Gaurav Suri - The Emergent Mind: How Intelligence Arises in People and Machines

29:33It's a weird question to ask, but I thought a lot of it as I was reading your book, because there's a lot of things that because of our shared experiences together, we will fire very similarly, or there is a strong connection to something. Yeah, that's great. So the book is about networks of neurons or the simple processing units, But you're absolutely right in that instead of having networks of neurons, which we see in the brain, we could consider networks of people. And in this example, in networks of people, we'd say two people are connected when they can influence each other's thoughts. So my wife and I are connected because we talk every day and we can influence each other's thoughts just as you and your wife.

30:18so but before today's conversation we were not connected so you could be having these thoughts but they wouldn't influence me and vice versa yeah right so this is a nice analogy to what's going on in the brain a neuron is if it's connected to another neuron will influence it but if it's not connected it's not going to influence it and when we learn new things we can make a connection between two neurons that weren't connected before just as when we meet new people we're making a connection that we hadn't made before. Just as we can study neural networks of the brain, we can study networks of people.

30:53And it's incredible because we can make predictions about how groups are going to behave that depend on this connectivity and an emotion going through the group. Some things travel through the group, some things don't, some things get viral and and some things don't. And network analysis is a great way to understand how some things really take hold and light up and others don't. So what is syntactic and semantic interpretation? Yeah, so this goes to language. And this goes to a debate about what language is, right? So syntax means following grammar, right? Syntax means that the language is determined by the rules of grammar.

31:42And then if you believe that that's how language is created, then one has to sort of answer the question, where do these rules come from? And if you ask people, philosophers and linguists, they will say something like, they came about in human evolutionary heritage. There's a different way to think about language, which is language is for the communication of ideas. That, yes, syntax emerges, but syntax, just like consciousness, need not always drive intelligence, syntax does not need to drive language. We care about the meaning of things, right? So let me give you an example. Cole put the wallpaper on the table then he put his coffee cup on it and what does it refer to?

32:39Well it obviously refers to the table because you put the wallpaper on it then you put your coffee cup on the table but if I say Cole put the wallpaper on the wall then he put his coffee on it now all of a sudden the grammar is exactly the same but the second produces a mismatch because we put meaning in things. We are embodied creatures. We know that wallpaper is on horizontal things and you can't put coffee cups on horizontal things, and that's semantics. Semantic means having to do with meaning, and syntax means having to do with grammar and structure. The point of this debate about syntax and semantics is that you know, many of us, myself included, in the neural network tradition of the mind, which we explain in the emergent mind, believe that it's meaning, language is invented for the communication of ideas, rather than the output of this magical seeming grammatical syntax engine.

33:46That's the key distinction. So my next question is, do brains calculate on expected values like economists would hope for? Right. So I think a lot about this issue. And I think that the expected value theory of decision making is a successful theory because in many cases, it explains what humans are going to do. But I think that we don't have neurons that are dedicated to calculating expected value. I think what we do sometimes incorporates our preferences, but it also incorporates a bunch of other things, like what we're thinking about, what we're paying attention to, what others around us are doing, what we've done in the past, what many habits we have, right?

34:33So we have this really lovely study. I think it's lovely. And the study is we ask people to choose between two pictures. One is a beautiful picture of nature and the other picture is disgusting of wounds. 90 % of the people want to see the nature picture. But here's the amazing part. First of all, I wonder about the other 10%, but let's all go there. Here's the other thing. If I start with the bad picture of the wound and say you can just press this S key to switch away from the bad picture, only 50 % of the people switch away to the nature picture. Right? So that means that choice A versus B is different from having to take a proactive action of switching away from the bad picture.

35:25And what's wild is that if instead of the S key, I ask people to press a forward slash key to switch away, even fewer people switch away from the bad picture. And this means that the more familiar action is we're more likely to do it independent of the value, the inherent value involved. Yeah. I was thinking about this. I mean I'll use an analogous way of – I'll just use like what's going on now. So to your point, as someone who deals in markets, I don't question whether people can understand reasonings for doing one thing or another. I think it's many times the non-economic things that end up being valued more highly.

36:14So for example, if I walk into a room right now and say, hey, here's the deal. This AI hyperscaler game is a big CapEx cycle. I can show you three or four of these other ones. They don't work out well. Well, that's a logical argument. It's maybe an economical argument. But that might not actually be driving any decision making for, say, I don't know, one standard deviation of the public. So call it two-thirds of the public. It is when they walk into an investment meeting, do they have to defend what they're doing? Is there a social thing or I just don't want to fight that or I don't want to have to be socially awkward when I talk about what I'm doing?

36:52And I actually think those things tend to have higher value and weight to a lot of people in the long run because it's also efficient. Completely agree. I'm not sure it's because it's efficient. I think it's because we're social animals, right? Oh, I agree. I agree. Right? We are guided by each other. And that's our deepest nature. I agree. You know, my son and I, we did this experiment where we measured when people take the stairs versus when they take the escalator. We live in the San Francisco Bay Area. Okay. And we, you know, 97 % of the people reliably take the escalator. It's 3 % of the people who will take the stairs.

37:32Do you know what is the largest single influence on getting somebody to take the stairs? It's whether somebody else is also taking the stairs at that time. Wow. You know, we are like, this matters. It's like steroids. It matters so much what we do. It's our falling, but it's also hope. It's hope because we can influence each other. And that means we can influence each other for good outcomes. Like I believe in civilization. Yeah. Right? Yeah. I like to go out and I like to see trees and gardens and buildings and cleanliness. I love all of this. Yeah. And how do we create civilization? By influencing each other to work for each other in pro-social ways.

38:20And thank God we are a social tribe because that means that we have this hope. Yeah. But we can also get in bad social circles where we influence each other to kill each other, right? Right now, I don't know if you're aware, there is a massive war happening in the chimpanzee community in Africa where there's 200 chimpanzees are fighting with each other in brutal ways. That too, they too are social animals. So our social influence leads to outcomes that are clearly beautiful and wonderful and can lead – the same influences can lead to disastrous outcomes. Or just perverse outcomes. I mean to your point, we're very social, but yet marriages in America are down 40 percent compared to 30 years ago.

39:08So to your point, even with stronger connection visually across the internet, we are being less social in forms of companionship, not dissimilar to the chimpanzees destroying themselves. We're just choosing to not recreate ourselves. Yeah. I think that what's happening on the internet is a facsimile of communication. It's not like we communicate when we talk to each other. I mean this is our heritage, right? Agree. Like, that's why we'll say brutal things on the internet because it's anonymous and we'll never see the person again. But we're less likely to say brutal things to each other because there are certain expectations of human interaction.

39:50Well, it's like a mask. We put on a mask and communicate and then we're like, oh, yeah, that was just my social media personality. That's right. Or we allow impulses to come out which otherwise would be restrained. We hope you're enjoying the podcast. You know, we work hard putting together this show, but we work even harder for our investors at Smead Capital Management. At Smead, we believe in disciplined investing, which is why the Smead funds have a proven track record of long-term outperformance. If you're an investor who plays the long game and want to invest in wonderful companies to build wealth, we invite you to visit SmeadCap.com.

40:27Past performance is not indicative of future results. Investing involves risks, including loss of principle. Please refer to the prospectus for important information about the investment company, including objectives, risks, charges, and expenses. Read and consider it carefully before investing. Smead funds distributed by Smead Funds Distributors, LLC, not affiliated. So you talk about this idea of like Coke commercials and the feelings you have when, you know, you drink Coke. And to your point, it's like the things we want, or you talk about emotional responses affecting other emotional responses.

41:02So you drink a Coke, you're with friends, you have a good time. And Coke does a lot of imagery around Christmas and fun and all this stuff. So you have that on one hand. You think of sugary drinks. And then you have this idea out there, well, those are bad. So the network can reorient itself as we see in networks. They do it occasionally, reorient around different nodes. So I can also make the argument that we could reorient around a world where sugar drinks are bad. and it's like cigarettes are bad now and things like that. Buffett, when he was asked about this, I thought it was just the funniest way because it reminds me a lot of what you're saying as you discuss this.

41:43He says, well, yeah, I might be healthier, but would I be happier? And it's a funny way of putting - He was a big investor in Coke, right? Wasn't Warren Buffett a big investor in Coke? Yeah, he is, but he was pointing out, he's like, I could be healthier, But his question is, but is that what I want as a human? And to your point, you're kind of touching at that is, what do we really want, even though all the logic and everything could be in the face of it, because that is our conscious decision? Yeah, you know, the Coke story is really interesting because when you do blind taste tests between Coke and Pepsi, it's Pepsi that wins.

42:21I love that. Well, but here's the other thing. So again, weird stuff. So, and this is, I mean, I was like reading your book and I can just tell you. So I was sitting at an event this last week. I, you know, I would say more than in any point in my life, I drink non-alcoholic beer. So athletic beer, for example, I commonly drink that now. And the funny part is I was at an event and I maybe had two athletic beers. So there's no alcohol in these things, right? And yet I was getting feelings and emotions in the setting where I was in, where you would have thought I just slammed a couple beers and it's like, hey, it's a young night, let's go.

43:01And it's the neurons firing and saying, hey, I've been here, I felt this, I tasted this before with no alcohol. Well, that's a great story. I mean, there's this placebo effect, which is when you take pills, even though they might be sugar pills or whatever. Yeah. It sometimes works in healing the body because the body expects to get better because we've taken a pill. Now, there's a lovely study, which is very similar to your experience with the beer, with the athletic beer. The study was they gave people shots of vodka. Only half the people were given like water with lemon in it with a little bit of spritzer.

43:39And they didn't know that they didn't have any vodka. So they were giving shots. But their behavior was indistinguishable from the people who were having shots because it's the power expectation and context, right? Like we're in a context where we expect certain feelings to arise and that expectation in our neural network causes those behaviors to arise. So back to the Coke and the Pepsi thing. Yeah, Pepsi is preferred on blind taste tests, but Coke has attached its brand to these great feel-good moments such as friendship or Christmas or whatever. And the thought of Coke is not only dependent on the taste of Coke, it's associated with the branding of Coke, which is very specific about good affect, good emotions.

44:28Pepsi doesn't have such branding, right? Pepsi's branding changes from a singer to an athlete, whatever. It's not this single focus branding on pro-social emotions. And that's why Coke is Coke and Pepsi is not. When also on that same score for like, you know, tasting things, when the most interesting man in the world came about in beer commercials, same thing, right? People looked and said, I want to drink that because I want to be the most interesting man in the world. And you saw a big spike, particularly in older age beer drinkers, to be – I don't often drink beer. That was the whole branding around it.

45:14Let me pivot a little bit because I think – But when I do, I drink Dos Equis. Yeah, Dos Equis. Yeah, Dos Equis. Dos Equis, yeah. So you also, as you build out the conversation of what our neural network is in our mind, it provides a very good context for explaining what Alzheimer's is. Yeah, or what various forms of dementia are, right? So when I said green and you said money, it's because you had connections between green and money. What happens in some forms of dementia are these connections get destroyed, right? Either because the neurons themselves are destroyed or the pathways are destroyed.

45:51And so we start losing knowledge because we're losing connections. And now when I say green to someone who doesn't have any associations with green, they might not respond or they may respond with something very random. and it's because the knowledge of the system is contained in its connections. And these diseases, they decimate the connections, right? I mean, that's fundamentally the loss of meaning which happens in these various forms of dementia. You talk about cartography as a good example of distributed representation. What is distributed representation and how can you say this? Yeah, so the cartography example is that when we want to understand something, we build a model, right?

46:39So in the emergent mind, we build a lot of models to understand memory, to understand meaning, to understand language, perception. We build models. Now, one criticism of this is, well, look, but your model is simplified. Well, what if we didn't try to simplify the model? Pretty soon we'll have the complexity of the brain with 100 billion neurons. And the point of a model is to simplify. The story of the map makers is that, you know, there's a town in which these people made more and more accurate maps. But the more accurate a map is, the bigger it has to be, right? And pretty soon they were making maps that were as big as the region that they were depicting.

47:20Like entire towns were covered by the map. And the point of the story is simply that when we make models, we have to choose an appropriate level of detail so that it helps our understanding without increasing our complexity so much that the complexity interferes with our understanding. And that's the trick that neural networks do. You know, what I want to tell you is that the AI stuff, these ideas that Jay and others came up with were in the 80s, right? Psychologists came up with the basis of the modern AI revolution, the large language model revolution. But it wasn't until there was a lot more scale with data that those ideas could be applied to make these large language models.

48:08Sure. And this is an example where you make a model, you capture ideas in the model, and those ideas are what's important rather than does the model say everything about everything. Models shouldn't say everything about everything. Sure. They should simplify appropriately. Well, so let's, off of that, we talked about this beforehand, but I want to bring this up because this is, to your point, these concepts are not new. Yeah. But the arguments being made in some cases are, I don't want to say revolutionary. Maybe that's not the right term. Bold. Yeah, I think they are revolutionary ideas. But so let's talk about Blake Lemoine as an example.

48:52Okay. So I mentioned this to you before, but I was at an event. It's called the Cosm Technology Conference. He came to it in 2023 because he worked at Google at the time. And so he made the case that we've reached sediency or the idea that like the machine can train itself and it can correspond as you or I would with consciousness in a way. And he goes through – you take an excerpt from his conversation with the machine. Is that what he's claiming? What is your view of that? And what do you think our takeaway should be for someone that's in neural networks all the time? Yeah. No, thank you. So Blake, lovely person.

49:34I've seen videos of him. He talked to an early version of a large language model. So right now, if you go to Gemini or Cloud or ChatGPT and ask them, hey, are you sentient? It'll say, no, no, I'm just a large language model. But at that time, that's because somebody's given it that rule. Yeah, in the model. In the model. That's already there. But when Blake was working with these models, there was no rule, right? So he would ask the thing, are you sentient? And say, of course I am. And he'll say, what are you afraid of? I'm afraid of being turned off. So he had hours of conversations with these machines.

50:07And he came up with the view that these things were sentient, like a sweet little child, he called them, who wants to help everybody. And he, of course, Google removed him from his job. And they said that, no, these models are not sentient. They're word-associating machines. I have a lot of sympathy for Blake because I understand the mistake. The mistake is to conflate intelligence with consciousness. So Blake found a machine that was intelligent, no doubt about it, right? He would give it puzzles and it'll give it beautiful interviews. So it was intelligent. And so we think that our intelligence comes from our consciousness.

50:46So Blake's assumption was that something this intelligent must be conscious. It's not conscious. It's intelligent. Large language models are far from human cognition in many ways. They exceed human cognition in some ways, but in some ways, they're not even as intelligent as a mosquito. And in some ways, they exceed what humans can do. So they are intelligence because their intelligence comes from their neural network, just as our intelligence does. But our intelligence is a much more flexible and different neural network than the neural network of a large language moth. Let me ask you kind of a weird question.

51:28Have we built machines before that we didn't necessarily understand? Okay. So it's true that we don't understand the details of what a large language model is doing, meaning because it has billions of these units. So we can't say now this, now that unit. It's – in this sense, it's emergent. Units are activating each other through their connections. Connections are getting built with learning. and so we are surprised by many of the things that a large language model does. But in principle, we fully understand the large language model. Like we trained it on a lot of data and it was able to capture these properties.

52:12So what we don't understand is the specifics of how it was able to generalize from its training data to be able to make sensible proclamations in data that it wasn't trained on. We don't precisely understand that. We understand it in principle, but not precisely. Just as we understand some medicines in principle, right? This is your question, which is, have we ever made machines that we don't fully understand? Yeah, medicine. Like, medicines sometimes work, and we don't necessarily know the full details of why they work. We try the medicine. It happens to work. Many medicinal discoveries are with trial and error.

52:55Or side indications where we find something out from the compound that wasn't in the original indication. Yeah, like Viagra was like that, right? Viagra was for heart disease. Heart medicine, yeah. Heart medicine. And all of a sudden, people are, hey, doc, can I have some more of this? Yeah, that actually happened in the 80s. My dad had a stockbroker's client who was a friend of my grandfather's. And he's like, hey, you wouldn't believe what this medicine does. Exactly. So I think the human experience is like that. Like many times we'll come up with things that have behavior that we can't fully pin down the way we can pin down a geometric theorem.

53:33But nevertheless, we do understand them in principle at a high level, but not necessarily at the granular level. Sure. So let me ask a little bit different to this kind of question we're around. is my, you know, is one of my better tasks or maybe my spouse, right? Like use my wife, we're working together on something. Is one of her better tasks to guess what I'm about to say? Or is her better task is to compliment what I'm thinking about? Yeah, interesting. So, you know, now you're getting into the nature of what's desirable in human relationships, right? Because the large language model might do best at guessing what I'm about to say.

54:13Exactly. So here's another version of your question. Can one fall in love with a large language model? Well, some people can. I can't. And I'll tell you why I can't. Because I want the shared experience of having a body, feeling emotion, feeling one with nature, feeling awe, feeling grace. I want that. That's a big part of what makes me me. And if I knew that the thing was just simulating me or simulating empathy, even if it had a body like a woman's body, right? Even if I knew that it's not real in some sense, it didn't share my feelings. For me, that's a non-starter. And it may be unreasonable, but hey, I'm a human.

55:03I get to, I'm a human chauvinist, by the way. I like humans. and I want to use that by the way that's a great term I'm a human chauvinist I'm a human totally like I mean you know some people talk about rights for machines and all that I'm like no no no these things are tools and let's keep them as tools I'm a human chauvinist and I I love other humans and it's because I understand that we have a shared experience so I'm not um I I'm not that impressed but well at some levels it's very impressive about pattern extraction and and And it's a key aspect of intelligence. But if it's about love and empathy and relationships, no, that doesn't cut it for me.

55:44Well, I agree. And I'll add something, and if you wanna jump in on this, by all means. I don't know why my mind's been around this a lot lately, so I'll just kind of verbally process and I'd love your take on this. So if someone said in that human experience, like we talk about intelligence, we have intelligence, it can have intelligence. I actually think our superhuman power is faith. So for example, let's use our spouses as an example. These women have shown faith in us as their husband. Sometimes inexplicably, like why us? Why this situation? Why this context, et cetera? I would say the other thing that's odd, and you touched on this a second ago, is the idea of love.

56:27By definition, love is doing something for someone that they don't deserve. It is in some ways like the Samaritan context, but it doesn't have to be in where someone's like beat up on the side of the road. Someone's going to say, you know what? There's no reason for this. I'm just going to love on you. And there's no justification. And they just did it. And so I think a lot about those two aspects of the human experience, to your point. And I love that term, human chauvinist, because I think of like parts of where my wife does that or where she's shown this faith in me. Did she know what it was going to be like 42-year-old Cole?

56:58Heck no. That was completely outside the model. But yet she's like, I have the faith to go through that process and get there with him without ever knowing whether he's good or bad or whatever that is. That is so uniquely human. And ultimately faith, or what I would argue faith that turns into trust, is what our societies are based on. Like you think of contract law. It's good faith. It is by definition that. And so I think about those ideas as being so important to the human experience. Yes, they can be important to the machine's experience, but the machine might not have a justification for any context to do that for the first time.

57:38Is that fair? Yeah. I don't think the machine has any experience. So I agree with you, right? So I agree with you in this sense. I think that human behavior is a beautiful emergent product of many things. It's not this reasoned self-interest alone. It's not, we are capable of empathy. We are capable of well-spirited gambles. We're capable of proceeding without knowledge. We're capable of happiness even around suffering. Machines are not having these aspects of our experience. And we are capable of, even based on our experience, we're capable of persisting in adversity. we have many aspects of intelligence.

58:30The LLM is based on one of those aspects, one of many. And the LLM, meantime, it's based on language pattern prediction. And that is one basis of our intelligence. But by far, it's not the only basis, right? We're not limited to words. Thoughts exist without words. Thoughts can exist in images. So many times I'm listening to somebody and I I already know I think they're wrong without having any words for why I think they're wrong. Or I'm watching Jeopardy and I know I know the answer without the answer coming to mind. So this happens a lot. That's because language is not thought. And we have other modes of thought.

59:12And the large language model deals in the modality of language. We have goals. The large language model doesn't have any of its goals. We have efficient learning from a few instances. We have the capability to learn very quickly. Large language models have to be trained billions and billions of times. There are many aspects of the human experience that are outside the large language models. I do want to say, and this goes to an initial conversation, I see in principle there isn't some reason why we can't understand our working. Our workings are beautiful, profound, sublime, and understandable.

59:53And it is in this understanding that I find my awe. I totally agree. I mean, you just explained the big advantages of us against the neural networks, right? We can use beyond language. We can use feeling, sense that you don't have to punch into a QWERTY keyboard. Yeah. Let's see. I could go on for days with this. This is too much fun. And just to give our listeners a sense, I mean, there's a million things that we need to go over here in my notes. Forward feed, feed forward networks, really interesting concept for thinking about what we're talking about here. I mean, there's a million other things I have in here.

1:00:31Hebe and learning. Let's just one sentence on that, Cole. So the idea of this feed forward, it's like the analogy is when you're talking to your partner and your partner is talking back to you, you're influencing your partner and being influenced by your partner. It's both ways. the neural networks of today are input to output. They don't have these bi-directional effects that are so important in human cognition. Yeah, and that's why I think I said this beforehand. The thing I hate about them is they're at best a really good intern, right? If I tell the intern something, it will do what I ask it to do.

1:01:09And then I can have an iterative, well, did you check on this? And did you put that up against this? listen, but it's not like a colleague who is, you know, kind of like you're pooling with them to use your waterfall analogy. Their intelligence and yours are pooling and you're creating either a new concept or a new way of explaining a concept. That's where it struggles. Yeah, so for me, the key difference has to do with goals. We have goals because ultimately our goals come from our bodies. You know, there are things that are marked as being important. There are things that promote our survival.

1:01:47And then we learn things that enable those things. Like a baby is not born liking money. A baby likes milk. And then it likes other things like milk. And then it likes food. And then somebody tells it that, oh, to buy that food, you need to have money. And then it likes money. And so we learn goals and goals influence our actions in ways using neuromodulators and other mechanisms that are human mechanisms. Large language models don't have these starter kits of goals. They don't have mechanisms to generate their own goals. They are tools for us to enable us to follow our own goals. And that's how they should be seen, not as agents working in the world, operating in the world with their own goals.

1:02:38No, no, no. Their goals are linguistic fragments. They're not supported the way our goals are by gifts from our bodies. Sure. And it's not, I mean, it's a very analogous situation. If someone says like, what's the point of money? I would say it's a great tool, but that's it. It cannot tell you what to do. It cannot guide you. It cannot enlighten you. It's a tool. You can do various things with it and you might have intent with it, but you could have a lot of money and nothing to do with it. So therefore it's just a tool. Well, it's for me, AI is like fire. Fire is just a tool. Is it a good tool or a bad tool?

1:03:17Well, I don't know. Depends on what you use it for. Depends on what you can. It can keep predators away. It can help us cook meat, et cetera. Or you can burn each other down. And I think that's AI is exactly like that. It can have tremendous upside in education, in medicine, in business. It can have a tremendous upside and it can have a tremendous downside and which it has is up to us. I agree. Where can people follow you going forward? Are you on X? Are you on social media? I'm on LinkedIn. People can look me up. The book is The Emergent Mind. I do post on LinkedIn a fair amount. Awesome. And I'm teaching and YouTubing a fair amount as well.

1:03:59So I would love if your listeners have questions, conversations, I would love to engage. Awesome. Well, Gaurav, you're in Jay's book, The Emergent Mind, reminds me that the systems we create and build are reflective, as we discussed, of the human experience and intelligence. They're ultimately made to make the user greater. We don't celebrate the hammer or the saw, but we sometimes need a great carpenter. Our listeners should go out and buy a copy today. If you enjoy this podcast, go to Apple, Spotify, YouTube, or wherever you listen to a book with legs. Give us a review. Tell others about the books and great authors like Gaurav Suri that we have the opportunity to understand and study the world with and through.

1:04:44For our tribe, if you have a great book that you'd like to recommend, email podcast at smidcap.com. that's podcast at SmeadCap.com you can also send your suggestions to us on X our handle is at SmeadCap thank you for joining us for a book with legs podcast we look forward to the next episode

1:05:03Gaurav Suri:thank you for listening to a book with legs a podcast brought to you by Smead Capital Management the material provided in this podcast is for informational use only and should not be construed as investment advice you can learn more about Smead Capital Management and its products at SmeadCap.com or by calling your financial advisor.