We Were Wrong About Our Minds—And AI

Hybrids, cyborgs, chimera: combinations of living material and designed, engineered material. There's an enormous option space of diverse bodies and minds. We are going to be living with these things. There is no doubt.

Somebody said to me once, "Well, I make these language models. I know exactly what they do. I program." I said, "You don't even know what bubble sword does. It's six lines of code and we found new things that nobody had seen. If we can't predict what bubble sword's going to do, you have no idea what this other thing is going to be."

I mean, let's ask the question: What is evolution really doing?

The following is a compilation of interviews from my Forbes reporting on Michael Levan, one of the most disruptive developmental synthetic biologists alive today. Levan is widely known for creating Xenobots, the world's first synthetic life forms. His research has uncovered intelligent agents within our own bodies and revealed that goal-directed intelligence can spontaneously emerge in nonbiological systems. His work challenges fundamental assumptions about evolution, about hidden agency in algorithms, and about what came first: life or intelligence. Levan's discoveries suggest that mind emerges much more fundamentally and more unexpectedly than traditionally recognized, with profound implications for longevity and health, and even what it means to be human, evolving alongside artificial intelligence.

"How are you?"

"I haven't in a very long time."

"I know, it's been a while. Yeah, I saw, I saw, I was thinking of you when I saw that that great piece on Dennis."

"Oh, I'm so glad you liked it because of course, there's a lot of blowback on saying anything is purposeful. I just want to level set for everybody where we're at because there seems to be this huge gulf of awareness between what's happened in the past 20 years, and particularly in your lab, because I think you're at the forefront of pushing these boundaries and redefining things, things you've discovered, observed, and created in your lab."

"Yeah, we make synthetic beings like Xenobots and Anthrobots. Xenobots are made of frog skin. Anthrobots are made of adult human tracheal epithelium. They have all kinds of interesting new behaviors that their origin tissue may not have had."

"So when you're making an Anthrobot and you're calling it synthetic, but it is biological, right? Is that right?"

"Yeah. Oh yeah."

"You're taking these cells from a human cadaver and then you're creating a new organism with it through, is it an AI-generated design?"

"First of all, it doesn't have to be a cadaver. It could be just donations from a biopsy, from a tracheal biopsy from a living patient. What we do is we capitalize on the competencies of these cells to pull themselves together into a functional form in new environments. The cells self-assemble into a creature with novel behaviors and form that we did not give it. So we use AI to try to discover. This is in collaboration with Josh Bongard's lab at UVM. We use it to discover new ways to prompt the cells to do new things. As a translator, you have these tools, including AI, that help you communicate your goals to those of the cellular collective."

The cellular collective are groups of cells in your body that act as a scalable intelligence team whose problem-solving power grows with its size. These collectives use shared goals and communication to solve problems, adapt, and accomplish tasks like building organs or regenerating tissue, showing an increasingly sophisticated collective intelligence beyond any one single cell.

"The thing that drives all of these, all these discoveries, is this idea that you really can't assume what level of intelligence or cognitive capacity a system has; you have to do experiments. And also the idea that tools from behavioral and cognitive science, having to do with the study of memory, the study of goal-directedness, problem-solving in various spaces, is much broader than we typically think of in terms of brains and activity in this three-dimensional world."

"So, so, correct me if I'm wrong. The things you're seeing in these cells is cognitive processes like sensing, perception, memory, different kinds of learning, decision-making, goal setting and pursuing, and communication."

"All, all of those, including my my favorite, which is um, creative problem-solving, finding new ways to get to your goal. When everything changes, the environment changes, your own parts change, everything changes, and you find new ways to get your goals met. That is a really non-controversial definition of intelligence: being able to use the tools you have in new ways to solve problems you've never seen before. That is not a a weird way to define intelligence, right? I think we could all agree that that's pretty much intelligence. And we see those all the time in unconventional beings navigating weird spaces that we're not used to thinking about. This is not just, you know, philosophical fluff or um, you know, this idea where you think there's a spirit under every rock. Where the rubber meets the road is simply this: Can you take the tools that people use to study memory, learning, goal-directedness, problem-solving in behavioral cognitive sciences, and can you apply them to the kinds of things I'm talking about, cells, tissues, molecular networks? And the answer is yes. That is what we've been doing. That is how you know that this is functionally real and useful because it lets you do something that you otherwise would never do. We do things like training molecular pathways using optogenetics and memory rewriting to permanently change the anatomy of various creatures. That is how you know you mean this literally is by being able to use the exact same tools as a behavioral and social scientist."

"So you're using behavioral and social sciences applied to the molecular level."

"Sure. Yeah. Yeah. I'll give you a simple example. So a gene regulatory network or a pathway is just, it could be small. It could be five or six or more genes or proteins that all regulate each other. They just turn each other on and off. So you can imagine this kind of network with arrows, right? Where each thing turns each other on and off. You could say, 'Well, I don't see any magic there. It is deterministic. It is simple. It's transparent. I'm going to assume that this thing is is dumb and it's mechanical and I'm going to treat it the way I would any mechanical system.' And the implications of that are that if you want to change its behavior, you have to rewire it. This means, in biological terms, it means gene therapy to change the ways that the different proteins interact with each other and so on. Or you could say, 'Actually, we don't know that, and the way to discover where something lands on the cognitive spectrum is to actually do experiments and see what happens.' And so one thing that we've been doing is taking the simple behavioral training assays and you can choose the different nodes in that network and you can say, 'Okay, this is going to be my unconditioned stimulus,' where when I trigger this one protein, this other thing, which I'm going to call the response, turns on, right? And then there's another node, we're going to call that a neutral stimulus because if I trigger this one, that one does not turn on. And then I'm going to do Pavlovian conditioning, right? Like Pavlov's dog. So what I'm going to do is I'm going to tweak these two nodes together repeatedly, and what we find is that if you do that with the right network and the right choice of nodes, if you tweak them repeatedly, eventually just the neutral one is sufficient to turn on the response, whereas it didn't do that before. That's the story of Pavlov's dog, you know, and the bell and the and the meat. I mean, it wasn't really a bell, but whatever. The idea is that the network has learned to associate a neutral stimulus with the actually salient stimulus. And this happens. You don't need a cell for this. You don't need any of the machinery of cells. Just a small collection of chemicals wired appropriately will already give you five or six different kinds of learning: sensitization, habituation, associative learning, and so on."

This means that basic chemical systems, before reaching the level of complexity in biology where we typically think life and intelligence emerges, these basic systems are capable of a variety of different categories of learning. Other researchers have shown that chemical droplets can solve mazes, demonstrating a sort of primitive intelligence. Understanding just how much of our natural world has varying degrees of cognitive competencies opens up whole new ways of relating to and negotiating with reality. A reality where intelligence emerges at a much deeper level and in far more varied systems.

The reason it's important is that if you have that mindset, that you're willing to say that the degree of intelligence of something is an experimental question, it's not, you can't just decide that from philosophical pre-commitments. Then what that means is, in this particular case, when you realize that, "Hey, I can train this thing with stimuli," it gives you an entirely new way to use drugs. All the problems that we have with drug discovery in biomedicine, where a certain drug will, first of all, have all kinds of different effects in different patients. It has side effects, then then things stop working, right? You take it for a while and all of a sudden it stops working. So drug habituation. Then you find out that, yeah, actually your molecular networks in your body have a kind of memory. It's exactly the kind of memory, but functionally, that we study in behavior. And that if we understand this, we can actually do interesting things with pulsing drug exposure in particular ways to either create or break those memories. So this is hugely potentially important biomedically, but you only get there if you're willing to make that hypothesis that this thing might actually be embeddable to the way you would normally train a complicated animal with a brain.

"So cells have a goal set in the future and they know what it is."

"We know they know what it is because they stop when they get there. This is part of a goal-seeking system: that when it gets the goal, then the activity, all the hard work stops. So then you then you know it knows where the end point is because you're right. Of course, in order to have goals, you have to have a recorded set point. You have to have a stored memory of what it is that you're trying to get to."

"This is absolutely right."

"Sorry. A stored memory or um, a prediction, like something that hasn't happened."

"Yes. It's a, it's not a memory in the sense that you've already been there. It's a memory in the sense that it's an encoded state of affairs that isn't true right now. So yes, so prediction is exact, is is fine. It's that kind of mental time travel where you can think about something that isn't true right now. And yes, I mean, absolutely, the cellular systems can have anticipation and predictive capacity, active inference, all that stuff."

Essential to intelligence is making predictions about the world and adapting based on those predictions. So active inference means that instead of just passively receiving information, your mind is actively trying to confirm its guesses about reality. As an intelligent system, you are constantly making predictions about the future, about what you'll see, hear, and feel next. And when those predictions are wrong, you either update your mental model of reality or act in ways to change your environment to make reality match your predictions.

"So we said, 'Okay, if all that is going to be true, then exactly as you are hypothesizing, somewhere there needs to be a recording of that anticipated endpoint, right? So what is that? Well, what is it in the brain? Well, in the brain, presumably, it's patterns of electrical activity.' So fine, let's look for it. So we looked for it, and sure enough, what did we find? We found a distributed bioelectrical state that we can see directly. What I mean is you can use a voltage-sensitive fluorescent dye, you soak the worms, and you see that there's a particular beautiful pattern that we have now decoded that says, 'Build one head. Build exactly one head.' That is the planarians. I'm not saying anything about consciousness. What I'm saying is that is the functional set point of what the heck are we building? We are a bunch of cells, we've been injured, we need to build something. This is the information that tells us what it is that we're building. How do we know that's the information? Because we went in and we rewrote it. And guess what we did? We altered that electrical state to say 'two heads.' And if you do that, that's what the worms build: a two-headed worm. Because you have altered their memory of what a correct worm is supposed to look like. And if you keep cutting them, because it's a memory, it sticks. Once they've learned that the right thing is two-headed, then they stay with it. And you can keep cutting them, and it's two-headed. And we have not changed the DNA at all. There is nothing wrong with the hardware of the system. Now, when you have these these memories, sometimes the agent that reads out those patterns might be confused about which interpretation it is. And we created worms that are confused about whether their pattern means one head or two heads. And every time you cut them into pieces, every piece makes an independent decision as to whether they're going to be one-headed or two-headed, right? We'll call them cryptic worms because the pattern can't quite decide. Every piece individually thinks, 'No, that says one head. No, that says two-headed,' right? You would never do that on this on on the previous paradigm. Then you say, 'Okay, so the other thing in cognitive science is that these goal-directed behaviors, they have a certain stress associated with not matching those goals.' And we know in psychiatric settings, we know that one thing you could use is anxiolytics. You can use compounds that will actually affect the motivation you have to meet goals. What does that look like when applied to a morphogenetic agent?"

A morphogenetic agent is an entity like a cell that processes information and problem-solves to create a functional form.

"We have hallucinogens. We have things like psilocybin and LSD that actually distort your ability to have cognition about the information that's distorted. What does that look like in morphogenesis? Can you make a morphogenetic agent hallucinate? And we've done some of that. All of these things you would never in a million years do these experiments under the standard paradigm. We've also shown that if you cause an eye to develop on the tail of a tadpole instead of the front of the head where they've been for millions of years, those animals can see perfectly well. That eye connects to sometimes the spinal cord, sometimes the gut, sometimes to nothing at all. And those animals can see. And it doesn't require new rounds of selection and mutation and all of that stuff to radically change that sensory-motor architecture and let them still be functional because all of these systems fundamentally are ready to reconfigure themselves in novel scenarios."

"Where does that come from? Is it in the D, like where is that programming, that scheme?"

"Are you asking about their specific form and function, or are you asking where does the ability to solve problems come from?"

"I guess if you're calling it the ability to solve problems, what I'm wondering is where does the capacity to be able to grow a working eyeball on your tail, where is that coming from?"

"Bodies consist of a set of nested modular agents. It's what I call a multiscale competency architecture, meaning it's got components that solve various problems and they solve these problems by delegation. So they they trigger downstream modules to do other things. You don't micromanage it. You don't have to tell every cell what to do. You don't have to tell every gene what to do. The the cellular collective has certain competencies. And when you trigger those, it will go ahead and make an eye. And what will happen is that if it's in a weird location, it will adjust. It will adjust size and shape, whatever it needs to be. Where all of this comes from is the fact that biology works kind of the opposite to how we do our computer technology. In our technology, the bottom layer is extremely stable. So you absolutely want your data to to stay still. You want fidelity of the data and everything on top of that is built on the idea that the layer below you is functional and it's operating properly. We expect our hardware to operate properly. In biology, it's not like that at all. In life, you are guaranteed that everything will change. Not only will the environment change, but your own body parts will mutate through long periods of evolution. Nothing stays the same. And so I think what evolution exploits is this hierarchical architecture. And specifically, a lot of effort has gone in not to finding specific solutions, but actually to finding policies for cooperation between molecular networks, between cells, between tissues that enable them to not assume much of anything about what happened before and to just solve problems and pursue specific goals in the current moment."

"And when you say cooperation and you're using words like intelligence, memory, problem-solving, are you using those all literally or is that metaphorical?"

"I'm using them, of course, metaphorically because every science uses its terms metaphorically. All of these things are just our models of what we see in the world. They are not the the real thing in and of itself. But but but also I'm using them quite literally because I mean the same functionally, I mean the exact same thing that people mean when they refer to, let's say, animal behavior."

"So is parts, parts with agency, is that sort of your definition of life?"

"Um, I, you know, uh, I I don't spend too much time trying to define life. I think cognition is a much more interesting and bigger category than life. But if I had to define life, I would say..."

"Wait, sorry. What do you mean it's bigger? Cognition is bigger than life. Like, how, how do you parse, how do you take it apart?"

"My definition, if I had to say what life was, my definition of life would be: we call life systems that are really good at scaling their cognitive light cone. What I mean by the cognitive light cone is the size of the biggest goal that a system can have in terms of space and time. So if you think about all sorts of creatures, for very simple things, cells and ticks and, you know, dogs and humans and whatever else, you can sort of put them all on one continuum by asking, 'What are the biggest goals that the system can possibly pursue?' Your dog is never going to be able to care what happens three weeks from now in a town 10 miles away. You can't, the size of the goals, right? Human cognitive light cones are enormous. They're bigger than our lifespan. Lots of us are working on things that may or may not happen long after we're dead. Bacteria and other things, their cognitive light cones are tiny. It's only the cell itself. It only cares about what happens there and everything outside, whatever, right? And then they have a short time horizon and so on."

A light cone in physics defines the boundary of what's physically possible for cause and effect events in spacetime, limited by the speed of light. It's shaped like a cone. Because as time marches upward and space spreads out horizontally, the paths that light can travel form a cone-shaped boundary or limit. Only events within this cone can affect each other because nothing travels faster than light.

Levan borrows this term in physics and applies it to how intelligent systems are able to scale their goals and capabilities. A cognitive light cone represents the boundary of events in time and space that an organism or any intelligent system can perceive, remember, or leverage to guide its actions. It's the scope of awareness: how far ahead, behind, or around itself an agent can consider events or information when making decisions. So, humans may be defined less by our biology than by the unusually broad scope of our cognitive light cone, with edges that we can't see beyond, but where societies and cultures may extend these parameters into collective memories and foresight.

"So that's what I think life is. Life, I think, is what we call evolutionary systems that are good at scaling their goals. But if you draw this kind of Venn diagram of what's, what's the set of all things that are alive and all, is the set of all things that are cognitive? I think life is a subset of the things that are cognitive. I do not think, uh, that things that we typically call alive are the only things that are amenable to the tools of studying cognition. And, uh, I think cognition was here first, and, uh, life as a means of scaling it up, we came after."

"And what does evolution do to the cognitive light cone?"

"What evolution does is provide ways to scale up, to increase the size of the cognitive light cone of systems. So we call alive things that are very good at scaling it. Rocks are not scaling it because the cognitive light cone of a rock is exactly the same as that of its individual parts. Hasn't scaled anything. It's exactly the same. It's just a pile of materials. But in a bacterium or a cell, the individual components project those larger goals into a new space. And individual cells that have metabolic goals, they have gene expression goals, things like that. When they get together into a cellular network like a tissue and organ, they're able to pursue these really grandiose goals like 'build a limb,' right? So the cells of a salamander limb have this enormous goal in anatomical space. They want to traverse to the place that that corresponds to having the right number of fingers, the right size, the right shape. If you amputate that limb, they know they've been deviated. They will come right back. They'll work really hard to get back there and then they'll stop. So the collective has this enormous grandiose goal. The individual cells have no idea what a limb is or what fingers are. And if individual cells get disconnected from that collective, that cognitive light collapses immediately. It becomes very small. The cognitive light cone is also the border between you and the outside world. So this is cancer. This is what cancer is. It's basically a biophysical, bioelectrical, informational disconnection between a cell and the collective. At which point the cognitive light cone shrinks to the level of an amoeba. You're just an amoeba at that point, and all your goals are little tiny goals, and the rest of the body is just environment to you, just external environment, not you anymore. So that boundary between you and itself and world can shift during development, during evolution, and then during cancer, it constricts again."

"So when it comes to intelligence, you're investigating cancer as a framework where a network of intelligent cells, one of them becomes disconnected and devolves into a single-cell organism instead of part of the network. And you're trying to rehabilitate that cell into rejoining the network of information. Is that right?"

"Yeah, that's a, that's a perfectly fair summary. I mean, basically, the collective is smart enough to store very large memories of what it's building. Let's say nice skin, muscle, whatever, you know, what organs, whatever it's building. And when you disconnect from that, you are no longer able to be part of that commitment to that journey in morphospace. You have a tiny little cognitive light cone that's just about protecting and optimizing your own local conditions. Every piece has its own ability to learn, its own agenda, its own little goals. And managing all that gives you access to this incredible intelligence spiral, but it also makes you susceptible to defections. Sometimes your parts don't agree with what the whole wants to do. That's dangerous, at least for the for the collective. But those are really, you know, that's that's how you get to high levels of agency is you use these kinds of things to scale up the competencies of parts."

"So you're taking cells from a human, and these cells have memories, and you're giving them new memories."

"So the memory work is not published yet. Make any claims there until it's out and peer-reviewed."

"Can I ask you for cellular intelligence, which is what is required to do this, right?"

"How do you do that without memory?"

"I'm not saying they don't have memory. I'm saying you can't say that they do or they don't until you've done an appropriate test. So, we we are doing all of those, and we actually have a bunch of really interesting data on this. You know, that will come out later. But but for right now, what I'm saying is if you want to know how intelligent something is, make a hypothesis about what space it's working in. Make a hypothesis about what its goals are. Put a barrier in. You see how much ingenuity it has in overcoming that barrier. We see this again and again in living systems. If you care to look, it's it's everywhere. We don't even have really the beginnings of yet of a solid science of guessing what goals new creatures are going to have."

"And we're using these words without necessarily saying conscious, like these are memories, learned, all this problem-solving, but we don't necessarily know that has ties to consciousness."

"Let's talk about that. That's a separate question. I have not said anything about consciousness yet in this conversation. That's a whole separate question."

"Does a cell have purpose?"

"So, for any of these questions, we can't just answer it sitting here in our chairs. We have to do experiments. If I had to guess, if I had to bet $10 right now, I would say probably not, but we don't know that. Uh, it may well. It definitely has goals. So, so cells and tissues for sure have goals. Um, but we don't know what the metacognitive capacities are. I mean, I think, uh, it's entirely possible."

"When I hear problem-solving, I'm trying to understand how I could separate that from some form of awareness of a future state."

"You, you, you shouldn't. Problem-solving is absolutely has awareness of a future state. The distinction I was making earlier was this: there is having an encoded future state that you're working really hard towards, and that is absolutely part of problem-solving, and everything has it down below, even below the cellular level, versus there's the human level of awareness, which is, 'I am able to verbalize that I have awareness.' It's a circular, it's a circular looking back on yourself and saying, 'Wow, I have this goal. When did I acquire this goal? Why do I have this goal? Is it good for me in the long term? Is it good for society?' That's a whole separate kettle of awareness than the basic, simple, 'This is my goal. I know what it is. I'm going to work like hell to get there.'"

"So if I were to strip down consciousness just a basic awareness of..."

"Well, I haven't, like, that's why I want to get rid of it. If I want to be able to understand this and just in terms of just awareness, like let's not take all the baggage that comes with consciousness, like just these, there's an awareness, may not be anything like ours, but there's got to be in order."

"Let's come up with a definition for awareness that works here. A useful definition for awareness that works beyond this kind of human second-order meta-awareness is any information that the system has access to to to drive behavior. That's awareness. Awareness is simply access to information. More broadly, you are aware of of an event happening or a state or a goal when you have access to that information and you can use it to drive your behavior."

"Everything you just said about an awareness of that information and how it applies, is that something you think cells..."

"Yeah. Oh, absolutely."

"Okay. Yeah."

"I think, I think this gets down to definition because I think a lot of people when they say purpose or consciousness, um, they think of like a really, like high-level self-awareness."

"Exactly."

"Yeah. And I, I don't..."

"Yeah, I agree. I agree with that."

"Yeah, I agree with that. Well, I think so, so we don't, we don't really have a good vocabulary to talk about all this kind of stuff, but I think it's very clear that a lot of people use the word, and I mean, I don't talk about consciousness much, but they use all of these things: cognition, consciousness, goals, and what they're thinking about is their own extremely high-level human version of it. And and by the way, humans can have unconscious awareness. For example, you're reading something and every time you come across the letter Q, they do a little puff into your eye, but you can't, it's, it's below your awareness and you can't see it. After after a little while, every time you come across the letter Q, your eyelids are going to shut in anticipation of the puff. Are you going to argue that you were never aware of the puff? You consciously were never aware of the puff. But if your body didn't have awareness of that event, you would not now be trained to close your eyes when you see the letter Q. Your body was absolutely aware of this information. Even though you are in my left hemispheres that are having this verbal chat, we were not aware of it, but your body most certainly was aware of it. But it's a non-verbal awareness. It's an awareness that your verbal system has no access to. Your language center has no access to that information."

"But do individual cells have inherent cognitive capabilities, problem-solving, self-organized?"

"Yes. Yes. People have found individual cells can, they can anticipate, they can, they can learn, they can anticipate future events. So they have a predictive capacity. They certainly have the same kind of morphological problem-solving that cellular collectives do. Yeah. Sure they did. Yeah, we have a new preprint up now that shows that the causal emergence metrics that people use in integrated information theory, I mean, they use it to quantify consciousness. I I don't know if it's really consciousness, but there is something very special about these causal emergence measures like Phi and so on."

Traditionally, physics assumes that if you know all of the microscopic details, every atom, every neuron, you can predict everything, at least in principle. But in complex systems like brains, economies, or ecosystems, the relationship between all of the parts and the way they interact can make the system as a whole generate new, irreducible causes and effects. This is causal emergence. One theory of consciousness, Integrated Information Theory, IIT, says that consciousness itself may be a form of causal emergence where neural relationships give rise to this higher-level phenomenon that has greater causal power and richer information than the micromechanics of our individual neurons.

"Gene regulatory networks, never mind the whole cell, but just the chemical network itself. As it's learning, its causal emergence goes up. So the reality of it over and above its parts already, that's there in the molecular networks. You don't even need cells for any of this."

"How important is networks and collections of networked things to intelligence?"

"It's incredibly important because I'm not aware of any examples of intelligence that's not collective intelligence. That's that's some kind of monolithic, indivisible. I mean, we, we don't, we've never seen that. Every example of intelligence, and in fact, if you try to think of, you know, how can you do learning if you don't have parts that join together and that change, you know, all intelligence is collective intelligence."

"How does that impact our current theories of biogenesis, origins of life, and evolution?"

"I mean, let's ask the question: What is evolution really doing? I I don't think that what it's doing is searching the space of all possible phenotypes, meaning outcomes, you know, all the different protein shapes or whatever. That's what we typically think it's doing. What I actually think it's doing is searching through a set of pointers into Platonic space. The Platonic space meaning the where wherever the laws of mathematics come from and so on."

What is the non-physical realm of Platonic space? And why do so many serious people think it might actually exist? Because abstract truths like mathematics seem objectively real, perfect, and unchanging, yet exist nowhere in the messy physical world, even though they seem to structure it. Take the number four. You can have four apples and four chairs, but the number four itself isn't contained in any one of those apples or chairs. It's an abstract pattern or truth that applies across examples. You can't change that truth. 2 + 2 can't suddenly equal 5. It has to equal 4. That gives the number four a sense of being objectively real. Even though you can't touch or see it directly.

The mathematical conviction is a bit stronger. See, you can certainly be wrong in mathematics. I don't know of any mathematician who hasn't been wrong at some stage. And yeah, you can certainly be wrong. But there is somehow that the, the wrongness is in one's own failings. And it's not in the absolute Platonic world that you're exploring.

Levan thinks that anatomical shapes function in a similar way, where an individual organism expresses an anatomical body shape by developing into it, just as a group of four apples expresses the abstract number four by adding up to it. But you still might be thinking, "Okay, this sounds abstract and conceptual. Why would anyone believe in this space?" Because in physics and chemistry, things just react. Particles collide, atoms bond, but there's no goal or targeted future outcome. Hydrogen and oxygen don't try to make water; they just do when conditions are right. Biology is different. Cells behave as if they have a specific end state they're driven to reach. DNA provides the instructions for life's building blocks, but nothing in DNA defines a living being's final complex architecture. DNA is more like a parts library than an engineering blueprint. Yet, cells reliably build eyes, limbs, and organs to precise, extremely functional specifications, even when they're scrambled or damaged. They can tell when something's wrong, and they fix it. So Levan and other systems biologists posit a Platonic space, a realm where anatomical goals called attractors exist and toward which our cells, these tiny cognitive agents, are drawn as they collaborate and problem-solve to bring our future physical forms into being in ways that can't be explained by local physical forces that act only in the present.

"So we, we like to think about goals, memories, preferences, and problem-solving in other spaces. And that helps you discover various things."

"When you say space, because I think a lot of us take that seriously, literally a space like a space of all possible forms. Is that what you mean literally, or is there something else you're..."

"I do not think there is a fundamental difference between the three-dimensional space that we're all used to and these other problem spaces that creatures live in that are hard for us to visualize. So an anatomical space is a very high-dimensional space, and species cluster in that space. What happens during embryogenesis is that you have to navigate that space because you start off as a single cell and you're in some region of a space, and then you sort of wander, and then eventually you end in that region of that space that corresponds to the correct shape. And if you interfere along the way, for example, you cut them in half and suddenly drag them, that's an effect, dragging them to a different part of the space. Or you scramble their craniofacial organs, as we did in in frog embryos, which again drags you to a a crazy part of that space that you've never been before in evolution. You still have to find your way to where you're going. And that's navigation. That's the ability to navigate that space despite new, unexpected things that are happening to you."

"So, is that like a, a non-static Platonic form of a creature?"

"So, I, I, I think yes. If we start to ask where do these forms come from, I think you very quickly get into the idea of Platonic space because with with Xenobots, with Anthrobots, with all of the stuff that we make, the typical answer to why does a creature have a specific shape, a specific behavior, you know, the standard answer is, 'Oh, eons of evolution, of course, was selected to do these.' Well, with these novel creatures, there are no, there, there is no history of of of selection. There have never been any Xenobots. There have never been any Anthrobots. Your tracheal cells have never had the ability to run around and fix neural defects. I I think this three-dimensional space that we're seeing now is a construction. I think it's a construction of our particular nervous system that is really used to physically moving around. The whole point is to physically move around in this space. And we have a cognitive system that gives priority to this space. We live at a particular rate of speed. And our evolutionary firmware is optimized to detect intelligence as movement through three-dimensional space. We have eyes and ears. And we mostly think about how things navigate three-dimensional space. Imagine for a moment if we had evolved with an internal sense organ that picked up your parameters of your blood chemistry. Kind of the way we do taste, right? So imagine if at any point you could feel what your blood chemistry was like. Let's say you had five new senses that picked up potassium level, the sodium level, pH, CO2, I I don't know, five other senses. I think if we had those senses, we would have absolutely no problem visualizing that your liver and your kidneys are intelligent agents that navigate in a space, solve problems, and help you stay alive. And they are some kind of a symbiant that traverses that space and keeps us healthy throughout the day by dealing with all kinds of stuff that we throw at them. We would know their intelligence direct. We would we would feel it directly the way we do now when we see, you know, monkeys open boxes and do smart things. So, it is it is just the lack of imagination on our part that makes it hard for us to notice embodiment in other spaces. Our bodies are full of agents that live and strive and suffer and solve problems and learn in these physiological spaces, gene expression spaces, and so on. We, we find these things hard to visualize because all of our behavioral and cognitive systems are pointed into this three-dimensional space. I take these other spaces as real as as I do three-dimensional space."

"What is a symbiant?"

"Oh, sorry. Symbiant is like, you know, sometimes species live in symbiosis where they live together and one species is doing something helpful for another, right? So one way to look at our various organs is as a highly competent symbiote whose job it is to navigate a specific space to keep us alive."

Multiple studies have been investigating transplant recipient reports of receiving more than just the donor organ. With one study of transplant recipients revealing that 89% experience personality changes regardless of which organ was transplanted, researchers are investigating whether "the donor's organ is capable of storing memories or other personality traits." Anecdotal reports like these have been around forever but were routinely dismissed as psychosomatic. With a broadening understanding of the nature of intelligence, these reports are now being interrogated to understand whether our parts are themselves living beings with memories and traits.

"Well, so, so let me try and um, and break this down because my original understanding of evolution is there's five mechanisms: there's natural selection, which is survival and reproduction; sexual selection, like reproduction of the sexiest; and mutation, changes to the DNA; genetic drift and genetic flow. But then, and all of these affect the DNA. And then you've got all this evidence, and so do many other labs at this point, of cellular cognition. How does that play? Like, it it seems like that would have to be a mechanism, an additional mechanism that we might not be able to trace be because it doesn't, uh, affect the DNA, because you're showing you can do all these things like change a whole creature's shape and its lineage without affecting the DNA. How do we, uh, account for its impact on evolution?"

"It's different than a mechanism. It's different than the question of whether or not it's DNA. I mean, the question of whether it's DNA is just a question about the scratch pad. Where does the information get recorded? And yes, we've shown that in addition to DNA, there are also bioelectric circuits that persist and that can be used as a scratch pad for these things. And it's not because, uh, you don't need to, um, negate the the importance of DNA or it's not specifically because of biometrics, although biomectrics are a huge part of that intelligence. It, it, it's, it's much more fundamental than that. It's, it's that the process of turning genes into outcomes is itself a problem-solving competency. It's an intelligence. It is not a dumb mechanical mapping. And that makes all the difference for evolution. Standard development lulls us into a false sense of boredom. Basically, we, we tend to think that, all right, well, this is the only thing this knows how to do. And the story of evolution feeds into that by saying, 'Of course, it is because there's a history that led up to this, and I'll tell you exactly what the history was.' That's the whole point of evolution. It's designed to have great specificity between the past environments of something and what it looks like and what it acts like now. It's not about whether it's in the DNA or really what whether it's a blueprint. It's about the fact that what is passed along are a series of really good suggestions that the collective intelligence will follow under normal circumstances. But it has way more plasticity than that. And when called upon, it will not only find new ways to reach the same goal. If it can't do that, it will find new goals to reach, which is basically, you know, the definition of intelligence."

"When you say it's more plastic, are you saying it has more agency?"

"So this is important. I'm saying that the process of morphogenesis, of forming bodies and minds, has agency and intelligence. But I want to be really clear. I'm not saying this as a philosophical claim or a linguistic claim or, you know, we're going to redefine what agency means or or or we're using these words because it's so beautiful. Like a lot of people say to me, 'You're right, you know, it's such a beautiful process. Of course, it has to have it.' Not because it's beautiful. Not because it's complex. Not because it's reliable. No, you can't, you can't attribute agency because of any of those things. That all that stuff cheapens the the the real argument I'm trying to make. What I'm saying is that is not why we attribute agency. Why we attribute agency is because if you take tools that are normally used to study highly agential things, meaning cognitive and behavioral science tools, and that means physical tools in the bench, but it also means conceptual tools, things like active inference and, you know, perceptual multi-stability and all that stuff. If you take those tools, you will do better science and make better discoveries when you apply them to morphogenesis. Then you will if you take the assumption, and let's be clear, it is it is just an assumption that everybody seems to have bought into as as a fact. Then then if you take the assumption that no, I'm only going to stick to the tools suitable for simple dumb mechanical machines, chemical machines. You can't, you can't assume any of that, right? This is a science. This is not a linguistic or philosophical endeavor. And you simply cannot assume in either direction. You have to do the experiments. So, so this is important because I, I think, um, I, I think this has implications beyond just the sort of infighting I see between between, um, evolutionary biologists and atheists and theists and and all that. The reason I care about this and why I think what you're doing is so relevant for where we're going and why I want to talk about what the hell a hybrid, a chimera, and a cyborg is, is I didn't care this much about evolution until I started to learn about cellular cognition and cognition in parts and cognition that isn't necessarily affecting DNA and seeing that it things can evolve that don't have biological substrates. And I think that's incredibly important for what we're seeing with intelligence and machines and how we're going to understand our own future. And I wanted to understand from you what this means for us as humans, how this has impacted your understanding of what it means to be human, where you think we're going, and what we need to be prepared for or understand about ourselves, where we came from to understand where we're going, and how we're going to co-evolve with AI."

"Wow. The simple, simple, short, short question, right? Uh, certainly the questions of evolution and how we got here are interesting, but I, I think we need to be very careful with this idea that stories of evolution and how we got here and what we are and things like that."

to let them constrain what we can be in the future. I'll give you a very simple example. Um, you know, uh, there's this movie X-Machina, and and this is this has been this has been in science fiction for decades. The guy has had an encounter with with this android. Uh, she's very convincing. So convincing, in fact, that he starts wondering very naturally, "Wait, am I an android?"

Now, an interesting question. Obviously, he has never taken developmental biology or biochemistry because he should have been thinking that all along, like, "Wait a minute, definitely I'm made of some parts, like, what's going on?" But okay, I guess that question never occurred to him. So now he sees a sander and he's like, "My god, what if I'm a robot too?" So he's standing in front of a bathroom mirror and he takes, um, and he takes a razor and he's, he's like, cutting his arm to see if there's, if there's wires in there.

Now, the reason this works as a movie scene, it's very tense, you know, the music is playing. What the reason? The reason it works is because everybody knows, the director knows, that everyone in the audience understands that if he opens his arm and finds cogs and gears, he's going to be super upset. Why is he going to be upset? Because he's, he's going to say, "My god, I'm not real. I don't have real agency. I'm just a robot. This is all been, my whole life has been a lie. Um, that's it." You know, "I don't know what to do anymore." Okay?

And and I get emails from people who say, "Wow, I read your papers. Now I understand that I'm a big bag of cells. I, I don't know what to do with my cells anymore, right?" Like, this is real. This, this stuff is very real. And so the, and so, so to me, this is an amazing, amazing, uh, kind of, uh, kind of thing because think about, think about what this means. This guy has whatever, 30, I forget how old he is, 30 odd years of primary experience with his, you know, with his inner thoughts, his his emotions, the values that he's been, you know, fighting over, the, um, uh, you know, his preferences and his and love and all of these things, right? And he's going to set them aside in, in, in a blink of an eye because of some kind of story that he thinks exists that will tell him that if you have cogs and gears, all of that you experienced is not real because cogs and gears can't do that. Wires can't do that. But the wet, squishy protein stuff that, by the way, came about by the meanderings of of cosmic rays hitting, you know, hitting egg cells through millions of years and also will return to the dust momentarily, like that stuff can do it. These frames that we have, we have bought into. I mean, I think, I think, I think this idea, this, this idea that that what, that something you know about your physical material is going to radically change the way you see yourself and the and the future, um, and what you're capable of. The fact that we've all, um, bought into that is the most effective piece of propaganda I've ever seen. I don't know when it, it shows up, but everybody tends to believe it and and they're willing to overthrow. I mean, I mean, Descartes, many, many problems with Descartes, but this, this part I think he got right, where he was saying, you don't know anything. What you know is your primary, you know, your primary perception. Because think of the other way it could have gone, like, right? It could have been, you open up your arm, you, you find out that there's cogs and gears, you go, "Amazing! I just learned that cogs and gears are perfectly sufficient for this incredible, rich inner life I had. Yep, moving on. You know, going to go call my friend." Like, all good. So, right? That's that's the other way it could have gone. Why is it that, why is it that we suddenly think we've learned something very disruptive about ourselves and our potential, instead of learning something moderately, only moderately disruptive about cogs and gears? I mean, who knew? Who knows what cogs and gears can do? No, nobody does.

Um, what I want to be clear is that whatever turns out to be the story of our past, I don't care if we were or weren't agential in the past, or if the DNA did this or that. Right now, sitting here, we all know we have the potential to to reinterpret that story, to do, to do amazing things going on into the future. And we should not let those stories constrain us. None of us can control the next thought we are going to have. We don't, we don't pick our next thought that happens to us. However, by effort, and that means education, meditation, anger management classes, you know, relationships, whatever you're going to do, you have the ability to change the distribution of future thoughts you're going to have. You can have better thoughts, kinder thoughts, you know, what you know, more interesting thoughts, whatever, by steps that you take right now.

So, this is what I'm, what I'm saying is that in the past, our bodies and the mind were shaped by, you know, pressures of the savannah and the meteor strikes and who, all this stuff that doesn't care about any of the things we care about. But looking into the future, we now have the ability to be more intentional about what that looks like. And we should do that both in our personal life and as a society. Of course, we're going to make mistakes. But I mean, you really think we're going to do worse than than random cosmic rays and meteor strikes? I, I suspect we can do better than that. Even if we, you know, don't do a perfect job of it. I, I think we can do better.

>> What are hybrids, chimeras, and cyborgs?

>> Hybrids and chimeras are, we can take frog cells and axolotl cells and and put them together and you get a frogalottle. They collaborate perfectly well. You can take mechanical parts and put them onto a body and you'll get a cyborg. So these are humans with implants and various other kinds of things, you know, whether it's mechanical limbs or or brain implants or whatever. You can have hybrids, which is kind of the opposite, which is mostly a robotic body with some biological tissue might be a brain, might be cells, might be a slime mold that are incorporated into that material. And people have made hybrids where, you know, slime molds or mice or something are running some kind of a robotic body, right? And the reason they're important is for, is for two reasons. One is because when people say, "That's a machine, I'm a real human." Yeah, that's easy to tell when it's 100% one or 100% the other. But what I think we don't want is for you to look at your neighbor and say, "I know you've had some brain implants, but, you know, if, if you've got more than 51% of of like mechanical stuff in your in your head, I'm not going to treat you with the same respect and compassion. You're a machine now. You're some sort of hybrid with 49% living cells." Like, this is what we're facing. It isn't the, the issue is going into the future is not going to be about language models sitting in the server somewhere. The issue is we are all going to change both biologically and, uh, and technologically. And we need to get over this idea that there's some kind of hard category between living things and machines that we are going to try to base legislation, behavior, compassion around this, around this thing, right? So, so that's why that's why those are important.

>> Our mind blindness, do you think it could be an evolutionary adaptation?

>> I don't know if it's an evolutionary adaptation, but I certainly think that up until now, we have not been under any pressure to see those other minds. You are advantageous if you can eat the other animals and not worry about it and and in fact save your energies for very mundane things and not ponder the trainability of the weather or any of these kinds of things. But those days are over. And and we can see this in terms of the kind of crises that we're going through, including ecologically and and socially. We are to the point where we really do need to understand what it means because of AI and biotechnologies. We've come to the point where, no, now it's pretty crucial. And evolution did not prepare us for this. That much is clear.

>> And you say you often get asked where this bottoms out.

>> Yeah. Because >> it's not clear. >> Right. So I, I get asked two things. One is, how do you avoid the slide towards panpsychism? And my point is, I'm not trying to avoid any slides. Wherever the data lead, that's where we slide. I'm not trying to avoid it. Straight up, I am a kind of panpsychist, not the usual kind, but a different kind, but still, yes, it's fine if that's where we slide, that's where we slide.

>> Panpsychism is the view that consciousness or mind [music] is a fundamental feature of reality that exists everywhere, even in particles. I don't see goal setting anywhere else in physics, and maybe I'm just not seeing it. I don't see anything with behavior like agency and behavior, or is that wrong?

>> A couple things that's interesting about agency in physics. First of all, the tools that you use have to have a certain kind of, in in electronics, you would call it an impedance match. It has to have a kind of resonance with what you're looking for. In physics, all you use are very low agency tools. You use rulers, voltmeters, the cameras. These are very low agency kinds of things. Well, guess what they find? They only find low agency phenomena. You, you need a mind to see other minds. You need some degree of mind to see some degree of other minds. And physics is what we call that branch of science that uses low agency tools to see low agency stuff. So, it's not a surprise that physicists don't generally talk about agency because that's what the science is about. Except, but having said all that, you might be asking, does the spectrum of mind go all the way to the bottom? Where does it stop? Right? When we look at the kinds of things studied in physics, meaning the very small and very simple systems, we necessarily have to scale down our expectations. I'm certainly not arguing that simple physical systems are going to have the same hopes and dreams that you and I have. Okay? But by definition, my my story is about scaling. My story is about a continuum of scaling. So when you look at the left side of that spectrum, you are not going to see the kinds of things you see on the right side of the spectrum. You have to scale down your expectations. If you scale down your expectations, you can ask, okay, what does a minimal goal look like? What does the smallest possible, the dumbest, simplest version of a goal look like? And I argue that it looks exactly like what's called least action principles. Everything in motion from a falling rock to an orbiting planet follows the path where action is minimized. Nature acts as if it's solving an optimization problem. That's what physicists call the principle of least action. It's a fundamental principle of reality. This is like the work of Karl Friston and other people like that. You can use the exact same mathematics to describe what's going on in those extremely humble systems with very minimal capacity, not zero, but minimal capacity to reach those goals. You can use exactly the same mathematics as you use in active inference in human brains and other kinds of brains.

>> So again, active inference is how your brain builds and updates its mental model of the world to anticipate the future and avoid destabilizing surprises. [music] Instead of just reacting, your brain predicts when you might get hungry, cold, or in danger, and it acts to keep you alive. Leven's pointing out that the same mathematical framework underlies both active inference in human brains and the principle of least action, one of the most fundamental principles of the universe.

>> Now, I asked Chris, is it possible to have a universe in which least action principles don't exist? Could we have a true zero? What would it take to have a world with a true zero of cognition? And he said, and so I'm leaning on him 'cause he's the physicist. He said that basically the only way that could happen is in a world where nothing ever happens, purely static world. So, so at this point, and again, this is not something that everybody agree that most people [music] agree with. I think in our universe, there is no zero. I think that our universe begins with a a little bit of a a hop onto the continuum. And that what life is really good at doing is scaling up the cognitive liones of parts, those tiny little extremely minimal degrees of protocognition, scaling them up and up. But I don't think there is a zero in our universe.

>> So fundamentally, there is an ultra-base cog, a base cognition, like super, super minimal, base cognition fundamentally in the universe, and then we happen to be in this little corner, and perhaps there are others where it's just scaling through biology.

>> Yeah, I think we call biology the process of the cognition scaling up from from these very simp, which is which is why I'm not super into definitions of life or anything like that. I think cognition and the scaling of cognition is much more interesting than trying to set a boundary for what we call life.

>> Do you think maybe one of the things we have to get our heads around is that a goal can be to take to grow to a physical form, a physical shape?

>> Yes, we need to really understand that the notion of, and I mean this is again, prior to, I don't know, the 1940s, this would have been really hard. But we have cybernetics now. We understand that goal-seeking machines are not magic. We, we now have a mature science of goal-directedness. There's a lot more work to do, but we now know it doesn't have to be magic. And goals, so this is really critical because this gets to the AI thing too. Goals can exist in many different problem spaces. So when people say, "You need embodiment for, let's say, intelligence." Yes, you do. But that embodiment doesn't have to happen in three-dimensional space. It doesn't mean that you're a robot that runs around in 3D space. You can also have a perception-action loop in all of these other spaces, and anatomical state space, and physiological space. We, we are fixated on three-dimensional space because that's where our sense organs are pointing. But we need to get beyond that.

>> Perception-action loops [music] are feedback cycles where you sense, decide, act, and sense again. It drives adaptation as agents react [music] to their environment and adjust their behavior. Leven saying this process isn't limited to beings with sensory organs operating in physical 3D space. Cognitive agents can have perception-action loops [music] in other contexts where they sense, evaluate, and act to maintain or achieve goals within environments that are so foreign to us, we're effectively blind to them, even if they're happening inside our technology or inside our own bodies.

Did this all sort of start with you, like the understanding of these issues when you started to see cellular cognition and intelligence? Is that sort of the beginning of your worldview?

>> For me personally, >> for me this started when I was about six years old and it started because this was, um, I, I had asthma and I would have these attacks and we didn't have any access to medicine where I was.

>> Where were you?

>> Uh, Moscow in the USSR. This was 1974-75. So the deal is that your airway starts to close up. When your airway closes up, you can't breathe. You get nervous. And when you get nervous, it closes up more. So there's a feedback loop. And because we didn't have proper meds, the only thing there was is to distract the kid so that it calms down. My dad would take the back off of this. We had this ancient TV set that had vacuum tubes and everything. And he would take the back off so to distract me from the asthma attacks. So he was trying to find ways to distract me so that I forget about the fact that I can't breathe. It would take a couple hours for it to pass and we would just sit there staring at the vacuum tubes glowing and look at the front. The cartoons are coming out the front, but in the back, there's all the stuff that makes it happen. I would ask him, "So clearly these things are not random. How, who, who knew how to put this together?" And he said, "There's people called engineers that do this." I said, "Can anybody learn to do this?" And he says, "Yeah, you can learn to do this." And I said, "That's amazing." And so I wanted to do that. That's when it really hit me that you go outside and you see insects and animals and you say, "Okay, so what's the same and what's different? Who put them together? And why doesn't the TV seem to care what happens, but they seem to, where does that difference come from?" I was interested in the stuff like extremely early and trying to understand what made the difference.

>> You're doing that while you're trying to breathe.

>> Correct. The good news is that you forget that you can't breathe because if you're interested, and then luckily I was interested in this stuff, but if you're interested, your mind goes to that, not to the, "Oh my god, I can't get the breath in." [music] So, [music] so you've also found unexpected cognitive behaviors in simple computer algorithms. [music]

>> We looked at unexpected problem solving and goal-directedness in the most, in the simplest things we could possibly find. What computer scientists call a sorting algorithm. It's six lines of code. It's completely deterministic, completely transparent. Computer science students have been studying this thing for decades. Everybody thinks they know what it does. We ask the question behaviorally, if you challenge it with new problems, what does it know how to do? What does it want to do? And we found new competencies that are completely not obvious from the algorithm itself. And if something that minimal and that dumb can do interesting problem solving that we cannot anticipate, what are all these more complex things doing? And this is, this is where this, this goes into AI. People focus on this idea that, okay, here's an algorithm and I could see what it's doing. It's, you know, completing the next word in a sentence or whatever. That's not real intelligence. Well, yeah, that's the algorithm you see it doing. And maybe even that's the algorithm you wrote. That doesn't mean that's actually what it's doing. And we, we have a very poor still understanding of minimal both physical systems and algorithmic systems where we project our, you know, this is our story of what we think it's doing, and we forget that that's just a story. And there are perhaps other better stories that might capture new aspects of the intelligence.

>> So the algorithm is bubble sort. Most people don't know what bubble sort is, but if you can explain just what that is and the three incredible things that you discovered.

>> Here's okay. Uh, the, the views that I have get pushed back from from two directions, right? The, the, the reductionist, um, mechanist, computationalist kind of people say, "All, like, all of it, including us, are are machines. And why are you trying to paint these cognitivist terms on these things?" Then you've got the organicists who go in the opposite direction. They say, "Look, we know life is special. It's magic. And you, by putting computers and and other things on that same spectrum, are cheapening the magic of life. You're going to bring this mechanist metaphor to all of life." And I'm saying, "No, no, I'm going in the exact opposite direction. What I'm pointing out is that perhaps even the machines are not properly encompassed by our simple mechanical stories. Maybe nothing in the universe is a dumb mechanical machine. And maybe those stories of things that only do what the algorithm says. Of course, they're not sufficient for humans or for living things. They are also not sufficient for machines." So, how would you prove that? So, I wanted some shock value to this. And so I try to, uh, I try to find the dumbest, simplest, and, and, by the way, you'll see in the next couple months, we'll have a couple more papers on something even simpler. I, I thought I found the simplest possible thing, but apparently I didn't. There's, there's something even simpler that works like this, which is amazing. But I tried to find the dumbest, simplest thing that I could say, "Look, here, even this thing has surprises that you didn't see coming." Not just complexity, not just unpredictability, but things that would be recognizable to a behavioral scientist. Okay. So, so >> agency >> a a degree, a a degree of a degree of agency. Okay. A degree of agency. So, so, so what I, so what I chose was something called sorting algorithms, of which bubble sort is just one example. There are many sorting algorithms. Sorting algorithms are a very simple set of steps. Okay, there's only a few of them. There's only five or six steps. They are completely deterministic. There is no randomness, no deviation from the rules. And what they do is they take a jumbled up, um, array of numbers that's disordered and they slowly, by following the rules, they will slowly swap everything around so that eventually it's it's ordered from from from the smallest number to the largest number. That's it. People have been studying these sorting algorithms for decades, probably 60 years or more. Every, uh, undergrad in computer science and CS 101 studies these sorting algorithms. Tons of very smart people have been looking at these things, but with the idea in computer science, the idea is this thing does exactly what you asked it to do. So, we looked at it from a slightly different perspective and we asked, "What might it be doing that we didn't ask it to do?" We figured out a way to let the pressure off of the algorithm a little bit without changing it. That's very hard, actually, but, but we found a clever way to like put the, take the pressure off of the algorithm a little bit. And what we find are novel behavioral competencies. For example, it does delayed gratification. It can actually desort the array for a while in order to recoup better gains later. The algorithm does not ask it to do that. Okay, there's nothing in the algorithm for that. And the other thing that it does is this thing called clustering, which is basically a side quest. It's like, "Yeah, I'm going to sort the numbers for you. All right, but while I'm doing that, I'll do some other stuff that isn't incompatible with what I asked, what you asked me to do, but neither is, so, so it's not forbidden by the algorithm, but it's not prescribed by the algorithm either." It's like when you force a kid to do certain chores, they will do them, but at the same time, they'll make a game of it or they'll do some other stuff that kind of, these are intrinsic motivations. And these intrinsic motivations show up even in the simplest, most deterministic machines that we did not anticipate.

>> So, so I keep hearing everyone talk about we have to stop or, or, um, slow down the charge to build superintelligence because at that point, we won't be able to anticipate what it's doing. And when you're finding these behaviors in such simple algorithms, these simple six code, uh, what does that mean for what we're building right now with AI?

>> Okay, the extra things that it's doing, these side quests are not particularly related to the main thing we ask it to do. They have nothing to do with sorting. And so that suggests that when a system like a machine is doing things that are not the things you've asked it to do, you can't just look at what the algorithm itself is doing. What I mean is when people say, "Oh, these language models, they might be agential and whatever." When you're looking at the language it's using, that's a complete red herring. You forced it to speak. That's not the interesting part. The interesting part, and where I think the agency comes from. So here's my crazy view on this. There, there are people who say these things have no cognition whatsoever. There are other people, computationalists, who say they have cognition because of the algorithm that they're following. I have a third and crazy position that I've, that I haven't heard anybody say before, which is, they, they may well have have cognition, but in spite of the algorithm, not because of the algorithm. They are not conscious because you've made them speak. The language is is irrelevant. But perhaps completely, maybe, maybe a little irrelevant, maybe completely irrelevant. Asking these things whether they're conscious, whether they feel, whatever. Total red herring. It does not matter. It's not that it's not the things you forced it to do that that are at at play here. It's the things that it might be doing despite what you forced it to do in the empty spots between the things that you've asked it to do. And we are very bad, as a community, at noticing those. Nobody noticed it. How come nobody noticed it in the, in the dumb sorting? Nobody noticed this in in decades because we have these blinders on. We assume that that our formal model tells the whole story. It doesn't do that for us. It doesn't do that even for these simple things. So my point is this. I don't believe that language models, I don't believe they have a human mind. I don't know what kind of mind they have. But I'm quite certain we need to investigate it with scientific methods. And I'm quite certain that, uh, just watching the language output is not how you do that. The language, the language that they do could be a complete red herring as far as what they're actually doing. What do they actually want to do? So my point is, if you, we have now a way, an actionable scientific research program to see what does, what, what are the intrinsic motivations of these things? What else do they want to do besides the thing you're forcing them to do? And that is absolutely critical, I think, to understanding alignment and everything else.

>> The implications seem massive, and I don't know anyone really looking at this.

>> Yes, I, I think that's, I think that's right because the field is dominated by those two views. Either it has nothing because algorithms are faking it, and it has something because algorithms are the real thing. And I'm saying both of those are wrong. Both of these are off-ramps to understanding what's going on. You have to take seriously, it's not the algorithm at all. It's not the materials. Just, just like for us, it's not the materials. It's not the algorithm. It's not the evolutionary history. This is where again, you get into the Platonic space because you say, "Well, then where the hell do these things come from?" And you say, "Ah, now we have a way to map out the space from which these are pulled."

>> Can you spell out the implications?

>> Okay. Never, never mind today's AI and the thing we have now. Let's just think more deeply about the space of possibilities and diverse intelligence. So all of the different future synthetic biological kinds of organisms, hybrids, cyborgs, chimeras, combinations of living material and designed engineered material, different kinds of AI architectures that much more than today's language models will be based on biological principles that we discovered. There's an enormous option space of diverse bodies and minds. We are going to be living with these things. There is no doubt they will have different degrees of intelligence and different degrees of agency. And talking about these language models is a distraction from this much bigger question because it's very easy to say they're not embodied. Although even that, I think, is wrong in a couple of different ways. I actually think they are embodied.

>> How? We are a kind of body for them because we do things. We move massive amounts of energy and money and everything else based on what they do or don't say. So we are a kind of body for them. These things we build, these software AIs, yeah, they don't roll around in the world. Some of them now do, but most of them don't run around and interact with physical objects, but they do navigate other problem spaces that are very hard for us to visualize. Somebody said to me once, "Well, I make these language models. I know exactly what they do. I program them. I, I know what they do." I said, "You don't even know what bubble sort does. It's six lines of code and we found new things that nobody had seen for decades. If we can't predict what bubble sort's going to do, you have no idea what this other thing is going to be capable of. We, um, we really need a lot of, a lot more humility about the idea of what matter does when we say this is just a material, this is just a machine, and what simple algorithms do because I don't think that what they do is necessarily just what we think they're doing. And we already have plenty of examples of that.

>> What do you say to the argument about autonomous AI systems that goes, uh, people worried about AI? It's, it's, it's ridiculous because they don't have drives. They don't have the same drives we do. So they're not going to try and dominate us 'cause that's a human drive, and they're not going to try and destroy us 'cause they're machines and we program them. Clearly they don't have human drives. I mean, not clearly, they probably don't. They're not human. [clears throat] But do you think that that, uh, is a frivolous way of dismissing this, considering we don't know what they could develop?

>> I think, uh, there are a lot of logical leaps there. First of all, things that are not human, don't have human drives, can still be extremely dangerous, right? There's, there's no reason to think that just because something doesn't have a human-like mind. Okay? I mean, viruses are autonomous in many ways and they are extremely dangerous, and they don't need a lot of agency to be dangerous. I mean, we've been making systems with autonomous, autonomous behavior for years. It doesn't mean they have to be human-like. It doesn't mean they have to be highly intelligent in order to be dangerous. Now, I do agree that it's likely that the kind of intelligence we're dealing with here, it's not similar to to human intelligence, but I don't think that's important in the slightest. I think that there are so many different ways of being intelligent, many of them dangerous, and many of them not dangerous, and many of them deserving of compassion and whatever. Being like a human is a bad criterion here.

>> Interesting. Lately, I worry that misunderstandings about evolution could obscure how and how quickly minds emerge and evolve. 'Cause I was talking to people who are doing origins of life, abiogenesis, and evolution. Those are different disciplines. The evolutionary people, I kept saying, "When we find out how life emerges, won't that impact your theory?" Because right now, we don't know. And the people in origins of life, they're not buying gene-centrism.

>> There's no getting around it because because they're working at a time period when genes didn't exist. So it's, it's inevitable. This, this bothered me for for for decades. This bothered me. Just imagine planaria. They are immortal. They don't age. At least the asexual forms don't age. They're cancer-resistant. They are incredibly regenerative. You can cut them into, you know, 275 pieces, I think is the record, right? They are also also something something weird. Um, they are incredibly resistant to transgenesis, meaning that for any other type of animal, you can call the stock setter and you can get a fly with curly wings or a mutant C. elegans with some kind of weird behavior or something, or chickens with funky toes and mice with kinky tails and things like that. You can get these genetic lines. There are no genetic lines of planaria. There is no abnormal worm that you can get as a mutant line, except for our tube-edited form, and that's not genetic. So why is that? Right? And especially worms do this really interesting thing. The way they reproduce is they tear themselves in half and then they regenerate. Each side regenerates. So that means unlike us, if you have a mutation in your body, your children don't inherit that mutation, right? But in planaria, they do. Any mutation that doesn't kill the stem cell gets amplified as the animal regenerates into the next generation. So their genome is a complete mess. For 400 million years, they've been accumulating mutations. Isn't it a scandal? I, I think it is, but no, you know, nobody ever talks about it. You certainly don't, don't, I've never heard about in any biology class I've ever had that the animal with the noisiest genome actually is the most immortal, cancer-resistant, uh, and, and, you know, damage-like. It should be the exact opposite. If, if the genome is what you need to make a proper body, it should be the exact opposite. The animals with with the cleanest, most protected genome should be the ones that have these properties. What's going on with planaria? It drove me crazy for decades. And I think we finally have begun to understand it. I mean, it's, it's an incredible, uh, it's an incredible thing. And so, and so here's, here's, here's what I think is going on. And this is where the intelligence ratchet and that scaling up comes from. Let's, let's just do a simple example of a, of a tadpole. In tadpoles, if you deviate the mouth, as, as I said with these craniofacial things, if you, if you deviate the mouth off to one side, it actually will eventually come back to where it needs to be. So imagine an animal like that. We call that competency in terms of solving a problem. You're in the wrong region of anatomical space. You need to get back to the right region. You can solve it by moving your mouth where it needs to be. Imagine that you have a mutation in this kind of organism. Pushes the mouth off to the side, but there's some other beneficial effect of that mutation in an animal that did not have that competency. It would die because the mouth is off. You can't eat. You would never be able to explore the beneficial consequences of that other mutation. But let's say you're in a competent organism. The mouth comes back. By the time you come up for selection, selection can't tell. I mean, you look pretty good. And do you look good because your genome was amazing, or do you look good because the genome is actually so, so, but everything got fixed? Selection doesn't know. So that means, can't tell. So that means selection, if you're a competent organism at fixing these things and adjusting to circumstances, it means that selection has a bit of a hard time seeing exactly which genomes were actually the best. They're masked by this behavioral competency. It's the reason it's the same way humans do. We have glasses, we have canes. You know, if your genetics aren't great, we can, we can make up for all that stuff. That's, so this is, we didn't invent any of that. Evolution's been doing that for a long time. And so your fitness, you know, if you can have glasses and if you have some metabolic disorder, you stay away from specific foods, you're good. You can have offspring and and it's fine. So selection then can't see your genetics very well and so it has a hard time improving based on the genome. What ends up happening is, and we've done simulations of this. We have a couple papers showing the computer, um, simulations of this process. What happens is that it spends all of its time perfecting the competency because it can't, it doesn't, it can't see the structural genome, but it can improve the competency. Now, what happens as soon as you improve that competency? Well, that makes the problem even worse. Now you really can't see the genome. And so you end up with this spiral where the, where the competency goes up, the, the, the quality of the genome flattens out, but it doesn't matter anymore. And so now, look, look what happens if you take that to its logical conclusion. If you let that feedback loop run all the way forward, I think that's what planaria are. And I think salamanders are sort of almost that way, and and mammals below that, and like C. elegans, you know, the nematodes are maybe on the left of that spectrum because they're very, um, kind of cookie-cutter organism. What happens there is, is basically a system that assumes the hardware is unreliable. You know, from the start, your genome is noisy. Your hardware is going to be unreliable. What you do have is an algorithm that has been honed over 400 million years to be able to get to the same goal, no matter what your hardware looks like. It is this idea that we are not going to assume our hardware works. We are going to assume there's going to be noise and junk. And what we're going to do is have an algorithm, an error-correcting intelligent algorithm that can cobble together a good solution. Then, in that case, of course, it's hard to make transgenic planaria. They ignore these kinds of genetic changes. It's very hard unless you target the actual competency process itself. And that's what we did with our biomimetic targeting.

>> What's the role of [music] bioelectricity?

>> Here on Earth, bioelectricity is an amazingly useful way that evolution has found to scale up the intelligence of individual cells. There are other modalities, and no doubt in other environments, there are other pieces of physics and other policies that that do that. But at least here on Earth, that bioelectricity is the cognitive glue that binds individual cells towards goals, towards larger scale anatomical goals.

>> Bioelectricity is the electrical signals that flow through living tissue. It appears [music] to play a much larger role in cognition than previously thought. Instead of thinking [music] being something that's confined to brains and neurons, electrical patterns across different bodily tissues appear to contribute [music] to memory, decision-making, and information processing, making cognition more of a whole-body electrical phenomenon [music] in all cognitive creatures from humans to worms.

>> I think planaria basically hold the secrets to all the big questions because they're telling you what happened. This, this feedback loop where intelligence rises because you can't trust your substrate. That is fundamentally what evolution is facing. In evolution, you know, the future is not going to be quite like the past. Then not only will the environment change, but your own parts will change. Your own, all your stuff will get mutated. There's many examples I can give you about creatures coming into the world in really weird configurations and figuring out a way to do it. It's not because this is some crazy new capacity they have. It's because all life faces this problem. I'm in the world. I can't really count on what happened before. All the memories I have, genetic memories and behavioral memories, are up for reinterpretation at any given moment in time. And so I need to tell the best story I can tell right now with the parts I have. I need to play the hand I'm dealt. So this is that biological ratchet that eventually gives rise to behavioral intelligence. It's exactly using the tools you have in new ways to solve problems. That's what we call behavioral intelligence. So I think that all of the story of evolution and what happened before genes and before traditional evolution is the story of scaling of basil, very basil competencies that exist in really simple things. So people are now studying minimal matter systems. So, you know, systems of three chemicals that make these droplets that navigate mazes and they do all this crazy stuff. It really doesn't take much. So the reason the two-headed worm is important because we started there by asking about what's evolution really doing and what's the role of the genome. If you ask what determines the number of heads in a planarian, you might initially think it's the genome, but just like the electrical network in your brain is processing memories about, you know, the layout of your apartment, remembering your goals, making plans about where to go. Those types of processes did not just suddenly spring into action when brains came on the scene. Brains adapted all of that from ancient evolutionary usage of electrical networks to process information about shape. And if you mess with it, what you can do is confuse it so that it will go to other areas instead of the area where it's supposed to go. Now, guess what those other areas correspond to? Many of them correspond to other species. So, you can take the exact same hardware, no genetic change whatsoever, just manipulate the electrical network as you could in the human brain and and it will end up in a different attractor in that space that corresponds to another species. So what, what you get are planaria with round heads, with flat heads, with brain shapes, with distributions of stem cells exactly like existing other species. You can also actually chase it into regions that don't correspond to any species and you can get crazy-looking things that don't even look flat or don't look like planaria. But that hardware is perfectly happy to visit these other, these other attractors. But altering that memory is now permanent. That memory is is there in the electrical circuit. It propagates across the pieces and they will forever, you've made a new line of animals that has a different anatomy than the previous ones. So, so just imagine we take these and we would never do that, but imagine you take these animals, you throw them in the Charles River here. Couple of decades later, some scientists come along, they scoop up some samples and they go, "Oh, look, a one-headed form and a two-headed form. Cool. A speciation event. Let's see what gene mutated to make this happen." That's not where it is. They, they will find nothing. Now, species is a funky thing. If you define the word species according to the genetics, then you automatically preclude this kind of result because you say, "Well, that's not a new species." If you start off from the very beginning by saying that species are only defined genetically, you're never going to find this. And if you do find it, you're not going to think this has anything to do with with other species.

>> So, they become deformationations.

>> You can, well, you can say it's a deformation, but every species is a birth defect compared to the parent species, right? Every new species looks wrong relative to the parent species. So, is it a birth defect or is it another species? Well, that depends how you look at it. And depending on how you look at it, that determines what you do next. And so, and so this is, this is really important. Two, two-headed worms were first described, made by a different method, uh, around the turn of the last century. So, 1905 or something. Uh, between that and 2009, when we first recut them, to my knowledge, nobody had recut them again to see what would happen. 100 years. Why? Because this, this goes back to our initial conversation about scientific metaphors. Because if your metaphor is that the genes determine the outcome, well, it has normal genetics. Why would you bother cutting them? You know what's going to happen when you cut that crazy second ectopic head off. Then it'll just go back to normal because it has the normal genome. Nobody had this idea that, well, maybe it's a memory. And so maybe the memory stays. To my knowledge, uh, they have never been recut. And so this is important that that's what, that's that's an example of how thinking in in these different ways leads you to new discoveries. These metaphors, all, all of our scientific structures are metaphors, but they enable or constrain different things.

>> That makes sense because I don't know how you do this work or any of this work without having an attached reason or speculation of where it came from. Like, even through process of elimination, you've got to have some way of thinking about where this came from in order to do this kind of work, right?

>> Yeah, I completely. I mean, the standard, well, let's look at what the standard alternative is. So, so the standard alternative is the current paradigm of emergence and complexity. So the idea is, and, and it's true, I'm not saying this doesn't happen, but, um, what, what people are focused on for for for decades now is this idea that in many cases, when you have very simple rules, let's say cellular automata or or, you know, like the game of life, things like that, when you have very simple rules, you let them go, often you get a very complex outcome. And so the paradigm is, okay, simple rules give rise to complexity. And, uh, maybe in some cases, we can, we can predict.

What's going to happen? In some cases, we can't. But that view, this idea that it's completely feed-forward, that you just crank, you turn, you turn the crank, the rules act, and eventually something will happen. That has a very specific implication. That means that if you want to make changes in what happens, you have to go at the, you have to go at the very beginning. You have to change the rules, right? The initial conditions, the initial configuration of whatever you're dealing with. You have to, that's where you have to act. In the biology case, it's the genome. This is why everything is going towards genomic editing, you know, things like that. You have to change the low-level rules.

Now, the problem with this is that going backwards, it's very easy to go forwards. You can do a simulation and you can say, "Okay, with these rules, this is what happens." But going backwards, if I want this other thing to happen, what should the rules be? That is incredibly difficult, and in most cases, it is completely intractable. It's not, these kinds of processes are not reversible in that way.

So this is why I think that a lot of the current excitement about CRISPR and genomic editing, all of this is going to plateau. And it's going to plateau because after we've solved a lot of the single-gene diseases and things like that, then the question is going to be, "Okay, well, now I want to make these large-scale changes. What genes do I edit?" We have absolutely no clue. And, you know, there are ways to discover that in certain cases, of course. That's what developmental biologists and bioengineers do. But in a large scheme of things, that is an incredibly difficult way to deal with complex systems.

I don't think that's what biology is doing. And I don't think that's what we as bioengineers in the future are going to be doing. Because I think the main thing here that comes out is not just emerging complexity. I think what comes is emerging cognition.

Ideas that challenge core assumptions central to a scientific paradigm are a gamble where you run a high risk of being wrong while needing institutional buy-in to fund the experiments that could, if you're proved right, upend your peers' worldview and research programs. The stakes are even higher when your experiments [music] challenge our perceived monopoly on intelligence.

So, like, how was your idea to test this stuff? And you say it's been tested before, but how was that received? How did you get buy-in from people you needed to do these experiments?

I don't have buy-in to do these experiments. If I waited for buy-in on these ideas to do experiments, we wouldn't have done a single experiment. There is no buy-in before you do the experiment. If you're going to get grant funding, you have to couch the experiment in very boring, very conventional terms so that it's not scary, so that it's not freaking people out. You don't come out and say, "I want to blow up this paradigm." You describe it in a way that looks like incremental science, good incremental science. You may or may not get buy-in. I mean, I have no idea what the level of buy-in nowadays is to these ideas. I think among professional scientists, it probably isn't super high. I think a lot of people are very resistant.

But where it's getting some very good buy-in is the young scientists, right? So, most of my outreach is targeted to PhD students and postdocs and, you know, and below. These are the people that are not, they're not yet fixed their research programs. They're not funded for the next 20 years. They're not yet mentally on any kind of railroad tracks that they have to stay on. And so that's where I hope the buy-in comes. But looking for buy-in to these things before doing experiments is not going to work.

And this is a common reaction that I get. So often times I'll give a talk and people will say, "Wow, the results are amazing. The new capabilities are amazing. We're glad you're going to the clinic with all this stuff, but I sure wish you'd stop talking about all this philosophical business. You don't need it. Just do the experiments. Like, don't do all this philosophy stuff. Don't talk about cognition. Just do the experiments and do the biomedicine."

And that is completely wrong in the sense that we wouldn't have done any of these experiments if we didn't have these ideas. Why hadn't anyone else done the experiments that we've done? Precisely because they were using a framework that does not facilitate these experiments. It may facilitate other experiments, but the things that we've done would not have been done if we didn't have this particular weird way of looking at it. And so I think the philosophy is actually really important here.

Well, I'm sorry, I'm a little confused. Is it even philosophy? Because it sounds like, like, how, what's the other explanation for what's happening there?

When I say philosophy, I don't mean some kind of airy-fairy, impractical thing. To me, philosophy is being conceptually clear in the questions that you're asking and specifically looking for the blind spots in your framework. So, science is done by taking a framework and using that framework to make new discoveries. Philosophy is the part where you say, "Okay, given this framework, what would I not see? What are the things that I'm missing if I use this framework? What kinds of things that are crazy according to this framework might actually be part of reality that I would miss? And if I wanted to see them, what would I do?"

This is certainly not a new idea with me. Alan Turing was very interested in mind and different implementations, right? So, he talked about machines that can think and he was interested in computation and reprogrammability. He wrote a paper on the self-organizing chemistry in embryonic development. Why is that? Why would a guy who's interested in computers and mind and artificial intelligence be thinking about chemicals in the early embryo? And I think he had the same insight, which is that the formation of the body and the formation of the mind are not separate things. They probably have a lot of important symmetries. And that the kinds of things that we study in cognitive science, memory, decision-making, learning from experience, preferences, navigational competencies within some kind of problem space, these are incredibly generic things, especially all through the universe. You don't expect every intelligent being in the universe to have a cortex like ours or to look anything like us. So what are the general, at least I don't. So, what are the general principles here? And so this is why you want to take those kind of tools and apply them to other things.

So you said a couple things there. A, this is not the first time anyone's thought to do this. Uh, given the extreme reaction I get from even scientists, is this something that's accepted in developmental biology, synthetic biology, and possibly not other departments? Because I think a lot of people would think, "Oh, it's crazy to assume cells might display cognitive behavior."

Oh. Oh, I certainly don't mean to say that this is accepted in developmental biology or molecular biology. Absolutely not. This is not accepted almost anywhere. But I should make the claim very clear. My claim is not that you can assume they have these things. That isn't my claim. You can't assume anything. My claim is that what you should be doing is good science, which is making hypotheses and testing them. What I don't like are frameworks and definitions and dusty old categories left over from pre-scientific times that prevent you from doing new experiments.

If you assume out of the gate that looking for cognition outside of brains is some sort of category error, this is a common thing that philosophers will charge us with as a category error. Well, guess what? Categories are not these things that are given to us by God and then we need to stick with them no matter what. Categories have to change with the science. Categories should be enabling. They shouldn't be constraining. So I don't buy any of this category error business. And so if you have this background assumption that you shouldn't do that, then you would never do the kinds of experiments that we and others have done where you take those tools, apply them, and actually make progress and discover new things. It's not an assumption. It's a recognition of a symmetry in thinking that leads you to new experiments.

And I suppose I want to break down what a category error is. Is it, um, you're challenging the assumptions that things fall into these particular categories?

Well, a category error, it's meant to be a specific misuse of a term applied to a category where it doesn't fit. The idea is that you would take a term that's widely used in one field and then use it for something that's obviously not the kind of thing that it's meant to apply to. That's what philosophers mean when they say category error. So if you say, let's say, a terminology of having goals, what does that mean? Well, if it means anything, at least the non-controversial version of it is that humans and maybe other animals have goals. And we can talk about, there's a whole philosophy and literature and so on around what it means to have goals and purpose and things like that. So, if you try to apply it to things that aren't brains or organisms, philosophers might say, "Well, that's a category error. You're taking a word that has a certain meaning in a certain domain and now you're misapplying it in another domain where it has no business."

That kind of gatekeeping assumes that from a philosophical armchair, you can tell what these categories pick out. And I'm saying, no, no, science leads you. You have no idea what the term actually means. I mean, of course, colloquially people use it for whatever. But we as scientists and philosophers, we're held to a higher standard. We can't just assume that the colloquial definition is the good one. We have to ask ourselves, what's the most useful definition? And the most useful definition for all of these words is one that lets you do new experiments, discover new things, and get to new capabilities.

So you were like, "I'm going to anthropomorphize because that is an assumption." I don't want to put words in your mouth, but is that sort of, and I want to define anthropomorphizing as taking human characteristics and putting them on things that are thought not to have those characteristics?

Let me give you a different definition. Here's how I hear anthropomorphizing. It's the unwarranted assumption that humans have a bunch of magical properties that you shouldn't look for anywhere else. That's anthropomorphizing. When somebody says you're anthropomorphizing, what they really mean is we have certain human magic, poorly specified. We don't know what that is. No one will ever tell you what it is, but we have some kind of human magic that you should be really careful not to see elsewhere because otherwise we lose our special place and it makes me feel uncomfortable. That's what anthropomorphic really is.

Do you find people are still trying to say, "Well, then that's just mechanical and it's not real learning. It's just what machines do." Is that the standard reaction you get?

That depends from whom. So, a vast number of people will say, "Yes, it's mechanical, and so am I, and so are you." If you zoom in far enough into a human, any kind of human behavior, guess what you see? You don't see fairies underneath. You see chemistry and you see physics. And so the problem is, in both cases, the issue is, did you pick the right formalism? Because taking the molecular perspective and saying it's a machine when you're dealing with a human, congratulations. You found a perspective that basically makes all the interesting things about being a human invisible. I mean, of course, there are bad perspectives, and you found them. Great. On the other side, it's exactly the same thing. If you want to take chemical networks and treat them as machines, again, fantastic. But you've also missed a bunch of much more interesting things that you could get to, which, by the way, have biomedical relevance, than if you were willing to entertain other hypotheses. It's all about finding the right perspective.

If you just list the set of discoveries, people are going to say, "We can explain all this with traditional means." And that may well be true after the fact. It's a true and totally vacuous explanation of what happened. Looking backwards, it is always possible to tell a reductive story of biochemistry. To Monday morning quarterback and look backwards and say, "Well, I could have done it." Well, you could have, but why didn't you? Because certain framings facilitate certain kinds of experiments. You wouldn't do certain experiments if you weren't thinking about it a specific way. And if you really are stuck at, you're really intent on holding together that worldview, you can always add epicycles to sort of make it make sense. It's not whether you can shoehorn these observations into a view that you like. The question is, did your view help you discover these? Do they imply this research program after the fact? You can always tell a reductive explanation after the fact. But that's not the same as using it to discover new capabilities going forward.

Interesting. I wonder [clears throat] too, because I read recently that actually we've only mechanized all of physics in the last 200 years, and before that, through all of history, there was not that assumption placed on it that it's that distilled.

Yeah. I guess the question is, what, right? So, this is the thing you have to get rid of assumptions. And there's no point in having assumptions here. So we used to have for thousands of years, we had the assumption that there is a full-fledged spirit under every rock. Okay, there's a mind that you could talk to in every rock. So that was an assumption. Then the pendulum swung the other way, and we made the assumption, actually, there aren't minds anywhere except for in brains. And then some people said, actually, there aren't any anywhere. And so there are people who believe that. And all of those are equally bad. They're equally bad because they're assumptions. None of this should be an assumption. We should not believe that we know these categories. We do not understand the terms that we're using: machine, human, mind, organism. None of these things are given to us. We have to do experiments. When you do experiments, you find out that this term actually isn't what I thought it was, because the better way to use the term is in a way that applies to more things and gets us to more discoveries. This should all be an experimental science. These are not about assumptions.

So, is the outdated assumption just like, we think this is how the universe works? This is what we can say about this field. This is what applies. And in chemistry, you're not going to have behavior?

Yeah. To put it another way, it's a question of which kind of conceptual tools are applicable to your model. The standard way people think about this is the laws of chemistry, the rules of genetics, those are the things I am allowed to apply to this model. These other things, with cognitive science, I'm not allowed to apply. That's, I wouldn't even think of applying it. That's the standard model. Which bag of tools are you bringing?

And just a quick question, the reason isn't just arbitrary, we're not using these tools. It's we don't think they apply, right? There's no way that could apply to this scheme.

The assumption, yes, the assumption is that that is a category error, that it would not work, that it's just completely, completely different.

So for people who are not necessarily doing the experiments, so they're not invested in that way, but they want to have some attachment to their worldview. I want to give people something to, be able to go, "This is the evidence we see." And if you're a scientist, we have to deal with this and reconcile this. This is what we can kind of infer logically by process of elimination. Is we kind of have to accept this, and this is what's still speculation. Is there a way to compartmentalize these into these three sections?

Yeah. Although that stuff is extremely plastic because things that were speculation a year ago have now been shown. Some biologists will say, "Look, the best stories are told at the level of chemistry. You are positing all this extra stuff, and it's too like a weird, incorporeal thing that you're talking about." I like to talk about molecules, even though a lot of that stuff is metaphorical, too.

So, think about what this means in terms of a computer. You have a computer. The computer certainly obeys the laws of physics, right? There's no question it's not breaking the laws of physics. Now, if somebody were to say to you, "What's this? What's this mystical algorithm you're talking about? There's no algorithm. There's Maxwell's equations that tell the electrons where to go. That's it. It's a piece of physics. The physics does what it does. No algorithm. What are you talking about? Where is this algorithm?" And it's not that that's not one perspective to take on the computer, but if you were hiring for your software company, would that be the person you hire? Would they ever code anything?

In fact, I wrote a funny, at least I thought it was funny, little blurb on my blog about an interview of a guy who comes for a software position and he says, "I just read Robert Sapolsky's book. Absolutely, there is nothing but the laws of physics, and every event has a physical cause." And so you say, "Yeah, but you're still going to code for us, right?" You mean code? There's no code. The electrons do what they do. I'm, you know, we'll just see what happens. The universe unfolds, and isn't it beautiful? Okay, great. Does that person get a job? That view, it's not wrong, exactly, but it's sterile as far as what do you do next?

So, computer science is already committed to this idea that yes, there's the physics, but there's also this thing, and it's mystical, except it isn't, because we know how to write it. We know how to manipulate it. And that's what I'm talking about here. It is completely practical because it gets you to the next discovery.

And if you hadn't speculated about that, that would have been in the dustbin of history. Do you think speculation? I've been talking to so many scientists who are like, "It's very dangerous to speculate." And I understand why you don't want to run off on tangents or I don't know why.

Of course, you want to run off on tangents.

Okay. So, everybody's acting like, of course, you shouldn't do this. But at the same time, I'm like, well, if you don't, then you don't cut off a worm head again and find out, oh, it grows back too. And oh, this is actually going to lead to cancer research. And that, if you hadn't come along and tried that, we might not be on this other path. And I'm just wondering, what the hell is happening to us that we're just getting like, a hundred years later, 200 years later, we pick up this thing again and then we go, "Oh, this is a big deal."

Yeah, I mean, to be fair, so, so first of all, yeah, absolutely. I think if you're interested in doing new science, you should absolutely go off on tangents. That's how you discover new science. But we also need to understand that the vast majority of new tangents don't pan out, right? So, there are way more bad ideas than there are good ideas. And just, you know, a lot of good ideas are crazy, but there's a ton of crazy ideas that are actually not good. So, you're NIH, you have a certain pot of funding. What percentage of that funding do you want to fund for crazy ideas? I mean, definitely not zero, but it's probably also not 100% of it because, right, because most of them are, in fact, not going to pan out. So, there's some kind of balance. And every scientist, in their own life, has to make a decision for them. Are they going to do incremental work where they know where they're going? They know exactly what's going to happen. It's pretty safe. Or how much of their time are they going to take these leaps? I personally like the leaps. That's me. Some people do and some people don't.

But one thing for sure, we have in science is that it's not a sort of monotonic gain in wisdom. There are tons of good information and good ideas that are ignored and are not taken forward. Sometimes rediscovered, sometimes not rediscovered. That's a real problem. That is something that I'm very excited about language models and AI in general helping us with. Because regardless of their degree of autonomy, what they probably will be able to help us with is to have and to find ideas among an incredible historical body of work that nobody can actually know. No individual person can actually remember everything that was found. So I think they're a tool. I think, actually, for science, they're going to be a really interesting tool to help make sure that things don't get irretrievably lost. But science loses a lot of stuff along the way.

And based on your experiments at this point, it sounds like, and correct me if I'm wrong, that you could potentially do a whole lot more. Like, if you can explore all of anatomical space, you could potentially actually cure not just manage disease. You could regenerate limbs. You could do...

That's my claim precisely. And I've published smaller versions of this, and I'm working on a big thing now on future medicine. All of the things that biomedicine has done so far are one corner of the space focused on micromanaging and forcing the hardware, as amazing as it is. And lots of amazing medical applications have come. But it has many limitations in terms of cells fighting back and having side effects and drugs stopping working because you're micromanaging symptoms, because you're trying to clamp down symptoms instead of getting the buy-in of the cells that their goal is to build something or repair something.

So cell training, molecular pathway learning, rewriting goal states, managing stress distribution among subsystems of the body, all of these things are a completely different way to manage the health state. And I think it's going to be transformative for cancer, for birth defects, for regenerative medicine, for aging, for degenerative disease, not to mention bioengineering and making new things that didn't exist before. It's going to be a completely different frontier. And to do that, we do have to, on some level, whether or not we want to talk about philosophy, we have to look at your evidence and actually accept what you're seeing.

I don't think anybody needs to accept anything. Let's put it this way. All I want is for people to be committed to outcomes, not philosophical preconceptions. The rest will take care of itself. I could stop work tomorrow if people were really willing to commit that they want specific kinds of outcomes and they're willing to explore whatever it takes to get there. Then you wouldn't really need me after that. That would take care of itself. The problem is that there are huge corners of the space that are inaccessible because they are taboo. And not in some weird conspiracy way. Nobody's actually trying to block anything from anybody. It's just we've had several hundred years of thinking about things in a very particular way, which is extremely limiting. We need to break through that way. And if you want to make these new discoveries, you try new tools. I mean, that's all there is to it.

So, you're a pragmatist about this?

I'm a pragmatist about everything. Yeah. Hey. [music]