What is Intelligence? Brains, Evolution, and A.I. with Max Bennett

Max Bennett is the author of 'A Brief History of Intelligence: Evolution, AI, and Five Breakthroughs That Made Our Brains.'  In this episode Andrea Hiott and Max discuss his journey from working in AI technology commercialization to embarking on a five-year passion project that led to writing the book. He explains the motivation behind studying the discrepancies between human brain functionality and AI systems, resulting in several published papers and collaborations with neuroscientists. The discussion delves into intelligence before brains, as Max argues against the notion that intelligence is exclusive to brains, highlighting the evolutionary history and complexity of intelligence across different life forms. They cover various breakthroughs in understanding intelligence, including steering, reinforcement learning, and the emergence of language as a tool for sharing mental simulations. Their conversation touches upon the implications of language for human intelligence, its role in facilitating complex planning and cooperation, and the feedback loop it creates for refining thought processes. It also gets into the potential of AI in continuous learning and the challenges in achieving a system that learns in real time (like the human brain!)

Max is at: https://www.abriefhistoryofintelligen...

00:00 Welcome and Introduction to A Brief History of Intelligence
00:14 The Genesis of a Passion Project: AI, Neuroscience, and a Side Project Turned Book
02:05 Exploring Intelligence Beyond the Brain: From Single Cells to AI
05:48 The Evolutionary Journey: Steering, Movement, and the Dawn of Intelligence
13:24 From Simple Beginnings to Complex Systems: The Roomba's Evolutionary Echo
15:40 Envisioning AI's Potential: Education, Healthcare, and Home Assistance
20:10 The Evolution of Intelligence: From Steering to Reinforcement Learning
28:24 Decoding Brain Algorithms: A New Approach to Understanding Intelligence
28:36 The Complexity of Brain Functions and the Power of Algorithmic Thinking
30:53 Exploring the Neocortex: Unveiling the Mysteries of Brain Structure
33:20 Simulation and Mentalization: The Breakthroughs in Cognitive Mapping
33:59 The Evolution of the Neocortex and Its Impact on Mammalian Intelligence
49:53 The Role of Language in Human Evolution and AI Development
53:55 Continual Learning and the Future of AI: Challenges and Opportunities
5

Transcript

Andrea Hiott: ​[00:00:00] Hello everyone. Welcome to love and philosophy beyond dichotomy, which started as research conversations that I was trying to have across disciplines because I felt there were just so many things I wanted to explore that I was told I shouldn't explore if I was studying neuroscience or if I was studying philosophy or if I had this or that particular passion or love.

So I just decided to open myself up to all of it and have all of these voices gathered together. In one space and try and find the patterns that connect us all. So that's what this is. It's part of my life work and research, but it's also something I just want to put out there and share with you and invite you to comment on and share your side and perspective and position. Uh, today we have [00:01:00] Max Bennett, who's the author of A Brief History of Intelligence, evolution, AI, and the five breakthroughs that made our brain.

And if that sounds like a lot, it is, but Max is so articulate, such a good speaker, as you'll hear. He's just very good at explaining things, but he's also a really amazing writer. He brings together a lot of different information so whether you're a neuroscientist or a philosopher or just someone interested in this, you will find something in this book because he really walks through What a mind is what cognition might be starting with worms with the nematodes and going up through the possibilities of what

artificial intelligence might really mean and if it actually has something to do with this organic Minded intelligence and what he finds is very interesting I'll let you at least listen to the conversation to find out what all that is, [00:02:00] but I do recommend the book and Yeah, most people who listen to Max or read him want more and more of it.

So we did have a second conversation this conversation is already on YouTube and I'll post, I'm going to post a second one with this one too. So after you listen to this conversation, you can also watch it there on the YouTube channel.

 We do get into some more personal things in the second episode. In this one, I was just so excited about some of the ideas relative to steering, as you can imagine, since I'm thinking about navigation and cognition often. so we talk about that and this these wonderful developments as he describes the five breakthroughs. I hope you enjoy this conversation let me know what you think and join us on YouTube too. There's a lot of extra stuff there. It'd be great to see you there and hear your comments, and as everyone knows, I'm a bit of a a recluse relative to social media, but we do have some social media platforms up and, or not platforms, what am I talking [00:03:00] about?

I don't even know the terminology. We do have some accounts out there. Try to find them, it really helps if you like us and review in all of these different spaces.

Even though it's hard for me to ask you to do that, I do want all the guests to be heard and I want them to feel their time, talking and having this conversation is, is worth it. And that they can really share something with a wide audience. So I really would appreciate it. If you go to Apple or Spotify or YouTube or anywhere and give us whatever it is, we need those thumbs up and hearts and stuff like that.

Gosh, anyway, here's Max enjoy. 

 hi, Max. It's so nice to meet you. Thank you for being here today. 

 Thanks so much for having me. 

 So we're going to talk about your book, A Brief History of Intelligence, Evolution, AI, and Five Breakthroughs That Made Our Brains. Um, this seemed like it was a kind of passion project almost like going back to [00:04:00] school, for something different.

Andrea Hiott: Is that kind of the right approach? Because you started or you are working in AI. Maybe you can give me a little intro into how this happened. 

Max Bennett: So I, um, I've spent most of my career. in commercializing AI technologies. So taking sort of, established AI tools and bringing them to industry. Um, so my last company, BlueCore we helped really large brands and e commerce companies implement AI to help personalize websites, make, marketing communications more, personalized.

There's a bunch of sort of segmentation tooling we built. And so the, the first. Interaction I had with the wandering about neuroscience and its relationship to AI was when trying to bring these AI systems to real world use cases and seeing the many sort of perplexing places that they keep falling short.

Um, and then all of, you know, this occurred before the [00:05:00] big craze of large language models, which, which obviously adds a lot more intrigue to that whole question, but I started this process of just trying to understand some of the Just for my own edification. What's the discrepancy between how the human brain works and how AI systems work.

And that sort of just unraveled into this, you know, five year long passion projects that, you know, led to several published papers and collaborating with neuroscientists that became mentors of mine and eventually culminated in the book, but yes, it was it started as just a side project. 

Andrea Hiott: Yeah, well it's a wonderful book and I really respect that you actually went through this whole process of publishing papers on this and refining it.

Um, you can really tell in the book, it's really tight. You can tell you've been getting a lot of comments and yeah, so it's, it's a wonderful resource. So I'm going to start with kind of the annoying questions first. Um, More just kind of provocative, but, I wonder if you see intelligence as beginning with brains, because, you know, we're focused on the brain here, and, the first breakthrough is steering, which is where [00:06:00] we're going to start, but I just wonder, at first, like, are we talking about intelligence beginning with brains, um, and, of course, you know, what a brain is, is a very blurry concept, even, uh, in the book, we're not talking about human brains, of course, so,, I just want to throw that out there at first.

Max Bennett: No, it's a great question. I think undeniably intelligence has existed way before brains. Um, you know, you can see by any definition of intelligence, whether it's, you know, solving problems to achieve a goal or learning from experience both of those features of intelligence. an entity's behavior eminently applied to a single cell.

Um, cells have ways in which they learn from experience and cells surely show the ability to react to their environment to solve problems. So it's by no means the case that intelligence began with brains. But obviously there is a unique interest we take with the intelligence instantiated in brains, um, because it seems to be that the [00:07:00] most sophisticated or at least, you know, almost tautologically human like, obviously the human like intelligences that we're interested in recreating are instantiated in brains.

Um, so it is, it is an exploration in the history of intelligence. As implemented in brains, but the first chapter does go into sort of the history of intelligence prior to brains, um, because some of the key ways in which life itself works does set the stage for the evolutionary pressures that drove the evolution of the first brain and why the first brain emerged and what that brain did.

Um, but it's absolutely the case that brains are not the only way in which you can implement intelligence. 

Andrea Hiott: Yeah. And you do a wonderful way of letting the brain emerge from that. I was also just kind of wondering in terms of your interest in AI, if it sort of starts with studying the brain it's kind of interesting to think, Of artificial intelligence, not only connected to the brain.

Max Bennett: Yeah, there's, there's a big schism that has existed between sort of, naturalists and the technologists at multiple times throughout the [00:08:00] history of innovation, which hinges on this question, how much of innovation should, or has been inspired by nature. And of course, in all of these types of debates, the right answer is, in the middle.

Um, both sides are right. So on one hand, there's people that think in the right way to, um, build really successful technologies in many cases is to learn from nature. Um, so, you know, we knew that flight was possible because we looked at birds and the structure of a bird wings gives us lots of insights about how flight works.

Um, but the counter by sort of the, the anti naturalists would say, sure, but Planes don't fly by flapping their wings. Um, so clearly we didn't directly copy nature when we build planes And so of course the answer is a middle ground which is nature can give us existence proofs of certain abilities And nature can give us way points for the fundamental features that we might want to use Um, but that doesn't mean our mission should be to wholesale copy what [00:09:00] nature, um, has created.

And I think that also applies to neuroscience, which is it's not to say that our goal is to recapitulate a human brain, you know, Adam for Adam. Um, but it is to say that I think there are insights left. still uncovered in how brains work that AI can still learn from. Um, and that is an open debate.

There are people that would argue against that, but I, I believe that there's still lots to learn from neuroscience. 

Andrea Hiott: Yeah, maybe we can move into that a little bit in a different way. But, um, so the first breakthrough is steering, which I just love so much that you start there and it's sort of, I'm not sure if you'd agree with me, but seems to orient through the remaining breakthroughs.

Um, could maybe talk about that a bit, but so we have movement and navigation and then it's this introduction of, of steering that we started to talk about, um, intelligence in a certain way. So would you like to kind of give a little. Summary of that, but I want to move into the AI too and I want to talk about the symbolic [00:10:00] AI Versus the behavioral AI sort of schism.

Max Bennett: Yeah, so animals before Our ancestors that had brains. So there were animals that existed prior to the first animals with brains. Um, and they were probably, we don't know for sure, but probably most akin to today's sort of sea anemones and coral, um, which are these radially symmetric. Jellyfish like creatures that don't have brains, but they have a nerve net.

Um, so they have sort of a web of distributed reflexes and a key feature that grouping of animals called Cnidarians is that they don't really find food by navigating towards food. There are a few exceptions that probably independently evolve, but if we use model organisms for probably our ancestors back then, it seems likely to be the case that they didn't navigate towards food or navigate away from danger.

Um, there, Their strategy was sitting in place [00:11:00] and filter feeding, so waiting for food particles to pass through them and they would sort of grasp the food. And when we go from that ancestor around 650 million years ago to the ancestor that had the first brain, We don't only see the emergence of the first brain, but the emergence with a whole variety of, of body changes that seem very well aligned with movement.

So the very first bilaterians, as they're called, because they don't have radial symmetry around a central axis, but bilateral symmetry across the central plane, like humans do, all mammals do, all vertebrates do, or most do. Um, and so, The simple version of bilaterians, which are a model organism of the first bilaterians, are like nematodes, small worms like flatworms, etc.

And these creatures have a centralized cluster of neurons in their head. And what, what these animals notably do relative to their radially symmetric cousins is that they're constantly moving around. [00:12:00] Um, and what's perplexing about these creatures is although they're moving around and they successfully get to food and avoid dangerous areas, their sensory organs are woefully simple.

I mean, if you look at a nematode, they have no eyes. They can't render image of the external world. They have a handful of sensory neurons around its head that just activate in the presence of certain things. So one neuron for detecting salt, another for detecting food smell, other types of food smells, etc.

But if you look at the algorithm of how it finds food, it's actually a really simple one, something called taxis navigation. Where, um, it takes advantage of a simple property of the way that smells diffuse in water, which is if you have a food source in one location, chemicals will plume from that food source, such that the closer you get to the food source, the higher the concentration of these chemicals.

And so, What evolution stumbled upon is you don't actually need to see anything in the world to find [00:13:00] food. You can just have a very basic algorithm of keep going forward if the concentration of food smells are increasing, and turn randomly if the concentration of food smells decrease. And by implementing that algorithm, you can get to food.

And so if we look at simple bilaterians like nematodes, their whole brain By and large seems to be effectively implementing this taxis navigation algorithm and the key feature of that is that it categorizes things in the world into good and bad. So it has hardwired neurons for detecting a good thing, which triggers forward movement and hardwired neurons for detecting quote unquote bad things, which triggers turning.

And, and, you know, one reason why it's so interesting that we started so simple is because evolution imposes this key constraints, which is you cannot redesign things from scratch. So the reason why the whole motivation for taking evolutionary approach to understanding the brain. Is that, because evolution is constrained, understanding the prior structure of things gives us [00:14:00] information about the subsequent structures, because it has to build on itself.

It reminds me of the 

Andrea Hiott: plane. You mentioned the planes earlier, and I think, yeah, um, yeah, you tell this great story. Antidote or , I'm not sure what you would call it in the book about what if people from the 1890s came to the 1990s thinking about the plane. So Rodney 

Max Bennett: Brooks had this great quote, which I have in the book and he's Rodney Brooks, MIT is a big, and the founder of iRobot, he's a big proponent of, you know, behavioral AI, which is, you know, really trying to start from very simple structures and build our way up.

And so he has this quote, I'm going to botch the exact wording of it, but the idea is if you took someone from the 1800s and you wanted them to understand how to create artificial flight and you put them in a time machine and, and let them play with a plane. For an hour. Um, would they actually come back identifying the correct features of the plane to recapitulate?

And so he [00:15:00] argued that they almost certainly would not. So they would be like, Oh, look at these seats and the plastics. Um, that's what we need to invent for flight. Not realizing that fundamentally all they needed was a certain structure. and surface area of a wing. And so I think one could argue with, um, how true our inability to decipher the key features of a plane.

But the essential point he's making is if you start by looking at something really complicated, like the human brain, you're at high risk of Misinterpreting features of that complexity that emerged, versus if you start simple and you have end to end systems that function like a paper kite or plane, and then you start adding to it over time, eventually it makes it easier for one to understand the complex thing at the end of that journey.

And that's sort of one key motivation for why I think these evolutionary stories that I tell in the brain can be really insightful for understanding the human brain. 

Andrea Hiott: Yeah. It feels very close to the approach of your book in a way. Yeah. [00:16:00] Taking that approach. It feels like a good kind of illustration of that.

And also we were talking about steering. So, um, as I understand it, you know, like, let's think about that in terms of AI. So this is going back to the beginning of what we can think of as intelligence and it's a beautiful story you tell in the book and the way you lay it out. And that we can almost think of, that we can think of steering as valence or, or almost zero ones or something it's really building a kind of algorithm for, um, how to survive in a way through this steering and movement.

 and then when we think about AI you talk about the Jetsons, which is interesting. There's this kind of way in which you're framing, or humans have thought of AI as. Um, wanting to create something that will be like us, or take care of us, so, if, if you start with like trying to create the brain, that would be like trying to create the jet plane in a way, um, but instead someone like Rodney Brooks decided to start more with what, where you started, with like the steering.

So maybe you could tell that story of how that kind of turned out in terms of his, [00:17:00] for, well, the first commercially viable robot, I guess. 

Max Bennett: Yeah. So Rodney Brooks, was the inventor of the Roomba. And that was the real first successfully successful robot that we see in our homes and what is analogous.

Between the Roomba and the very first bilaterians is its algorithm for moving around the world. So, while other sort of roboticists and still today we're working on these humanoid like robots that have fine motor skills and can walk around and have maps of our homes and eventually we'll, we'll build that.

Rodney Brooks decided to start with something the same way that evolution did, which is what if we create a robot with almost no sensory organs whatsoever? Um, and so it doesn't create. A rich map of the world. Um, but if we implement a really basic algorithm, which is pretty much exactly what taxes is, it just moves around randomly until it hits a wall and keeps moving around.

Um, we can actually achieve the goal we want, which is it will eventually get to all [00:18:00] parts of the floor and clean things up. And so by simplifying the problem the same way evolution did, you know, he found something that works quite well. Of course, he found an application of it, um, which is vacuum cleaning, where you can have that type of simplicity.

If we want robots to wash our dishes, you know, that's insufficient. We need, we need more complex things, but, but yeah, I think there, there's 

Andrea Hiott: first step towards Rosie, I think Rosie, the Jetsons robot, which I think if anyone knows who that is, remembers that or has seen the Jetsons. 

Max Bennett: Exactly. Yeah. So it's the first, I think there's something, um, satisfying about the fact that the sort of first commercially successful robot step in that direction had a lot of common features with the very first step in the evolution of the human brain.

Andrea Hiott: Did you have a Roomba by any chance? 

Max Bennett: I actually don't have a Roomba, but I have lots of friends with Roombas. 

Andrea Hiott: Yeah, I've seen them. Um, just kind of a side note, but this idea of the robot or the AI, a lot of the examples I've [00:19:00] seen have to do with taking care of humans or, or something like that.

 Do, do you think, I mean, you say this the first step toward Rosie, but is that what you meant in a way?

Um, those are the kind of robots or AI that we haven't built yet. So I just, I just wanted to see what you thought about that if that's the kind of AI you would like to see, a Rosie or a doctor or a caretaker for the elderly or something like that. 

Max Bennett: Yeah, I think there's, there are certain applications of, so AI is an interesting technology because unlike a lot of technologies, it's not clearly universally good.

So you know, if someone invented a cure for cancer, I think most would agree that's universally a good thing. It might create secondary effects like overpopulation, et cetera, et cetera. Yeah. But the proximate effect of that technology, everyone can universally say is good. Um, you know, if we invented fusion.

Cold fusion, you know, that might have secondary effects. We, we don't like, but the proximate effect seems clearly good, which is we now have more access to energy. We can democratize that, et cetera. [00:20:00] AI is more akin to technologies like, you know, nuclear energy where it can also be applied in dangerous ways, but it can also give us access to, you know, cleaner energy, depending on your.

Proclivities on that. So I like grounding the discussion AI not only in the scary things it can do, which I think is worth a discussion, but also what are the most like amazing things that an AI system could do that adhere to human values? And I think there's a few that stand out. One education is an example where, um, I think there can be some really magical applications for people having, you know, tutors that they can always ask questions to that help them go at their own pace and their own learning style.

I mean, there's so many. Students that get left behind because we can't in a, even a classroom of 20 people, you can't teach the same way to everyone. Everyone has different learning styles. So, um, [00:21:00] either we're, you know, we have to choose a style and then the people that don't have that style will get, you know, a worse experience.

So having this one on one coaching that people have a trusted advisor that they don't feel embarrassed to ask questions to that learns their style and teaches them over time, I think can be just a real, amazing thing for people learning. 

Andrea Hiott: Yeah, I find that so exciting. You know, the Consilience Project and um, 

Max Bennett: I don't know that.

Andrea Hiott: Okay. Well, they've written about that, but actually I wanted to ask you, were you homeschooled or did you have a tutor of some kind or something? I'm just, 

Max Bennett: I was, I was not homeschooled, but I, I grew up with a working single mom. So I had a lot of time by myself reading books. 

Andrea Hiott: Cause your, your mind seems to work in a, I mean, you definitely can deal with a whole lot of information and present it clearly.

So I was just wondering. Um, but yeah, so those are the kind of, no homeschooling, but yeah, 

Max Bennett: that's, that is an example of an AI. Another good one to add is, um, medical use cases. I mean, you know, I think if you live in certain areas [00:22:00] of the world one might like, you know, New York or places and another ones, you probably have access to really good, you know, world class healthcare, but in a lot of places in the world, if you wanted to get access to, you know, in the U S the best cancer place, or one of them is called Sloan Kettering.

You know, you would have, it would be very difficult for you to get in front of a Sloan Kettering cancer doctor. And if you did, you'd get maybe 15 minutes of their time. And so the prospect of building an AI system to democratize access to high quality medical information and advice to for largely free to anyone in the world is like, what an amazing possibility that could do for people that just don't have access to that type of information.

So I think, and then another one would be what you were getting at originally where. Um, you know, help around the home. I think there's some cases of this, which is more adhering to human laziness, like the value of a remote so I can push buttons and not get up and change things on my TV. Maybe someone could have 

Andrea Hiott: helped [00:23:00] your mom deal with, I don't know, how many children did she have?

Only one. Okay. Only one. Deal with you. And all the housework and everything else as a single mom, maybe. 

Max Bennett: Right. Exactly. Yes. Um, so those types of things could be great, but also helping elderly, I think is an amazing you know, having someone that's around I mean, there's a huge loneliness crisis and I think.

You want it. One wants to be careful that we don't use robots to replace human interactions. Um, but that doesn't mean, you know, having someone around that is human like to support people could do a lot of good. Um, so I think those are three areas where I think AI can really adhere and align to human values.

So when thinking about the future we're trying to create and working backwards from, and I think it's useful to ground ourselves in use cases like that. 

Andrea Hiott: Yeah, it just, it reminded me reading your book. It feels like we've gotten away from thinking of it like that in a way with all the LLMs and stuff that we'll get into in a minute, but okay, just to go back, so there was [00:24:00] steering and then we go into reinforcement, which I guess is, I was thinking it's more like instead of parts, we're now, Thinking of patterns in terms of intelligence, is that a fair way to start to summarize what's happening in this area?

So a 

Max Bennett: key, so a key change that occurs between the first bilaterians and the first vertebrates, um, which is where you see reinforcement learning. Emerge is that early BS detected things in the world with individual neurons. So they did not detect things as patterns. So for example, if you look at an nematode, the way it learns to navigate towards salt is there is a individual neuron that's responsive to that, that triggers forward momentum in the presence of salt.

Um, it is not that. It detects a molecule in the air that activates a specific pattern of five of 300 neurons, and it learns that this smell equals these [00:25:00] specific five neurons of 100. So objects in the mind of a nematode is directly mapped to individual neurons. In a vertebrate, you see the emergence of A structure called the cortex, not the neocortex, which we see in, in mammals, but something just called the cortex.

Andrea Hiott: Sorry, I have to stop you. Objects in the mind of the nematode. It's not even that they have a mind. It's like direct, right? It's like the, the neuron is, there's no sort of cognitive map, spatial map kind of thing going on. 

Max Bennett: Correct. Yes. So if one, yes, yes. So if one thinks about a mind as this type of cognitive map, then yes, then a nematode would have no cognition level.

It would, it's just a direct mapping between a stimulus and a reflexive forward or turning response. 

Andrea Hiott: I would still think of it as cognition, but just. Yeah, it's not, it's not working with looking at its own pattern, so to speak, but go ahead. 

Max Bennett: So, um, in a vertebrate, you see the structure called the cortex, which can learn about objects as patterns of neural activations.

So what that [00:26:00] means is. And we do this all the time with visual objects, so I recognize a letter, not because it activates a single neuron that maps to the letter A, but because my cortex is deciphering a pattern of activations in my retina as being the object A. And so that Ability of pattern recognition.

We see in fish and reptiles. Um, and so we're very confident in early vertebrates. That type of ability existed. Um, so I do think there, as you said, there's a really interesting sort of dividing line between recognizing things. They're single neurons. Um, and then recognizing things as patterns and there's this huge dimensionality expansion.

If you can recognize things within a pattern of neurons, um, just as a, to sort of quantify that a little bit, if you have 20 different individual neurons in a nematode, it's only capable of recognizing 20 different things. I'm slightly oversimplifying because what actually happened in evolution, which [00:27:00] is interesting is these neurons in nematodes over 600 million years of additional evolution.

Um, Just became multimodal neurons. So there is not one neuron that activates for salts. The salt neuron actually has all of these crazy new types of protein receptors for just, so it just activates for lots of other good things. So one neuron doesn't mean one thing the way evolution empowered nematodes to recognize different stuff is their, their neurons became really multipurpose, but in vertebrates, we have, we don't have that we have neurons in our eyes neurons in our ears neurons in our nose that are special purposed, um, for detecting certain kinds of things.

And then deep in our brain, we decipher the pattern of activations to mean. to mean things. Um, so yeah, there's a really cool dividing line, but invertebrates, it's not just pattern recognition because we also see the ability to learn arbitrary actions. Sequences of arbitrary actions in response to patterns, which is something that we don't see in nematodes.

Um, and that builds on top of this foundation of [00:28:00] valence or what people in AI would call reward. That's the word that AI researchers prefer, um, where there's other structures in a vertebrate brain that learns what previous behaviors did I take that led to me getting this reward? Um, and then how do I make sure I take those types of behaviors again?

Andrea Hiott: Which you see as sort of the main thing that we, or that life got from steering in a way, right? 

Max Bennett: Yeah, I think, um, in AI there's this big Open question called the reward hypothesis, which is how much of intellectual behavior can be understood as optimizing for a reward function. And the story told in the book would suggest from an evolutionary perspective, intelligence didn't start at least brain based intelligence did start with something akin to reward, which is like valence.

Um, however, Um, as we'll get into later breakthroughs. I don't think it means that all human behavior is derived from reward just because the evolutionary origin is reward. [00:29:00] And there's a distinction we can, we can draw there. But 

Andrea Hiott: yeah, I think that's why steering works better because that. I mean, I know you're saying it is kind of linked to reward in the beginning, but I think the overall idea of what you're saying can be generalized.

Max Bennett: Yeah. I think what's cool too. Yeah. A hundred percent. What's cool on that also is in AI, when we think about reward, we think about this as this like scalar function of just like a number, um, but in, um, In evolution, what we see, the reason why I don't, I agree with you, I don't think breakthrough one should be called just reward is there's no, there's no distinction between reward and the action because it's all just one.

So it's just, fundamentally what's happening is the system's categorizing things and the good or bad for the sole purpose of making a reflexive choice as to whether to go forward or back, it's not doing it to induce some reward function for any other purpose. It's only with vertebrates that we see that sort of be folded inward into the [00:30:00] hypothalamus, which then becomes this sort of brain structure and vertebrates that does measure the valence of things in the world and then is used by these other structures, um, to learn in a way that is akin to what AI researchers refer to as reward.

Andrea Hiott: I think this is also another thing I really like about the book is I, I would say you're talking mostly from like MARS level two or from algorithms and, but you're showing that they scale and nest together in all these really interesting ways. I don't, I think you would agree with this, but you can also correct me, feel free, but it, it feels like, um, That's happening with all of this.

Like we can, we're seeing how things scale algorithmically, which means in a way that parts and the details can change a bit, but these overall patterns stay the same, kind of like how you just you described the nematode neurons as being almost like multi dimensional in their way of being, but because I kind of see something like a cognitive map or a spatial map [00:31:00] or these collections of neurons doing that too, similar algorithm, but as kind of now groups of neurons.

So I don't know what you think about that, but am I right? Are you, I mean, of course I know you're mostly working with this algorithmic level, but is it, do you see a kind of scaling going on? Okay. 

Max Bennett: Yeah, I think that's spot on. I, um, I actually in the talks I've been doing recently, I start with talking about the motivation of the book is Mars level two, which is I'm so, so you're spot on.

And I think in reflection, I think it would have, it's a little technical because a lot of, you know, average people don't know what David Marr three levels are. But I think, um, It is important to grounds sort of the motivation of the book is I'm not as interested in the implementation details of exactly how biophysically these things work.

And I'm less interested in sort of the computational level of like spatial mapping is associated with this region of the brain, but I want to know, like, what are the underlying [00:32:00] algorithms being implemented? Um, and I think that. Has two benefits. One benefit is perhaps more obvious, which is if we know those algorithms, it makes it easier to port over to Silicon and then try to re implement them into computers, but then too, I think it's actually a much more useful.

And, um, fruitful way to reverse engineer how the brain works at the algorithmic level, because I think there's been so many, as we've seen in the last few decades, challenges trying to, you know, assign functions to brain regions is like a, you know, a dangerous business because what we've, the more data that we see, the more we realize that functions aren't associated with individual brain regions, they're distributed.

And so I think thinking about things from an algorithmic level, um, actually gives us. It makes it easier to understand what's happening in the brain. 

Andrea Hiott: Exactly. And that speaks to why this isn't a functional decomposition or that doesn't really work. Um, but you can still look at the parts without decomposing them or even composing them.

It's [00:33:00] just, it's a scaling, but I feel like that algorithmic way that you're using algorithm isn't some, some people who. are maybe listening to this, think that means a really, it really is really technical and math based. And it is. But it can also, I think, at least in my reading of this or my understanding of intelligence, we can think of it, um, not only as something extremely static and that doesn't move, that it is also like a kind of recognition of, of patterns.

You don't have to use exactly the same math to use the same algorithm maybe. What do you think about that? Yeah. 

Max Bennett: Yeah. I mean, I think the word, I mean, the, the. What I Mars people interpret Mars is three levels slightly differently. But what I mean by algorithm is not that we need the explicit sort of, um, sort of pseudo code of what's happening.

But I think if we understand conceptually what is being performed, what informational process is being performed to go from sensory [00:34:00] input into An entity like a human to the motor output we see, um, that will give us a much richer understanding of how the brain works and what features want to apply an AI.

And I think what I'm advocating for, what I think level two would advocate for is going. deeper into what's happening than simply saying that, you know, planning is implemented in frontal cortex and visual perceptions implemented in sensory cortex. I'll give one example to motivate this, which I think is the standard one, but is the one that originally got me excited about Mars level two.

Um, and maybe some of your listeners are familiar with this, but it used to be the case that we would look at the neocortex, which is all of the folds of the human brain. Um, so when you look at a human brain, most of what you see is neocortex, which is the foldy stuff that surrounds it. And based on lesion studies and damage studies, it seems like, Different regions of neocortex do lots of different things.

So the back of your neocortex seems to be where we [00:35:00] recognize visual objects. Cause if it gets damaged, people would come blind on the sides. We have auditory cortex gets damaged. You don't recognize speech. There's regions for language. There's a ring around the top for. motor planning, where if it gets damaged, you become paralyzed, et cetera, et cetera.

So through a lens of Mars Level 3, so computation or function, we would say, cool, um, it looks like motor skills live here, and visual perception lives there, and we sort of assign these functions around. But then we run into a problem. Where we realize in reality, language doesn't just live in the neocortex.

It also is implemented in deeper structures, working in interplay with each other. Um, and then we, we realized that visual object perception doesn't only happen in the visual cortex happens in other places like the amygdala. So, okay. So that complicates things, but then we also see an opportunity for. An amazing unification.

When we look into. how the neocortex seems to be functioning, we realize the neocortex looks the same everywhere. [00:36:00] And so that gives us this incredible constraint of how can the same structure be doing visual object detection in one area and motor planning in another area. And so if we go one level deeper, we can start saying, well, there seems to be some underlying Conceptual quote unquote algorithm being implemented.

That's capable of doing these different functions. And so by going one level deeper, um, in some ways, it actually simplifies things. What was originally 20 different functions may in fact to be only a handful of algorithms. And I think that's part of what. is tantalizing about trying to understand what algorithms are.

I find it really 

Andrea Hiott: exciting. And I find also, I mean, we won't get into it, but it could work about trying to understand different kinds of brains or different kinds of species, because you can also understand that without needing to be stuck with these very specific parts, which we already do in a way you do it in your book, like showing kind hippocampus would be and certain other creatures.

But I think [00:37:00] this makes it, you know, a bit easier. But since I mentioned the hippocampus, um, I'm really interested in when you got interested in the hippocampus, when you started thinking about Tolman, now we can move towards simulation and mentalization, the third and fourth breakthroughs. Um, this is very fascinating to me, something I think about a lot.

So yeah, I'd like to kind of lead into that and how So we had this kind of, um, movement steering through the world, then this way patterns are being recognized and created and generated. And now we start to understand how those, um, become patterns that can be recalled in terms of something like memory or imagination or something.

Max Bennett: Yeah. So with the emergence of the first mammals. We see, which is, you know, depending on where one clocks that geologically approximately around 150 million years ago, we see the emergence of a new brain structure called the neocortex and what's. Really interesting, actually, is how similar a mammal [00:38:00] brain is to a fish brain.

Um, they have all the very basic structures there. The fundamental distinction is you see this new structure called the neocortex grow out of this sort of older cortex. And, Um, the hippocampus, the amygdala and the olfactory cortex within mammals actually looks very similar to the overall cortex of non mammalian vertebrates.

So what it suggests is sort of the hippocampus amygdala or the what's called the cortical amygdala. Um, and the olfactory cortex are sort of evolutionary remnants of the original cortex. And then in between, nested in between them, we have this neocortex, which does some magical stuff as an aside. And you can see this just under our microscope because the older cortex has three layers and the neocortex has six.

So, the hippocampus has three layers. Olfactory cortex has three layers. You get the idea. So there's this big open question of like, why did the neocortex evolve to which there's no. Um, agreed upon answer, but [00:39:00] I think the evolutionary story suggests something that is not the consensus. So I think most people, when they try to understand the function of the neocortex focus on its ability to sort of recognize objects, because that's the best understood feature of the neocortex and mammals.

We have a part of the brain, as we said, that can recognize visual objects, sounds, et cetera. But what's odd about that interpretation is it's not clear at all that. Simpler mammals are really at all better in a meaningful way than non mammalian vertebrates at object recognition. Um, fish can readily recognize human faces, even when rotated in 3d space.

They can do this in one shot. They rec, they can maintain those memories for long periods of time. 

Andrea Hiott: I can't let you just go over that. It's so cool. I mean, I love that you use fish having studied neuroscience. I don't think I ever read a fish study and you have so many, and it's so cool. That one or. So the fish can actually recognize a photo.

That's, it's wonderful. 

Max Bennett: Yeah. The it's, you're absolutely right. There's [00:40:00] such a, under fish are really under research or really non mammalian vertebrates are from a neuroscience perspective, really under research. Primarily because a lot of the research funding goes to brain research is motivated by, um, Um, sort of curing neurological disease in, in, in humans.

And so we are, you know, want brains that are as similar to us as possible. And so, so much of it goes into rodents. But for, for the purposes of algorithmically trying to reverse engineer how our brains work. Um, I think non mammalian vertebrates are perhaps should be of equal importance because in that we can reverse engineer a more primitive foundational sort of template for our ancestral brains.

Andrea Hiott: Which you were about to demonstrate with this whole idea of simulation, but. 

Max Bennett: Yeah. So, um, when we try to understand the algorithm being implemented in the neocortex, um, we To to pair it with this open question of why did the neocortex evolve doesn't seem it's it's hard to motivate an argument that it [00:41:00] was for better object recognition, but there's been a ton of research into the underlying algorithm.

And again, there's no consensus, but a lot of sort of neuroscientists have rallied around this idea of predictive coding or generative modeling, which is this notion, which we do see an AI now. Where the neocortex seems to be self supervised, meaning it's trying to predict its own input and train itself over time to be better and better at predicting its own input by virtue of building its own model.

of the external world. And so there's a lot of people that write about this using different language, but it's largely the same concept, which is you know, the, the neocortex builds a sort of internal model of what's happening and it uses that to predict what's going to happen next. And so most people think about this internal model for the purposes of recognizing things.

So I can real time recognize an object when rotated because my neocortex has a model of that object but there's another feature of generative models that are less appreciated, [00:42:00] um, which I argue is the real reason the neocortex evolved, which is in order for a generative model to be self supervised, it needs to be generative, by which I mean it needs to be able to simulate or generate its own data, because what it, the way it learns is it tries to predict the data and compares its predictions against its own input.

And so, a key feature of that is you're able to turn off the input and start exploring an internal world without the input actually occurring. And so, one part of that is, of course, recognizing things, um, but another part of that is being able to simulate the world as it is not. And so, when we look at the sort of comparative psychology of non mammalian vertebrates and mammals, um, with the exception of birds, which very clearly independently evolved simulation we see a pretty strong set of evidence that mammals can do this type of imagining the future, re rendering the past, imagining other alternatives.

And we do not see evidence that fish and turtles and other types [00:43:00] of reptiles, again, with the exception of birds can do this. And so, when we go back and look at the, sort of, neocortex, I think there's a lot of really good evidence that really what this enabled our ancestors to do is imagine the world as it's not.

And this, of course, gives us episodic memories. It enables us to plan our, even our, Our hand movements. One reason why, if you just look at a squirrel and compare it to a lizard walking around a tree, I mean, it's we don't, it's so, um, we don't think about it because we're used to seeing squirrels jump around from tree branches and used to seeing slow moving lizards.

But there's a very clear intellectual capacity difference that we're observing, which is the fact that a squirrel can look up at a bunch of branches and without any practice. Identify exactly how to place its hands to run across. It's just, you know, it's amazing. Like the, we, the roboticists desperately want to have something like that.

Andrea Hiott: Yeah. When you really start looking at the way life is, it's can be, yeah. It's like, well, how is this [00:44:00] happening when you try it? It's mesmerizing. 

Max Bennett: Yeah. 

Andrea Hiott: Yeah. 

Max Bennett: So, so, um, so simulation, yeah, it gives us imagination. It gives us fine motor skills et cetera, et cetera. 

Andrea Hiott: You study with, or you, I think you, you talk Is it Redish?

Reddish? Do I say it right? David Radish. Yep. Um, just really quick. So this idea of Tolman and the cognitive map was like, controversial because people cognitive versus behaviorist. How could the rat be doing something like planning or the squirrel? It was not really accepted, but then, um, the studies, I guess it was the early 2000s sort of solidified for most of us that when a rat, um, kind of stops and looks like it's thinking and you you can see this actually in this activity of the brain. So this is what you say is the kind of the next breakthrough, right? Is this mentalizing or simulation, which then becomes mentalizing.

Max Bennett: So, so just Build on the story you're describing, you know, even though we could see rats pause and look back and forth and make a choice, [00:45:00] um, you know, there's, There's no proof that is actually imagining it's different options. And it was David Reddish and his PhD student, um, who actually recorded neurons in the hippocampus of a rat, which you can see it modeling space and then when it pauses and looks back and forth, you can literally watch a rat's brain, re rendering different paths and making a choice, which is just such an incredible study that I think.

Is so, um, you know, there are still a few holdouts, which I don't, it's hard for me to understand their arguments, but there are still some people that stay true to the argument that 

Andrea Hiott: only 

Max Bennett: humans can imagine the future. 

Andrea Hiott: But 

Max Bennett: if you look at that evidence, I find that very hard to argue, given that we can literally watch animals doing that.

Um, and it's such a, it's such a big breakthrough because In the absence of being able to plan what we're doing ahead of time, Um, it's very hard to flexibly respond to the world. And this is why we s this is one of the key things that are trying to recapitulate in AI systems, so that the fine motor skills of the [00:46:00] squirrel derive in large part from its ability to imagine what it does before it acts.

So what's happening, um, almost certainly really quickly in, in a squirrel's brain when it looks out on different paths as it considers a few options and it makes a plan of where it wants to place its, its hand or paw, um, as it's moving. And we know we can see even cats do this when it looks at platforms, um, and you record the motor cortex of a cat when it's eyes no longer gazing at the platform.

It's motor cortex stays active until it fulfills the action and then stops. So, and you can observe through introspection yourself doing this. If you're if you're walking through on a hike and you're running really fast, um, if one introspects and what's happening, you'll look ahead of yourself and you'll plan where you're going to place your feet.

Um, but you, your mind is focused on the next three steps, but somehow you hold in your mind. Before your feet finishes the action exactly where in the world, you're going to place your foot as you're running to make sure you don't trip on branches. And in that [00:47:00] simple act, what you're, what you're, what's actually happening is you're doing what these early mammals did.

You're not planning, you know, your career path. You're not imagining years in the future, but the simple act of just making plans a few seconds ahead of time is the algorithmic foundation for, for doing that bigger type of planning. 

Andrea Hiott: Yeah, it's really cool. I mean, I don't know what you think about this, but to me, it seems again, that kind of scaling continuous algorithmic thing where we're navigating, we're steering and we're developing these scales of, of how to do that.

So the mentalizing or simulation first, um, these cognitive maps seem to me like still ways of choosing the best. Direction or, or valence in a way that it's just developing different kinds of ways of being in different kinds of landscapes. So mentalizing is something like more like metacognition or, or, or theory of, of mind.

That's the fourth breakthrough. And then leading in to, to speaking. Um, I wonder what you think about that. Do you see these as also like, Ways we're still finding to better steer and navigate. That gets a little [00:48:00] metaphorical, but actually kind of literal too. 

Max Bennett: So I think this is a really, you're touching on like a really foundational open question both in neuroscience, cognitive neuroscience, cognitive psychology, and AI, which, which goes back to this reward hypothesis, which is how much of intelligence can be understood as fundamentally optimizing for a reward function.

And here's a take on that. So even if the original set of actions that a creature takes. starts with basic reflexes. Um, and then there's this sort of reinforcement learning system that learns based on what activates positive valence neurons to repeat those actions. Um, even if that's sort of the foundation of an entity, which would adhere very directly to the reward hypothesis, you're just taking actions that maximize reward.

Um, that doesn't mean all behavior in the end is reward for the following reason. Um, [00:49:00] one theory for how sort of the self exists. It's in, in mammals, um, is the neocortex or at least frontal cortex models and animals own behavior, which starts as just fundamentally reward driven. But what it tries to do is infer an intent of why I'm doing certain things and then tries to drive an animal's behavior towards fulfilling that intent.

So this speaks to what Carl Friston refers to as active inference, which is this idea that. That frontal cortex is a model of yourself that infers who you are and what you want to do based on observing your behavior and then tries to drive you towards fulfilling that. And so, for example, when an animal wakes up and is thirsty.

Um, what, you know, older vertebrate structures will do is it'll want to motivate actions that historically have led to the satisfaction of reward in the presence of thirst. But what the frontal cortex might do [00:50:00] is it might say, okay, when I tend to be thirsty, here's behaviors I tended to do. It doesn't mean that it's optimizing for reward.

It's trying to predict the best way to fulfill a goal. What it thinks the animal will do. So in simple animals, this ends up likely also optimizing for reward because it's learning the self model on a system that is reward driven. But that algorithm is not inherently optimizing for reward. That algorithm is optimizing for sort of the self supervised generative model that you see in the neocortex.

And so that's where. You get some interesting effects. Um, I'll give you a more direct example of this to be less conceptual about it. Um, there's a very known behavior amongst humans called the, um, cognitive dissonance effect. And then the a very famous study on the cognitive dissonance effect is you, you have a person read some words aloud to get admission to like sorority or fraternity club for like, they have to read these words [00:51:00] aloud for 30 minutes.

 So they read some words to a group, and at the end of that, they join some reading group, or they watch a video of a group they're about to be a part of, and then they leave, and a day later, or something, they review how happy they are they're a part of the group. And so the video people see is exactly the same, the only difference is one group has to read really embarrassing words out loud, um, and another word has to read really innocuous words out loud.

In other words, One admissions criteria is slightly more stressful than another admissions criteria. And a very bizarre effect is the people who had to struggle more and suffered more to get part of the group claim that they're happier and enjoyed being part of the group more. And so this is a direct violation of the reward hypothesis, which is, I suffered more to be part of this group, um, And so, and hence I like it more, but it makes perfect sense from an active inference model of the world, which is which, which, you know, especially in the world of mentalizing, where we have this not only model of our own behavior, but model of our own simulation.

So thinking about our own sort of [00:52:00] thinking processes, because what you would infer is, well, because I went through that painful process and joined the group, my model of myself would suggest I must really like the group. It would only make sense that I did that if the group was actually good. So the algorithm of active inference would self report that I like the group highly, more so than it would report if it was actually an innocuous thing.

So, so this is where, although that behavior's originally modeled on a human that from a young age at first is only reflexively sort of doing things, probably based on something adhering to the reward hypothesis, more and more behavior, as you get older, gets controlled by parts of the brain whose algorithm is not fundamentally about reward.

So anyway, very long winded way, just to say that it's, it's a big open question, how to model behavior between reward and sort of more sophisticated types of algorithms. But my general point of view on this is that evolutionarily it did start with reward, ontogenetically, meaning as [00:53:00] we grow from infants, it probably starts with reward, but as adult humans, not all behavior is fundamentally driven by reward.

Andrea Hiott: I agree with that. I think we can still think of steering and navigability without, as we were saying before, it doesn't need to be linked. To reward, especially not in the scientific way that we've learned to reward. And it doesn't even have to be linked to only one conception of self. I mean, this would be a whole other conversation, which we can't have 'cause we have limited time.

But I think that scalability and that algorithmic way of looking at it as a kind of navigation or movement. Um, it doesn't need to be stuck to reward so much. , but I want to get into the last one, which is, you know, speaking is the last, Breakthrough. And I think you talk about this as like a super human power, this, because it allows us to transfer.

So if we kind of stay with this idea of finding better ways to be in the world or move through the world or, or navigate or whatever, then I mean, language is like, wow, now, as you explain in the book, we can almost read each other's thoughts. We can, [00:54:00] because we can. Transfer them through this kind of system three ish medium.

 So I I want to hear your thoughts on that a little bit and try to link it to where we are now, because in this moment we have an explosion with the open AI and, all this large language models and so on. Also, you mentioned Karl Friston, which makes me think of active inference AI.

Andrea Hiott: In the book a lot, you talk about the, the catastrophic forgetting and how AI can't learn in the moment. So I wonder if you see in terms of, speaking and language, if we might with large action models, large world models, active inference AI, get away from those constraints that you present early in the book, I know that's a lot, but you can do whatever you want with it the next few minutes.

Max Bennett: So, yeah. So, um, So on the first point about the power of language, so, , it was fascinating to go through the journey of writing this book and just chronicle and weave through these past and still sometimes outstanding debates amongst, you know, very smart, cognitive psychologists and [00:55:00] AI researchers and neuroscientists where they're, you know, people of equal, um, sort of, knowledge can still come to different conclusions about something. So, um, so with language, there's historically been a discussion. Although most people fall in one camp of this between how much of intelligence comes from language itself. In other words, is language a tool for thinking versus, you know, most of our intelligence is not language based, but language is sort of bopped on as this ability to communicate.

And so most people were in that latter camp. Um, and still are in that latter camp. Noam Chomsky is the most famous sort of advocate for languages thinking. And so the story in the book very clearly comes down on the side of language is not fundamental, is not the foundation of our thinking language evolved as a sort of add on that enabled us to communicate.

That's not to say that language is not useful for our own thinking and there's feedback loops. Surely. [00:56:00] Um, but the evolutionary story told in the book provides evidence that clearly leads us towards the conclusion that human and animal intelligence is not first and foremost, or at least human intelligence, not first and foremost, grounded in language.

Um, but language is a tool for us sharing something fundamental about at least mammalian intelligence, which is we can render inner simulations of things and language enables us to share the results of those simulations. So, as a. More tangible example. Um, if me and you were, you know, hunting for boar in the African Savannah, you know, 100, 000 years ago what With language, we could say, okay, here's the plan.

I'm going to go right. You go left. And when I whistle three times, I'm going to run out and then you take out the net and you'll catch the boar and then we'll be able to eat tonight. And if we, if we catch that boar, that is only because we could share in a plan, a mental imagination of a future thing, and then make sure that we synergize that plan together, [00:57:00] no other animals capable of doing that so flexibly.

And so that is, I would argue, and many others would argue is the, is the fundamental. Sort of ability that language provides us. There's a secondary effect, which is it feedbacks on itself, which is when I use language to write things down, I can keep going over the ideas I write down and realize fundamental flaws in my own thinking and how I'm not.

You know, putting things together correctly in minor simulation. And in that way, it does improve thinking. Um, but I, you know, writing only appeared way after language evolved. So the evolutionary pressure was clearly something different. Um, and to your second your sort of second set of questions around continual learning problem, I do think this is a huge area of open research where it's not clear.

If we already have the ideas, they just haven't been scaled up. Or there's a fundamental change that we need in artificial neural networks, or even [00:58:00] something more fundamental on that in order to get these types of networks to learn continually as they get new information. And so there's so many different research streams here.

Um, there's a ton of work going into just trying to augment large language models with persistent memory stores and trying to sort of what I would argue Hack the experience of continual learning on the existing paradigm and big open question, how far that will go. It's possible that will bring us to something that feels like it's continual learning, even though under the hood, it's not.

Um, or there's a whole nother group of people, which is where I'm spending more of my time, because I find it at least more beneficial. Interesting, which is trying to reimagine the algorithm, the learning algorithms of neural networks to adhere towards or to produce the types of continual learning performance that we see in, in mammals.

Um, but yeah, more to come on that. 

Andrea Hiott: Yeah. I have to let you go. I didn't even get to ask you about your motivation and the love aspect, so.

Max Bennett: Maybe a future session. 

Andrea Hiott: Yeah. A future session, but [00:59:00] thanks for the time thanks for the work you're doing. Thanks for the book. It's wonderful.

Everyone should read it. 

Max Bennett: I really appreciate it. It's wonderful being here. 

Andrea Hiott: It's great. Be well.