Ep34 "What is intelligence?"

Published Nov 13, 2023, 11:00 AM

What would it be like to have a much lower or higher IQ than you currently have -- for example, to be a squirrel or an advanced space alien? This week's episode is about intelligence. What is it and what is its history and future? Join Eagleman on a whistle-stop tour of several schools of thought about what intelligence might mean in the brain.

What's it like to have a much lower IQ than you currently have or to have a much higher IQ? Welcome to the inner Cosmos with me David Eagleman. I'm a neuroscientist and an author at Stanford and in these episodes we sail deeply into our three pound universe to understand why and how our lives look the way they do. Today's episode is about intelligence. What is it What would it be like to have the intelligence of a mosquito, or a horse or a squirrel. What would it be like to presumably understand only very basic things right around you, not doing sophisticated simulation of the future like we do as humans. And what can we say about the present and future of intelligence that is artificial? Okay, so let's start with this question of what is it like to have a different level of intelligence? I see people post this question sometimes online on forums like Quora. Someone will write, I have an IQ of sixty eight, what is it like to have a higher IQ? Now? First of all, I think this is an amazing question because it acknowledges that not everyone is having the same experience on the inside, and so someone is taking the time to ask, what would it be like to have what is called a higher intelligence level. What's the experience of that? Now, the interesting thing in life is that we can't run a control experiment on our own experience of the world, and so whatever IQ you have, you sort of have just that one experience of reality. But to get at this, let's start by thinking about what it would be like to have a much lower IQ than you do now. One way to get at this is to ask the question of what would it be like to be a squirrel. I'm choosing a squirrel just because I was watching one in my backyard yesterday, and I'm watching him run along the top of the fence and climb up the tree trunk and find a little scrap of food and look around nervously. And a squirrel's cerebrum has about one hundred million neurons, while ours has about one hundred billion neurons, so a thousand times more. Now, size isn't everything, which we'll return to a little bit. Presumably the issue is the algorithm that's running. But we can watch the behavior of lots and lots of squirrels over lots of time, and it certainly doesn't seem like they're having the kind of capacity for thought that we are. So I was watching this squirrel, and I thought, what would it be like to be able to jump around from branch to branch, but have no hope, presumably of ever discovering that force equals masstimes acceleration, or for that matter, or not even ever being able to discover that e equals mc squared, or just basic things like how do you build a chair, or how do you think about the competition between Amazon and Netflix, or just the idea that the telephone lines they're running on top of are carrying megabytes of information flowing as zeros and ones as one member of our species communicates to another, or that more broadly, we have airwaves and fiber optics carrying a flow of zetabytes information in a massive ocean of data around us. For the squirrel, none of this exists. For the squirrel, none of this is comprehensible. It's thinking about its acorn and where it hid its last one, and it's thinking about safety and maybe about mating. As far as we know, or as far as we can tell, which of course has its limitations. The squirrel is not ruminating on a play it saw last night. By the squirrel equivalent of Shakespeare and what it means about the aspirations of a monarch and the cruelty inherent in the competition for power. It's not thinking about how to get to the moon or how to build the next vaccine. Now, it's not to say that there aren't specialized kinds of intelligence. Every move that the squirrel makes along the tree branch is very impressive. There is no way that I can could hope to stick landing after landing like that from the branch of one tree to the next to the next. But the squirrel is capable of doing that. But even though it performs this incredible ballet in the face of gravity, it's presumably never going to get to the point where it can characterize gravity in equation form to understand how it should move if it were on a different planet, or to conceptualize gravity as curvature in the fabric of space time. Although humans are capable of getting that in high school, presumably squirrels of any age would find the concept well beyond what their brains could even hope to have a flicker of. And when you look around the animal kingdom around us, we find lots of creatures presumably occupying very different levels of intelligence. So a bat has let's call it, ten million neurons, and again it's not the size but the structure that matters.

But presumably they are not writing the equivalent of bat books or building a little bat Internet where they can capture for eternity everything that every generation of bats before them has learned.

A fish has only about one hundred thousand neurons, a house cricket has fifty thousand neurons. A common fruitfly has only about two five hundred neurons in the equivalent to its brain. So if you were a fruit fly, you simply don't have the notion of seeing the moon in the sky and thinking, okay, that's an orbiting sphere, and I'm going to derive a plan with my fellow flies to get there by building new technologies that we can fit our little bodies inside of so that we can survive in low oxygen. And what if you're a mosquito, with your little mosquito brain, all you know is the mad attraction to certain odors which indicate a warm blood and animal, and the pleasure of slipping your proboscis in and satisfying your thirst with the warm liquid. You presumably don't even have the concept of blood, that it has plasma and cells that are specialized for grabbing oxygen, and all kinds of useful machinery for defending the host animal, and so on and so on. So I've thought about these issues for years, and I think there's a way to understand what it is like to be so limited. The way that the mosquito might look at the human, if it could even have a concept of a human, how might look at the human and think, oh, my gosh, I can't believe they understand all that stuff, and they do it all at once. And the reason we can understand what it's like to be limited is because we are up against problems all the time that we're just smart enough to recognize, but not yet smart enough to solve. Take the origin of life, this is a massively difficult problem. I worked with brilliant scientists like Sidney Brenner and Francis Krik at the Salt Innstitute, who worked on theories of the origin of life. And these were among the smartest biologists of the twentieth century. And even they were like little fruit flies when trying to tackle that problem. They constantly were aware of the enormous gaps in whatever story they were hoping to put together, because we just don't have much of any lasting data from the past three point eight billion years, and we're talking about how trillions of atoms might come together in just the right way over time to form things that can self replicate. It's the kind of problem that when you really start to reach your arms down into it, you realize that even very smart human brains just aren't equipped for a problem of that size. Or just take something like thinking about the cosmos with one hundred billion galaxies in it and one hundred billion stars inside each of those galaxies, and uncountable numbers of planets rotating around those stars, and then trying to picture or answer whether there is life elsewhere in the galaxy and what it would look like. It's clear that our human brains aren't so good at grocking numbers like that. Even though we can estimate the numbers and we can use words to talk about them, we're not really capable of understanding them. And it's the same thing when we study neuroscience. Of course, we've got in the ballpark of eighty six billion neurons and each one of those is connected to so many of its neighbors, about ten thousand, that if you took a cubic millimeter of brain tissue, there are more connections in there than there are stars in the Milky Way galaxy. This thing that we're facing on a daily basis this thing the brain that tens of thousands of people on the planet's study, this three pound organ that we have completely cornered. It is so vastly complex that there is no way for a human brain to understand itself. And in neuroscience we have foundational problems that we can't answer, like why something feels good? Take an orgasm? Why does an orgasm feel good? We can, of course tell the evolutionary story, which is that it benefits the species to reproduce, so it is advantageous for it to feel good. But the question from a neurobiology point of view is how do you build a network that feels anything. Take a big, impressive artificial neural network like GPT four. It can do incredibly impressive work by taking some prompt and written language and generating words that would statistically go with that prompt and so on. It's mind blowing how well it appears to do. But GPT four presumably can't feel pain or pleasure. There's nothing about one of its sentences that it generates that it appreciates as hilarious or tear jerking. It doesn't have any capacity to feel concerned about its survival or demise. When the programmers turn the computer off. It's just running numbers down a long, complex algorithmic network, and that's it. So how do we ever come to feel something? This is perhaps the central unsolved question in neuroscience. It's usually summarized as consciousness, and specifically the hard problem of consciousness, which is to say, why does all this signaling moving through networks of cells feel like something? How could you ever program a computer to feel pain or detect some wavelength of electromagnetic radiation and as purple, or to detect some wavelength of electromagnetic radiation and experience it as purpleness, or to enjoy the beauty of a sunset. These are totally unsolved questions in neuroscience, and presumably there are whole classes of problems that we are not even smart enough to realize our questions that we could be asking. So, despite the incredible pride filling progress of our species. Our ignorance vastly outstrips our knowledge, and that affords us just a little bit of insight into the limitations of our brains, the glass walls of our fish bowl, and lets us even very roughly imagine what it would be like to have the intelligence of a squirrel. And these are the things that I was thinking about when I wrote a fictional short story called Descentive Species in my book Some and so I'm going to read it here, and then I'll come back to the question of intelligence. In the afterlife, you are treated to a generous opportunity. You can choose whatever you would like to be in the next life. Would you like to be a member of the opposite sex, born into royalty, a philosopher with bottomless profundity, a soldier facing triumphant battles. But perhaps you've just returned here from a hard life. Perhaps you were tortured by the enormity of the decisions and responsibilities that surrounded you. And now there's only one thing you yearn for, simplicity that's permissible. So for the next round, you choose to be a horse. You covet the bliss of that simple life. Afternoons of grazing and grassy fields, the handsome angles of your skeleton and the prominence of your muscles. The peace of the slow flicking tail, or the steam rifling through your nostrils. As you lope across snow blanketed planes, you announce your decision. Incantations are muttered, a wand is waved, and your body begins to metamorphose into a horse. Your muscles start to bulge, A mat of strong hair erupts to cover you like a comfortable blanket in winter. The thickening and lengthening of your neck immediately feels normal as it comes about. Your carotid arteries grow in diameter, your fingers blend hoofward, your knees stiffen, your hips strengthen. And meanwhile, as your skull lengthens into its new shape, your brain races in its changes. Your cortex retreats as your cerebellum grows. The homuncular smelts man to horse neurons, redirect synapses, unplug and replug on their way to equestrian patterns, and your dream of understanding what it is like to be a horse gallops toward you from the distance. Your concern about human affairs begins to slip away, Your cynicism about human behavior melts, and even your human way of thinking begins to drift away from you. Suddenly, for just a moment, you are aware of the problem you overlooked. The more you become a horse, the more you forget the original wish. You forget what it was like to be a human, wondering what it was like to be a horse. This moment of lucidity does not last long, but it serves as the punishment for your sins. A Promethean entrails hecking moment, crouching half horse, half man with the knowledge that you cannot appreciate the destination without knowing the starting point. You cannot revel in the simplicity unless you remember the alternatives. And that's not the worst of your revelation. You realize that the next time you return here with your thick horse brain, you won't have the capacity to ask to become a human again. You won't understand what a human is. Your choice to slide down the intelligent sladder is irreversible, and just before you lose your final human faculties, you painfully ponder what magnificent extraterrestrial creature enthralled with the idea of finding a simpler life, chose in the last round to be come a human. So in the story, I try to give a way to think about the possibility that something could be much smarter than us, that we are not at the top of the latter but maybe somewhere in the middle, and that to some other creatures in the universe, we would appear to be like the squirrels are to us. It certainly could be that in this vast cosmos there are intelligences that are so much higher than ours that we lack even a good imagination or vocabulary to paint these creatures in the same way that presumably the squirrels would be unable to give a reasonable description of us and what we're up to. Maybe these extraterrestrials can understand the entirety of cosmic evolution by the time they're in second grade, and they can keep in mind the trillions of animal species on this planet and all the other planets, and keep track of all the interactions and therefore understand the biological history and future of a planet at depth. While we mostly just use the word evolution to capture something that we can't comprehend at a deep level. Now, let me put on the table that I'm completely uncompelled by the claims that there are UFOs, or nowadays they're called UAPs. But I have a very smart friend named Kevin who told me the other day that he has no problem believing that that's true. Now, I'm not defending his position, but his stance was simply that if you imagine the aliens are much smarter than we are, then the particulars of what we're looking for, some Morse code signal or some take me to your leader's sign, that's actually the wrong thing for us to be looking for. Because if we imagine some civilization that is, say, totally differ from us and three million years ahead of us, and they are to us as we are to the squirrels, it's certainly not difficult to imagine the possibility that we are simply not smart enough to construct a good model of them and therefore even recognize them. Now, you might assume that if they're much smarter than us, then they could dumb themselves down to communicate with us the way that we sort of know how to talk with a child at a child's level. But our ability to model lesser intelligences is still pretty terrible. I mean, you still have no idea how to go out in your yard and communicate with squirrels. Just try having a meaningful conversation. Good luck. We're so much smarter than a squirrel, but we have no idea how to plug into their neural networks. Or just go to a zoo and try to have a conversation with a panda bear and commune unicate to him. Take me to your leader, or do that with a camel or a dolphin. You get the point, which is that just because you are smarter doesn't necessitate that you know how to talk to these other animals. And this is the situation that we could hypothetically be in with extraterrestrial civilizations. That we are here even though we don't recognize that they are there, because we don't even have the capacity to imagine them, and they have no meaningful way to communicate with us, because our needs and desires are so different from what they can even understand. Now, this lack of communication across species or across planets is really brought into relief when we consider that intelligence is not one thing, but there are many different behaviors that we might put under the umbrella of intelligence. To this point. In nineteen seventy four, the philosopher Thomas Nagel wrote an essay called what is it Like to Be a Bat? Because fundamentally, being a bat is a pretty different experience than being a different human. If you are a blind echo locating bat, you emit chirps in the dark, and you receive back echoes of your chirps, and you translate those air compression waves into a three dimensional picture of what is in front of you. You make a mental map of your surroundings this way. So Nagel asked this question of what it's like to be a bat in the context of consciousness, as in, given that we have such a different sensory world, is there any way that we could understand what it would be like to be in such a different way of detecting and sensing the world. But this same question could be applied to what we're thinking about here, which is intelligence. Intelligence in the context of a bat allows the bat to navigate around and find food, and talk with other and adapt when the conditions change. But it's hard to directly compare it to human intelligence because they have traits and adaptations that are very sophisticated in their own way. Like I said, with echolocation, they're creating this three D map in their space. They're using auditory information in real time, and they can have such precision that they can detect an object as thin as a hair and fly around, that they can figure out the size and shape and speed of objects like a little moth flying around, so they can zoom in on it and grab it. And they also have sophisticated social behavior, but presumably about different social things than what we care about. And we know that they do all kinds of problem solving, but it strikes me that it's really difficult to know what sorts of problems they solve, because some of the problems are so foreign to us that we don't even know how to think about them. So all this leads us back around to the main question for today, which is what is intelligence? How do we define it?

Well?

As it turns out, this has not been an easy question for scientists, and it has come with lots of debate. And this is one of those things where we all have an intuition about what we mean by the word. But the trick from a neuroscience perspective is how do you rigorously define it and therefore, how do you study it? When we talk about intelligence, let's say just human intelligence, what are we even talking about. We all have a sense of what an intelligent person is, but what is happening in their brain that is different from someone else who you might think is not so intelligent. How do giant networks of individual neurons, billions of them manipulate information that you've taken in before and simulate possible futures and evaluate those and throw out all the information that doesn't matter. And do people who are intelligent store knowledge in a different way, Maybe not categorically different, but just perhaps in a way that's more distilled or more easily retrievable. So these are the kind of questions we're facing now. The first thing to appreciate about intelligence in the brain is that size does not seem to matter. Andre the giant had a brain volume that might have been eight times the size of yours, but he was probably not eight times smarter than you. In fact, what is so remarkable is that brains that are enormous, like in elephants, and brains that are very tiny, like a little mouse brain can both tackle very complex problems like foraging for food and setting up a home and mating and defending itself against predators. The Spanish neuroscientist Santiago Romoni Cajol, like many neuroscientists before and after him, was really struck by this thought, and he had this beautiful comparison of large and small brains to large and small clocks like big ben and a wristwatch both tell the time with equal accuracy despite the size difference. So all this is to say that when we stare at brains, this secret of intelligence is not immediately obvious just from looking at the brain. Now, when we look across species, we can see what we might mean by intelligence. For example, good problem solving skills. Some primates are really good at using tools, like orangutans, while well other primates like bonobo's are really good at social intelligence. If you look at purposes, you find that they are better problem solvers and do much more than say other swimmers like catfish. And when we examine humans we see that somebody can be a genius in one domain but quite bad at another. Rinaldo is a genius at soccer, but he might not be so great at differential equations. I recently saw a video of a kid who can do a Rubik's cuban about three seconds, but he's autistic and therefore is not particularly good at anything involving social interaction. So how can we put a measure to what we are talking about here? About a century and a half ago people started working on the question of how you could quantify this. The British scientists Sir Francis Galton was one of the first that I know of who said we should be able to measure intelligence. So is that you could quantify it by measuring things like the strength of someone's eyesight and hearing, or the strength of their grip. So that approach didn't last long, but by nineteen oh five, two scientists, Alfred Binet and Theodore Simon built the Simon Benay test as a way of quantifying some number for intelligence. And then in nineteen sixteen here at Stanford University there was an educator named Lewis Turman who developed the test more and he renamed it the Stanford Bena test, which you might have heard of because it is still used now. These sorts of tests allow us to put a number on something, but we still know what exactly we're measuring. The psychologist Charles Spearman was intrigued by this question, and he made an observation, which is that if you do well on one task, something like verbal skills, you tend to also do well at other tasks, like spatial skills, and so these things correlated, and he speculates that there was some sort of general intelligence involved here, and so he used the letter G for this idea of a general intelligence factor, like a general skill set of the brain. And other researchers noticed this correlation also between very different sorts of tasks like memory and perception and language and solving new problems and pattern recognition and a whole bunch of others, and so it still wasn't clear what intelligence is, but it's clear that these things correlated, and so in nineteen twenty one researcher wrote that while it is difficult to define precisely what intelligence is, tests tested. So some people felt that there's one thing, this G, that underlies lots of different skills, and others felt that maybe these are completely separate things and intelligence is not one thing. So that's the debate that got rolling over a century ago and it remains an unsolved issue. The fact is that if intelligence were just one thing, you might expect to sometimes see a small bit of brain damage where someone loses skills across different types of intelligence, or with the introduction of brain imaging some decades ago, we might be able to see a single small network becoming active even with very different problems. But interestingly, this is still unresolved because some researchers ask participants to do very different kinds of tasks like verbal and perceptual and spatial things while their brain is getting scanned, and they find that all of these tasks lead to activity in an area called the lateral frontal cortex, and so it might be interpreted to support the unitary intelligence hypothesis because you're seeing one area becoming active even when people are doing different kinds of tasks. But on the other hand, we're always faced with the problem that our current brain reading technology only lights up areas where there's a lo a lot of activation, and it doesn't catch the areas that are more diffuse where the real detailed action might be happening. And there's also an issue that highly intelligent people find particular tasks less challenging and so they often show less activity in the frontal cortex, not more. And so it may be that even with our terrific technology, it's still a little bit too crude to tell us what intelligence is by simply going around and looking for a spot or a collection of spots in the brain. This is in the same way that you're not going to look at chat GPT and say, ah, what makes it intelligent? Are these few nodes here out of the billions of nodes. Instead, it's a function of the whole of the activity running through the enormous system. So it may turn out that intelligence is not going to be captured by a single brain area or even a system. For all we know, it might not even be about neurons, but about what's going on at the molecular level inside of neurons, which means we might just be looking for some correl it at the level of neurons. Now, all that speculative, but I just want to make clear that often in neuroscience we are like the drunk looking for the keys under the street light because the lighting is better there, even though we dropped our keys over there. Our technology has its limitations, and we often gravitate towards the street light and ask if we happen to be able to find the keys there, And sometimes that strategy works and sometimes it doesn't. Now, part of the challenge in asking what intelligence is is that the word probably tries to hold up too much weight by itself, because what we call intelligence is almost certainly made up of multiple facets. For example, some people break this down to analytic intelligence like you use in math problems, or creative intelligence like writing a caption for a cartoon, or practical intelligen just like how to operate well in the world. So one question is whether these different categories of intelligence truly represent different things with fence lines around them, or whether they're underpinned by the same mechanisms in the brain or overlapping mechanisms. But the problem is even trickier than that, because even within any of these categories, we still have to answer questions like how knowledge gets stored and retrieved, how it can get restructured, how it can get erased, and so on. So the question of what intelligence is has attracted scientists throughout the ages to propose all kinds of different answers, none of which may be mutually exclusive, but they're all different angles on answering what it is when somebody is intelligent. So let's look at some proposals. One proposal is that intelligence has to do with squelching distractors. Technically, this is called resolving cognitive conflict. So for example, let's say we're playing the Simon Says game, where I say, Simon says, look to your left, and then you do it. But let's say I say lift your arm, but I don't preface it with Simon says. Then what you're supposed to do is override your reflex to lift your arm. This is an example where you'd have cognitive conflict. So the way neuroscientists study this is, for example, by using something called a three back task. So imagine you're watching a series of faces getting presented on the screen. So first you see Tom Cruise, and then you see Beyonce, and then you see Taylor Swift, and then you see Anthony Hopkins and so on. Your job is simply to say, when you see a face that matched the face that you saw three faces ago, in other words, three back. If you then see Emma Thompson and then Taylor Swift again, you'd say, yes, Taylor's matched what I saw three faces ago. But if you see Zendaia and then Jennifer Lawrence and then Zendia again, that's a distractor because her face was only two a go. And so you're supposed to hold your tongue or specifically not press your button. So to perform this task requires not only a small window of working memory, but you have to squelch distractors. You have to squelch faces that matched what was two faces ago, or four faces ago or five faces. You can only hit the button when the face matches what was three ago. Now, you run this test on a whole bunch of people with different levels of G generalized intelligence score, and what you find is that people with a high G are better at the task, in large part because they don't respond to the distractors. When you do this in brain imaging, you find that particular areas come online, like the anterior singular texts and the lateral prefrontal cortex, and these areas seem to be necessary for overriding the cognitive conflict. So that's one idea for what intelligence is, but other studies suggest no, it's not about conflict resolution. Instead, intelligence is about how many things you can hold in working memory. So, for example, our visual memory can only hold let's say three or four objects in mind at any given time. So let's imagine that I show you some colored shapes like a green triangle and a red circle and a blue square, and then a moment later, I show you a similar image, and I ask you where any of these shapes are colors different? And you can probably do this for three or four objects. But as it turns out, some people are only able to retain the information from one or two objects, and other people can hold more, let's say five objects, And so some people have suggested that that is really related to intelligence, with the idea being that critical reasoning depends on how many things you can hold in your working memory. If you can hold more things in your head at any one time, you'll be better able to manipulate things for solving problems. So again, people have done brain imaging with EEG and fMRI and found a little area in the posterior paridal cortex that seems to give a memory bottleneck and correlates with what different people can hold in mind. Now it seems likely that working memory capacity won't be the final unlock to the question of intelligence, but it probably plays a role. So what other ideas are there? Well? As it turns out, people in the late nineteen nineties got excited about the idea of forming associations in the brain. And there's a particular type of receptor in the brain called an NMDA receptor. Don't worry about the details here, I'll link a paper on the website. But you can genetically engineer this receptor in a mouse and show that the mouse can link things more strongly, like this light predicts food, or this is the location where some reward is located. So a scientist named Joe Chen and his colleagues at Princeton engineered a strain of mouse to have more of this NMDA receptor subunit. And this hit the news at the end of the nineties because these mice called Doogie mice after the TV show Doogie Howser MD, which was about a really smart kid. These Doogie mice outperformed normal mice in recognizing things they had seen before, or swimming their way through a pool of milky water to remember where a hidden plat form was. Now, this news made a real splash when it came out because the idea was that wow, we've just invented intelligent mice. But we do have to ask whether we think the doogie mice are more intelligent just because they can do these laboratory tests better. After all, intelligence is more than simply nailing down associations, and the other thing to keep in mind is that all animals have to balance the things they know against exploring new possibilities. This is known as the balance between exploitation and exploration. The reason animals have to balance this is because the world changes and you never know exactly how and when it's going to change. So if you are an animal who's used to finding worms under the green rocks, you want to spend some of your time exploring under the blue rocks and the red rocks too, because you never know when things in the world are going to change. So the googie mice seemed to be more about exploiting knowledge that they learned and less about exploration. But that's not necessarily a good thing. It depends on what happens with the world. So just forming stronger associations is probably not going to be the full answer to what intelligence is. Now, there's another pathway we can sniff down when we're looking for the root of intelligence, and that is the Eureka moment. That is what happens when two concepts suddenly fit together. Like I remember the moment when I was a kid when I learned that fog is just the same thing as a cloud, but it's low to the ground, And it was a physical sensation for me to have these two concepts fit together. Or if you're a detective, you might have a bunch of clues on your desk and then suddenly, aha, it all coalesces into a narrative because all the facts fit. Now, what has just happened in your brain? And how does your brain know and alert you that a fit has been achieved. This is the restructuring of information. And I just want to make clear we are nothing like a computer that takes in files of facts. Instead, we're always structuring and restructuring information. Now. One of the places we can see that is when a monkey learns a task. What you'll notice is that you can't tell the monkey the rules of the task. They have to figure it out themselves by doing it over and over and getting reinforced with let's say, juice in their mouth or something like that. Over hundreds of trials, and monkeys can learn this way and they can get better through time. Their performance just rises like a shallowly sloped line. But if you give the same task to an undergraduate, something very different happens. They'll try a few things and then they'll suddenly get it, and they're performance jumps up. Suddenly they have an Aha moment, they have a Eureka. Now, this observation implies that humans are doing something that monkeys can't. Perhaps this has to do something with restructuring knowledge, or perhaps the human student gets to try out lots of hypotheses and evaluate them and then restructure things accordingly. But whatever the issue is, this certainly seems to play a role in what we think of as intelligence. And it also suggests that animal models of intelligence are going to be too limited for some of the forms of sophisticated reasoning that we care about. And I'll give you another thing that we might look for. What if intelligence is about the ability to make good predictions about the world. In previous episodes, I've talked about the internal model, and I've emphasized that the only reason the brain builds an internal model is so that we can make better predictions about the future. So emulation of possible futures is a giant part of what intelligent brains do. As the philosopher Carl Popper said, this is what allows our hypotheses to die in our stead. My friend and colleague Jeff Hawkins has emphasized this for a couple of decades, that we only have memory in order to make predictions. So the idea is that you write down things that happen to you that seem salient, and you use those building blocks to springboard into possible futures. As Jeff puts it, intelligence is the capacity of the brain to predict the future by analogy to the past, and we can find lots of evidence for that in examples of brain damage, where people lose the ability to store memory and as a result are unable to simulate the future. So this whole memory prediction framework almost certainly plays a role in intelligence. But there are a lot of unanswered questions here. For example, there are a huge number of possible future moves. How does the brain simulate them all? Perhaps an intelligence simulator saves time by developing tricks so that you don't have to simulate everything. So there are lots of proposals and possibilities for what intelligence is in the brain, and probably there are many other possibilities that we haven't even begun to explore or know how to explore. So I want to pose a question about intelligence, and this one is really important, and that is the question of why do we have lions in zoos? After all, a lion is so much more powerful than you are. A lion can easily kill a human. It has these razor sharp claws, and its body is all muscle and speed, and yet we put lions in zoos. How well, there's only one thing we have over lions, and that is intelligence, and intelligence enables control. We don't brute force the lion into the cage, we don't wrestle a man. Demand Instead, we do things like set up traps, or develop chemicals that happen to interact with their neurochemistry and put them to sleep, and then we package that into a syringe and use explosives to launch it really quickly down a metal barrel so it punctures their skin. All of these things are moves that the lion cannot possibly predict because it couldn't possibly conceive of them. And that's what makes so salient. Our contemporary discussions about AI, because often when someone is thinking about the question of whether AI could control humans, they think about physically manhandling us with robots. But that seems really unlikely because it's so hard to build physical robots. You're constantly tending to the toilet of the robot machinery, You're trying to keep all the pieces and parts together and not have a wire pop somewhere. But the important concept to get straight is that for AI to control humans, they don't need brute force. Why because intelligence enables control. Could we imagine a scenario in which the AI does something that we can't predict because we can't possibly conceive of it. Sure, And the interesting part is that there's a whole space of scenarios that we can conceive of and write science fiction novels about, But there's also the space of the unknowns Now, I'm not suggesting that modern AI is going to move in that direction, because at the moment it's just doing very sophisticated statistical games and it doesn't have any particular desire for power. But I think for sure things are going to get strange as we grow into a world with another intelligence, one which has read every single book and blog post ever written by humans, and knows every map that we've ever made, from streets to chemical signaling, and can create a video of any new idea, and can simulate new combinations of machines and fractions of a second. So this is the reason it's important to understand what intelligence is when we talk about artificial intelligence now. Earlier this year, I published a paper about how we might meaningfully assess intelligence in AI, and I discussed this in episode seven. In other words, how would we know if some artificial neural network like chat GPT we're actually intelligent versus just computing the probability of the next word based on a slurry of everything humans have ever written. Well, for sure, it is just computing the probability of the next word. But the surprise has been all the stuff that we didn't expect it to be able to do. With this straightforward statistical prediction model, it does more than it was programmed or expected to do. So that has left the whole field with a question of whether simply having enough data gives us something that is actually intelligent or whether it just seems intelligent. So in that previous episode, I proposed that the tests we currently have, like the Turing test, are outdated as a test for meaningful intelligence. Why because the Turing test can already be passed and it still doesn't tell us really what we need to know. And it's the same with other tests that have been proposed in the AST, like the Loveless test, which asks whether computers could ever be creative, and all it takes is a few seconds with mid journey or chat GPT to see that that landmark is also in the rear view mirror. So what I've proposed is not about moving the goalpost. It's about fundamentally asking what is the right test for a meaningful sort of intelligence. So what I suggested is that we will know if a system has some real intelligence once it starts doing meaningful scientific discovery and puts all the scientists out of business, because scientific discovery is something that requires a meaningful level of intelligence. And I'm not talking about the type of science that's just piecing together things in the literature, although that's of course very useful. I'm talking about the type of science where you think of something new that doesn't already exist, and you simulate that and you evaluate whether this crazy model you just came up with would give a good understanding of the facts on the ground. So, for example, when Alfred Wegner proposed that the continental plates were drifting, that gave a totally different explanation for all kinds of data, including the fact that South America and Africa seemed to plug into each other like puzzle pieces, And it gave an explanation for mountain ranges and so on. And he simulated what would be the case what we would expect to see if this were true, and he realized it made a good match to the data around him. Or when Einstein imagined what it would be like to ride on a beam of light and this is how he derived the theory of special relativity, or when Charles Darwin came up with a theory of evolution by natural selection by thinking about all the animals that weren't here. I suggest that these are the kind of things that humans can do that represent real intelligence, the kind of intelligence that has made our species more successful than any other on the planet. So is modern AI intelligent in this way? As of this recording, there's no simple answer to this. There are arguments on all sides that generative AI has actually reached some sort of intelligence or that it hasn't. But it's not easy at the moment to come to a clear conclusion on this. And although AI intelligence might not be quite the same thing as what we have, I suspect it's going to matter a lot for us to better understand what human intelligence is made of, so we can understand when AI grows up to be the same or better and why. And I suspect that the simple existence of AI is going to help us think through these problems, because we're going to try things and get over our naive assumptions about what intelligence might be. For example, from at least the nineteen fifty onward, the old way of trying to build artificial intelligence was to give a computer a giant list of facts. You explain that birds have wings and beaks and feathers, and they fly, and then maybe you have to teach it that there are some exceptions to the rule, like ostriches or penguins, and you keep giving it these rules and structure. And that approach never worked, and the field of artificial intelligence descended into its winter. So what we learned from that is that intelligence is probably not a series of propositions, but rather it's stored in a very different way, for example, a giant cascade of information in vast networks. And so studying intelligence that is artificial, that's what's going to sharpen our focus on intelligence that is evolved. So let's wrap up. As you know, if you've been a listener to this podcast, I'm obsessed with the way that we all see the world from different points of view, not least because we have subtly different genetic details in our brains from person to person, as well as different life experiences which have wired up the circuitry. And as a result, we also have different intelligences that allow us to see the world differently and sometimes with more or less clarity. And what we've done today is looked at the complexity of what seems like a simple question, what is intelligence? We know that there are differences between species and even within members of any species, but we don't always know how to capture that. And the fact that we can address the question but after one hundred years still not come to a clear answer probably indicates that the word intelligence simply holds up too many different things, different skills, whether that's the squelching of distractors or the number of things you can hold in memory at any given moment, or the formatting of information or making associations, or the ability to simulate possible futures. It seems to me that one of the most meaningful tests for the intelligence of our species will be this. Will we be able to define and understand intelligence before we create it and perhaps get taken over by it, That will be the true test of the intelligence of our species. Go to eagleman dot com slash podcast for more information and define further reading. Send me an email at podcast at eagleman dot com with questions or discussion, and I'll be making more episodes in which I address those. Until next time, I'm David Eagleman, and this is Inner Cosmos.

Inner Cosmos with David Eagleman

Neuroscientist and author David Eagleman discusses how our brain interprets the world and what that  
Social links
Follow podcast
Recent clips
Browse 98 clip(s)