated AI gets, the less we will have to do for ourselves. Now, it can write speeches/letters/etc for you, so soon you we won’t have to be articulate at all. My evidence lies in the fact that, since the Internet explosion and the sophisticated search it gives us, we have a world of information virtually at our fingertips. Doesn’t seem to have helped much in improving the intelligence level of the general population.
Or maybe it will turn out that Penrose is right and there is something beyond computation to consciousness. And similarly to the way true randomness cannot be produced algorithmically so designers have used things like lava lamps, they will keep us around as a source of whacky ideas…
Nick Bostrom said there are basically 4 paths to super intelligence.
Speed based - You have a machine intelligence that is as smart as an intelligent human, but operates thousands or millions of times faster. So you’d have thousands of years of thinking in a month.
Quantity based - Earth has 8 billion human minds in it. If there were infinite parallel universes and we could communicate with them, then we could have 100 trillion human minds. We could distribute labor more efficiency among those 100 trillion human minds, increasing the rate of progress.
Quality based - This is true ASI, which is an ASI which is smarter than us the way that we are smarter than a cockroach
Biotechnology based - This isn’t really superintellience, its more advances in bioengineering (gene therapy, embryo selection, neural engineering) that could increase IQ. Bostrom wrote a paper once discussing how embryo selection could raise cognitive abilities pretty dramatically, likely far beyond current human comprehension.
It sounds like your type I is speed based and your type II is quality based.
I do not see why a type II would be impossible. The idea that primates designed by natural selection, with our 3 pound brains that are limited to biological materials would be the peak of cognition that the universe is capable of doesn’t really make sense. At the end of the day we are primates with 86 billion neurons, of which 16 billion are cortical neurons. Our brains consume about 300 calories a day, which is about 20 watts. In theory, a race of primates with brains that have 400 billion neurons and 320 billion cortical neurons, but that had the same bodyweight that we do, would likely have cognitive abilities we can’t fathom the same way a squirrel cannot fathom nuclear energy.
But even if all we ever have is speed based, quantity based and biotechnology based, that is still going to dramatically increase the rate of progress and advancement. Trillions of AI witht he cognitive abilities of nobel laureates, combined with humans that
I agree, but I think some would argue that humans have reached some cognitive generality threshold that squirrels have not. Maybe one would make an analogy to a Turing machine, which can solve any computable problem, even though there are more sophisticated computing models out there. If true, then humans might be slower than a superintelligence but could, in principle, get there eventually.
I don’t think that line of argument is too convincing and agree that a Type II is possible–that there are problems out there that really are beyond human cognition that a superintelligence could solve. But I can’t identify such a problem, and it’s possible that many such problems require superintelligence to even comprehend the problem (the squirrel doesn’t know about the precession of the perihelion of Mercury).
Really, the claim that it’s impossible is mostly founded on human chauvinism, IMHO. It’s asserting that humans are so very special that nothing can ever be smarter than us. When in reality, we are just the dumbest possible creature that can build a technological civilization, given that none of our less intelligent ancestors did.
We’ve gotten as far as we have because so far, we’ve been able to break up problems that were too difficult for us into smaller bits that we could solve. But it seems plausible that there are problems that can’t be broken into simpler pieces.
We can get some flavor for the possibility by looking at special cases like chess. I don’t think the experiment has been tried, but if you assembled a team of Grandmasters and gave them a month to make a move, I think the consensus is that they’d still lose to the best chess computers. So that points in the direction of a Type II at least for narrow problems. Not sure if the same is true of Go or other games.
Agreed. Maybe the chess problems get into this. They require a kind of gestalt understanding of the board that can’t be factored into parts.
Unless SuperAI has a Theory Of Everything that allows it to simulate anything to whatever desired accuracy?
This is an interesting question. I myself am already just a little sceptical of the heavy reliance on computer simulations in current science.
As the saying goes: if your only tool is a hammer, every problem starts to look like a nail.
And if your tool is a computer, every problem starts to look like an algorithm…?
Machine learning is a bit different than just being able to process a lot of things at once; it can be trained to identify stuff in ways that are just beyond “does X meet some threshold” type stuff. You can take a video stream and train the machine learning AI to say… identify people in sets of photos, and it’ll be able to do it very well. Where it gets really useful is that you can train it on truly huge data sets, and also set them up to self-train to some degree. What’s really interesting is that we’re not 100% sure how they’re learning. We feed it a bunch of pictures of clouds, and then tell it which ones are cirrus, cumulus, etc… and then feed it another set of pictures of clouds, and it’ll identify them, but we don’t know what it used to identify them internally. Is it the shape? Contrast? Average luminance of the image? Some combination of them? Something else entirely? We just don’t really know.
As far as I know, they’re not even able to do it at all. They’re intended to basically interpret a text string, retrieve “pertinent” information about it, and then produce an output in natural language using that information.
So when someone types “What is the difference between a Samsung Galaxy S23 and an Iphone 17?”, the LLM will interpret that sentence and look in its own set of information sources about what the differences are, and construct a sensible and grammatically correct answer about that information.
But they don’t have the ability to look at a data source within its set and interpret the validity of its data. If one says that the S23 is the phone for intelligent people, and iPhone users are intellectually stunted cretins, LLMs don’t have any ability to evaluate that- it’s just another data point, so to speak. As far as it’s concerned, any data it’s got is equally valid.
That’s one way to look at it; I think it is more that they’re intended to be more like a search engine that uses actual language as its I/O methodology, and many people will confuse the ability to produce authoritative-sounding output in correct language as more than it is.
I keep seeing this line of thought and am confused. We do know how, we just don’t know the values that the individual weights and biases get set to as the network adjusts them. We CAN know, we just choose not to as there are so many of them. I can absolutely put print functionality all over the neural network code, but I’d need a LOT of paper.
We use the term “black box” because the network is setting the values, not the programmer, but we absolutely do know how it works. There aren’t magical bytes of memory in the computer that only the CNN can access. You can play around with Tensorboard if you want to see some of what’s going on inside the network.
Are you using the term “how they’re learning” in a way that I’m misunderstanding?
If you truly want to see how they work at a low level…
Almost a decade ago while I was still working, my BI team sometimes got jealous of what my data science team was doing. I bought them all this book, which is actually a pretty good primer in how image recognition works. You don’t require any prerequisites as it teaches you basic Python as well as the limited amount of Calculus that you’ll need to understand it all. You want to know how they work, buy that book and go through it. At the end, you’ll have your very own neural network that you built from scratch, versus calling a library with a bunch of parameters.
An interesting possibility in this case is that the AI decides to categorize the clouds in an entirely different way? It ignores the labels we assign (cirrus, cumulus, etc) and finds or creates an entirely different classification model. Not limited to clouds, of course.