WEBVTT - Ep147 "Can we engineer human thought?" with Tom Griffiths 0:00:05.120 --> 0:00:07.440 AI seems like it burst out of the gate a 0:00:07.480 --> 0:00:10.920 few years ago, But is it actually the latest chapter 0:00:11.119 --> 0:00:16.040 in a three hundred year trajectory to turn thought into math? 0:00:16.680 --> 0:00:19.400 Can the mind be captured with equations? 0:00:19.960 --> 0:00:23.400 Why do current AI models need petabytes of data but 0:00:23.480 --> 0:00:26.759 a child can learn from just a few examples. Why 0:00:26.760 --> 0:00:31.440 does AI have jagged intelligence, meaning it looks brilliant in 0:00:31.480 --> 0:00:36.080 one moment and then it does something totally nonsensical. In physics, 0:00:36.440 --> 0:00:39.479 we have various laws, like the law of gravity or 0:00:39.520 --> 0:00:43.320 the laws of motion, And today we're joined by cognitive 0:00:43.360 --> 0:00:47.080 scientist Tom Griffiths from Princeton to talk about whether we 0:00:47.159 --> 0:00:55.400 are moving towards nailing down laws of thought. Welcome to 0:00:55.400 --> 0:00:58.880 Inner Cosmos with me David Eagleman. I'm a neuroscientist and 0:00:58.920 --> 0:01:02.400 author at Stanford, and in these episodes we sail deeply 0:01:02.440 --> 0:01:05.960 into our three pound universe to understand why and how. 0:01:05.800 --> 0:01:07.880 Our lives look the way they do. 0:01:23.920 --> 0:01:27.000 One thing that distinguishes Homo sapiens from all our cousins 0:01:27.040 --> 0:01:30.119 in the animal kingdom is that we watch the world 0:01:30.120 --> 0:01:34.160 around us and we try to abstract patterns from it. 0:01:34.600 --> 0:01:37.679 For example, you might watch the way that a stone 0:01:37.760 --> 0:01:41.039 falls to the ground and maybe you see a tree 0:01:41.120 --> 0:01:44.200 branch fall, and maybe you see a glacier and one 0:01:44.240 --> 0:01:46.959 day a huge wall of ice falls off it, and 0:01:47.040 --> 0:01:51.000 pretty soon you start seeing an underlying similarity to the 0:01:51.040 --> 0:01:54.600 way that things move. And eventually someone very very smart 0:01:54.680 --> 0:01:58.360 comes along, like Isaac Newton and summarizes all this in 0:01:58.600 --> 0:02:02.080 the law of gravity. And then the same smart guy, 0:02:02.160 --> 0:02:05.600 Newton comes up with the three laws of motion. And 0:02:05.640 --> 0:02:08.560 then another smart person is Einstein. He figures out the 0:02:08.919 --> 0:02:12.840 conservation of mass and energy, which seems to be another 0:02:13.040 --> 0:02:16.919 ironclad law, and then we have the laws of thermodynamics 0:02:16.960 --> 0:02:20.320 and electrostatic laws, and all of this speaks to the 0:02:20.360 --> 0:02:24.079 great success that we've had as the species in figuring 0:02:24.120 --> 0:02:28.160 out the lowest level of code that's running in the universe. 0:02:28.880 --> 0:02:32.120 But for most of human history, the concept of a 0:02:32.600 --> 0:02:35.840 thought has felt like the most intimate thing we experience 0:02:35.960 --> 0:02:38.560 and the least tractable thing to study. 0:02:39.200 --> 0:02:41.840 What a thought is and how it occurs. 0:02:42.280 --> 0:02:46.200 That seems to live in a different category of mystery 0:02:46.240 --> 0:02:49.560 from how an object falls. Why, Well, it's because the 0:02:49.639 --> 0:02:53.600 thought pops into your head and somehow it carries memory 0:02:53.680 --> 0:02:57.760 and expectation and language and often a feeling. But it 0:02:57.840 --> 0:03:02.440 feels vaporous and private. It feels like the one thing 0:03:02.960 --> 0:03:07.880 that will forever escape formal description. But what's interesting is 0:03:07.919 --> 0:03:11.680 that for centuries people have tried, there's always been a 0:03:11.800 --> 0:03:13.399 deep human urge. 0:03:13.120 --> 0:03:16.480 To ask whether thought has laws to it? 0:03:16.560 --> 0:03:19.720 In other words, does the mind have principles that you 0:03:19.760 --> 0:03:23.760 can write down? Does reasoning have a grammar to it? 0:03:24.120 --> 0:03:27.960 Can you describe intelligence in a language that's precise enough 0:03:28.360 --> 0:03:31.440 that once you understand the rules, you can begin to 0:03:31.480 --> 0:03:35.800 build with them, like build artificial intelligence. Most of us 0:03:35.840 --> 0:03:38.400 are old enough to remember that this question of AI 0:03:38.960 --> 0:03:42.840 once lived in philosophy seminars and math departments, but now 0:03:42.880 --> 0:03:45.320 it's sitting at the center of our economy. 0:03:46.280 --> 0:03:48.200 Okay, So what is thought? 0:03:48.440 --> 0:03:52.640 Can we capture it in formal systems like laws or equations? 0:03:53.040 --> 0:03:57.600 Do different parts of intelligence come from logic, from learning, 0:03:57.720 --> 0:04:03.840 from uncertainty, from memory, from prior knowledge, from living inside bodies, 0:04:03.920 --> 0:04:06.120 from living inside our cultures? 0:04:06.640 --> 0:04:08.040 From the particular. 0:04:07.600 --> 0:04:11.120 Constraints of being a human animal with a short lifespan 0:04:11.520 --> 0:04:15.320 and limited bandwidth. Our guest today is someone who lives 0:04:15.400 --> 0:04:19.240 right at the intersection of all these questions. Tom Griffiths 0:04:19.320 --> 0:04:22.880 is a professor at Princeton, where He directs the Computational 0:04:22.960 --> 0:04:27.680 Cognitive Science Lab and the Princeton Laboratory for Artificial Intelligence. 0:04:28.160 --> 0:04:31.719 He has spent years asking how minds work through the 0:04:31.839 --> 0:04:35.760 different lenses of math and computation and learning. And he's 0:04:35.760 --> 0:04:38.919 the author of a wonderful new book called The Laws 0:04:38.960 --> 0:04:43.719 of Thought, which traces the long history of thinkers asking 0:04:44.080 --> 0:04:46.719 are their rules to this? Can we understand what human 0:04:46.760 --> 0:04:50.839 thinking is? In his book we get the lengthy arc 0:04:51.040 --> 0:04:55.839 of minds trying to understand mind. This begins millennia ago 0:04:55.920 --> 0:05:00.520 with Aristotle, who wondered whether logic itself could be math matized, 0:05:00.960 --> 0:05:05.520 and Tom follows the trail through the architects of symbolic reasoning, 0:05:05.960 --> 0:05:09.240 through the birth of computation, through the rise of neural networks, 0:05:09.640 --> 0:05:13.680 through the realization that probability theory might serve as a 0:05:13.800 --> 0:05:17.719 language for our beliefs about things. Along the way, in 0:05:17.760 --> 0:05:20.880 his book, a picture emerges that there may not be 0:05:21.320 --> 0:05:24.440 just a single tool for capturing their mind, but instead 0:05:24.440 --> 0:05:27.760 there are different ways of trying to tackle the problem, 0:05:28.120 --> 0:05:32.520 and each one sheds light on a different aspect of cognition. 0:05:33.240 --> 0:05:36.360 So we're going to talk about ourselves human minds, and 0:05:36.440 --> 0:05:40.160 we'll talk about AI what kind of intelligence is this 0:05:40.440 --> 0:05:47.480 and what is missing? Here's my interview with Tom Griffiths. 0:05:48.160 --> 0:05:50.320 As soon as you turn thought into math, it becomes 0:05:50.320 --> 0:05:52.480 something that machines would be able to do. And so 0:05:52.600 --> 0:05:57.160 our modern AI systems are really a consequence of, you know, 0:05:57.200 --> 0:06:00.520 that thought that people were having hundreds of years ago, 0:06:01.080 --> 0:06:04.160 of being able to turn thought into something that can 0:06:04.160 --> 0:06:05.800 be expressed in mathematical terms. 0:06:05.920 --> 0:06:07.440 And so one of the things that I loved about 0:06:07.440 --> 0:06:09.400 your book, by the way, is that you really tell 0:06:09.560 --> 0:06:11.640 stories of all the thinkers. 0:06:12.240 --> 0:06:14.880 You dive into the lives, you tell them with real color. 0:06:15.160 --> 0:06:17.440 If you were going to start with one thinker that 0:06:17.480 --> 0:06:19.160 you think is the most important, who would that be. 0:06:19.400 --> 0:06:21.160 There are a couple of people who have this sort 0:06:21.160 --> 0:06:23.960 of enduring influence throughout the book. One of them is Leibnitz, 0:06:24.040 --> 0:06:26.680 who kind of started this enterprise in some sense. He 0:06:26.760 --> 0:06:30.360 was really trying to take the idea of logic as 0:06:30.400 --> 0:06:33.280 expressed by Aristotle and turn it into math, but ultimately 0:06:33.279 --> 0:06:35.960 failed in doing that. But along the way he also 0:06:36.120 --> 0:06:38.799 discovered the calculus, which turned out to be really important 0:06:38.800 --> 0:06:41.440 when people wanted to make neural networks that could learn 0:06:41.520 --> 0:06:44.480 from data. It turns out that the trick for doing 0:06:44.560 --> 0:06:46.840 that is actually a trick that lad had figured out 0:06:46.880 --> 0:06:50.599 all that time ago. And then another key figure here, 0:06:50.760 --> 0:06:53.320 as might be suggested by the title of the book, 0:06:53.400 --> 0:06:57.800 is George Bull, who was a nineteenth century mathematician. He 0:06:57.920 --> 0:06:59.600 was a school teacher for most of his life and 0:06:59.680 --> 0:07:01.600 did a lot of like serious math on the side 0:07:01.640 --> 0:07:03.680 instead of you know, had a big effect on the 0:07:03.720 --> 0:07:08.080 history of mathematics. But he was really the person who 0:07:08.160 --> 0:07:11.560 then first solved that problem that Leibnitz had posed. And 0:07:11.600 --> 0:07:15.480 in addition to the impact that that work had, he's 0:07:15.560 --> 0:07:19.000 also the great grandfather of Jeff Hinton, who was one 0:07:19.000 --> 0:07:21.400 of the people who played an important role in developing 0:07:21.440 --> 0:07:24.040 these algorithms for learning from your own networks. And so 0:07:24.200 --> 0:07:26.240 you could make an argument that without Boole we would 0:07:26.240 --> 0:07:29.480 be a fair way back from where we are today. 0:07:29.920 --> 0:07:30.080 You know. 0:07:30.200 --> 0:07:33.160 Interestingly, when most people think about Boole, they only know 0:07:33.320 --> 0:07:38.360 about Boolean numbers. They know about zero and one binary numbers, 0:07:38.680 --> 0:07:41.520 and that's essentially the extent of the think. But he 0:07:41.560 --> 0:07:43.760 was quite celebrated in his life right even though he 0:07:43.840 --> 0:07:48.440 was a headmaster and not formally involved as a professor. 0:07:48.560 --> 0:07:51.000 Am I correct about this? He nonetheless was quite recognized 0:07:51.000 --> 0:07:51.800 as a mathematician. 0:07:52.400 --> 0:07:55.080 Yeah, he became a university professor later in his life, 0:07:55.120 --> 0:07:57.080 but spent most of his life as a teacher and 0:07:57.120 --> 0:08:01.080 a head master. But yeah, he won a gold medal 0:08:01.080 --> 0:08:04.080 in mathematics from the Royal Society. Was a very prestigious award, 0:08:05.080 --> 0:08:08.960 and you know, was this amazing person who was having 0:08:09.000 --> 0:08:11.880 these high level correspondences with the leading mathematicians of the 0:08:11.960 --> 0:08:16.360 day while holding down his job running a small school. 0:08:16.960 --> 0:08:17.640 Yeah. 0:08:17.800 --> 0:08:22.840 Now, in the book, you essentially use three different frameworks. 0:08:22.920 --> 0:08:26.440 What phenomenon does each framework explain? 0:08:26.560 --> 0:08:27.960 Unusually Well, the. 0:08:27.840 --> 0:08:29.880 Three frameworks I talk about in the book are what 0:08:29.920 --> 0:08:32.360 I call rules and symbols, which is what we've been 0:08:32.400 --> 0:08:34.800 talking about. This kind of like approach that stems out 0:08:34.840 --> 0:08:37.120 of logic, where the idea is that you're going to 0:08:37.200 --> 0:08:39.719 be able to write down some rules that characterize the 0:08:39.720 --> 0:08:41.960 structure of thought, and by following those rules, you end 0:08:42.000 --> 0:08:47.319 up with interesting consequences. The second approach is networks, features 0:08:47.320 --> 0:08:47.840 and spaces. 0:08:47.920 --> 0:08:48.079 Right. 0:08:48.120 --> 0:08:50.640 This is neural networks, which you can kind of think 0:08:50.640 --> 0:08:54.160 about as a system for doing computation when you start 0:08:54.200 --> 0:08:56.839 representing things as points in a space. Right, So if 0:08:56.840 --> 0:09:01.600 you start to think about you know, every object that 0:09:01.640 --> 0:09:03.360 you could see in the world is not being something 0:09:03.400 --> 0:09:05.920 that's described by rules, but being described by a location 0:09:06.040 --> 0:09:08.560 along some dimensions. You need to have a way of 0:09:08.600 --> 0:09:11.400 talking about how to map between those spaces and your 0:09:11.440 --> 0:09:14.120 all network solve that problem. And then the third is 0:09:14.960 --> 0:09:20.439 probability and statistics. And probability theory is really powerful because 0:09:20.720 --> 0:09:24.120 it is the complement to logic, where logic tells us 0:09:24.160 --> 0:09:26.120 how to go from things that we know to be 0:09:26.200 --> 0:09:28.959 true to other things that we're equally certain or true. 0:09:29.360 --> 0:09:31.680 Probability theory tells us what to do when we're uncertain. 0:09:32.160 --> 0:09:34.920 So if we get some information we want to draw 0:09:34.920 --> 0:09:37.319 a conclusion, but we're not able to draw that conclusion 0:09:37.360 --> 0:09:40.559 with perfect certainty, Probability theory tells us how to do that, 0:09:40.920 --> 0:09:44.000 and it tells us how to combine our sort of 0:09:44.400 --> 0:09:49.320 background beliefs, the other sources of information we have our 0:09:49.360 --> 0:09:52.120 biases in with the data that we see in a 0:09:52.120 --> 0:09:54.160 way that helps us to explain how it's possible to 0:09:54.240 --> 0:09:56.440 learn from small amounts of data. And that's one thing 0:09:56.480 --> 0:09:59.199 which is still something that discriminates human learning from the 0:09:59.280 --> 0:10:01.120 learning that's done by AI systems today. 0:10:01.520 --> 0:10:03.560 Okay, great, so we're going to dive into each of 0:10:03.600 --> 0:10:06.560 these three lenses. But just before we do, do you 0:10:06.640 --> 0:10:11.680 see the AI conversation today over indexing on one of 0:10:11.720 --> 0:10:13.160 these lenses over the others. 0:10:14.760 --> 0:10:17.520 I think there's a lot of emphasis on neural networks, 0:10:17.559 --> 0:10:21.240 which are fundamentally the sort of engineering technology which is 0:10:21.280 --> 0:10:25.360 making possible the creation of our chatbots and the other 0:10:25.520 --> 0:10:29.600 sort of big AI systems that are deployed. I think 0:10:30.200 --> 0:10:34.360 that potentially misses out the importance of these other threads 0:10:34.559 --> 0:10:37.679 right where. One thing that's important to remember is that 0:10:37.720 --> 0:10:40.680 those neural networks are being trained on what is essentially 0:10:40.720 --> 0:10:43.200 a system of rules and symbols. They're being trained on 0:10:43.720 --> 0:10:47.480 human language, which is symbolic and rule like in various ways, 0:10:48.160 --> 0:10:50.680 and they're being trained on code, which is even more 0:10:50.720 --> 0:10:54.240 symbolic and even more rule like, And those things together 0:10:54.280 --> 0:10:56.559 provide some of the substrate for developing the kind of 0:10:56.600 --> 0:10:59.280 intelligence that they demonstrate. And then the way that they're 0:10:59.320 --> 0:11:03.079 trained is by learning to predict the next token, right, 0:11:03.120 --> 0:11:05.080 the next word or part of word, based on what 0:11:05.080 --> 0:11:07.760 they've seen so far. And that way of training them 0:11:07.800 --> 0:11:12.040 is actually using probability theory. So that's a probabilistic problem 0:11:12.040 --> 0:11:14.079 because you're making a guess about what the next thing 0:11:14.120 --> 0:11:15.920 is going to be based on the things that you see, 0:11:16.200 --> 0:11:18.520 and so that's an important ingredient in their success as well, 0:11:18.559 --> 0:11:22.000 is that they're essentially learning to approximate a big probability distribution. 0:11:22.360 --> 0:11:25.079 So let's dive into the first one, rules and symbols. 0:11:25.280 --> 0:11:28.600 So take us back to the original urge. Why did 0:11:28.679 --> 0:11:33.559 early thinkers believe that this could be used to explain thinking. 0:11:35.200 --> 0:11:38.240 I think a lot of the draw of rules and 0:11:38.280 --> 0:11:41.760 symbols was that that really was, in some way what 0:11:42.120 --> 0:11:45.400 mathematics was to people, right, So Leibniz, part of the 0:11:45.400 --> 0:11:47.400 reason why he wasn't able to solve this problem of 0:11:47.440 --> 0:11:50.160 figuring out how to turn thought into math is that 0:11:50.679 --> 0:11:53.560 what he thought math was, or the kind of math 0:11:53.600 --> 0:11:55.440 that he was trying to use to solve that problem, 0:11:56.000 --> 0:11:59.080 was really arithmetic, right, And arithmetic was kind of like 0:11:59.120 --> 0:12:01.560 the model that they had for a mathematical system. So 0:12:01.600 --> 0:12:04.120 you can think about ideas being added together or subtracting 0:12:04.120 --> 0:12:07.080 one idea from another, and really thinking about the operators 0:12:07.120 --> 0:12:09.160 that you're using as being the things that are sort 0:12:09.160 --> 0:12:11.280 of coming from this familiar mathematical language. 0:12:11.360 --> 0:12:13.640 And so I think part of. 0:12:13.640 --> 0:12:16.240 The reason that we end up with that approach is 0:12:16.280 --> 0:12:19.120 because of the kind of math that has been successful 0:12:19.120 --> 0:12:22.920 in other settings, right where we need to do arithmetic 0:12:23.040 --> 0:12:25.320 to you know, that's a good description of certain. 0:12:25.160 --> 0:12:26.320 Kinds of things that human minds do. 0:12:27.720 --> 0:12:30.120 Google had the insight that you needed a different kind 0:12:30.120 --> 0:12:32.640 of algebra in order to describe thought, and then that's 0:12:32.679 --> 0:12:36.200 what leads to modern mathematical logic. But it's still in 0:12:36.240 --> 0:12:39.560 this kind of symbolic language, although Gooole also talked about 0:12:39.559 --> 0:12:42.840 probability theory as being important for capturing languages as well. 0:12:42.920 --> 0:12:45.360 So I think it's really more about what are the 0:12:45.440 --> 0:12:48.640 kinds of mathematical systems that it was sort of straightforward 0:12:48.640 --> 0:12:51.199 to formalize, and that gave us something that we could 0:12:51.240 --> 0:12:53.600 try to map thought onto. And that's what we do 0:12:53.679 --> 0:12:58.320 as scientists is often taking mathematical systems that mathematicians have 0:12:58.360 --> 0:13:00.840 defined for us and then saying, oh, I think this 0:13:01.000 --> 0:13:03.800 mathematical system maps onto the thing that I want to understand, 0:13:04.200 --> 0:13:06.400 and so trying to establish that correspondence and not just 0:13:06.520 --> 0:13:08.199 then allow us to derive its consequences. 0:13:09.160 --> 0:13:12.840 So speaking of rules and symbols, So thinkers like Newl 0:13:12.920 --> 0:13:17.040 and Simon, they popularize this idea of goals and sub goals. 0:13:17.480 --> 0:13:21.920 What did that viewpoint get exactly right about human problem solving. 0:13:23.440 --> 0:13:26.760 So now we're fast forwarding a bit right from we 0:13:26.840 --> 0:13:30.800 have Boule figuring out the structure of logic. That turns 0:13:30.840 --> 0:13:32.480 into you know, lots of people then sort of turn 0:13:32.520 --> 0:13:34.840 that into a sort of mature theory of logic. You 0:13:34.920 --> 0:13:39.160 get aalenteering kind of turning this into a theory of computation, 0:13:39.480 --> 0:13:41.960 thinking about what an abstract mathematician is doing when they're 0:13:42.000 --> 0:13:43.880 doing something like logic, and thinking about how you can 0:13:43.920 --> 0:13:48.240 make a machine do that. And then we have people 0:13:48.520 --> 0:13:51.560 starting to realize that, you know, as digital computers are 0:13:51.559 --> 0:13:55.560 being developed, maybe those provide a good model for how 0:13:55.600 --> 0:13:59.880 thinking works in general, and then trying to use a 0:14:00.000 --> 0:14:03.200 computer as a sort of foundation for you know, thinking 0:14:03.200 --> 0:14:05.280 about things like how people might solve problems. And so 0:14:06.040 --> 0:14:09.800 Alan Ewele and Herbert Simon were influential cognitive scientists who 0:14:10.600 --> 0:14:14.679 did exactly that. They had this idea that maybe there 0:14:14.760 --> 0:14:17.400 is a way that you could make computers smarter by 0:14:17.440 --> 0:14:20.720 using insights from human cognition, but also get a better 0:14:20.800 --> 0:14:23.160 understanding of what humans are doing when they're solving problems 0:14:23.200 --> 0:14:25.400 by using the sort of ideas that come from things 0:14:25.440 --> 0:14:29.360 like computer programming, and so they set up you know this, 0:14:29.640 --> 0:14:31.240 you know, when we're trying to solve a problem or 0:14:31.240 --> 0:14:33.360 prove a mathematical theorem or play a game of chess, 0:14:33.720 --> 0:14:35.520 they set this up as a problem of searching through 0:14:35.560 --> 0:14:40.600 a tree of possibilities, where what you're doing is making choices, 0:14:40.880 --> 0:14:42.480 and then each of those choices gives you a new 0:14:42.520 --> 0:14:44.240 set of choices, and each of those choices gives you 0:14:44.240 --> 0:14:46.360 a new set of choices, and the hard thing is 0:14:46.880 --> 0:14:49.160 finding a path through those choices that leads to the 0:14:49.160 --> 0:14:51.160 point that you want to end up at. And so 0:14:51.720 --> 0:14:54.120 that's something where you can take inspiration from how human 0:14:54.120 --> 0:14:57.720 mathematicians solve problems. You can take inspiration from the kind 0:14:57.760 --> 0:14:59.960 of you know, tricks like working backwards from the end 0:15:00.080 --> 0:15:03.320 towards the start. Right, Those were principles that they were 0:15:03.360 --> 0:15:05.280 able to use to try and explain these aspects of 0:15:05.360 --> 0:15:07.800 human cognition as well as making the machines work better. 0:15:08.040 --> 0:15:09.960 Okay, but then one of the things that happened is 0:15:10.000 --> 0:15:13.120 that at least one of these attempts had ballooned into 0:15:13.200 --> 0:15:17.320 twenty five million rules. And so what does that teach 0:15:17.400 --> 0:15:20.520 us about the shape of human intelligence. 0:15:21.960 --> 0:15:23.800 This rules and symbols enterprise. 0:15:24.040 --> 0:15:24.160 Right. 0:15:24.240 --> 0:15:27.000 The sort of appeal that this had was that maybe 0:15:27.040 --> 0:15:27.520 one day. 0:15:27.400 --> 0:15:29.560 You could just write down all of the rules that 0:15:29.560 --> 0:15:31.720 you need to write down, and then you've characterized how 0:15:31.760 --> 0:15:34.320 intelligence works. Right, So it's just a matter of getting 0:15:34.400 --> 0:15:37.440 enough rules in a way that's very reminiscent today, right 0:15:37.520 --> 0:15:40.920 of you know, the way that our modern AI systems 0:15:40.920 --> 0:15:43.520 are being made is by training them on more and 0:15:43.600 --> 0:15:46.640 more data, right, feeding in more and more language. There 0:15:46.680 --> 0:15:48.520 was a hope that you could just like, yeah, like 0:15:48.800 --> 0:15:50.760 document all of the rules that you need to capture 0:15:51.080 --> 0:15:54.360 the structure of human knowledge. And so that led to 0:15:54.960 --> 0:15:57.240 you know, companies being started to try and engage in 0:15:57.280 --> 0:16:01.880 that enterprise, ultimately I would say, unsuccessfully, but giving us 0:16:01.920 --> 0:16:06.440 some kind of characterization of like particular subsets of human knowledge. 0:16:06.520 --> 0:16:08.640 And so I think the thing that came out of 0:16:08.920 --> 0:16:12.280 that enterprise was revealing that maybe you need something more 0:16:12.360 --> 0:16:16.600 than just rules, right, that maybe thinking about logic as 0:16:16.640 --> 0:16:19.480 a basis for our model of intelligence was missing something. 0:16:20.040 --> 0:16:22.240 It's an approach that worked really well for certain kinds 0:16:22.280 --> 0:16:26.200 of problems like doing arithmetic, playing games or chess, but 0:16:26.240 --> 0:16:28.520 it didn't work very well for other kinds of problems 0:16:28.640 --> 0:16:31.520 like figuring out what you're seeing in the world, or 0:16:31.880 --> 0:16:34.360 actually learning language or these other kinds of things. 0:16:34.400 --> 0:16:37.080 And so this is what leads to your second lens, 0:16:37.240 --> 0:16:40.800 which is neural networks. And you talk about these as 0:16:40.880 --> 0:16:44.080 having you know, a boom and bust history. So, first, 0:16:44.120 --> 0:16:46.720 what happened in the last decade that allowed them to 0:16:46.760 --> 0:16:48.880 turn into the dominant paradigm. 0:16:49.320 --> 0:16:53.440 The big breakthrough in the last decade was really about 0:16:53.760 --> 0:16:56.680 being able to make bigger in neural networks that could 0:16:56.720 --> 0:17:01.840 be trained on more data in a way that could scale, right, 0:17:01.880 --> 0:17:05.159 and so bigger here means what these are. An artificial 0:17:05.240 --> 0:17:08.840 neural network is a set of units that are communicating 0:17:08.840 --> 0:17:13.120 with one another. They're communicating along weighted connections, a sort 0:17:13.160 --> 0:17:15.639 of you know, imagine like how neurons are connected in 0:17:15.680 --> 0:17:17.840 your brain, and those neurons are connected to one another 0:17:17.880 --> 0:17:20.359 and sending each other signals. An artificial neural network is 0:17:20.400 --> 0:17:23.719 basically simulating that kind of structure inside a computer. And 0:17:23.760 --> 0:17:26.679 so for a long time, the sort of the history 0:17:26.680 --> 0:17:30.040 of neural networks has been one of people figuring out 0:17:30.119 --> 0:17:33.880 how to make bigger neural networks work. So the very 0:17:33.920 --> 0:17:37.800 first you know, learning neural networks. They had a learning 0:17:37.840 --> 0:17:40.719 algorithm that worked for one layer of weights, and then 0:17:40.760 --> 0:17:42.560 there was a breakthrough in the nineteen eighties that meant, 0:17:42.600 --> 0:17:44.280 now you had a learning algorithm that could work for 0:17:44.359 --> 0:17:46.600 multiple layers of weights, but it didn't work for very 0:17:46.760 --> 0:17:48.920 deep neural networks with lots of layers of net weights 0:17:48.960 --> 0:17:52.160 because it I can sort of explain the technical reasons 0:17:52.200 --> 0:17:54.159 behind it, but you know, sort of like the basic 0:17:54.160 --> 0:17:57.760 algorithm didn't quite work. And so the big breakthroughs of 0:17:57.840 --> 0:18:00.000 the last you know, ten to fifteen years have been 0:18:00.080 --> 0:18:03.480 about coming up with ways to take those algorithms and 0:18:03.520 --> 0:18:05.440 actually make them work for neural networks that are bigger 0:18:05.480 --> 0:18:07.679 and bigger and deeper and deeper, that are able to 0:18:07.840 --> 0:18:11.840 easily learn more complex functions and can do so from 0:18:12.160 --> 0:18:14.879 massive amounts of data in a way that means that 0:18:14.920 --> 0:18:18.000 they're able to discover sort of complex relationships between things 0:18:18.000 --> 0:18:19.960 that are necessary to produce intelligent behavior. 0:18:20.280 --> 0:18:24.080 And so, what are these neural networks capture about cognition 0:18:24.760 --> 0:18:29.200 that symbols missed, especially in terms of things like similarity 0:18:29.240 --> 0:18:32.040 and fuzziness and graded concepts. 0:18:32.560 --> 0:18:35.000 Fuzziness is a really good way of describing it. It's 0:18:35.040 --> 0:18:39.080 that you know, if you ask somebody, you know, whether 0:18:39.119 --> 0:18:41.800 something is a piece of furniture, they're going to say, 0:18:42.119 --> 0:18:43.840 you know, if you show them a chair, they'll say, yes, 0:18:43.920 --> 0:18:46.879 definitely a piece of furniture. If you show them a rug, 0:18:47.440 --> 0:18:51.119 they'll say, yeah, maybe a piece of furniture. Right, it 0:18:51.119 --> 0:18:52.960 doesn't sort of fit with our you know, week sort 0:18:52.960 --> 0:18:56.240 of have a prototypical idea of what furniture is, which 0:18:56.240 --> 0:18:58.479 contains things like chairs and tables and ottomans and these 0:18:58.480 --> 0:19:02.119 other kinds of things, and then rugs and treadmills, and you. 0:19:02.160 --> 0:19:04.040 Know, like these are things that maybe. 0:19:03.760 --> 0:19:06.320 You're in this category, but maybe an't right. And so 0:19:07.160 --> 0:19:09.520 we need to have a way of thinking about concepts 0:19:09.520 --> 0:19:11.760 that's not just the sort of yes or no, true 0:19:11.800 --> 0:19:14.760 or false one or zero that logic would give us. 0:19:14.800 --> 0:19:16.840 We need to have something which has that fuzziness in it. 0:19:17.160 --> 0:19:19.399 One way of getting fuzziness is by thinking about a 0:19:19.480 --> 0:19:22.760 concept in terms of points in space, right where you 0:19:22.760 --> 0:19:26.399 could think chairs are here in one location, rugs are 0:19:26.440 --> 0:19:29.040 here in another location, and maybe what it is to 0:19:29.040 --> 0:19:30.439 be a piece of furniture is to just be in 0:19:30.440 --> 0:19:32.560 some part of that space, and how close you are 0:19:32.600 --> 0:19:34.280 to that part of the space is like how good 0:19:34.320 --> 0:19:36.560 you are as an example of that kind of furniture. 0:19:37.359 --> 0:19:39.800 And so as soon as you think in those terms, 0:19:39.800 --> 0:19:42.160 you have a new problem, which is with our rules 0:19:42.160 --> 0:19:44.400 and symbols. We knew how to do computation, we knew 0:19:44.400 --> 0:19:47.000 how to describe thinking. Thinking was a matter of applying 0:19:47.040 --> 0:19:49.240 the rules and seat of you know, repeating that process. 0:19:50.040 --> 0:19:52.640 But we don't have a way of doing computation in spaces. 0:19:52.720 --> 0:19:54.439 And that's what youral networks give us. So you can 0:19:54.520 --> 0:19:58.760 kind of think about a space corresponding to the activation 0:19:58.920 --> 0:20:00.639 of the units inside this neural network. 0:20:00.680 --> 0:20:03.040 How much you know, how much input. 0:20:02.720 --> 0:20:05.600 Each neural unit in that neural network is receiving, and 0:20:04.920 --> 0:20:08.600 how much of a response it's making that characterizes some 0:20:08.680 --> 0:20:10.639 kind of space. And then neural network gives us a 0:20:10.640 --> 0:20:13.359 way of mapping from the inputs that it's getting to 0:20:13.440 --> 0:20:14.080 some output. 0:20:14.200 --> 0:20:16.040 So you could put in you know. 0:20:16.000 --> 0:20:18.239 Your picture of a chair, and it maps that to 0:20:18.240 --> 0:20:20.560 some point in space, and then it put sort of 0:20:20.560 --> 0:20:23.440 produces out an output the corresponds to, yes, this is 0:20:23.480 --> 0:20:26.080 a piece of furniture. And because those outputs can now 0:20:26.119 --> 0:20:29.199 be continuous values, you can capture the fuzziness and other 0:20:29.280 --> 0:20:31.360 kinds of things that you want for your concepts. 0:20:31.640 --> 0:20:34.960 And so, in what sense are these modern systems, these 0:20:35.040 --> 0:20:38.360 artificial neural networks learning, and in what sense are they 0:20:38.400 --> 0:20:42.720 doing something that's maybe categorically different from how children learn. 0:20:44.520 --> 0:20:48.320 This is a fundamental question, right, That's the kind of 0:20:48.320 --> 0:20:50.440 thing that we cognitive scientists think about a lot, and 0:20:50.800 --> 0:20:53.080 I think that AI researchers are starting to care about 0:20:53.119 --> 0:20:55.359 a lot too, which is, you know, what are these 0:20:55.440 --> 0:20:59.720 sort of meaningful differences between human minds, human brains and 0:20:59.720 --> 0:21:01.680 what we building in these AI systems or these sort 0:21:01.680 --> 0:21:07.240 of artificial brains. I think one very salient difference is 0:21:07.359 --> 0:21:10.399 the amount of data which is needed for a human 0:21:10.640 --> 0:21:12.639 to learn language compared to the amount of data you 0:21:12.640 --> 0:21:15.159 need to put into on neural network. So if you 0:21:15.200 --> 0:21:18.640 take a system like chat GPT, right, one of these chatbots, 0:21:19.000 --> 0:21:21.639 those systems are trained on the equivalent of something like 0:21:21.720 --> 0:21:24.920 five thousand to fifty thousand years of continuous speech. There's 0:21:24.920 --> 0:21:26.800 sort of massive amounts of data that are going into 0:21:26.800 --> 0:21:28.919 that system. So it's like on the order of a 0:21:28.960 --> 0:21:31.159 thousand or ten thousand times as much data as a 0:21:31.240 --> 0:21:33.879 human child might get in order to learn language. And 0:21:33.960 --> 0:21:37.280 the reason is that those artificial neural networks are really 0:21:37.359 --> 0:21:40.399 kind of like undifferentiated learning machines. You can take that 0:21:40.440 --> 0:21:41.960 same kind of neural network, you can get it to 0:21:42.040 --> 0:21:44.400 learn all sorts of different kinds of things. It works 0:21:44.400 --> 0:21:46.119 really well for learning language, but you can use it 0:21:46.160 --> 0:21:47.920 to learn something about vision or something. You know, you 0:21:47.960 --> 0:21:49.280 can sort of take all sorts of problems and give 0:21:49.280 --> 0:21:50.520 it to them and it can learn how to do that. 0:21:51.040 --> 0:21:56.000 And so as a consequence, they have what we call 0:21:56.000 --> 0:21:59.760 in cognitive science machine learning inductive biases. They're not biased 0:21:59.800 --> 0:22:04.120 to towards any particular solution to the learning problem, and 0:22:04.400 --> 0:22:09.399 human brains have stronger inductive biases for things like learning language. Right, 0:22:09.440 --> 0:22:12.200 we're sort of disposed towards certain kinds of things, which 0:22:12.200 --> 0:22:14.679 are human languages. The things that we call human languages 0:22:14.680 --> 0:22:16.880 are the things that we're disposed to learn. And as 0:22:16.880 --> 0:22:19.480 a consequence, you know, we're able to sort of narrow 0:22:19.560 --> 0:22:21.320 down the space of possibilities in a way that means 0:22:21.320 --> 0:22:22.680 that we're able to learn from less data. 0:22:36.960 --> 0:22:39.399 Okay, this makes a great segue to the third lens. 0:22:39.440 --> 0:22:42.880 So you talked about rules and symbols, and you talked 0:22:42.920 --> 0:22:45.560 about artificial neural networks. The third part of your book 0:22:45.640 --> 0:22:51.200 is about probabilities and statistics. So why did probability become 0:22:51.240 --> 0:22:56.000 an attractive candidate language for thinking about cognition? 0:22:56.920 --> 0:22:59.520 Probability there is a good way of answering certain kinds 0:22:59.520 --> 0:23:02.840 of why questions that we might have, right so, and 0:23:02.880 --> 0:23:05.520 the reason is that it's a way of characterizing how 0:23:05.560 --> 0:23:08.960 a rational agent should make an inference. So all the 0:23:08.960 --> 0:23:13.159 way back in the eighteenth century, British nonconformist minister, the 0:23:13.160 --> 0:23:16.440 Reverend Thomas Bays had this radical idea that you could 0:23:17.119 --> 0:23:19.800 talk about, you know, again, like take a mathematical system 0:23:19.880 --> 0:23:22.399 probability theory, which we would use for describing what happens 0:23:22.440 --> 0:23:24.639 when you roll dice or flip coins right, sort of 0:23:24.720 --> 0:23:27.400 you know, sort of language of gambling and saying, oh, 0:23:27.480 --> 0:23:30.240 in fact, that mathematical system might also be a really 0:23:30.280 --> 0:23:33.560 good system for describing how beliefs work. And so what 0:23:33.640 --> 0:23:35.399 he was interested in was if you think about a 0:23:35.400 --> 0:23:38.280 belief as you know, a degree of belief, right, you 0:23:38.320 --> 0:23:40.160 can say, oh, I think it's going to rain tomorrow, 0:23:40.600 --> 0:23:42.800 and I'll put put that on a scale which goes 0:23:42.800 --> 0:23:45.479 from zero to one, where zero is you know, not 0:23:45.520 --> 0:23:47.359 going to rain, and one is one hundred percent it's 0:23:47.359 --> 0:23:47.760 going to rain. 0:23:47.800 --> 0:23:48.200 Tomorrow. 0:23:48.280 --> 0:23:52.040 Right, That is a belief that you've expressed in the 0:23:52.080 --> 0:23:54.960 form of a probability. And now if you, you know, 0:23:55.119 --> 0:23:56.639 wake up in the morning and look out the window 0:23:56.640 --> 0:23:59.239 and you see gray storm clouds, you've got a new 0:23:59.240 --> 0:24:02.480 piece of data. You need to revise your beliefs, and 0:24:02.560 --> 0:24:05.119 probability theory actually tells you how to do that. It says, 0:24:06.320 --> 0:24:09.200 you know, for each hypothesis, right, so our hypotheses here 0:24:09.240 --> 0:24:11.040 are it's going. 0:24:10.960 --> 0:24:12.080 To rain or it's not going to rain. 0:24:12.440 --> 0:24:15.360 Right, You're going to modify that belief based on how 0:24:15.440 --> 0:24:18.639 likely the data is that you saw if that hypothesis 0:24:18.680 --> 0:24:21.359 were true. So, because gray storm clouds are more likely 0:24:21.440 --> 0:24:23.920 if it's going to rain that day, we should increase 0:24:23.960 --> 0:24:27.600 our belief that it's going to rain. And as a consequence, well, 0:24:27.680 --> 0:24:29.040 we'll end up with a number that's a little bit 0:24:29.119 --> 0:24:30.359 higher than the number we had before. 0:24:30.600 --> 0:24:32.280 And probability theory tells us how to do that. 0:24:32.520 --> 0:24:35.240 There's a principle of probability theory called Bays rule after 0:24:35.240 --> 0:24:37.680 the Reverend Thomas Beys that tells you how to take 0:24:37.760 --> 0:24:40.960 your original beliefs and then turn them into the beliefs 0:24:40.960 --> 0:24:43.680 that you get after seeing data. And that turns out 0:24:43.720 --> 0:24:45.600 to be exactly the tool that you need to answer 0:24:45.640 --> 0:24:49.560 these kinds of questions about how inductive biases work. Right, So, 0:24:50.160 --> 0:24:53.320 how is it that children are able to learn from 0:24:53.600 --> 0:24:56.880 less data than anural networks. Well, it's a consequence of 0:24:57.440 --> 0:25:00.000 you know, these things that we can describe using different 0:25:00.160 --> 0:25:03.399 probabilities being assigned to different hypotheses, whether hypotheses correspond to 0:25:03.400 --> 0:25:05.080 the structure of the languages that are being learned. 0:25:05.320 --> 0:25:10.320 And when people call humans irrational, what changes if we 0:25:10.440 --> 0:25:16.120 look at mistakes as resource limited inferences. 0:25:17.640 --> 0:25:20.280 This is one of the things that I explore in 0:25:19.760 --> 0:25:23.399 my own research is this question of how we should 0:25:23.400 --> 0:25:26.919 actually think about rationality for real agents, right, And this 0:25:27.000 --> 0:25:28.919 is relevant if you're building an AI system or if 0:25:28.920 --> 0:25:31.760 you're just trying to understand human behavior. So I think, 0:25:33.040 --> 0:25:35.720 like I said, probability theory gives us a characterization of 0:25:35.800 --> 0:25:38.679 what you should do as an ideal rational agent, But 0:25:38.760 --> 0:25:42.640 that assumes that you have infinite computational resources. We mirror, 0:25:42.760 --> 0:25:46.440 humans don't have infinite computational resources, nor do our AI systems, 0:25:46.560 --> 0:25:49.280 And so you can ask what should a rational agent 0:25:49.320 --> 0:25:52.280 do if they don't have all of the computational resources 0:25:52.280 --> 0:25:54.359 that you might need and then out of that you 0:25:54.400 --> 0:25:57.160 get the answer is that you know, you should follow 0:25:57.480 --> 0:26:02.320 an algorithm, follow us strategy makes sense given the resources 0:26:02.359 --> 0:26:04.280 that you have. That's what it means to be rational 0:26:04.280 --> 0:26:06.840 in those circumstances where you're sort of doing the best 0:26:06.920 --> 0:26:12.159 job you can of approximating probabilistic inference given those resource constraints. 0:26:12.200 --> 0:26:14.479 And so some of the things that people do when 0:26:14.520 --> 0:26:16.400 people do weird things, and we do lots of weird things, 0:26:16.440 --> 0:26:19.000 and we don't always follow probability theory, things that we 0:26:19.000 --> 0:26:22.320 can understand as us, you know, running into those resource 0:26:22.359 --> 0:26:25.200 limitations and then coming up with, you know, reasonable strategies 0:26:25.200 --> 0:26:26.719 for trying to approximate the right answer. 0:26:28.040 --> 0:26:29.080 So, if we look. 0:26:28.960 --> 0:26:33.800 At probability being the grammar of uncertainty, one thing we 0:26:33.880 --> 0:26:36.840 know is that our prior expectations matter. And one of 0:26:36.880 --> 0:26:39.280 the things I've been obsessed with and doing a lot 0:26:39.320 --> 0:26:41.000 of research on and talking a lot about on the 0:26:41.000 --> 0:26:44.200 podcast is the way that all of us drop into 0:26:44.200 --> 0:26:48.520 the world and our cultures influence us and our language 0:26:48.560 --> 0:26:51.160 and our moment in time and our neighborhood, and this 0:26:51.320 --> 0:26:55.760 leads to people being quite different on the inside. Is 0:26:55.800 --> 0:26:58.159 this something that you think about sometimes about how we 0:26:58.240 --> 0:27:03.760 develop our priors differently based on you know what where 0:27:03.760 --> 0:27:04.360 we grow up. 0:27:05.600 --> 0:27:08.679 Yeah, So priors is that Daysian language, right, for talking 0:27:08.680 --> 0:27:11.240 about the beliefs that you have before you get data 0:27:11.440 --> 0:27:14.840 that you then update into what we call posterior probabilities 0:27:14.840 --> 0:27:17.440 that are informed by those data. And so yeah, I 0:27:18.359 --> 0:27:20.720 spend a lot of my research time thinking about these 0:27:20.800 --> 0:27:23.920 kinds of questions of you know what are these sort 0:27:23.960 --> 0:27:28.880 of prior distributions for humans? How do we acquire good 0:27:28.960 --> 0:27:31.959 prior distributions for solving different kinds of problems. One thing 0:27:32.000 --> 0:27:34.520 there is that calling it a prior makes it sound 0:27:34.560 --> 0:27:36.840 like maybe it's something you're born with, but in fact, 0:27:36.880 --> 0:27:37.680 it just means. 0:27:37.720 --> 0:27:38.960 It's before you get data. 0:27:39.040 --> 0:27:41.439 Right, So when you're seeing that storm cloud in the morning, 0:27:42.040 --> 0:27:44.520 you had a prior probability from last night, and then 0:27:44.640 --> 0:27:47.560 that prior probability was informed by everything else that you know, Right, 0:27:47.680 --> 0:27:49.920 The priors are all of the biases and knowledge that 0:27:49.960 --> 0:27:51.880 we bring to bear when we're trying to make an difference. 0:27:51.960 --> 0:27:55.080 And so yeah, I think I think understanding that is 0:27:55.119 --> 0:27:57.280 a big part of the project of understanding human cognition. 0:27:57.600 --> 0:27:59.120 So let's zoom the camera out. 0:27:59.200 --> 0:28:01.639 We've talked about the three lenses that you describe in 0:28:01.640 --> 0:28:04.119 the book. Now, you also point out that we have 0:28:04.160 --> 0:28:08.199 a lot of constraints like finite lives in limited compute, 0:28:08.200 --> 0:28:11.760 and limited bandwidth, and so how do these constraints sculpt 0:28:12.040 --> 0:28:13.320 human intelligence? 0:28:13.800 --> 0:28:16.520 I think this is really important to just thinking about 0:28:16.560 --> 0:28:18.399 the moment that we're in where there's a lot of 0:28:18.400 --> 0:28:21.520 anxiety around AI, right, And I think if you think 0:28:21.520 --> 0:28:24.439 about intelligence as a kind of one dimensional quantity, you 0:28:24.440 --> 0:28:26.560 can kind of imagine that you know, humans are somewhere, 0:28:26.920 --> 0:28:28.320 our AI systems are somewhere. 0:28:28.359 --> 0:28:29.560 It seems like they're approaching us. 0:28:29.600 --> 0:28:31.560 Maybe they're going to overtake us, and then, oh my god, 0:28:31.560 --> 0:28:33.120 what is going to happen when that happens, Right, We're 0:28:33.160 --> 0:28:34.600 just going to become redundant. There's not going to be 0:28:34.600 --> 0:28:36.520 any jobs. Everything is going to fall apart. And so 0:28:37.160 --> 0:28:39.520 that's a consequence of having a particular conception of what 0:28:39.640 --> 0:28:41.880 intelligence is, which is this kind of one dimensional way 0:28:41.880 --> 0:28:44.040 of thinking about it. And I think there's a different 0:28:44.080 --> 0:28:46.200 way of thinking about it which gives us a little 0:28:46.200 --> 0:28:48.000 more flexibility and maybe a little more hope in the 0:28:48.040 --> 0:28:50.120 way that we think about what's going to happen with AI, 0:28:50.560 --> 0:28:55.840 and that is thinking about intelligence as being an adaptation 0:28:56.040 --> 0:28:59.240 to the kinds of computational problems that a system has 0:28:59.360 --> 0:29:02.520 either of or being trained to solve, right, And so 0:29:03.360 --> 0:29:06.240 for human beings, those computational problems are shaped by the 0:29:06.280 --> 0:29:08.440 constraints that we operate under. And a lot of those 0:29:08.440 --> 0:29:11.080 constraints come from our biology, right that we, as you said, 0:29:11.480 --> 0:29:14.880 have limited lives, have you know, limited compute resources what 0:29:14.880 --> 0:29:17.560 we can carry around inside our heads, have limited bandwidth 0:29:17.600 --> 0:29:20.480 for communication. We have to like make noises at each other, 0:29:20.800 --> 0:29:22.800 or wiggle our fingers or you know, somehow use our 0:29:22.840 --> 0:29:26.280 bodies to transfer data from one human mind to another 0:29:26.360 --> 0:29:30.440 human mind. It's very inefficient. And so those constraints are 0:29:30.520 --> 0:29:34.200 things that mean that human intelligence takes a particular shape, 0:29:34.440 --> 0:29:37.920 which is we're able to learn from limited data because 0:29:37.920 --> 0:29:40.080 we have to because we don't live that long. Right, 0:29:40.280 --> 0:29:44.120 You can't rely on getting five thousand years of language 0:29:44.160 --> 0:29:47.600 data or multiple human lifetimes of you know, chess playing 0:29:47.680 --> 0:29:51.400 or whatever it is. Right, you have to be good 0:29:51.400 --> 0:29:53.960 at using the resources that you have in ways that 0:29:53.960 --> 0:29:56.000 are efficient. And so that's kind of like deciding what 0:29:56.080 --> 0:29:59.160 to think about being able to recognize when a problem 0:29:59.200 --> 0:30:01.800 has a structure that you I've seen before being able 0:30:01.840 --> 0:30:04.280 to you know, sort of like become sort of automatic 0:30:04.320 --> 0:30:07.160 in using certain kinds of patterns of thinking and strategies 0:30:07.160 --> 0:30:11.120 for solving problems, really trying to make it as easy as. 0:30:11.000 --> 0:30:13.320 Possible for us to use the resources that we have. 0:30:13.800 --> 0:30:18.760 And then you need to develop capacities for trying to 0:30:18.800 --> 0:30:22.000 circumvent those bandwidth constraints in order to be able to 0:30:22.040 --> 0:30:24.320 do things that transcend what any individual human can do, 0:30:24.640 --> 0:30:30.600 and that means developing things like language writing societies LLCs, 0:30:30.920 --> 0:30:34.920 you know, all of the sort of libraries, right institutions, 0:30:35.240 --> 0:30:37.960 all of the theory of mind, right for reasoning about 0:30:37.960 --> 0:30:40.360 what someone else might be trying to communicate to you. 0:30:40.840 --> 0:30:43.120 All of this stuff is actually sort of like human 0:30:43.160 --> 0:30:46.920 stuff that's a consequence of these constraints. And so as 0:30:46.920 --> 0:30:50.800 we make AI systems that are smarter, those AI systems 0:30:50.800 --> 0:30:53.400 are in turn being shaped by what they're being trained 0:30:53.400 --> 0:30:56.280 to do and what constraints they operate under. But those 0:30:56.320 --> 0:30:58.760 constraints are different from the ones that humans have. They 0:30:58.800 --> 0:31:02.320 can you know, get more data, they can get access 0:31:02.320 --> 0:31:04.520 to more compute, they don't have bandwidth limitations. 0:31:04.520 --> 0:31:05.520 You can just copy. 0:31:05.240 --> 0:31:09.280 A you know, a state of an AI system across machines. 0:31:09.360 --> 0:31:12.440 You can use the same data to train multiple AI systems, 0:31:13.000 --> 0:31:15.280 and all of those things mean that I think, rather 0:31:15.320 --> 0:31:17.600 than being sort of on one axis where we're sort 0:31:17.600 --> 0:31:20.800 of thinking about better and worse, it's more that there 0:31:20.800 --> 0:31:23.560 are many axes that we can think about intelligent systems 0:31:23.600 --> 0:31:25.680 developing along, and we're just going to end up in 0:31:25.680 --> 0:31:28.080 a state where we have human intelligence and we have AIS, 0:31:28.520 --> 0:31:30.640 and they're going to be meaningfully different from one another, 0:31:31.160 --> 0:31:33.200 rather than things that are sort of directly competing in 0:31:33.320 --> 0:31:34.600 terms of the capacities that they have. 0:31:35.280 --> 0:31:36.640 Yeah, I agree with you on that. 0:31:37.280 --> 0:31:41.480 When you think about the way that humans beat machines 0:31:41.640 --> 0:31:44.400 on data efficiency, what do you think that means is 0:31:44.520 --> 0:31:47.320 missing architecturally from our AI systems. 0:31:48.880 --> 0:31:52.360 I think it's actually it's a great question, and the 0:31:52.360 --> 0:31:55.040 way I would express it is not in terms of architecture. 0:31:55.720 --> 0:31:58.280 So it's actually in terms of a different part of 0:31:58.320 --> 0:32:01.000 a neural network. So when we think about this problem 0:32:01.000 --> 0:32:05.520 of inductive bias, right, which is we're you know, what 0:32:06.160 --> 0:32:09.120 a system is sort of disposed towards learning. As I said, 0:32:09.240 --> 0:32:11.840 our neural networks, the way we normally set them up, 0:32:12.040 --> 0:32:14.040 are pretty weak inductive biases. They can learn all sorts 0:32:14.040 --> 0:32:17.320 of things. The inductive bias that a neural network has 0:32:17.560 --> 0:32:20.640 it is constrained by its architecture, but it's also constrained 0:32:20.640 --> 0:32:23.080 by where it starts out in the space of the 0:32:23.120 --> 0:32:26.680 settings of all of those weights. And normally the default 0:32:26.720 --> 0:32:28.360 is that you set up your neural networks so those 0:32:28.360 --> 0:32:31.000 weights start out really small, close to zero, and then 0:32:31.000 --> 0:32:32.960 they sort of grow away from that as it starts 0:32:33.000 --> 0:32:36.040 to learn how to do things. We've had success in 0:32:36.120 --> 0:32:40.719 taking neural networks that are architecturally identical but setting them 0:32:40.720 --> 0:32:43.720 with different initial weights in order to create an inductive 0:32:43.760 --> 0:32:47.000 bias that enables rapid learning. So we act to use 0:32:47.040 --> 0:32:49.720 a technique called meta learning, which is a method from 0:32:49.720 --> 0:32:53.440 machine learning where you take the same neural network architecture 0:32:53.600 --> 0:32:55.960 and the same initial weights, and you use it to 0:32:56.040 --> 0:32:57.960 learn to solve lots of different problems, like you can 0:32:58.040 --> 0:33:00.800 use it to learn lots of different languages, say, from 0:33:00.840 --> 0:33:04.000 limited data, and then you try and optimize the initial 0:33:04.040 --> 0:33:05.800 weights of the neural network to make it so it 0:33:05.800 --> 0:33:09.000 can learn all of those languages better using the same 0:33:09.080 --> 0:33:11.000 kinds of algorithms we use for training the weights of 0:33:11.040 --> 0:33:12.880 the neural network. When we have these giant data sets. 0:33:12.880 --> 0:33:15.560 You can instead use those algorithms to train the initial 0:33:15.560 --> 0:33:17.680 weights of a neural network for a small data set, 0:33:17.720 --> 0:33:20.760 for lots of small data sets, And when you do that, 0:33:20.800 --> 0:33:22.320 you end up with a neural network that has an 0:33:22.360 --> 0:33:25.400 inductive bias that makes it possible to learn from small 0:33:25.400 --> 0:33:26.960 amounts of data. And so that's the kind of thing 0:33:27.000 --> 0:33:29.920 we've been exploring in my lab is can we find 0:33:29.960 --> 0:33:33.080 a way of taking exactly these same neural network architectures 0:33:33.600 --> 0:33:36.000 and just starting them out in a different place that 0:33:36.120 --> 0:33:39.120 maybe aligns better with the kinds of things that humans do. Okay, well, 0:33:39.160 --> 0:33:40.880 this is a really good segue to what I wanted 0:33:40.880 --> 0:33:43.600 to ask you, which is, if you're looking at rules 0:33:43.640 --> 0:33:46.880 and systems is one sort of math to describe the mind, 0:33:46.880 --> 0:33:49.320 and artificial neural networks is another kind of math, and 0:33:49.360 --> 0:33:54.200 probability is another. What does an optimal hybrid look like? 0:33:54.320 --> 0:33:58.920 Given that no single, no one of these describes everything 0:33:58.960 --> 0:34:02.240 about what's going on with minds, So what does the 0:34:02.360 --> 0:34:04.120 hybrid for an aisystem look like? 0:34:04.200 --> 0:34:07.120 In twenty twenty six, the place. 0:34:06.920 --> 0:34:09.759 Where I end up in the book is saying that 0:34:10.080 --> 0:34:13.600 these different kinds of math really do all fit together 0:34:13.640 --> 0:34:16.239 in an interesting way, and in order to understand that 0:34:16.280 --> 0:34:18.560 we can talk about different levels of analysis when we're 0:34:18.560 --> 0:34:21.040 trying to make sense of an information processing system. So 0:34:21.920 --> 0:34:23.880 the most abstract level, this is an idea that was 0:34:23.880 --> 0:34:27.640 introduced by the computational neuroscientist David Marr. He said, the 0:34:27.680 --> 0:34:29.560 most abstract level is just thinking about the problem that 0:34:29.560 --> 0:34:32.279 the system is solving in its ideal solution, right. And 0:34:32.960 --> 0:34:36.279 I think logic and symbolic systems and probability theory give 0:34:36.360 --> 0:34:39.239 us a good way of characterizing the kinds of problems 0:34:39.280 --> 0:34:43.279 that minds have to solve, right They you know, probabilistic 0:34:43.280 --> 0:34:45.960 inference because we have to make these uncertain inferences. And 0:34:46.000 --> 0:34:49.400 then logic as a way of characterizing the kinds of 0:34:49.440 --> 0:34:50.960 things that are in the world that have this rich 0:34:51.000 --> 0:34:54.480 structure of you know, like a sort of combinatorial structure 0:34:54.520 --> 0:34:57.400 that you get from from having symbols and rules that 0:34:57.440 --> 0:35:00.160 combine together with things like language and dance and all 0:35:00.160 --> 0:35:02.400 of these you know, structured things. Even you know, if 0:35:02.400 --> 0:35:03.960 you look at trees, you can see they have like 0:35:04.120 --> 0:35:06.759 recursive structures that are expressed in them. Right, So these 0:35:06.840 --> 0:35:08.600 kind of occur in nature and are important to be 0:35:08.600 --> 0:35:11.640 able to understand. And then at the level below that 0:35:11.840 --> 0:35:17.160 you have how the system solves those problems, right, like 0:35:17.480 --> 0:35:20.279 what algorithms it might use, what representations it might use, 0:35:20.400 --> 0:35:22.920 And then below that it's you know, how that's actually 0:35:22.920 --> 0:35:25.960 implemented in some kind of physical system, right, and artificial 0:35:25.960 --> 0:35:27.799 neural networks give us a kind of story at those 0:35:27.880 --> 0:35:31.720 levels where we can think about them as being a 0:35:31.760 --> 0:35:35.439 good general purpose system for learning to approximate the things 0:35:35.440 --> 0:35:37.799 that probabilistic in front tells you to do, and learning 0:35:37.840 --> 0:35:40.560 to approximate the structure that's contained within those symbolic systems. 0:35:40.960 --> 0:35:45.160 So I actually think, you know, the kind of story 0:35:45.200 --> 0:35:47.319 that we have right now that's emerged out of these 0:35:47.400 --> 0:35:50.120 advances in AI is actually a pretty good story for 0:35:50.200 --> 0:35:52.759 how we could think about human minds working. The thing 0:35:52.800 --> 0:35:55.279 that's missing, most important thing that's missing is this kind 0:35:55.320 --> 0:35:59.279 of aspect of inductive bias, right where we haven't been 0:35:59.320 --> 0:36:01.440 able to capture what human inductive biases are like in 0:36:01.520 --> 0:36:03.799 machines and so that you have these meaningful differences that come. 0:36:03.680 --> 0:36:04.080 Out of that. 0:36:05.440 --> 0:36:08.600 But it's not a bad place for thinking about how 0:36:08.640 --> 0:36:11.080 these pieces might fit together to give us an explanation 0:36:11.120 --> 0:36:12.279 for how it is that mind's work. 0:36:12.680 --> 0:36:18.399 So along these lines, which AI benchmarks feel to you misleading? 0:36:18.520 --> 0:36:21.080 And how would you make better benchmarks? 0:36:22.640 --> 0:36:26.239 So, in general, I'm not a huge fan of benchmarks, 0:36:26.360 --> 0:36:29.279 because I think benchmarks are useful as an engineering tool, 0:36:29.880 --> 0:36:33.560 but I, as a cognitive scientist, don't just want to 0:36:33.600 --> 0:36:36.120 know how well something is doing something. I want to 0:36:36.120 --> 0:36:38.960 know how it's doing that thing right and how it 0:36:39.040 --> 0:36:41.719 might be sort of messing that up right. So when 0:36:41.800 --> 0:36:46.160 we are designing experiments as cognitive scientists, we don't just say, oh, 0:36:46.200 --> 0:36:48.120 here's one hundred math problems. Go do with one hundred 0:36:48.160 --> 0:36:50.720 math problems and we'll get a score. We say, let's 0:36:50.840 --> 0:36:54.480 choose a set of math problems so that which answers 0:36:54.640 --> 0:36:57.960 people give us tell us about the misconceptions that they 0:36:58.000 --> 0:37:00.319 have in a way that we can then diagnose, oh, 0:37:00.440 --> 0:37:02.080 you know, this is why this person is thinking this 0:37:02.120 --> 0:37:04.160 particular thing. And so I think there's lots of room 0:37:04.200 --> 0:37:06.520 for coming up with better ways of evaluating our AI 0:37:06.560 --> 0:37:09.600 systems that look more like cognitive science experiments. We're really 0:37:09.640 --> 0:37:12.719 targeting understanding what's going on rather than just trying to 0:37:12.719 --> 0:37:15.480 get some brute sort of you know, performance score. 0:37:16.320 --> 0:37:19.920 Okay, good, And you have talked about curiosity as a 0:37:19.960 --> 0:37:25.800 computational problem. So how do you think about what curiosity 0:37:25.960 --> 0:37:30.320 is and how we might measure real curiosity in a machine? 0:37:30.680 --> 0:37:34.400 What problem is curiosity trying to solve? Yeah, this is 0:37:34.440 --> 0:37:36.000 this is a good question. You can you can ask 0:37:36.040 --> 0:37:38.480 this kind of question that we call rational analysis. Right, 0:37:38.520 --> 0:37:42.640 if you have a system that's solving a problem, you 0:37:42.680 --> 0:37:44.000 know what's what's the problem? 0:37:44.000 --> 0:37:44.800 What's the ideal solution? 0:37:44.960 --> 0:37:48.279 Okay, So for curiosity, we've argued and this is work 0:37:48.320 --> 0:37:49.399 with wretched debate. 0:37:49.400 --> 0:37:52.200 Who is that UCLA That. 0:37:53.800 --> 0:37:56.960 One way I think about curiosity is that you're trying 0:37:56.960 --> 0:38:02.760 to find things that are good in increasing your long 0:38:03.120 --> 0:38:06.880 run probability of being able to solve problems in the future. Right, 0:38:07.000 --> 0:38:10.120 So you know, it's sort of like you want data 0:38:10.200 --> 0:38:13.840 which for which the derivative of your total knowledge is 0:38:13.960 --> 0:38:18.960 high relative to that particular data point. And so that 0:38:19.040 --> 0:38:21.800 explanation captures some of the things that happen in human cognition, 0:38:22.000 --> 0:38:25.239 where you know, in some circumstances, we're interested in the 0:38:25.280 --> 0:38:29.239 newest thing, something we've never seen before, Right, But in 0:38:29.280 --> 0:38:31.560 a lot of circumstances, those things aren't the things that 0:38:31.600 --> 0:38:34.520 grab our attention. It's more things that maybe we've seen 0:38:34.520 --> 0:38:38.640 a few times, and you know, we just sort of 0:38:38.680 --> 0:38:41.319 noticed that they're starting to occur. If something happens once, 0:38:41.360 --> 0:38:43.000 you're just say okay, that was weird, and you sort 0:38:43.000 --> 0:38:45.160 of dismiss it. But when something happens a few times 0:38:45.200 --> 0:38:47.279 and it's unfamiliar to you, you say, okay, maybe I 0:38:47.320 --> 0:38:49.520 need to figure that out. Right, And something that happens 0:38:49.520 --> 0:38:51.120 all the time, you're not that curious about. That's just 0:38:51.120 --> 0:38:52.520 the thing that happens all the time. And you can 0:38:52.560 --> 0:38:55.120 explain that by thinking about this sort of derivative, right, 0:38:55.239 --> 0:38:58.920 where if something just happens once, you shouldn't be interested 0:38:58.960 --> 0:39:00.480 in it. Because it just happened once, it's probably never 0:39:00.480 --> 0:39:03.160 going to happen again. If something happens a few times, 0:39:03.719 --> 0:39:06.640 that's a clue that it's probably going to happen again 0:39:06.640 --> 0:39:09.160 in the future, and you've not seen it enough to 0:39:09.200 --> 0:39:11.719 actually know what's going on, And so paying attention to 0:39:11.760 --> 0:39:13.520 that is good in terms of that derivative of your 0:39:13.520 --> 0:39:16.759 future knowledge. And if something happens a lot, then it's 0:39:16.760 --> 0:39:18.680 probably happened enough that you know something about what's going 0:39:18.680 --> 0:39:21.160 on and it's not that interesting, right, And so so 0:39:21.280 --> 0:39:23.120 that sort of sweet spot ends up being around the 0:39:23.120 --> 0:39:25.359 things that are sort of like just happening to your 0:39:25.480 --> 0:39:27.719 enough times that you're starting to realize, oh, this might be. 0:39:27.640 --> 0:39:28.880 A thing that I need to pay attention to. 0:39:42.800 --> 0:39:45.880 If you hadn't been on one capability that's going to 0:39:46.120 --> 0:39:50.319 unlock a broader intelligence, unlock a jump to that. 0:39:50.360 --> 0:39:51.240 What's your candidate? 0:39:52.160 --> 0:39:54.319 I actually think the biggest obstacle at the moment is 0:39:54.400 --> 0:39:59.120 more about generalizability of intelligence rather than any specific capacity, right. 0:39:59.239 --> 0:40:03.920 And so people in the AI world talk about jagged intelligence, right, 0:40:03.920 --> 0:40:06.720 the sort of phenomenon where you have an AI system 0:40:06.760 --> 0:40:08.799 that can do something that's really smart and impress you, 0:40:09.280 --> 0:40:11.480 and then five minutes later does something that's really dumb 0:40:11.520 --> 0:40:13.839 on a problem that's like right next to it, and like, 0:40:13.880 --> 0:40:15.479 if it's able to solve that first problem, it seems 0:40:15.480 --> 0:40:17.160 obvious that it should be able to solve the second problem. 0:40:17.160 --> 0:40:19.120 And you're just like, what happened? You know, why did 0:40:19.120 --> 0:40:19.799 it go wrong there? 0:40:20.200 --> 0:40:24.680 And so that lack of generalizability is also a consequence 0:40:24.680 --> 0:40:27.680 of these kinds of inductive biases, right, So these human 0:40:27.719 --> 0:40:30.719 inductive biases that steer us towards a solution and let 0:40:30.840 --> 0:40:34.640 us learn from limited amounts of data, they constrain the 0:40:34.680 --> 0:40:36.160 kinds of solutions that we find are The kinds of 0:40:36.200 --> 0:40:37.719 solutions that we find are the ones that are sort 0:40:37.719 --> 0:40:40.640 of like generalizable at least to us, right. They are 0:40:40.640 --> 0:40:42.239 things that kind of make sense where if someone's able 0:40:42.280 --> 0:40:43.239 to do one thing, they'll be able to do the 0:40:43.320 --> 0:40:45.799 other thing. And because the AI systems are approaching these 0:40:45.840 --> 0:40:48.440 problems just in a completely different way from a different 0:40:48.440 --> 0:40:50.800 starting point and then getting tons of data that's allowing 0:40:50.800 --> 0:40:52.919 them to sort of approximate what the human solutions are. 0:40:53.120 --> 0:40:55.600 But they're coming at it from another angle. That's the 0:40:55.600 --> 0:40:58.160 thing that makes them jagged. It's not that they don't 0:40:58.160 --> 0:41:00.400 have sort of these same compatible inducted bis is that 0:41:00.400 --> 0:41:03.560 we have that are informed by having evolved in certain 0:41:03.640 --> 0:41:06.279 environments and having had experience of the world, and you know, 0:41:06.440 --> 0:41:08.000 all of these other things that are part of what 0:41:08.040 --> 0:41:11.239 it means to you know, sort of learn anything as 0:41:11.239 --> 0:41:15.080 a human being. And so because they are coming with 0:41:15.160 --> 0:41:17.640 this different set of inductive biases, they're very influenced by 0:41:17.680 --> 0:41:20.560 their training data, they end up doing things that are 0:41:20.560 --> 0:41:24.800 sort of inscrutable to us because they are, you know, yeah, 0:41:24.840 --> 0:41:26.720 like coming at these problems in a way that doesn't 0:41:27.400 --> 0:41:29.759 make sense to us. You know, from the starting point 0:41:29.800 --> 0:41:30.680 that humans come from. 0:41:30.960 --> 0:41:33.960 After writing this book, what do you think we understand 0:41:34.080 --> 0:41:36.960 now about minds that we didn't understand let's say, a 0:41:37.040 --> 0:41:37.879 decade or two ago. 0:41:39.239 --> 0:41:43.480 So it's funny because when I have taught this material 0:41:43.680 --> 0:41:47.680 for you know, twenty twenty years at this point, I 0:41:47.800 --> 0:41:51.200 normally start my cognitive science classes saying, you know, welcome 0:41:51.239 --> 0:41:53.279 to cognitive science. This is going to be different from 0:41:53.320 --> 0:41:55.480 your other science classes. Normally, when you take a science class, 0:41:55.480 --> 0:41:57.160 someone is going to stand up and say, Okay, here's 0:41:57.200 --> 0:41:58.520 all the things that we figured out. Here are the 0:41:58.520 --> 0:42:02.319 answers to the questions. And in cognito science, it's more 0:42:02.320 --> 0:42:04.640 that we figured out how to get better at asking 0:42:04.680 --> 0:42:05.240 the questions. 0:42:05.239 --> 0:42:06.359 We haven't answered them. We don't. 0:42:06.360 --> 0:42:08.200 It's not like you have a consensus across the whole 0:42:08.200 --> 0:42:10.520 field about what those answers look like. And so I 0:42:10.520 --> 0:42:13.040 think that's important that we're still very much And this 0:42:13.120 --> 0:42:14.759 is what got me interested in cognres science in the 0:42:14.760 --> 0:42:17.120 first place. You know, still a field that has deep 0:42:17.160 --> 0:42:20.640 mysteries and lots of opportunities to learn and discover interesting things. 0:42:21.200 --> 0:42:25.560 But I think over the last ten years, like so, 0:42:25.640 --> 0:42:28.560 as I was working on this book, I wrote the 0:42:28.640 --> 0:42:32.919 first chapter, and I had that disclaimer in the first 0:42:32.960 --> 0:42:37.520 chapter and said, okay, look, you know I'm not bromising 0:42:37.560 --> 0:42:40.080 you answers. Well, well, we're going to see if we 0:42:40.120 --> 0:42:42.839 can get a good handle on the questions. But by 0:42:42.880 --> 0:42:44.440 the time I got to the end of the book, 0:42:44.520 --> 0:42:46.680 right after that sort of process of working on it 0:42:46.680 --> 0:42:49.160 for years, I felt like things that actually, you know, 0:42:49.360 --> 0:42:51.480 me going through the process of writing it and exploring 0:42:51.480 --> 0:42:53.120 all these things and thinking about how they fit together, 0:42:53.440 --> 0:42:56.040 but also just where the field was, you know, having 0:42:56.040 --> 0:42:58.399 moved forward, I actually started to feel like, actually, these 0:42:58.400 --> 0:43:00.480 things do fit together in a way where you can 0:43:00.480 --> 0:43:03.040 see the glimpses of what answers are going to look 0:43:03.080 --> 0:43:05.040 like in a way that I think really wasn't there 0:43:05.040 --> 0:43:08.200 ten years ago. And it's that story of Okay, we 0:43:08.239 --> 0:43:11.200 sort of know what the goals are, right, we know 0:43:11.239 --> 0:43:13.880 what the right mathematical systems are for describing what intelligent 0:43:13.880 --> 0:43:16.520 system should be doing. You have these ingredients of symbolic 0:43:16.560 --> 0:43:19.920 systems and probablistic inference, and we've discovered that in fact, 0:43:19.960 --> 0:43:22.760 you can get a remarkable way just using these artificial 0:43:22.760 --> 0:43:25.680 neural networks to learn to approximate those things, and so 0:43:26.560 --> 0:43:29.839 that demonstration I think has shown first of all, that 0:43:30.440 --> 0:43:36.120 language is a extremely good substrate for intelligence right in 0:43:36.160 --> 0:43:38.319 a way that I think people had not anticipated before 0:43:38.400 --> 0:43:41.520 large language models, and that you can make big neural 0:43:41.520 --> 0:43:44.799 networks that can learn to approximate really complex probability distributions. 0:43:45.360 --> 0:43:48.080 And so it gives us some of these ingredients for 0:43:48.360 --> 0:43:52.080 seeing how what originally worth three very different views of 0:43:52.120 --> 0:43:54.719 the mind might start to fit together to make something 0:43:54.760 --> 0:43:56.239 that's a little bit more of a unified hole. 0:43:57.440 --> 0:44:00.839 Excellent, And when you wrote the book what struck You 0:44:00.960 --> 0:44:04.920 is the most beautiful idea in the whole quest, in 0:44:04.960 --> 0:44:07.480 the whole history of this gosh. 0:44:07.560 --> 0:44:10.960 Okay, I mean, I'm a big probability for theory fan, 0:44:11.120 --> 0:44:16.560 so going to you're gonna get me endorsing Bays rule, 0:44:16.560 --> 0:44:19.040 which I really do think is like it's it's when 0:44:19.080 --> 0:44:21.000 you learn it, take it a probability class. It's just 0:44:21.040 --> 0:44:23.359 like it's just a dumb principle of probability theory. But 0:44:23.400 --> 0:44:26.279 when you make this move of saying probability theory isn't 0:44:26.360 --> 0:44:30.360 just about dice and cards, it's about you know, beliefs, 0:44:30.640 --> 0:44:34.440 it suddenly becomes a very deep and insightful sort of principle, 0:44:34.480 --> 0:44:37.640 And in the book, I also show probability theory kind 0:44:37.640 --> 0:44:40.480 of subsumes logic, like everything that's a valid logical inference 0:44:40.520 --> 0:44:43.040 is also a valid inference in probability theory. Probability theory 0:44:43.080 --> 0:44:45.879 just kind of extends the surmountics of logic to these 0:44:45.920 --> 0:44:49.200 cases of uncertainty. So to me, I think that's a 0:44:49.320 --> 0:44:50.840 that's a that's a big one. I kind of like 0:44:50.880 --> 0:44:51.520 that's where I live. 0:44:51.640 --> 0:44:55.080 Yeah, excellent, And and somebody, if we do have a 0:44:55.480 --> 0:44:58.520 mature physics of thought, let's say fifty years from now, 0:44:58.560 --> 0:45:01.400 what is that change from us in terms of education, 0:45:02.040 --> 0:45:04.040 in terms of the way we build machines. 0:45:05.000 --> 0:45:07.640 So I think this is this is exactly where we 0:45:07.680 --> 0:45:11.400 can go, right, which is, once you figure out the 0:45:11.440 --> 0:45:13.880 scientific principles of a domain, you can start to think 0:45:13.920 --> 0:45:16.680 about how to do engineering right. So like you know, 0:45:17.040 --> 0:45:19.680 when you're an engineer and you go to engineering school, 0:45:20.000 --> 0:45:22.480 you take physics, right, and then you learn in your 0:45:22.480 --> 0:45:24.799 physics class what these principles are, and then you take 0:45:24.840 --> 0:45:27.280 your applied engineering classes, which are like taking those physical 0:45:27.320 --> 0:45:29.040 principles and telling you how to build a bridge right 0:45:29.080 --> 0:45:32.239 and explaining that you know not in terms of heuristics 0:45:32.760 --> 0:45:34.520 for what makes a good bridge, but in terms of 0:45:34.600 --> 0:45:37.520 those fundamental physical principles. So I think that's a thing 0:45:37.520 --> 0:45:40.640 that's incredibly exciting here is that as we start to 0:45:40.719 --> 0:45:45.440 converge on what these laws of thought look like, it 0:45:45.480 --> 0:45:48.359 gives us the opportunity to do a much more sort 0:45:48.400 --> 0:45:54.719 of science based form of engineering applied to human cognition, 0:45:55.000 --> 0:45:58.480 thinking about how do we make an optimal you know, 0:45:58.680 --> 0:46:02.239 sort of learning environment, how do we support human decision making. 0:46:02.280 --> 0:46:03.840 That's something that I work on in my lab is like, 0:46:04.239 --> 0:46:08.480 how do we put computation into human environments to overcome 0:46:08.520 --> 0:46:12.440 whatever computational constraints we have as individual decision makers and 0:46:12.480 --> 0:46:17.160 help us make better decisions. And how do we understand, 0:46:17.320 --> 0:46:21.359 you know, the kinds of things that people are doing 0:46:21.360 --> 0:46:23.319 in a way that allows us to then sort of 0:46:23.360 --> 0:46:27.160 like make suggestions about, you know, how they might do 0:46:27.239 --> 0:46:30.160 them better. Right, And so I think there's a there's 0:46:30.160 --> 0:46:32.680 a lot of potential for you know, sort of human 0:46:32.760 --> 0:46:35.440 upside as we start to be able to answer these 0:46:35.440 --> 0:46:36.400 scientific questions. 0:46:41.200 --> 0:46:44.480 That was my interview with Tom Griffith's. To quickly summarize 0:46:44.520 --> 0:46:49.120 his framework, Tom sees three major scientific approaches that all 0:46:49.120 --> 0:46:53.000 try to capture the mind. You've got rules and symbols, 0:46:53.440 --> 0:46:57.240 you've got artificial neural networks, and you've got probability theory. 0:46:57.600 --> 0:47:01.239 These very different approaches, and each which one has delivered 0:47:01.280 --> 0:47:05.759 something a little different. Rules and symbols give us language 0:47:05.920 --> 0:47:10.040 like machinery where pieces can be assembled and reassembled into 0:47:10.080 --> 0:47:15.759 complex ideas. Artificial neural networks they give us graded concepts, 0:47:15.880 --> 0:47:20.160 meaning ideas can be fuzzy, and probability theory gives us 0:47:20.160 --> 0:47:24.160 a language for dealing with uncertainty. Now, what's interesting is 0:47:24.200 --> 0:47:28.000 that human minds seem to traffic in all of these modes. 0:47:28.120 --> 0:47:32.719 We use structured symbols, we also use graded concepts. We 0:47:32.760 --> 0:47:36.480 also revise our belief as new evidence comes in. And 0:47:36.560 --> 0:47:38.400 part of that is that we move through the world 0:47:38.400 --> 0:47:43.040 with prior beliefs shaped by our history, our culture, our language, 0:47:43.040 --> 0:47:46.799 our neighborhood, our moment in time. So none of these 0:47:46.840 --> 0:47:50.719 models by themselves are the final answer. And what this 0:47:50.840 --> 0:47:54.120 means is that, like most scientific stories, this is one 0:47:54.160 --> 0:47:58.880 about humility. Tom's book illustrates how every generation arrives with 0:47:59.280 --> 0:48:03.319 some new formalism, some new piece of math, some new 0:48:03.719 --> 0:48:08.240 model that's powerful enough to illuminate an area of mental 0:48:08.280 --> 0:48:10.880 life and for a moment it feels like, hey, the 0:48:10.880 --> 0:48:15.600 whole mystery is finally collapsing. But then the spotlight widens 0:48:15.680 --> 0:48:19.160 and we see more terrain. So what I love about 0:48:19.200 --> 0:48:21.640 this conversation is that it can leave us with a 0:48:21.840 --> 0:48:25.120 sense of progress and a sense of wonder. At the 0:48:25.160 --> 0:48:29.239 same time, we feel a convergence of different fields, and 0:48:29.320 --> 0:48:34.280 we can also feel how large this subject remains. Cognition 0:48:34.440 --> 0:48:37.560 is still a field in motion, So let's look at 0:48:37.560 --> 0:48:40.799 the big picture. When the field of physics matured, we 0:48:40.880 --> 0:48:45.399 could then build bridges and airplanes and power grids because 0:48:45.440 --> 0:48:50.520 we had firm principles to build on. So once the 0:48:50.640 --> 0:48:55.640 laws of thought come into clearer view, what becomes possible 0:48:55.840 --> 0:49:00.520 for education and for decision making, and for rules that 0:49:00.600 --> 0:49:04.839 help us reason more effectively. So here we are at 0:49:04.880 --> 0:49:08.000 a very cool moment in history where the old dream 0:49:08.640 --> 0:49:13.040 of formalizing thought has escaped to the library and shown 0:49:13.120 --> 0:49:17.439 up in everyone's laptop. The big thinkers of centuries ago 0:49:17.920 --> 0:49:21.239 could sort of squint and see the outline of the project, 0:49:21.840 --> 0:49:24.839 and now we're living much more squarely right in the 0:49:24.880 --> 0:49:27.799 middle of it. If there truly are laws of thought, 0:49:27.840 --> 0:49:30.960 they're going to teach us about our machines, but more importantly, 0:49:30.960 --> 0:49:35.399 they're going to teach us about ourselves, because although it's 0:49:35.400 --> 0:49:39.320 sometimes tempting to view the mind as a ghostly exception 0:49:39.440 --> 0:49:43.640 to the universe, the mind, presumably is part of the universe, 0:49:44.120 --> 0:49:48.800 and it is lawful and wondrous and discoverable, and every 0:49:48.880 --> 0:49:56.840 step towards understanding it enlarges the human story. Go to 0:49:56.880 --> 0:49:59.680 eagleman dot com slash podcast for more information and to 0:49:59.719 --> 0:50:03.520 find further reading. Join the weekly discussions on my substack, 0:50:03.800 --> 0:50:06.800 and check out and subscribe to Inner Cosmos on YouTube 0:50:06.800 --> 0:50:10.160 for videos of each episode and to leave comments until 0:50:10.200 --> 0:50:13.640 next time. I'm David Eagleman, and this is Inner Cosmos.