"Nothing in biology makes sense without teleology" by Michael Levin

If you want to download any of the primary papers, the data, the software, everything is free and open on our website, which is here. And then this is my own personal blog about what I think some of these things actually mean. So, okay.

I modeled the title of this talk based on this very famous quote by Dobzhansky called 'Nothing in biology makes sense in light of evolution.' And what he was referring to is a unifying principle, something that can be used to make sense of the biological world that we see. And I think this is really powerful. This idea of unification is really powerful. And I often think about the example of the electromagnetic spectrum. So prior to having a good theory of electromagnetism, we used to think that static electricity, lightning, magnets, light, we thought those were all completely different things. They certainly seem like different things. We had different words for them. The vocabulary that we used helped us to see them as different. And then what the unifying theory allowed us to do is first of all to see them all as one spectrum and to figure out that actually they all have something very fundamental in common. And the other thing that it allowed us to do was to make new technology that enabled us to operate in parts of that spectrum that our evolutionary history does not allow. So we're only sensitive naturally to this narrow space, but now we have technology to see all this. And the reason I bring all this up is that there's a large part of my efforts—even though our lab, we're about 43 people or so, and a lot of the practical output of our lab is things like interventions for birth defects, regenerative medicine, cancer, bioengineering, we do some AI as well—these are very practical things for modern regenerative medicine. But what underlies all this is my commitment to try to understand how minds exist in the physical world. That is, what does it mean to be a mind that operates in the physical universe? What is embodiment? And we try to bring some of those ideas into practical both computational and wet lab applications that, to whatever extent they're useful, help us discover new things that tells us whether we're on the right track or not.

So this business of unification—first and foremost what I'm interested in is understanding very diverse kinds of intelligence on the same scale. So we're talking about not just birds and primates and things like that, and maybe a whale or an octopus or something like that, but also very weird colonial creatures and swarms and synthetic new life forms and AIs, whether purely software or robotic, perhaps someday exobiological agents such as some alien life if and when we find it, and even something really exotic which I won't have time to talk about today, but agents that are actually patterns within media. And I'm not, of course, the first person to try for this. Here's Rosenblueth, Wiener, and Bigelow, who were trying to make a kind of spectrum of categories going all the way from passive matter through various cybernetic stages all the way up through human level second-order metacognition and so on. So I'm interested in this kind of thing, in understanding what do all minds have in common no matter what their scale or what their origin story or what their composition. For mainly two reasons. First of all, to help move forward practical applications in biomedicine, synthetic morphology, and so on. But also actually an improvement in our ethical systems, because I think the ones that we have are based on ancient philosophical and linguistic categories that actually pick out nothing real, no real distinctions in the world, and are going to get us into a lot of trouble as we start to share our world with increasingly diverse beings—you know, cyborgs and hybrids and all kinds of things, some of them are here now, some of them are coming. But all of them will need a very different kind of ethical system that isn't so obsessed with human-style brains.

So before I get going and show you a bunch of biology, here's an outline of my points for today. First of all, let's just get some metaphysics out of the way. When I talk about teleology, I do not mean anything related to quibbles of linguistics or philosophy. I don't think you can just sit back and make decisions about whether teleology does or doesn't exist. I think it's an empirical question. And the way that we treat these things empirically is this: all of these frameworks, no matter how weird they are—I don't care how weird the framework is, I don't care how unconventional, whether it's vitalism or anything else, I'm game for anything as long as it helps us generate new discoveries and new capabilities. I don't just mean look backwards at something that already happened and give some kind of account of what it was, because you can always do that at the lowest level. You can always tell an atomic story about anything that happened. What I mean is, does the framework help you do new experiments and discover new things that you otherwise wouldn't have discovered? That's it. That's my criteria for today. So cards on the table, that's what I'm looking for. And so that's the first thing.

Second, when I say teleology, I don't mean human-level purpose, meaning that it's a second-order kind of thing where I know what my goal is and I know that I'm pursuing a goal and all of that. I don't necessarily mean that, but it's on the spectrum. What I mean by teleology is the kind of thing that cybernetics does for us. It gives us a non-mysterious science of systems that are consistent with both the laws of physics and the ideas in cognitive science about systems that have goals. So cybernetics tells us about how systems can have goals, and computer science tells us how to study the ingenuity of these systems to meet those goals under different circumstances. So we have these sciences now, and we no longer have to freak out about goals being some sort of magical thing that science isn't going to be able to get a handle on.

So part of my argument—there's two main things I want to do today. First of all, I'm going to show you the data for a very simple argument. The argument goes like this: goal directedness and teleology, I'm using interchangeably. First, hopefully uncontroversial point, is that whatever goals are, they're the sort of thing that humans have. We all know we can represent goals in our minds. We can undergo actions to pursue them. We have some degree of ingenuity to meet those goals. Sometimes we succeed, sometimes we fail. But if goals mean anything at all, we as humans are an example of it. That's point number one. Point number two, we actually know something about the mechanisms that enable goal-directed behavior in brainy organisms like us. We know something about the electrophysiological networks in our brain that enable goal directedness. But remarkably—and I'll show you all this today—remarkably those mechanisms are ancient. They are highly conserved. They are not unique to brains and nervous systems. And there is a very deep and fascinating evolutionary homology, not just analogy, not just metaphorical thinking, but actual homology between cognition and morphogenesis. And this is seen most clearly in developmental bioelectricity, which is what I'm about to show you.

And so what I'm going to go through is a bunch of examples of biological goal-directed problem solving, mostly in the bio realm but also I'll show you some other examples that are different. But in any case, that's my argument: we have a non-controversial system, aka us, brainy animals, that clearly have teleological behavior. And what I'm going to point out is that the same mechanisms that underlie that are actually very widely distributed in the biological world, and also actually I think beyond that. But that's a topic for another day. So it's a very simple argument. But the other thing I want to say is that teleology, I think, is just the tip of the iceberg. If you're worked up about teleology, it's going to get much worse, because what I think is really the unifying principle of biology isn't just teleology, it's cognition broadly considered. So goal directedness is just the first step of a huge spectrum of cognitive skills, and that is the spectrum that I think underlies everything that we're interested in. Teleology is just kind of the first step. So this is part of a view that some people call cognition all the way down. And it's a reasonably extreme position in the emerging field of diverse intelligence, but that's where I am.

And so what I like to claim is this: I think that cognitive terms—meaning when you say something has goals, that it is intelligent, that it has memories, whatever, or that it can be trained, whatever your assessment is—I suggest that what those things are are interaction protocol claims. I don't think that these are essentialist facts about what the thing truly is in its nature, and that we can all argue about it for a little while and then just sort of settle on whatever the right answer is. I think that all of these are observer-relative interaction protocol claims. So what you're saying when you say that a certain system has or doesn't have some capacity is you're basically picking out a set of tools across a spectrum of very different kinds of tools—hardware modification in the case of mechanical clocks, resetting set points via cybernetics and control theory as Rob just pointed out, behavior science, psychoanalysis and psychiatry, whatever is going to be. We have bags of tools, and what you're saying is I'm going to try applying a particular set of tools, and then we can all see how well I did because it's an empirical thing as to how that turns out.

So my claim, when I talk about this in more conventional audiences, is I'm going to take the tools of behavioral and cognitive neuroscience, both the conceptual tools and the bench tools, and I'm going to apply them far outside the brain. I'm going to show you that the tools actually don't distinguish. And if the tools don't distinguish, what is it that we are trying to draw a boundary between? And the fact that this kind of thinking allows us to discover new capabilities by reusing those tools elsewhere, and the conventional model did not do that, because it suggests that those tools should only be used in the brain, that should tell us something. And so the model system that we use for a lot of this—I'm going to show you other things—we years ago started working on this model system, which is groups of cells as a collective intelligence. I'm going to define those terms momentarily. And what this collective intelligence is doing is exerting behavior by navigating anatomical morphospace, which I will also define. But in order to do that, this idea breaks a number of assumptions, but it has implications.

So let's look at that. The first thing to point out is that we and our goal-directed capabilities are a multiscale phenomenon. We're made of an agential material. We are not made of passive matter. The things that we are made of, on multiple scales of organization, themselves have agendas and they have learning properties and all kinds of other capabilities. Jamie Davies and I talk here about how you engineer with things like this. It's very different. Engineering with agential materials is very different than engineering with passive matter or even just active matter. So this is basically a multiscale competency architecture where all through the scales, from the molecules up to the collectives and groups, all of these things are, as it turns out, solving problems in diverse spaces, with very different degrees of sophistication.

Now what are these spaces? Well, we as humans, because of our own evolutionary history, are obsessed with three-dimensional space, primed to notice intelligent behavior, goal-seeking behaviors of medium-sized objects moving at medium speeds, so birds and mammals and things like that. But biology has been solving problems in other spaces for a really long time. So there's the space of gene expression, high-dimensional transcriptional space, space of physiological states, and anatomical morphospace, which is basically the morphospace of all possible configurations or shapes of a body. This is the one we're going to talk the most about today. But my claim is that just as obvious animals navigate three-dimensional space to meet their goals, we have all kinds of structures in our body that are navigating with similar kinds of skills navigating these other spaces that are very hard for us to notice. And until we have the right theory and the right instrumentation, which is what I think we're doing, just like with the idea of the electromagnetic spectrum, we're developing tools to port everything that works here, so behavioral and cognitive science, we are porting it to these other substrates, and that helps us visualize these things.

So I claim that teleology is a subset of the spectrum of intelligence. And what I mean by intelligence is this: William James gave a nice, very cybernetic definition of intelligence. He didn't talk about brains. He didn't talk about what problem space was involved. He said it's the ability to reach the same goal by different means. So this ability to reach the same goal, that's basically teleology—that there is a goal and that the system is going to exert effort to try to get there. But the different means is interesting. It means that every system is going to have some degree of the ability to get there when circumstances change, when you put barriers in its place and so on. And that's a really important part of the toolkit here. If you want to study teleology, you cannot do it from pure observation. You have to put barriers between the system and whatever you think its goal is, and observe the degree to which it can be clever by getting around whatever you just did to it.

So our hypothesis is this: for this portion, I'm going to show you that morphogenesis is a collective intelligence. We are all collective intelligences because we are made of parts. So all intelligence is collective intelligence, I think. And what it does is it exerts its behavioral competencies in anatomical space, in getting from being a single cell to whatever we end up being. It navigates that space, and I'm going to show you some really cool ways that it does that. So we're going to study this moving here from neuroscience—you have to go to a different time scale. So minutes instead of milliseconds, different problem space, morphospace instead of three-dimensional motion, but everything else is the same. The mechanisms are homologous. Many of the same algorithms, perceptual bistability, active inference, you name it. All of this stuff works really well. And in fact, I used to have my students do this by hand. And now we made a tool for people to do this. If you take any neuroscience paper and you do a find replace, and every time it says neuron, you just say cell, and every time it says millisecond, you say minutes or hours, you got yourself a developmental biology paper. And it works amazingly well because there is this deep underlying symmetry between the two.

So I'm going to show you here six, or however many there are, examples from our world of the components of intelligence. Let's start with goal directedness in a simple way: anatomical homeostasis. So here's a creature. This is an axolotl. It's an amphibian. This little guy regenerates its limbs, its eyes, its jaws, its ovaries, its spinal cords, portions of the brain and heart. And the way that works is here's the leg. If you amputate—or if you let them bite off each other's legs, which they do constantly—anywhere along this path, the cells immediately sense that they've been deviated from the correct region of anatomical space. They work really hard to build this thing, and then they stop. When do they stop? Well, which to me is the most amazing thing about regeneration, is that they know when to stop. Knowing when to stop because you've rebuilt the same structure, meaning you've gone back to the same region of anatomical space within some small error, is basically anatomical homeostasis. It is goal-directed because the system has a set point. If you deviate it from that set point, it will try really hard to get there. Okay, so this is the most simple version of this, and I'm going to return to this theme and show you lots of detail about how this actually works and so on.

The next thing I want to show you is the second part of this, which is more to the point that it doesn't just do the same thing that it always does to get to that point. It actually can get there by different means. And as I realized, I'll just say one example which I didn't include in the slides. When you amputate a frog leg—adult frogs don't regenerate their limbs, but we've learned to make them do it. And when they do, the leg they build ends up being a very nice frog leg, but it does not get there by the same path that the embryonic leg grows. It actually looks completely different, very plant-like, actually. Just very different. But here's a basic example. We know development is very reliable. So the egg always gives rise to the kind of thing it's supposed to build. But that's not why I'm calling it intelligent. The reality is that I can take this early embryo, cut it into pieces, and I don't get half bodies. I get perfectly normal twins and triplets and quadruplets and so on. Because in anatomical space, the same system can get to where it's going—this ensemble of goal states representing a normal human target morphology—it can get there from here, but also from here, and actually from all kinds of different starting positions, even avoiding local minima. It can get to where it needs to go. And again, this process of homeostasis, right? So half of it knows that the other half is missing. It will rebuild it as much as it needs. I actually think development is basically just an extreme version of regeneration. Instead of regrowing your limb, you're regrowing the entire body from one cell. All of these things, I think, are fundamentally just homeostasis.

And here's another example. This is something that we discovered a few years ago. If you—here's a tadpole. So here are the eyes, the nostrils, the brain, the mouth, here's the gut. This tadpole has to turn into a frog, so all of these craniofacial organs have to move, they have to rearrange—the eyes have to move forward, the mouth has to pull out, and so on. It used to be thought that this was somehow hardwired, that somehow the genetics encoded a set of precise movements for every structure so that you start out as a normal tadpole and you end a normal frog. So we decided to test this, because with all of these things you have to test them. You cannot just assume what the level of intelligence is. And what we did was we created so-called Picasso tadpoles. In Picasso tadpoles, everything is scrambled. The eyes on the back of the head, the mouth is off to the side. It's like a Mr. Potato Head where you've shuffled around all the organs, right? Just completely screwed up. Well, what you find is that these animals make very normal frogs because all of the different organs, starting in their crazy positions, will move through novel, unnatural paths to get to where they need to be. And in fact, they actually sometimes go too far and they have to double back. So the genetics does not give you a set of hardwired rearrangements. What it actually specifies is a system that can do an error minimization scheme. It knows what the set point is, and I'm going to show you how it knows in a minute. And it is able to reduce that error by whatever actions it needs to take. These things are moving in novel paths. And that, of course, raises the obvious question: okay, then how does it know what a correct pattern is? I'm going to show you that momentarily.

So I've shown you two things. I've shown you a basic homeostasis. I've shown you a more flexible kind of homeostasis. Now I'm going to show you non-local control. So the key thing to realize about this is that it is not just about damage. The point is not about injury response and just simply restoring what you have. It's a much deeper type of teleology that's really built into all of the aspects of this. And here's the experiment. This wasn't us, this was these guys in the '50s. You take a tail and you graft it in the middle of the flank, somewhere between the forelimb and the hindlimb. You just surgically graft it to the flank. Well, what happens over time is that this thing turns into a limb. It gets it in place. It actively remodels from being a tail to being a limb. Now notice something interesting. If you're a tail tip cell sitting at the end of a tail, locally there's nothing wrong. You are exactly where you're supposed to be at the end of a tail, and yet you turn into fingers. Why do you turn into fingers? Because the key aspect of these kind of problem-solving, goal-seeking systems is that in a collective intelligence where the large-scale system is solving a goal in its own space, it has to control its own parts and deform their option landscape so that they do things that match the higher-level goals. No individual cell knows what a finger is or what a tail is or how many you're supposed to have or where they're supposed to be. But the collective has a goal. The goal is a reasonably normal-looking axolotl. And in order to meet that goal, it has to propagate downwards all of the control signals to its various parts. Notice the parallel of this in cognition. When I'm talking to you right now, I am not worried about having to adjust all your synaptic proteins so that you're able to remember what I said. You're going to do that yourself. The beauty of these kind of multiscale problem-solving systems is that they take care of the underlying molecular events and they take care of whatever needs to happen in order for the higher-level system to meet its goal. And so that's why these tail tip cells start to become fingers. So it's not just about repairing damage. It's about harnessing local order to a global plan that belongs—it's a goal that belongs to the large-scale collective. It does not belong to any of the parts, to any of the cells, but it belongs to the whole, and it's a very large-scale goal.

So now notice how all of this differs greatly from the mainstream approach to these problems that you find in your developmental biology textbook. The standard approach, as I'm sure you guys know, is this feed-forward, open-loop kind of thing consisting of complexity and emergence, meaning that okay, we have gene regulatory networks, these genes turn each other on and off, some of them become effector proteins that diffuse or they're enzymes or they do things, and then there's all this local physics that happens, and then bam, so that this complex shape emerges. And what that word means is simply that when you do a lot of repeated operations on local components, out might come something complex. Now that's absolutely true, and we have lots of model systems for this—cellular automata, you know, fractals and so on. We know that by recurrent simple rules, you can get complexity. But that isn't the whole—that isn't how biology does a lot of these things, because there's a problem here. If that's what it is, if you want to make changes up here—let's say you wanted three-fold symmetry instead of bilateral symmetry or something—you know what this model is telling you is that you have to make changes down here, and you don't know what changes to make down here in the large majority of changes you might want to make. This is not reversible. It's an inverse problem that is not solvable. And in fact, what I'm adding to this is this idea that under this scheme there is no goal. There is no teleology because there is no goal. There's only feed-forward, turn the cranks, local rules. Out will come something complex or it won't. But there is no goal. The system is not going towards anything. I'm proposing a different model where yeah this happens but something else happens that's very important, is that there actually is a represented goal in the system, and I'm about to show you a picture of it, the data around it. There is a represented goal, and what the system will do if you deviate from this set point is turn on, both at the level of physics and genetics, all kinds of mechanisms that will try to reduce the delta and get you back to that goal. And if that were the case, that would be a huge and important thing for, for example, regenerative medicine, because it would mean that we don't need to solve this terrible inverse problem. We don't need to mess with the hardware necessarily. We just need to respecify the goal states and let the system do what it does best, which is try to meet the goal.

And so this is what we've been doing now for about 25 years, I guess, is to find—if this is going to be a goal-seeking system like your thermostat, it means that it has to represent the set point somewhere. It has to remember what the correct pattern is, and that means I should be able to find it, I should be able to decode it, and I should be able to rewrite it. Those are the three things. And if I can't do those things, then we have not shown that the system has goals, right? So I want to show you some of this data. And the idea again is that morphogenesis is a collective intelligence, and that its behavioral competencies occur in anatomical space. That's the kind of problems that it solves. So here's what I'm going to claim that bioelectric networks can do for us. And these claims parallel—you'll see exactly parallel—what is the deal in neuroscience. And I'm only going to show you, I only have time to show you a couple of these things. The first thing these networks do is they store memories. If you're going to have a homeostatic system that seeks goals, you have to know what your goal is. So you have to somehow store and represent the set point. So that's the first thing they do. Also, they have to do the computations needed to perform error minimization. So to actually navigate the space, then you also need to integrate all kinds of long-range information to assess the current state. What does my body look like? Do I have a tail in the middle or do I have legs or nothing? Or what does it look like in order to compare against the set point in order to initiate any kind of remodeling. And the final thing—and this is really an interesting function—is they have to serve as a kind of cognitive glue. So the reason that you and I are not just a pile of neurons, and that we know things that our neurons don't know, is that electrophysiology serves as a cognitive glue binding individual cells towards larger-scale systems whose goals project into new spaces—spaces that the individual parts don't have any access to. And bioelectricity does this in your body just like it does this in the brain.

So let's talk about something that you might think is a solved problem. Often times people think that developmental biology has already sort of figured out everything. The question is where do anatomies come from in the first place? So here's a cross-section through a human torso. You see incredible order. Everything is in the right place, the right size, next to the right thing. Amazing. But we also started life as a ball of embryonic blastomeres. Where was this pattern encoded? And then people often will say, well, it's in the DNA. Of course, it's in the genome. But we can read genomes now. We know that what the genome encodes is proteins. It doesn't directly specify any of this stuff, any more than the genome of termites specifies the structure of the mound, or the genome of a spider will tell you exactly what the web is going to look like. So the genome gives you the hardware. The genome tells you what computational hardware all the cells are going to have. But now we need to understand the software. How do the cells know what to build? How do they know when to stop? As engineers, we'd like to know, could we ask them to build something completely different? Same cell, same DNA. Could we ask them to build something different? If we get to it, I'll show you at the end that the answer is yes. We can build some crazy things that way. And for regenerative medicine, we'd like to know how do we tell them to repair things that are missing? And so all of these kinds of things—this kind of goal-directed behavior in the brain is handled by electrical circuits which consist of cells that have ion channels. These ion channels allow the cell to reach some sort of voltage potential. They have these electrical synapses that allow the voltage to propagate or not to neighboring cells. And that whole thing—the enormous network of these things—executes a kind of electrophysiological dynamics that neuroscientists think are the bearers of cognition.

So here's a living zebrafish brain that these guys imaged while the fish was doing some sort of cognitive task. And there's this idea of neural decoding, such that if we were to record this information from a fish or from a human, if we knew how to decode it, we would be able to read the thoughts, the memories, the goals, the preferences, all of those things, right? That is where all of these things are in us. So in the one example of goal directedness that pretty much everybody agrees on, that is humans. This is the mechanism that maintains it. This mechanism allows you to store memories as goals. It allows you to control your muscles to try to act on those goals, and it does all the computations needed to try to reduce the delta between where you are now and wherever you're trying to be. Fine. So this is the mechanism of well-accepted teleology. However, it turns out evolution caught on to this way before neurons and brains evolved. So basically every cell in your body has ion channels. Most cells have these electrical synapses. All of this idea of using electricity to coordinate and bind individuals into a larger-scale unit was discovered around the time of bacterial biofilms, which are already—this is beautiful work from Gürol Süel showing how bioelectricity was used in bacterial networks to organize the microbes into a larger-scale thing, a colony. So kind of a primitive multicellular creature. And what we decided to do was to basically steal everything from computational neuroscience and to use the same techniques to try to decode what the cognitive content of the body might be. We know what the brain thinks about. The brain thinks about moving you through three-dimensional space to reach your goals. But what do the biological networks of your body, which exist both evolutionarily and developmentally long before you have nerves, what do those networks think about? And it turns out that they think about shape. They think about moving your body through the anatomical morphospace as opposed to 3D space. And I actually think that that is where the brain learned all of its cool tricks, is by simply pivoting this much more ancient problem-solving, teleological system from morphospace into 3D space. And then of course speeding up things so that you can run around and eat other things and so on.

So we developed a bunch of tools, and I'll show you—the first thing we had to develop was a way to read, just like neuroscientists read in the brain, we had to read the electrophysiology of the body. And we developed tools—I mean basically voltage-sensitive fluorescent dyes and reporter proteins. Here's a frog embryo where you can see all the electrical conversations that cells are having with each other in real time. This isn't a model. This is real data. And then of course cells in culture, you can see the colors represent voltage. You can see the different voltages as the cells are making decisions about whether to be part of this morphogenetic mass or not, or to leave. We do a lot of simulations, multiscale simulations, starting with the ion channels that produce the gradients into tissue-level dynamics that help you explain all kinds of interesting facts about regeneration and so on. And then we keep going further and we try to merge this with ideas in connectionist machine learning and dynamical systems theory so that we can try to understand memories as attractors and pattern completion and all of these things—these sort of proto-cognitive kinds of things going on in electrical networks.

Let me show you a couple of real examples of this. First of all, this is an early frog embryo putting its face together. So the grayscale are the different voltage values. There's a lot going on here, but look at this one frame from that video. We call this the electric face. So long before the genes actually turn on to regionalize that face, here's where the eye is going to go, here's where the mouth is going to go, here are the placodes out here. This is the early, subtle information structure that tells the face—and then the other eye comes in slightly later—it tells the face where all the craniofacial organs are supposed to go. And if you move any of these bioelectric patterns, the gene expression will follow and the anatomy will follow. And this is how you can make a scrambled Picasso tadpole by changing the pattern that normally tells these things where they should start out.

I will point out that not only is bioelectricity a cognitive glue on the scale of a single embryo, meaning that it binds individual cells towards embryo-level outcomes, it also works on the next level. And so these are multiple embryos. So you see here one, two, three, there are three of them. If you poke this guy right here, there's an injury wave within seconds. This one finds out about it. And then this one, same thing happens here. And it turns out that bioelectrical communication among embryos within a group actually helps them resist teratogens, deviations from their set point, much better than singletons. So the collective—it's at least two layers of collective intelligence. Cells into a single embryo, single embryos into a collective that can solve problems that individual cells have trouble doing. So we have tons of this kind of data, watching and trying to decode what these patterns represent, just like neuroscientists try to decode from the brain. But more importantly than just watching them, you have to be able to rewrite them, because I told you we have to detect the memory and then we have to rewrite the set point. That's really critical.

So again, we took tools from the neuroscientists. We do not use any electrodes. There are no magnets, no frequencies, no electromagnetic fields, no waves, nothing like that. What we are doing is using the same interface that cells use to hack each other, which is we control the gap junctions, the electrical synapses. So we can control the topology of the network in cells and tissues, and we can also control the actual voltage by targeting the ion channels, whether it be through drugs or, for example, optogenetics. We can lay down light masks and put specific patterns of voltage down, and then we have some other tricks with neurotransmitters and so on. Same exact stuff that the neuroscientists do, except we do all of this outside the nervous system. You know, a fun story that Rob might appreciate—when I was a postdoc, I wanted to do this kind of thing, and I started writing to neuroscientists asking them for plasmids encoding all these different channels. And at first, when I did it, I was dumb enough to tell people why I was doing it. I said, look, I'm going to express these things in non-neural cells to guide decision-making and morphogenesis. And at one point, my postdoc mentor Mark came to me and he said, I just got a letter from some neuroscientist, he's warning me that his postdoc—that my postdoc—is unhinged, because he's writing me these crazy requests. And literally, he warned the guy that I might be dangerous. So this is the level of surprise that this engendered in neuroscientists, this idea that these channels are going to do anything else outside the nervous system, other than maybe some supportive housekeeping or something. It was considered a completely crazy idea. But what I did is I stopped telling them why, and I would just write to people and ask for these different channels that I was then going to use to control cell voltage in these settings.

And before I show you what happens when we do it, I want to show you a natural example. So these are just a bunch of cells, a natural example of cells hacking each other. So these are cells in 2D culture. The colors are voltage indicators. And what you see is something very interesting. This cell—and I'm going to play the video momentarily—this cell has a different voltage than all these other cells. And in the next frame, it just touches this little projection here. And as soon as it touches, boom. These guys have rewritten its electrical state. And watch. So here it comes. It's different. It's moving around. Now this one's going to reach out and just touch it. Bang. There. It's now one of them. And now here it comes. And now it's going to participate in this event. So the ability of cells to control this kind of electrical state—this is what it naturally looks like at the single-cell level.

And so what I'm going to show you now is the use of this interface. And that's what I think the biomediatrics is. I think the reason it's interesting is because it's this amazing interface to the teleology. And so here's my claim: goals exist if and only if you can rewrite them. If I can't rewrite—if I can't change the goal and have the exact same hardware execute on a totally novel goal—then I can't say that there's a goal. Okay, so this is back to my claim that this is not a philosophical debate. This is very practical. So this is my criteria. Can I hack these things specifically by rewriting goal states? So I'll show you an example. So here's an early frog embryo, and what we're going to do is we're going to inject some ion channel RNA that we chose. And there are many RNAs, many ion channels you could choose, as long as they set the voltage to the correct state. This is not specific to the molecular biology. It doesn't care which channel it is, doesn't even care which ion it is, as long as you get the voltage right, it works. And what we're going to do is inject that RNA into some cells that are going to be various parts of the body, for example in this case the gut. So I inject the RNA, and sure enough, these cells will form an eye. And I already showed you that there's a very particular voltage pattern that induces eyes. So here's the eye. If you section them, they have lens, retina, optic nerve. They have all the same stuff they're supposed to have. Now notice a few really interesting things. First of all, it tells you that this bioelectrical signal is instructive. We didn't just screw something up. We actually called up a coherent entire organ. So it's instructive for what happens. It's not permissive. It's instructive. Second thing, it's incredibly modular. You know, we don't have any idea how to build an eye. We don't know how to tell the cells which stem cells go where and what layers go where. We don't have any ability to do that. It's a very complicated process. We've discovered a very simple message, a high-level trigger that when the cells receive it, they will organize all the downstream stuff themselves to build that eye. This is what I was saying before about multiscale cognitive systems. Just like when I talked to you, I don't worry about the molecular biology of your...

...

brain. Same thing here. We gave it a voltage message that said build an eye and everything else happened after that. All the genes turned on and off. Every that the system takes care of all that. So it's very modular. Next thing I'll tell you is that if you look at the developmental biology textbook, it actually says that in invertebrates only the anterior neurodede. So this stuff up here is competent to form eyes. The other tissue can't do it. The reason they say that is because up until then, and this was I don't know 2010 or 2012 or so, everybody had been probing these things with what they call the master eye gene or PAC 6. And sure enough, if you only use PAC 6, right? These are the only cells that are willing to make an eye. But the competency problem wasn't the problem with the cells. It was a problem with us, the scientists, because we weren't using the right prompt. You have to, like with any intelligence system, if you don't see its intelligence, you know, it's a two-way IQ test. It might not be the system. It might be you. And so, we just weren't using the right prompt. If you use the right prompt, pretty much any region of the tadpole can form an eye. So, we got to, you know, that's a humility warning for us. And the final thing I'll point out about this is something cool. If you cross, if you take a cross-section through this eye, for example, the lens, the blue cells are the ones that we injected. What about all this other stuff? All this other stuff that are participating in this formation, they do not, we never touched them. We never injected them. They got recruited by the cells that we did inject in order to, and so basically these cells can tell there's not enough of them to finish the eye and they do a secondary recruitment to get these other cells to play along. Again, competency of the material. We didn't teach them to do that. They already knew that. And there are many other collective intelligences that do this. So ants and termites, right? When they encounter something that's too heavy for them to lift, what do they do? They call up the rest of their buddies to help them. The system self-scales to the task. Okay. Now, there's something quite interesting here. When we first did this, we noticed that if I inject my channel, I can get embryos with, so this is an expression of a gene called RX one, and RX one is an early eye marker. So you see, okay, here's the native eye, but here, one, two, sort of three, four, there's a bunch of these. You might think, okay, amazing. This embryo is going to have like five extra eyes, but you look the first day and there's, you know, five or six. You look the next day, there's only three. And you look the next day and there might be one. What's happening to this? It turns out that there's a battle of, you can say, you could say it's a battle of chemical signals, but I actually think the more interesting aspect of this is a battle of goals. It's a battle of world views because the cells we injected are saying to their neighbors, we should definitely be an eye. Help me build an eye like this. But there's a cancer suppression mechanism that basically functions this way. If you're a cell and your neighbor shows up with some sort of weird voltage, you try to normalize it. Okay? It's a continuous cancer suppression mechanism. So you use your gap junctions to connect and you just like wipe out their crazy voltage. So what happens here is that the patterns fight. They compete, and there's a pattern that says no no you should be skin, and these guys are saying no no no we're going to be eye, and sometimes the eye wins and sometimes the skin wins or whatever, and that's what, you know, this reminds us that for regenerative medicine kinds of applications you need to be able to speak the language, you need to be able to give these kinds of prompts, but you have to make the prompts convincing. This is not like rewiring a mechanical clock where you just sort of micromanage everything and put it where it goes. You are dealing with a decision-making system where you have to get the buy-in of the cells by respecifying their set point. And sometimes it'll take and sometimes it won't unless we really understand how to make it stick.

So I want to show you another wild example of rewriting this anatomical pattern memory. So these are planarians. These are flatworms. And one cool thing about these flatworms is that if you cut them into pieces, every piece will regenerate very reliably. So here's a worm. Here's a head. Here's a tail. I cut off the head, cut off the tail. This middle fragment here will give me one perfectly normal worm. So you could ask the question, how does this thing know where the head goes? Because it's got two wounds, okay? And back here, the cells back here need to make a tail, but the cells over here need to make a head. So they're like very close to each other. You can't tell from the local position whether you should be a head or a tail. You don't know because it could be either one depending on what the rest of the fragment is doing. So it's actually quite interesting. How do you know how many heads you're supposed to have? So, it turns out this thing has a particular voltage pattern that says one head, one tail. And what we ended up doing was figuring out how to change that voltage pattern. And this one, this is an old picture and it's kind of a mess, but the idea is that what we did was we targeted the ion channels to produce a pattern that says two heads instead of one. And when you do that, nothing happens at first. Here's a normal animal. It is anatomically normal. It is normal according to the molecular biology markers. In other words, the head marker is on in the head. It is not on in the tail. So these cells do not think they are a head. And everything sits quietly until you injure the animal and cut. Once you cut them, bang, that's when this latent memory becomes a recalled memory. And that's when they build two-headed worm. Now, this is not Photoshop. This is not AI. These are real animals. And so this is quite cool. This bioelectric map is not a map of this two-headed animal. It is a map of a perfectly anatomically normal one-headed animal. And it is a latent memory that only becomes activated if I injure the thing. But then it serves as the new set point. The new memory of what a correct planarian looks like. The new goal becomes active here and the cells will build and then they stop. Like any homeostatic process, once this pattern has been activated.

So two kinds of interesting things. A single normal planarian body can store at least one of two representations, one of two internal representations of a future goal state of what it's going to do if it gets injured. So that's a counterfactual memory. I think it's actually an evolutionary precursor to our brain's amazing ability to time travel, right? To have thoughts about things that are not happening right at this moment, either past or future. And I think that's what it is. So, if you want to know, at the very beginning when I told you that I have to be able to detect, to read and rewrite the pattern memory, and if you want to know what do goals look like in non-human systems, this is what they look like. You are literally looking at a memory that tells this thing how many heads it's supposed to have. Now, we didn't touch the genetics, and that's cool because good memories shouldn't require modification of the hardware. They shouldn't require genetic modification. They don't in the brain. And the same thing is true here. You take this animal, perfectly normal standard genetics. If you were to sequence it, you would see absolutely nothing different. But when you cut a two-headed animal, you continue to get two-headed animals. That electrical circuit stores the pattern stably. Once you've changed the goal, it remembers the goal and it will remember it, as far as I can tell, forever. We can, by the way, reset it back. We now know how to reset it back. And here you can see the first time I showed these data at a conference many years ago, somebody stood up and said these animals can't exist because this goes against a lot of thought about what the genome does, namely specify the number of heads you're supposed to have, and clearly that's not the case. And so now I bring these videos, although now with AI, I guess anything's possible. But this is what I've been showing people. No, those animals absolutely exist because the number of heads, like many other interesting things, are not in some way encoded in the genome. They are actually an active memory that is maintained in this case by a biological circuit. It is not just about head number. Amazingly enough, it is also about head shape. So we can take this planarian with a nice triangular head and cut off the head, perturb the biological network that guides the movement of the cells in anatomical space. So normally you're supposed to go to this area where this attractor here where data live. But that same hardware is perfectly happy to go to other attractors. It can build a flathead like a planaria. It can build a round head like an S Mediterranean. About 100 to 150 million years evolutionary distance between these guys and this guy. The insides of the brain of the head also change. So they have brains like these other species. They have a distribution of stem cells like these other species. Again, no genetic change, but if you change the pattern memory, they will use it to navigate to a different region of that morphospace where typically other species hang out. You can even go further and find out that actually the hardware is really plastic. I mean you can make things that don't look like planarians at all. You can make these weird spiky things and you can make round, you know, totally different symmetry type and these composite things. This latent morphospace is massive.

And there's a cool, this is one of my favorite examples of this kind of thing with plants. This is these acorns. You know, you would look at this and you would say, well, we know what the oak genome encodes. It encodes this. It encodes this kind of leaf. Well, along comes a non-human bio-engineer, this little wasp. And the wasp, because these circuits are so reprogrammable, the wasp can put down signals that will cause the plant cells to build something like this. This is called a gall. And you can see instead of this normal flat green thing, it makes these round spiky, you know, these incredible other constructions. We would have had no idea that this is even possible. Except that this thing found an example, and one of the things we're doing in our lab is trying to automate this process to crack the morphogenetic code so that we can call up structures at will. Yeah. So we can program them. So I want to talk about two examples of learning. Previously I showed you that we directly rewrote or edited the memory that guides goal-directed activity. We specifically changed those memories. But I want to look at the way the system has the ability to change its own memory. In other words, learning is the ability of systems to write their own memory engrams. And so the first one is at the molecular level. And I would just point out that we found, and you can see that in these papers, we found that even gene regulatory networks alone, so never mind the cell, the nucleus, any of that stuff, just the gene regulatory network itself, so a set of genes that turns each other on and off, and actually any kind of pathway that has elements that turn each other on and off, are able to show up to six different kinds of learning. So associative or Pavlovian conditioning, and also simpler things like habituation and sensitization and accounting to small numbers and things like that. So even the molecular networks inside of cells are able to do this, and we are building devices in our lab to actually take advantage of this for biomedical purposes like drug conditioning and things like that. So you can see, again my claim that the material is a multiscale agential material, that even below the cell level you already have some examples of learning which can then be harnessed by the higher levels of organization and underly more complex forms of goal directedness.

And then, to contrast that, a very large-scale example of learning, and this is called trophic memory in deer antlers. So, father and son team, Bubenik, did these experiments for about 30 years in Canada, in deer, in a herd of deer. I mean, talk about an inconvenient model system, it's just amazing. This kind of data set will probably never be gotten again. But what they found is that every year the basic pattern of the antlers remains constant, and the whole thing is shed and then they regrow a new rack. And what they found is that if you have a bit of damage at one particular location that breaks through the skin and etches into the bone, there's a little callus and it heals, and that's that for that year. And then eventually this whole thing falls off and next year they grow a brand new tine, brand new antler rack rather, that has an ectopic tine at that location. So there's great specificity here. And then this goes on for about 5 years. Every year it has this ectopic branch, and then eventually it gets smaller and smaller and disappears. So think about what this means, and how you might try to draw a molecular pathway model. You know the kind of models that are popular in cell biology textbooks where there's some arrows and some things binding some other things and so on. What possible pathway could you draw to explain that this thing somehow conveys to the cells, presumably the cells at the base of the scalp where the damage occurred, those cells store that information for months until it gets ready to build a new antler, and then it propagates that information back up to guide growth toward a specific outcome, a very specific outcome, next year. Okay. And so that ability to form three-dimensional memories of location in three-dimensional space in the structure is remarkable. We don't even have the beginnings of a model in molecular biology of how you would do this. But in neuroscience you can kind of imagine what kind of models around bioelectric networks remembering locations in space and so on. So this is another example. And this guy, apparently I was the only one who was writing about these things in the modern literature. And so when he retired and needed to get rid of all the antlers, he sent them to me. And so I have 13 boxes of these things that are, you know, everyone is labeled, you know, Lenny 1987, 1988 and so on, in terms of these multi-year experiments, just an unbelievably remarkable biological object that again shows how you really need concepts from these other fields related to cognitive science in order to be able to understand these kinds of phenomena.

So the next step, the next thing I want to show, after learning, is creative problem solving. So at first towards the default goal. So the system has a default goal, presumably conveyed onto it by evolution, that sets up a default memory pattern that's being used as the set point for the homeostatic activity. And we're going to look at different ways to get to that goal under novel circumstances. And the first example that I want to show you has to do with planarians. So we discovered this a few years ago. Here is a flatworm. You put the flatworms in a solution of barium. And barium is a non-specific potassium channel blocker. So it blocks all the potassium channels. The heads are very unhappy. You know, it's got neurons and so on. They're very unhappy about not being able to pass potassium, and the whole head explodes. Basically, their heads blow off within 24 to 48 hours, just gone. And then something amazing happens, which is that if you leave them in the barium, they eventually regenerate a new head. And the new head is totally fine in the barium. It's somehow barium adapted or insensitive. And we did an experiment to try to figure out how this was possible. And so we did RNA-seq to look at all the genes that were expressed in barium-adapted heads versus normal heads. And we did a subtraction and asked what was the difference. It's really only a handful of genes that are different. And so I'm always amazed by this example, and I visualize it like being inside a nuclear reactor control room and the thing's melting down and there's a million different buttons and knobs and so on, and figuring out which ones do I need to flip in order to solve my problem. I mean, think about what's happening here. They're under an incredible stressor. They're under a physiological stressor, one for which evolution did not really prepare them. Planarians don't see barium in the wild. And so they've got this novel physiological stressor, and they've got tens of thousands of potential actions they can take, those being genes that they could turn on and off. So those are vectors in transcriptional space. So they're looking for a solution, a behavior in transcriptional space, meaning turning certain genes on and off to solve a physiological stressor. And very rapidly they find it. They don't have enough time for any kind of random search or not even hill climbing and things like this, because it happens pretty fast and their cells don't turn over that fast. It's not like bacteria where everybody can try a different strategy and then most of them die and then somebody survives. The planarian cells don't turn over like that. And so they really need to find a solution in place. And by the way, this is the most recent data, if we do this in a collection of worms kept in separate dishes, they all find the same solution. So apparently there aren't, or at least not easily found, multiple solutions to this problem; they all roughly turn on and off the same set of the same handful of genes. So how is it that they navigate that transcriptional space to find a solution to a problem they probably have not seen before? Certainly these worms haven't seen them, but maybe the lineage has seen something similar. For example, maybe epileptic seizures may have some similarity, but certainly not exactly this. We have no idea how it finds so rapidly the vectors that it needs. And so I consider this a really interesting example of problem solving, of resolving a novel stressor using the genetic affordances you have. And in a, I don't know if they have some sort of internal self-model through which the cells can figure out what actions they should take to resolve specific stressors. I'm not sure. This is a big open question.

Another example that I really like is this. This is a morphogenetic one. This is a cross-section of a kidney tubule that's got a lumen and a bunch of cells making the actual tubule. And what you'll see is that there's maybe 8 to 10 cells working together to form this kind of structure normally. Now you can make polyploid newts, which is basically take the fertilized egg and do a procedure so that it ends up with more copies of chromosome numbers. So it has extra copies of the genetic material. The cells get bigger to accommodate the bigger nucleus, but the newt stays the same size. And this is Fankhauser's work, and that's amazing. And when you look to see how that could be possible, you see what's going on. It uses fewer yet larger cells. Okay? So the number of cells adjusts to the size in order to give you that same structure. Now, something really amazing happens when you make a really highly polyploid newt. I think these are like 6N or something like that. And the cells get truly gigantic, and then just one cell, what it will do is wrap around itself to still give you a lumen. There's a few amazing things about this. First of all, notice the adjustments and the adaptation. So they adjust the cell number to the new cell size, and when needed, they actually pick a completely different molecular mechanism to do so. So first of all, they're using cell-to-cell communication and tubulogenesis, but here in this case, it's cytoskeletal bending. It's just one cell. So there's this amazing example of top-down causation. You have the collective working towards an anatomical goal, and it's able to not only instruct the cells what to do, but it's able to actually pick different molecular affordances in its toolkit to get the job done. And this is a standard sort of example of IQ tests where they give you a set of objects and they ask you to do something with these, you know, use these objects to solve a particular problem, like it's a standard IQ test kind of thing. And imagine too what it's like to be a newt coming into the world, or actually I think this is almost every creature. Never mind the environment and the fact that you can't really know what's going to happen with your environment. You can't even trust your own parts. You don't know how many copies of your genome you're going to have. You don't know how big your cells are going to be. You don't know how many cells you're going to have. This is some other work by other folks who showed kind of an invariance to the number of cells that you start with. And you have to get the job done with the tools you have, regardless of what's going on, not only in the environment but in your own parts. Even your own parts are unreliable. And this turns out to be, biology as an unreliable medium turns out to be actually very important for the development of intelligence. I might mention that towards the end. But basically this is not only goal-directed, meaning that under all of these circumstances the system makes the same large-scale shape, but actually uses some very clever different ways to get there when circumstances change in a radical way.

And then I want to get to this final bit, which is this: the other thing about life and intelligence is that beyond the basics of being able to achieve a single goal and then stop, and then the next level. So that was basic homeostasis, and then I showed you a kind of second-order homeostasis where you've got, you're able to reach the same end but you can do so by different paths. And now there's something else, which is, if you can't reach your same goals, or if you're a novel creature that's never had the benefit of an evolutionary history that selects for specific goal states, what goals will you have? Do you have the ability to form novel goals and find new forms and behaviors that you can adopt? And so this is a kind of, I think this is progressively, I mean basically I organized this in terms of progressively more sophisticated cognitive kinds of capacities. And so even beyond creative problem solving toward default goals, let's look at some novel beings and this idea that life is actually good at picking new problems, not just solving old problems. And the first thing that I'm going to show you is xenobots. So, xenobots are self-organized little motile creatures. They're biobots as well, composed of frog epithelial cells. So, in the early frog embryo, we take these cells, we dissociate them, we put them by themselves here. They could have died, they could have crawled off, they could have made a flat monolayer like cell culture. Instead, what they do is they coalesce together and they make this. So they combine, and the way they combine is actually amazing. Take a look at here. Each one of these circles is a single cell. Okay. So here they are. Look at this little clump. It's kind of fun that it looks like a little horsey, and it sort of wanders over here using some motility mechanism we don't understand, and it's sort of sniffing around, and then eventually you get this little calcium flash of some sort of communication event going on here. And so all of these things, there's, if you could stare at this all day, there's all kinds of amazing patterns, but eventually they all coalesce into this thing, and then it has some interesting behavior. So the outer surface is covered with cilia, and it organizes the cilia so that they can swim. So here it is swimming through some particles. It can go in circles. It can patrol back and forth like this. You can make them into weird shapes like these donuts. They have collective behaviors or group behaviors. So they can, you know, these two are interacting. These are sitting there doing nothing. This one's going on some sort of long journey.

Here's one traversing a maze. So there's no flow here, no gradients. So watch, it takes the corner without bumping into the opposite wall. So it takes the corner, and then at this point it spontaneously turns around and goes back where it came from. Okay. So they have all these interesting behaviors after they self-organize. Interestingly, if you look at calcium signaling, they have some very lively calcium signaling. We have been, in collaboration with Josh Bongard's lab and Sarah Walker, so people like Thomas Varley has been analyzing the calcium flashing and using metrics that are used to analyze brains. So connectivity, that is information structures, but also in more recent work, some amazing studies of various causal emergence and other metrics that people use on brain activity to try to determine if you're looking at a human, somebody a pile of neurons, somebody with locked-in syndrome, or somebody who's anesthetized and so on. So the idea is to use these metrics from neuroscience to understand the collectivity of the system. So we've been using that, and it's really interesting. Stay tuned for that. Even though there's no neurons here, these are all epithelial cells, and yet those same metrics are revealing some very interesting information structures here. Now, one of the remarkable things about these is that while they can't reproduce in the normal frog-like fashion, they do something quite different. They solve it a different way. If you provide them with loose epithelial cells like here, they move around both individually and as a group, and they push all of them into little balls, and then they polish the little balls like this. And because they're dealing with an agential material themselves, meaning cells, these balls mature to be the next generation of xenobots. And guess what they do? They run around and they make the next generation, and these make the next generation, and so on. So we call this kinematic self-replication. It is kind of von Neumann's dream of a robot that goes around and collects materials from the environment and it makes copies of itself. This is something that as far as we know no other creature on earth does. We don't know of any other examples that reproduce by kinematic self-replication. And of course the cool thing is there's never been any xenobots. There's never been any selection to be a good xenobot. There's never been a selection for kinematic self-replication. And this is just something that they do. And I think it's really important, and there's a research program to help us get better at predicting these kinds of things, you know, to not be surprised and simply label them as emergent when we do get surprised, but to really understand what are the competencies of the material that we're working with that are not well explained or made predictable by knowing the history of selection under which the genome arose. Because of course, like with everything else I showed you, this is a wild-type genome. So when you ask what does a Xenopus laevis, which is the frog we're working with, what does that genome encode? Well, it encodes tadpoles and frogs, sure, but it also encodes this, and who knows how many other things that have very different lifestyles. And you get all of that as a kind of software byproduct in the biological hardware that that genome encodes.

And so we asked one other question recently, which was, these creatures with their new lifestyle, they have the same genome, but what does their transcriptome look like? What kind of genes do they express? And when we looked at the differential gene expression between xenobots and age-matched frog embryos, we found hundreds of genes that xenobots expressed that frog embryos don't express. Now notice, going back in the other direction doesn't make any sense because of course they're missing a whole bunch of genes because they don't have all the tissues that these guys have. So we ignored that. We just looked at what new genes do they express that normal embryos express at a very low level or not at all. And there's hundreds of them, and there are many interesting things. But one of the things we found in this drastically remodeled transcriptome is a cluster of genes that are normally related to hearing and the perception of sound and so on. And so we asked, could it be possible that they can hear? And it turns out, so we put a speaker under the dish. It turns out that when you play sounds to them with a speaker underneath the dish, they change their movement. In other words, during the sound they move in a completely different way, and then they go back to almost the same movement after it stops. So they can in fact react to sound. And so they have a very different transcriptome that is just a function of their new lifestyle. They can basically dip into their genome and change transcription based on their new lifestyle, not requiring any kind of change of the genome. And we don't know what the purpose here is, but for example, it is known that coral larvae react to sounds made by different kinds of coral reefs by the surf hitting the coral reefs, and they use this to guide their decisions of where they're going to go. So I don't know if that's the kind of thing these guys are trying to do, but this is just the beginning of understanding these transcriptional changes and why these lifestyles are driving specific genes to be expressed in a coherent manner.

The next thing I want to show you is another biobot called an anthrobot. Now, when you look at this, it kind of looks like something that would have come from the bottom of a pond somewhere. It looks like a primitive organism. If you try to guess the genome, based on those criteria, you would be very wrong. The genome here is 100% normal Homo sapiens. These are made of adult human cells, unedited, from tracheal epithelia. They basically self-organize into these motile little creatures. It's swimming here in a group of other cells. We tried to do this, like, I think it's trying to do the same thing that xenobots do, which is to collect these things into a ball, but it doesn't, it's, you know, it doesn't seem to do kinematic replication successfully the way that xenobots can. But in any case, and by the way, both for these and the xenobots, I'm not making any claims about their specific capabilities as far as cognition, because we're just now starting to study their memories, their ability to learn, their preferences, what problems they can solve, and so on. We don't know that yet. So not making any claims about it. But morphologically what's happened is that these cells have basically formed a new kind of creature that doesn't look anything like any stage of normal human development. It just has a completely different lifestyle. It doesn't look like anything that you would expect from human cells. It has a number of interesting properties. It can self-repair. So what you see here, after mechanical damage, the anthrobot actually can repair itself, and it will repair itself towards the new pattern. So in other words, the kind of homeostasis you see in limbs in amphibians and in planarian fragments, this thing does eventually repair itself to the new anthrobot shape. And they don't just repair themselves, they also repair other things in their environment. So if you grow a bunch of human neurons on a petri dish, you take a scalpel, you put a big scratch through it, here's a neural wound. The anthrobots, which are here labeled in green, will collect, and we call it a superbotic cluster. There's maybe a dozen of them collecting here. And what you'll see after about 4 days is that what they're doing is they're knitting across the gap. They're healing. This is what it looks like. They're healing the injury. So who would have thought that your tracheal epithelial cells, which sit there quietly in your airway for long periods of time dealing with mucus and air particles and so on, that they have the ability to comprise a novel self-motile little creature that by the way also knows how to heal neural wounds. And so, again, this is what it looks like. These little guys have cilia covering their surface. That is how they move. There are several different types. They have four different behaviors. They have discrete behaviors. And we can draw an ethogram of transition probabilities between their behaviors. And even more than xenobots, they have thousands, about 9,000 transcripts, so half the genome, that they express differently than their tissue of origin. So again, no genetic change, but in a new environment, they dip into their genetically provided toolkit and express thousands of genes differently. And who knows, we haven't even begun to scratch the surface of what they're actually capable of. So that plasticity, this idea that you're a bunch of tracheal epithelial cells in a matrix, you're not going to be able to make an entire human, but you can make something else, and you can make something else and then repair to that goal and express novel genes and exert other types of effects on your environment that we're not very good at guessing or predicting up front. So again, no history of selection to be a good anthrobot. There's never been any anthrobot. So I think that plasticity is part of the spectrum of intelligence that we're all talking about.

So I'm going to start to wrap up here, but I just want to mention a couple of other topics that we could talk about. The first is the intelligence ratchet. Basically the competency of the material. The fact that the material is an agential material that does teleological types of problem solving and also many other things makes evolution go completely differently. And you can see that in these two papers we did a lot of simulations and some analysis of what happens when you evolve over an agential material. And it basically ends up making this really interesting ratchet that just keeps cranking up the different types of intelligence. And so there are implications for evolution. We could also talk about a story of the cognitive light cone. We define the cognitive light cone as the size of the biggest goals that a system could pursue. So not how far it can sense or act, but the size of the biggest, in whatever space, the size of the biggest goal it can pursue. And you can see what's happening here with single cells having a tiny little cognitive light cone that only cares about this region of space and a little bit of anticipation potential, a little bit of memory. But when you gather into collectives, you can have these grandiose goals. So these things have the goal of building this limb, and again, you know, it's a goal because if you try to prevent them by cutting it off, it will keep rebuilding it. But this is all part of a story around the scaling of goals in teleology. So you can be a teleological system, but you can have little tiny, sort of humble goals, or you can have these very large grandiose construction projects as goals, and that way of thinking about it has some great implications for cancer therapeutics. And we're pursuing ways to specifically understand cancer as a dissociative identity disorder of this collective intelligence. When cells disconnect, you can actually use that as a cancer detection modality. And in fact, you can forcibly reconnect them despite their ankyrin proteins and so on, and basically prevent or reverse tumors. And so this is where we, you know, we did this in frog, and now it's in human cancer spheroids and so on.

And then the next thing I'll mention is what it is that drives this homeostatic loop. So one thing that you can imagine, and all of this is discussed here, is that stress and the sharing of stress among cells is part of the engine that keeps cells in this homeostatic cycle. So being away from your goals is stressful, and there are some interesting collective dynamics of how stress scales to enable larger and larger systems to pursue larger and larger goals. And then the final kind of part of this story is that the same issue with an unreliable medium affects both cognition and evolution. In other words, both memory engrams and genetic information have to be interpreted. They cannot be taken literally because of this bottleneck where information is squeezed down into this thin representation in latent space. And so this is learning, and then from here this is the now moment where information has to be creatively reinterpreted. And this, I think, is what provides this amazing plasticity. It's because from the very beginning of life the material could not take any information literally. It had to take them as prompts or suggestions. And evolution cranked on this part, which is the creative, confabulatory aspect of life that takes its information, whether those be your memories or those be your genetic material, and does its best to tell not the same story that was told before necessarily, but the most adaptive story it can tell going forward into the future. And all of that is described here. And again, I think this is part of that intelligence ratchet.

And the actual final thing I'll say is this: what we've been looking at today, all during this talk, is different aspects of intelligence in the living material. And the thing is that unexpected cognitive competencies like goal directedness, like different kinds of problem solving, is not just the province of complex biological systems. It turns out that even extremely minimal deterministic simple systems, and in this paper here we studied sorting algorithms, so things like bubble sort and so on that people have been studying for, I don't know, probably 60 years or more, and we found out that even in that system, if you look at it the right way, it's actually doing some very interesting side quests that are not anything that the algorithm is telling them to do. So in other words, even simple systems can provide unexpected navigation of various spaces. And so this means that we need to be very careful about assuming that large complexity is necessary for some of these things, or that quote unquote passive matter isn't doing some of the things that life is doing. In this regard, I think we are really bad at noticing cognition in unexpected places. And for this reason, I like looking for it in these very surprising, very minimal systems. And so I think we really need to get better at noticing it in molecular networks and other kinds of substrates. So I'll stop here and just thank all the people who did the work. So the postdocs and PhD students and our many amazing collaborators. And I have to do a disclosure. There are three companies that have spun out of our work that provide some funding for some of this research, and various other funders over the years to whom I am extremely grateful. And of course, lots of thanks go to the actual model systems which do all the hard work as we try to figure this stuff out. So thank you very much.