Watch Episode Here
Listen to Episode Here
Show Notes
This is a 1 hour 8 minute talk covering topics of collective intelligence in morphogenesis as a model system for thinking about the origin and scaling of cognition in diverse embodiments.
CHAPTERS:
(00:00) Diverse Intelligence Framework
(12:00) Morphogenesis Problem Solving
(22:00) Bioelectric Pattern Memory
(34:30) Exploring Latent Space
CONNECT WITH ME:
Transcript
Thank you so much. I very much appreciate this opportunity to share some ideas with all of you. What an amazing audience to speak to.
For today's talk, if you're interested in any of the details, the software, the papers, the data sets, everything is here. And this is a blog where I talk about more personal views of what I think these things mean.
I'd like to do an arc today that has three fundamental points. First, I'm going to talk about the field of diverse intelligence and the idea of an agential material that is problem-solving in unconventional substrates. I'm going to talk specifically about one type of cognitive glue which helps us to solve the scaling problem – that is, how emergent minds arise from the competencies of their components. I'll specifically talk about morphogenesis, development, regeneration, and cancer as a model system for understanding collective intelligence. After all of those examples, we'll come back to kind of the big picture of some thoughts about novel embodied minds and the origin of anatomical but also cognitive patterns. We'll specifically talk about emergent intelligence as distinct from just emergent complexity.
In my group, everything we do boils down to the use of collective intelligence, whether they be groups of cells or molecular networks and so on, navigating different problem spaces and trying to show how philosophical ideas in these kinds of questions can actually become therapeutics. I won't focus on that today, but probably two-thirds of my lab does things that are very practical and apply these ideas to biomedicine of various kinds.
So the talk will be in three parts. The first thing I'm going to do is introduce the way that I try to think about these things, and then we'll go into some examples.
This is a well-known old image called Adam Naming the Animals in the Garden of Eden. There are two things about this that are interesting: one that I think is deeply wrong and one that I think is profoundly correct. The thing that we're going to have to change is the idea that there are discrete natural kinds here. It's very easy to tell who's who. This is Adam, these are the various species, and we can spend our time on the "easy" problem of discovering what the natural competencies of all these different creatures are, what kinds of minds they have, and so on. That, I think, as I'll argue in a minute, we're going to need to change.
What is profoundly correct about this is that in these old traditions, discovering or naming something means that you've understood its true nature. Giving something a name is basically the idea that you've understood what it really is, and it gives you power over that thing. I think we are going to have to name all kinds of really unusual beings in the coming decades and learn to understand their true nature so that we can relate to them. I think that will become very important, and I'll show you some of those towards the end of the talk.
The first and most clear thing that we need to change about this is this idea that we are this well-demarcated human. There's lots of philosophy and ideas out there about what humans do and how what they do might be different from "machines" and so on. But we need to understand that we stand at the intersection of several smooth continua. One is this: both on an evolutionary time scale and on a developmental time scale, we are slowly and gradually changing from very different beings. So whatever you think is true of this modern adult creature, you have to be able to say where that came from and how did it either emerge or scale up from other things that were going on before then.
Now, with advances in morpho-engineering, synthetic biology, and biotechnology of various kinds, we also know that we can make smooth and gradual and perhaps drastic modifications both on the biological end and on the technological end. We are now able to introduce engineered components across the size scales of living systems, and whatever conceptual frameworks we have for dealing with embodied minds, they need to be able to apply to all of this stuff. It is not enough to be able to say things about standard humans.
I've been working on a framework whose goal it is to be able to recognize, create, and ethically relate to truly diverse intelligences regardless of what they're made of or how they got here. This means familiar kinds of creatures like primates and birds, but also unusual things like colonial organisms and swarms, engineered synthetic new life forms, AIs whether purely software or embodied in robotics, and maybe someday exobiological agents.
The requirements for any such framework that I like are that it has to move experimental work forward. It cannot just be philosophy; it has to actually lead to new discoveries, new capabilities, and new research programs. It has to help us improve or refine ethical frameworks for relating to the unconventional beings all around us. All the details are here, and I call it TAME, T-A-M-E, Technological Approach to Mind Everywhere. Obviously, I'm not the first person to try for something like that. Here's Weiner, Rose, Rosenbluth, Weiner, and Bigelow who tried for this kind of scale all the way from passive matter up to the sort of second-order metacognition that humans have in trying to understand how these things scale up.
What I like to think about is something I call the axis of persuadability. What this means is that I view cognitive claims as interaction protocol claims. If you think that something has a certain degree of cognition, what you're really saying is there's a bag of tools, maybe including rewiring, maybe from cybernetics, control theory, behavioral science, psychoanalysis, whatever it's going to be. You're somewhere along the spectrum, and there's a certain set of tools that are going to be optimal in relating to that system. Back here, it looks like prediction and control. Out here, it's some sort of bidirectional enrichment, friendship, love, whatever. But in any case, this is all about finding out what is the best way to interact with a system.
I think there are two key parts to this. The most important being that we can't simply have armchair feelings about where things land on the spectrum. We have to do experiments. We are sometimes strongly motivated to assume that something is somewhere along this continuum, but actually, as we found, there are many, many surprises. So we have to commit to the idea that you have to make hypotheses and do experiments as to where something is going to be on this. That's good. That means that this becomes an empirical problem where you can hypothesize a set of tools. Somebody else hypothesizes a different set of tools as appropriate. You both try it, and we all get to find out who had the better experience.
The other thing about this is that the problem space within which these systems can work could be very difficult for us to recognize. We as humans and as scientists are okay at recognizing intelligence in medium-sized objects moving at medium speeds, so birds and primates and maybe a whale or an octopus or something. We can sort of see what's going on in three-dimensional space. But life uses many other problem spaces: transcriptional space or the high-dimensional space of gene expression, physiological state space, and anatomical morphospace, which we will spend most of the time today talking about. These other spaces are spaces in which different types of agents can run perception, decision-making, action kinds of loops, and those are very difficult for us to visualize.
I think that if we had evolved with a primary sense of our blood chemistry, the way that we can taste and smell and so on, if we could sense a bunch of other parameters in our blood chemistry, I think we would have no trouble visualizing that our liver and our kidneys are intelligent symbionts that traverse these spaces daily and try to achieve certain goals and keep us out of trouble and so on. But it's very hard for us to observe behavior in these other spaces.
It's also the case that the distinction between these spaces is very artificial. Just to show you one example, this is a slime mold called Physarum polycephalum. We put a little bit of the slime mold here. There are three glass disks here. There's one glass disk here. They are completely inert. There's no food or anything like that. They're sitting on an agar substrate. What you will see is that here it sort of grows out in all directions. But what it's doing that you can't see during all this time is it's gently tugging on the medium and receiving back information, biomechanical information about the strain angle. In the end, it's able to figure out where the larger mass is, and it does this quite reliably.
So what it's doing is obtaining information about its environment. Up until here is sort of cogitation time, and then now it's made a decision and bang, there it goes. The thing about it is that this is all one cell. This whole thing is just one cell. There are no neurons. It's just one cell. But what's interesting for the current point here is that in this creature, morphogenesis is behavior. It's in the same space. We like to divide those things, but this thing behaves by changing its shape. This is literally its outgrowth. So what we see as spaces is not necessarily what the creature itself sees, and we need to take that into account. That just reminds us that this is not about humans, and scientists in particular. This is about all observers.
I like to think about bodies, and this is a poly-computing framework that Josh Bongard and I developed, where bodies are composed of numerous nested, interpenetrating systems that observe and try to hack each other constantly. We need to understand what each of these systems is seeing, what space they're navigating, and what their goals are.
Let's think about cognition from the beginning. We all start life here as a little blob of chemistry and physics. We all look at an oocyte, and many of us say, "Well, this thing just kind of follows the laws of chemistry. There is no mind. There is no cognition. It's purely mechanical." But what we know from developmental biology is right in front of your eyes, for many of these species, it will slowly turn itself from this kind of system to this kind of system, right? We all make this journey. At some point, we become capable of voicing this idea that we are more than quote-unquote "machines." So we need to understand where and how that happens, right? What is the scaling that allows this to take place?
Many people, once they start thinking deeply about this, find this disturbing because it really emphasizes the continuity of whatever it is that we are with very basic physical mechanisms. But at least we are a unified intelligence, right? We have this nice centralized brain. We're not like what Ricard Soay calls a liquid brain, a collective swarm intelligence like ants and bees and things like that. In particular, Descartes really liked the pineal gland because there's only one of those in the brain, and he felt that that unity was appropriate to the kind of coherent, unified perspective that we as humans enjoy. But what he was missing was access to good microscopy because if he had had that, he would've been able to look and find that there isn't one of anything. Inside the pineal gland is all of this, a huge number of individual cells. And inside of each of those little cells is all of this stuff, right? There's an incredible nested complexity.
To some extent, we are all collective intelligences, not just the ant colonies. All of us are made of parts, and our parts in particular are very clever. This is a single cell. This is an animal called the Lacrymaria. There's no brain, there's no nervous system, but you can see it has this amazing control over its morphology here. It's hunting for food in its environment, and it's very competent at single-cell level agendas. This is the kind of thing we're made of.
What we really need to do is to develop models of the scaling of competencies of our components into the kinds of things that can store and process goals and memories and preferences that we have that none of our parts have. In fact, the competency goes below the single-cell level. If you wonder how this thing is doing what it's doing, well, what it has is a bunch of biochemical networks. We've now shown that even very simple models of very simple molecular networks, for example, a five or six node gene regulatory network, can, if you actually do the experiment and not assume that they're hardwired and thus boring and mechanical, you can actually train them, and they will do six different kinds of memory: habituation, sensitization, even Pavlovian conditioning, by exposing them to different stimuli and watching some output nodes. In fact, it's all described here, and we've now built some devices to try to actually take advantage of that and train cells for applications in biomedicine, drug conditioning and things like that. So even inside of single cells, you already have, just from the virtue of the math alone, you don't need a nucleus, you don't need all this other stuff that cells have. Just these kinds of simple things can already do certain kinds of learning already. I'm sure this is just scratching the surface of what's there.
In our body is a kind of multi-scale competency architecture. We are not just nested dolls structurally, but at every level there are components that solve problems in their own space. That kind of architecture has some really exciting and interesting implications. In particular, I think it has not only for biomedicine, but it has implications for trying to understand what minds are and where they come from.
What I want to do next then is to spend some time talking about one particular example of an unconventional agent that navigates an unfamiliar space, which is anatomical morphospace. We will talk about morphogenesis as a collective intelligence, and I'll show you the problems that it can solve. We'll talk about bioelectricity as the communication interface to that intelligence and kind of talk about its properties as a cognitive glue.
I always thought it was interesting that Alan Turing, who was very fascinated with problem-solving machines and intelligence through plasticity and reprogrammability and those kinds of questions, had this amazing paper where he asked the question of how order in development arises from chemicals. I think that he was onto a very great truth, which is that the self-assembly of the body and the self-assembly of minds have a fundamental symmetry to them, that there are some very deep common things that underlie both of these things. So I thought that was interesting, that he was onto this a really long time ago.
Let's look into this. Where do anatomies come from? This is a cross-section through the human torso. You can see all of the incredible complexity here, all the organs the right size, the right shape, next to the right thing. Where is this pattern coming from? This is an early embryo. It's a jumble of embryonic blastomeres. Where does this all come from? You might be tempted to say it's in the DNA, but of course, we know that DNA doesn't directly encode any of this. What the DNA specifies is the tiny protein-level hardware that every cell gets to have. The DNA specifies the proteins. It is the physiological activity of that hardware that gives rise to this, and this pattern is no more directly in the DNA than the structure of these termite mounds or the precise shape of spider webs are in the DNA of those creatures. There are hardware specifications, and then there is the outcome of what that hardware does.
So we need to understand, first of all, how do the cell groups know what to do, when to stop, and how much plasticity they have. We need to understand how to communicate with them, because sometimes you may want to recreate something that was damaged or missing or degenerating. As engineers interested in A-life, we would like to know what else is possible. Given that same hardware, what else could they build? Is this the only thing that these cells can build? We like to think so because, you know, cats have kittens and dogs have puppies, but I'll show you some amazing plasticity where it becomes clear that there is maybe not universality, but certainly not hardwiring between what the cells are capable of and what they actually end up building.
If we think about running this all the way forward, what would it mean if we solved this problem? I think it's important to think about what an answer to these questions actually looks like. One way to think about this is as the anatomical compiler. Someday, you could be sitting in front of this machine, and you would draw the plant, animal, organ, or bio-bot that you want in whatever shape you want. So, not the molecular properties of it, but actually the anatomy. The thing we actually care about is its functional form. In this case, we've drawn this nice three-headed flatworm. If we knew what we were doing, we would have a system that took this and compiled it down to a set of stimuli that would have to be given to cells to get them to build exactly this.
Now, it's obvious why we need this, because if we had something like this, all of this would go away. Birth defects, traumatic injury, cancer, aging, degenerative disease would be a non-issue if we had the ability to tell cells what it is that we want them to build. But why don't we have this? Molecular biology and biochemistry have been advancing for many decades. Why don't we have anything remotely like this? I think it's because we have been thinking about this all wrong. This is not supposed to be a 3D printer or some other way to micromanage cell behavior. This should be a communications device. This should be something that allows you to translate your goals as the worker in regenerative medicine onto the goals of the cellular collective.
Specifically, where molecular medicine today is at, we're very good at manipulating the hardware, so which cells interact with which other cells, what proteins are made, and so on. But we're really a long way away from control of large-scale form and function. If somebody loses a limb, or there's a birth defect, or somebody wants a different shape, in the generic case, we have no idea how to do that. I think that's because biomedicine is still where computer science was in the '40s and '50s, where people think that the correct level of interaction is through the hardware. Think of all the exciting advances: CRISPR, protein engineering, pathway rewiring. All of those things are down at the level of the hardware like this. But what we haven't done yet is what computer science has done, which is take advantage of the reprogrammability and higher-level interactions with the system that does not require you to change the hardware.
I think the big stumbling block has been that everybody understands that biology's complex, but I actually think there's a lot more to it. It's not simply complexity that shows up, but actually it's emergent agency. I'll repeat this theme a couple of times as we go along.
I'm going to show you what I think are examples of a collective intelligence, and what I mean by intelligence, which is not to say that I think this is the one correct definition or that it encompasses everything that everybody wants to subsume under that term. But I like this definition because it's nice and practical, by William James: it's the ability to reach the same goal by different means. Some degree of reaching a goal in a space despite various things that might happen along the way, perturbations. I like it because it's very cybernetic. It doesn't say you have to have a brain. It doesn't say what the space is or what the goals are or anything like that. It's quite generic.
So we can think about all ranges of systems from two magnets, which if separated by a barrier will never come around and meet each other because they cannot go further from their goal in order to recoup gains later. They don't have this delayed gratification. Here's advanced Romeo and Juliet who have all kinds of long-term planning and all sorts of other tools to avoid physical and social barriers. In between, you've got your self-driving vehicles, cells, tissues that have some degree of capacity greater than this but probably smaller than that.
Okay, so let's get to it. What kind of collective intelligence do cellular swarms deploy? The first thing you notice about development is that it's incredibly reliable, so almost all of the time these cells produce exactly what they should. Also, there is a massive increase of complexity. You go from this system which is already quite complex, but it becomes much more so. Now, I want to be super clear that when I say morphogenesis has intelligence, I do not mean either the reliability of it or the increase in complexity. Just the fact that you get from here to here is not, in my framework, a sign of intelligence. What I'm talking about is problem-solving and navigation with unexpected scenarios.
You can easily produce some of those by cutting these embryos into pieces, or in fact rearranging the pieces, or in fact mushing multiple embryos together like a snowball. If you do that in mammals, you don't get half bodies. You don't get confused bodies. If you do it early enough in development, you get perfectly normal monozygotic twins, triplets, and so on. That is because the system, if you, for example, cut it in half, it immediately figures out that half of it is missing. It will regrow whatever it needs. So in this space, you can start off from different starting positions, avoiding local maxima, and end up within the correct ensemble of states in the anatomical space corresponding to a normal human target morphology. It can make up for quite a bit, and we could spend a whole hour just going through these kinds of examples.
Some lucky animals are able to do this throughout their lifetime. This guy is an axolotl, and they regenerate their eyes, their limbs, their jaws, portions of their heart and brain. If you amputate anywhere along the limb, you discover that the cells immediately spring into action. They build exactly what's needed, and then they stop. This is the most amazing thing about that process: it knows when to stop. What you really have here is an anatomical homeostatic system that when you deviate from the correct position in morphospace, it will work really hard to get back there.
Here's another example that we discovered a few years ago. This is a tadpole. Here are the eyes, the nostrils, the brain, the gut. These tadpoles are supposed to become frogs, and they have to rearrange their face in order to do so. Their jaws, their eyes, their nostrils, everything moves. It was thought that this is a hardwired process. After all, if every organ moves in the correct direction the correct amount, you'll go from a normal tadpole to a normal frog. Easy.
Laura van den Bergh in my group decided to test this hypothesis. What we did was create these so-called Picasso tadpoles. We scrambled these cranial facial organs, so the eyes are on top of the head, the mouth is off to the side. Everything is kind of mixed up. What you find is that these give rise to quite normal frogs because all of these organs will move in novel, unnatural paths, sometimes going too far and even coming back, until you get a normal frog face, and then they stop. So the genetics does not give you a hardwired set of movements. What it actually specifies is a highly flexible error minimization scheme. It gives you a system that can move through morphospace and recognize when the pattern is incorrect, get to where it needs to go, and then give you the correct pattern, and then stop.
That raises the obvious question: how does it know what the correct pattern is? Where is it going? What controls its navigation through morphospace? We started to add... So this is the standard view of developmental biology. It's very much feedforward and emergence. The idea is a focus on emergent complexity from low-level rules. The genes are interacting with each other here. They make some proteins that do things. Then there are some laws of physics, and then eventually, voila. Something like this emerges as a complex agent. We all know that there are lots of simple systems, from cellular automata up, that will give you complexity from very simple interaction rules. That's easy.
But what we actually see here is that this is not the end of the story. In fact, it's only the beginning. If this system is now deviated from this by injury, by mutation, by teratogens, whatever, then what happens is that mechanisms kick in both at the level of physics and genetics that try to get you back to where it needs to be. So now you see this pattern homeostatic system. It has a simple goal. It has simple navigational capacities, and this is not just a feedforward emergence of complexity. You actually have a state that the system works really hard to maintain against perturbations.
That has important specific implications. This idea of goals and arguing about whether these kinds of biochemical signal systems can be set to have goals are very practical. This is not just philosophy because if you buy into this idea that this is an emergent complex... that all of this is the result of emergent complexity, then the only game in town is to modify these genes and see what happens because doing this backwards is not reversible. Development is computationally irreversible, which is why CRISPR and some of those technologies have a real ceiling to them. When you want to make changes here, how do you know which genes to edit? This is generically a very difficult problem.
But if we're right and there is this kind of a homeostatic system, then there's a hope of a completely different strategy. Find the encoding of the set point, modify that set point—interpret it and modify it—and then let the system do what it does best. In other words, do not interfere with the hardware. Let the hardware do exactly what it does, but change the set point, and that requires you to find the set point and then decode the set point, and this is what we've been doing.
Now I'll just take up an aside for the moment to just tell you that it's actually even much richer and much more interesting than that, although this particular example I don't have any explanation for. But just to point out how much intelligence there really is, this is one of my favorite examples. This right here is the cross-section through the kidney tubule in a newt. Normally, there are eight to 10 cells that work together to form this structure. One thing you can do is you can make polyploid newts that have multiple copies of their genetic material, and when you do that, the cells get bigger.
So the first remarkable thing is that you can change the number or copy number of all the genes, and you still get a totally normal newt. How many chromosomes apparently is not an issue, and the size of the cells adjusts to the size of the nucleus. That's pretty impressive, but it gets better. The newt that you get with these giant cells is exactly the same size as the original newt, which means that there have to be fewer cells to make up the exact same tubule. So now the whole structure adjusts to the increased size of the cells.
Even more amazing, the final thing is that if you make the cells so gigantic that there isn't room for more than one around the structure, what they will do is bend around themselves, leaving a hole in the middle, and make the whole thing out of just one cell. Now look at what's happening here. There are a couple of interesting things. First of all, there's a kind of interesting downward causation where in the service of this very large-scale anatomical structure, you are using different underlying molecular mechanisms. Here, this is cell-to-cell communication. This is cytoskeletal bending. So this is very much a kind of intelligence where you are confronted with a problem you've never seen before, and what you're able to do is figure out how to use the tools you have to solve this problem.
There are numerous molecular components. Every newt you've ever seen in the wild is doing this, but in this remarkable case where somebody does this crazy process involving pressure to magnify the copy number and so on, they can figure out that there's a different way to get to that same goal and give you that same newt. So think about you're a newt coming into this world. What can you rely on? Well, you can't really rely on your environment because we all know environments change, but you can't even rely on your own parts. You don't know how many copies of your genome you're going to have. You don't know the size of your cells. You don't know how many cells you're going to have. You have to do something very interesting, which is, and I'll get to this at the end of the talk again, which is to creatively get to your goals despite the unreliability of even your own parts and in fact your evolutionary history. All the things that have happened in the past are not necessarily a great guide to what's happening now, and this idea has many, many implications.
All of these... I hope I've convinced you that morphogenesis is not just about complexity. It is about problem-solving and, in fact, creative problem-solving. So that requires us to understand how all of that emerges from the competent cells that make it up. What we're looking for here is a kind of cognitive glue. We're looking for policies and mechanisms that help us to overcome this scaling problem.
In neuroscience, we kind of know what's going on more or less. Here's a rat that's been trained to press a lever and get a reward. The cells at the bottom of the feet interact with the lever. The cells in the gut get the delicious reward. No individual cell has had both experiences. Who owns the associative memory? Well, the rat does. There's this emergent being that has memories that do not belong to any of its components alone, and we know that what holds that being together in the conventional kinds of contexts is bioelectricity. The nervous system uses this architecture to accomplish that amazing feat of taking a whole bunch of neurons and making a collective intelligence out of them.
There are these ion channels in the cell membrane that establish a voltage gradient across them. These voltage states may or may not propagate across the gap junctions, which are electrical connections. That kind of a system runs this sort of software, which this group here is visualizing in a living zebrafish. It is the commitment of neuroscience, generally, that we should be able to do neural decoding. That is, we should be able to read this electrophysiology and decode it and know what the animal is thinking, what memories it has, and so on. The idea is that all of its cognitive structure is in some way encoded in this electrophysiological activity.
It turns out that the utility of electrical networks for integrating and aligning subunits into higher-level computational structures was noticed by evolution long before brains showed up, actually around the time of bacterial biofilms is when this first got going. Every cell in your body has these ion channels. Most cells have electrical connections to their neighbors. So what we started to do was to borrow a lot of ideas from neuroscience and to ask, could we do a kind of neural decoding here, except in non-neural cells? Could we ask, well, what are these networks? We know what these networks think about. They think about moving your body through three-dimensional space. What do these networks think about?
It turns out there's a really strong symmetry here between development and cognition, and we exploited that in a couple of different ways. First, we developed tools. These are the first molecular level tools to read and write the information content of these electrical networks. Just to say it again, the claim here is that groups of cells are a collective intelligence whose behavior plays out in anatomical space. What these somatic ancient networks are doing is thinking about where you should be, where the configuration of your body should be in anatomical morphospace. Thinking about it that way allows us to deploy all sorts of tools from behavioral science and computational neuroscience and try to understand how this works.
We developed these imaging processes. Danny Adams made this video using a voltage-sensitive fluorescent dye, and you can see all the electrical conversations that these cells are having with each other, who's going to be anterior or posterior, left to right. We do a lot of computation. Alexis Pytak made this amazing simulator, and we try to integrate from the expression of these ion channels all the way through the electrical dynamics of the circuits to the large-scale patterns and what happens during regeneration, for example, to understand pattern completion, the way that these electrical circuits can actually restore missing information if something is damaged.
Just to show you what these look like, this is a time-lapse video, also made by Danny. This is the craniofacial morphogenesis of an early frog embryo. The face is getting set up here. This is one frame from that video, and what you're watching here, the colors or the grayscale indicates the voltage of each cell, the resting potential of each cell. These are not spiking the way that neurons spike. These are steady. Basically, to turn almost any study in neuroscience into a developmental biology paper, you can just replace the word neuron with the word cell, and where it says milliseconds, you just say hours, and basically everything else maps very nicely.
So this is one frame from that video, and you can see here that already we know here's where the animal's right eye is going to be, here's where the mouth is going to be, the placodes, all of this. As I'll show you momentarily, this is determinative of what these cells are going to do in the future as they build the frog face. So this is an endogenous pattern. It is required for normal face development. This is a pathological pattern where we inject a human oncogene into these animals. It will eventually make a tumor, but even before it does that, what you can see with this voltage map is that these cells have electrically isolated from their neighbors, they've disconnected, and now they're just amoebas. As far as they're concerned, everything else around them is just external environment, and we'll talk about that momentarily. This is the work of Brooke Chernet in my group.
Okay, so tracking those kinds of things is nice, but what's really important are the functional tools. How do we actually change that information to be able to now write directly into the cognitive system of this morphogenetic intelligence? We do not use any applied fields. There are no electrodes, there are no magnets, no waves, no frequencies. What we do is exactly what neuroscientists do, which is to hack the interface that these cells are normally using to control each other.
So we can open and close these gap junctions, so that controls the topology of the network, and we can open and close these various ion channels using drugs or optogenetics or any kind of those tools. So we can set the distribution of electric potentials, or we can set the topology of which cell talks to which other cell.
Now comes the real question. Having done that, what can we do with this? How do we prove that these electrical patterns are determinative of morphogenesis? How do we communicate new goals to this system? I showed you here this little voltage spot here encodes the message of building an eye. So one thing you might imagine is what happens if we recapitulate this pattern somewhere else? Here's how we did that, and this is the work of Sherry Au and Vaibhav Pai. We inject potassium channel RNA into a particular region, in this case, a region that's going to give rise to this gut tissue. What this does is set the voltage and establish a little eye, a little spot that says build an eye here. And sure enough, these gut cells immediately build an eye. These eyes have the same components—retina, lens, optic nerve, all that stuff—as normal eyes.
So now we learn a couple of things from this kind of experiment. We can make eyes, we can make hearts, brains, a few other organs. What you learn is this: First of all, these electrical patterns are instructive. They tell the cells what to build. They're functionally determinative. We are really dealing with the system that controls the behavior. Number two, it's extremely modular. This is like a simple stimulus that kicks off a complex set of responses. We didn't tell these cells how to build an eye. We have no idea how to build an eye. It's too complicated. What we said was like a high-level subroutine call that says, "Build an eye here." That's it. Everything else is encapsulated. All the rest of the activity is encapsulated under there.
Next, if you read the developmental biology textbook, they will tell you that only these cells up here, the anterior neurectoderm, are competent to make eye. That is because they're using a prompt called Pax6. Pax6 is a transcription factor. It's a biochemical signal that induces eyes. And true enough, it only works up here in the anterior neurectoderm. But this, I think, is an important point for all of us working in this field, which is that any estimate of the intelligence of a system is basically just us taking an IQ test ourselves. All you're saying there is, what have we figured out how to get the system to do? If you stick to these kind of biochemical prompts, it's true, the cells do not look competent to make an eye out here. But if you use a much better prompt that is more salient to the system itself, which is this bioelectric state, then you find out that no, actually, pretty much every region of the body can make an eye if correctly communicated with. That speaks to the importance of not assuming that we're right when we find limits in these kinds of systems.
The final thing that I think is pretty cool is that this is one of the competencies of the material. This is a cross-section of a lens sitting out in the tail somewhere of the tadpole. The blue cells are the ones we injected. All this other stuff is recruited by the blue cells to participate in this project. They know there's not enough of them to make a proper eye, so we instruct these cells, "Make an eye." They automatically do the rest by instructing enough of these other cells to help them complete the project, much like some other collective intelligences recruit their nest mates to carry heavier loads and so on.
This is just one example. I won't dwell on all the biomedical stuff. We have a regenerative medicine program where we try to induce animals that don't regenerate their legs, like frogs, to regenerate their legs using a very simple, very brief trigger of a voltage pattern that says, "Go to the leg-making region. Do not go to the scarring region." You can see eventually you get this quite nice leg. It's touch-sensitive. It's motile. This is where I have to do a disclosure because Dave Kaplan and I have a company called Morphaceuticals, which is trying to move this now to mammals and eventually hopefully to humans. The idea, again, is using these biodome delivery devices not to micromanage it—not scaffolds, not stem cell controls, nothing like that—but to really just convince the cells that this is what they should be doing. A simple, high-level signal at the beginning.
I want to switch and introduce you to a different model system in which we can really see, I think, something very interesting about the way that biological tissues use electricity to store memories. This is a planarian. It's a flatworm. It has some really incredible properties, including immortality, cancer resistance, and so on. We can talk about that in the Q&A if people have questions. What we're interested in here is the fact that if you chop these guys into pieces, every piece makes a perfect little worm. They're extremely regenerative.
We asked the simple question, in a piece like this, how does it know how many heads to have? Why does this wound form a head? Because if you actually look at a single cut, the cells on this side of the cut will make a head. The cells on that side of that cut are going to make a tail. They were direct neighbors before you cut them apart. How come they have these widely differing anatomical fates? How do they know what they should do and how many heads they should have?
We did some of this bioelectric profiling, and we found something interesting. This fragment actually contains a bioelectrical pattern that says how many heads. The depolarized regions tell you how many heads you should have and where they should be. What we were able to do then, and it's a little messy still, the technology is still being worked out, but what you can do is you can introduce two of these regions. And sure enough, when the cells consult that pattern, they say, "Oh, two heads," and they build an animal with two heads. This is not Photoshop or AI. These are real creatures.
Now, here's something very important. Two key things here. First of all, what we are doing here is literally reading the set points, pattern memories, of this collective intelligence. That is, my claim is that these cells work together to maintain a memory of where in anatomical space they should go. Using these techniques, we can... This is the neural decoding, except it's not neural. We are reading directly the information that encodes this kind of information. We know it does because if you change it, the cells build something different.
But there's something else here, which is this. This bioelectrical map is not a voltage map of this two-headed creature. This map is a map of this perfectly normal-looking one-headed animal. One head, one tail. The molecular biology bears it out. A head marker in the front, no head marker in the tail. Anatomically and molecularly, this is a perfectly normal animal. But he's got one thing that's wrong with him, and that is he's got a weird internal belief, not the animal, the cells, that a correct planarian should have two heads. The way you know this is that if you injure it, meaning you cut off the head and the tail, they will then go ahead and build these two heads.
It is a latent memory, because until you injure it, nothing happens, right? He just sits there being normal until he gets injured, and that's when you find out that, oh, by the way, he's got a completely different internal representation of what makes for a normal worm. So this is kind of a simple version of a counterfactual. It's a simple version of our brain's ability to have this incredible time travel, like mental time travel capacity, where we can think about and remember and anticipate things that are not true right now. This state is not true right now, but that is what is going to guide you if you get injured in the future. The same body can store at least two different representations of what I should do if I get injured in the future. So I think this could be a simple model for thinking about how you can store counterfactuals in this kind of collective and so on.
I keep calling it a memory because, for example, if we were to take these two-headed worms and cut them again and again and again, in perpetuity, no more manipulations of any kind, just plain water, what you'll see is that once changed, that idea of how many heads should you have stays. That electrical state is kept. So this has all the properties of memory. It's long-term stable. It is rewritable. It has conditional recall, which I just showed you, and it has some discrete behaviors, one head versus two heads. We can set it back eventually, but once you change it, it stays.
Here are these two-headed worms hanging out. You can see what they do. Remember that in all of this, there is nothing wrong with this hardware. We did not change the genome. We did not put in any synthetic... There are no transgenes here. This is purely a short experience in the change in the bioelectrical state of the cells, that then, from then on, remember that there should be two heads. So no amount of sequencing, no amount of characterization of the hardware, the proteins, the RNAs, none of that will tell you what's different here, because it is not at that level.
Now, realizing then that what we're changing here is the information that guides the movement of this agent through anatomical space, we can wonder, "Well, where else can we tell it to go?" This controls head number, but what else can we do? It turns out that the other thing that is available in this morphospace are different shapes of heads. There are attractors in this space that belong to different species that have different shapes of heads. So now the question is, can we get the same hardware? Here's Dugesia doratocephala, nice triangular head, little auricles here. Could we get these cells to build one of these other species and get them to visit a different region of that space?
It turns out you can. So you cut off the head, you perturb the bioelectrical gradient, and you can make flat heads like Epiphilina. You can make round heads like an S mediterranea. Not just the shape of the head, but actually the shape of the brain and the distribution of stem cells become like these other creatures. There are about between 100 and 150 million years of evolutionary distance between these guys and him. There's nothing genetically wrong, but these cells are perfectly capable of visiting these other regions in the anatomical space.
You can go further, and you can visit regions that are not used by any planaria as far as we know. So here's this crazy spiky form. Here are some cylindrical shapes. Here are kind of a hybrid shape. One of the cool things about exploring this invariance or symmetry between developmental biology and cognitive science is that we get to use all their tools. We can take anxiolytics. We can take hallucinogens, plastogens, all of these neat compounds that are used, and we can ask interesting questions like, "What happens when the morphogenetic agent hallucinates, when you distort its perception of its environment and its memory patterns?" Well, you can get fish, frogs of a specific species to build heads that are appropriate to a different species. You can get them to make zebrafish tails instead of normal frog tails, and so on.
The interesting thing is that we humans are not the only hackers who do this. There is a non-human bioengineer. There's this wasp that prompts these leaf cells to build this kind of crazy structure. This is not made by the wasp. This is made by the leaves. We would have never known in a million years that these flat green cells, which do this so reliably in every oak tree in the world, we would have never known that it's competent to do this if evolution hadn't enabled this guy to push them to do it. So now, of course, the question is, okay, well, it probably took a long time for this lineage to learn to do that. Could we use modern tools and maybe AI to start to exploit these competencies in a rational way?
The final piece of this, I want to come back. Now that you've seen examples of morphogenetic problem-solving, you've seen the bioelectric interface as both a cognitive glue and thus a way to communicate with it for biomedical purposes, regeneration. We can repair birth defects and some things like that. Now I want to come around and talk about kind of the big picture of what is all this telling us about different kinds of cognitive systems?
The first thing that I like to keep an eye on is what I call the system's cognitive light cone, and that is basically a way to estimate the size of the biggest goals that a given agent can pursue. If we make... This is almost like a Minkowski diagram. You flatten all of space on this axis. Time is here. You can draw not the reach, the sensory-motor reach of the system, but actually the size of the goal state towards which they actively work. Another way to put it is what kinds of states make it stressed when those conditions are not met? So you can think about what do ticks care about? What do dogs care about? How far forward could they have a goal? Spatially and so on, and then humans and whatever.
So one thing you can do with this is you can plot out the different sizes of different cognitive light cones for all kinds of systems, no matter what they're made of, no matter how they got here. What's interesting is to think about the scaling. What the biology teaches us is that a single cell might have a very small cognitive light cone, and all it cares about is its own parameters, pH, hunger level, some other things like that. So very simple, very tiny little goals. But together in a network, in particular, in an electrical network, but electricity isn't magic. There are many other modalities you could use. You could network them, which immediately widens the cognitive light cone.
Here's a simple example. Here's a single cell pursuing its own tiny little goals. Now evolution and development allows the systems to scale up to these incredibly grandiose constructions. All of these cells are working towards this massive set point in anatomical space. In fact, many of them will die doing it. That's fine. They're all committed to this... They're all aligned and committed to this cause. They're all going to work together. If you deviate them, they will figure out a way to, somewhat, some other way to do it. And they sort of maintain, try to maintain this goal.
That enlargement of the cognitive light cone, the goal here is huge. The goal state here is much smaller. That enlargement has a failure mode, and that failure mode is cancer. This is human glioblastoma. This is some work from a number of people in the lab, but this right here, Juanita Matthews, is studying human cancer. What happens here is that when cells disconnect from this network, they can no longer remember this enormous goal that they should be pursuing. They go back to here. They basically roll back to their ancient unicellular form of life. The rest of the body to them is just environment now. That border between self and world, which has to be set by any agent, is flexible, and it shrinks, and it becomes just down to the size of a single cell.
So these cancer cells are not more selfish. There's a lot of game theory, for example, that models cancer cells as uncooperative and selfish. I don't think they're more selfish. I think they have smaller selves. They just collapse that boundary between self and world has now shrunk. So now that weird, kind of philosophical idea has specific implications for a biomedical program, which is that maybe then for cancer, instead of having toxic chemotherapy to try to kill these cells, what if we try to simply connect them back to the rest of their neighbors so that they can rejoin this collective?
If you do that, so here we inject nasty oncogenes, K-RAS mutations, p53 mutants. Normally they make a tumor. Here, we've labeled the oncoprotein in red so you can see fluorescently. It's all over the place. This is the same animal. There's no tumor. The reason is even though we haven't killed the cells, we haven't removed or fixed... The oncoprotein's still there. But because the cell is connected to this larger network, it works towards what the whole system works towards, which is making nice skin, nice muscle, and so on. So, at least in some cases, and we've seen this in birth defects, we've seen this in regeneration, and now here in cancer, in software, you can override certain kinds of hardware defects. I don't mean all of them, but in this case, there's a real genetic problem that you would discover if you were to sequence this. What you wouldn't know is that there isn't actually a tumor because the network bends the option space for the components, and they will not make a tumor and metastasize. They continue their morphogenetic cascade.
So what we've seen here is, and we could talk about some other properties of this cognitive glue, like memory anonymization and stress sharing and some other things. But for now, what we've seen is the ability of this collective to store large-scale set points. Now we can ask, just at the very end, where do these set points come from? Who sets them? How do they get set? We've looked here at new ways to test intelligence in unfamiliar spaces. We've looked at new ways to communicate goals to these systems, not to micromanage them.
Now for the last couple of minutes, I want to address two questions, and then at the end, if anybody's interested, we can also dig into some evolutionary implications. The first is, why are they so plastic? Why are these animals so incredibly plastic? With the exception of a few things like nematodes, like C. elegans, where every cell is numbered and they're pretty paint by numbers there, but most other creatures appear to be incredibly plastic. Why is that? And the question of where do these goals come from.
I want to just point out something interesting about the information flow in living systems both on the cognitive scale and on the evolutionary scale. You can think about the fact that at any given point in time, you don't have access to the past. What you have access to are the engrams, the memory traces that past events have left in your brain. So you can think about your own memories as instances of communication, messages from your past self. The same way that you communicate laterally with other creatures at the current time, you also receive communications from your past self, and you leave messages for your future self.
It's an interesting way of thinking about it because what's happening here is that all of these past experiences are being compressed because you're learning and you're abstracting patterns. You don't remember every single microstate that you've experienced. You compress them into a compressed representation. That representation is the engram that you get at any given moment. Now your job is to reinflate that representation into some meaningful data structure that is actionable right now, for whatever you need to do next. Like with any other message from another agent, you are under no obligation to interpret that message in exactly the same interpretation with which it was sent, right? At any given moment, it's up to you how you're going to interpret. Not only is it up to you, it is actually required for you to be creative on this end, because while this process, the encoding process, is somewhat algorithmic and deductive, you're throwing away a bunch of detail and a bunch of correlations. In order to decompress it, you now have to be creative. You now have to ask, you don't know what the original meaning was, but you have this information. What are you going to do with it?
Not only does this happen in cognition, this also happens in evolution, because there is fundamentally, as I showed you with that newt, this kind of autoencoder kind of architecture here where there's this bottleneck that requires this side to be more creative is true for the body as well as the mind. Everything gets squeezed down into the egg. The materials, including the genes, but also a bunch of other stuff in the egg, and then that has to get reinflated. As I showed you with the example of the newt and some other things I'm going to show you momentarily, that is not a hardwired process. You do not have to do it exactly the same way as past generations have done it. That leads to an interesting intelligence ratchet. All of that is described here if you're interested in that.
Here are a couple of consequences of this idea that, much like with that newt, you really can't count on the past. You can't take it too seriously. What you have to do is interpret what you have in the best way that you possibly can given whatever's happening now. The fact that you're dealing with this unreliable medium, a fundamentally unreliable medium where living things in evolution, not only is the environment going to change, but you're going to change. All your parts are going to be different. You're going to be mutated. There's going to be all sorts of things happening. You cannot take the past literally.
Here are some examples of this. This is a caterpillar, and these caterpillars can be trained. They have this brain suitable for doing what they do. They're taught to eat leaves on a particular color background. This is the work of Doug Blakiston actually. Then they have to become a butterfly. You go from a creature that lives in a two-dimensional world to a creature that's... so this is kind of a soft-bodied robot that has to move in a very particular way because it's soft. This is a hard-bodied, needs a hard-body kind of controller that flies, and so on. It has a different brain. During the metamorphosis, the brain is massively refactored. Many of the cells, if not most of the cells, are killed off. The connections are broken. It reassembles a new brain. These guys still remember the original information.
But think of something interesting here. Yes, there's the obvious question of how do memories survive this incredible refactoring? But the deeper question is around the fact that the exact memories of the caterpillar are of absolutely no use to the butterfly. The butterfly does not move the way this thing does. They don't want leaves. Butterflies want nectar. So you have to... the memories cannot be just kept. That's not enough. You have to translate them. You have to convert them. This is your bottleneck. You have to convert them into something that makes more sense at this point, that's adaptive. So the idea is to emphasize salience, not fidelity of the original information, because you know everything is going to change anyway. You can't rely on keeping the original information exactly intact.
In planaria, you can watch the information move across tissue. If you train them to eat liver in a particular bumpy environment, you then cut their heads off, which is where their brain is. Their centralized brain is here. The tail sits there doing nothing. It regenerates a new brain, and at this point, you find out that actually that information has now been imprinted on the brand new brain, and they now remember where to find it. So the information can move across the body, but it can also move across radically different bodies, as I just showed you, which requires a lot of reinterpretation.
We see that in these kind of cases where if we make a tadpole that has no eyes and we stick an eye on its tail, these animals can see perfectly well. We know because we have a machine that trains them in visual assays. These eyes make an optic nerve. The optic nerve does not go to the brain. It goes maybe to the spinal cord and ends there. Maybe it goes to the gut. But that's it. And it's fine. Why does this work out of the box? Why does this animal with a radically different sensory-motor architecture not require new rounds of selection and modification in order to be able to see in this crazy configuration? I think that's because the standard tadpole, even though we couldn't see it under standard circumstances, never assumed, was never hardwired for specific facts anyway. It's basically a sense-making system that emerges from scratch.
Here's the final thing I want to show you. I'm almost done. We started to ask, okay, how does this plasticity work out for creatures that have never been here before? So we tried to make some novel multicellular beings. Here's a frog embryo. Here are some epithelial cells, cells that normally would make the outer skin surface of the animal. We liberate some of those from the rest of the animal. They could have done many things. They could have died. They could have spread out, gone away from each other. They could have made a flat monolayer. But instead, what they do is they come together and make this cool little thing. The flashes that you see are calcium fluorescence. What they make is something we call xenobots. Xenopus laevis is the name of the frog, and we think this is a bio-robotics platform, so xenobots.
What they're doing is they're autonomously moving because they have little hairs that used to wash the mucus down the side of the animal. But now they're using it to row against the water. So they're swimming along. They can go in circles. They can patrol back and forth like this. They have group behaviors. Here's one navigating or traversing a maze. Here it goes. It's going to take a corner without bumping into the opposite wall, so it takes that corner, and then here for some reason it spontaneously turns around and goes back where it comes from. So a wide range of behaviors. If you look at the calcium signaling, you see some very interesting things. We're analyzing these now. I'm not going to make any conclusions yet, but we're analyzing these now using the same tools that neuroscientists use to analyze calcium signaling in the brain. Very interesting, but remember, there are no neurons here. This is just skin.
This remarkable capacity they have is to reproduce. Now, they can't reproduce in the normal froggy fashion because we made that impossible. But if you give them a bunch of loose skin cells, that's what this white stuff is, then both together and individually, they collect into little piles, they polish the little piles, and the little piles mature into the next generation of Xenobot. Guess what they do? They run around and do exactly the same thing, which makes the next generation and the next generation. So this is kinematic replication, and it works because the material they're working with is also not passive pebbles, but it is itself an agential material. So you get this thing that's kind of like von Neumann's dream of an agent that goes around and makes copies of itself from parts it finds nearby.
Where did that come from? There's never been any Xenobots. There's never been any selection to be a good Xenobot. The frog genome, we tend to think, has learned to do this throughout development and has this developmental sequence, and then eventually it makes this, and this is what it's learned to be a good frog in a froggy environment. But if you liberate these cells from the influence, from the hacking by the other cells that force them to be a boring two-dimensional, bacterial sort of suppressing layer around the animal, they actually have their own very different life. They can move. They do this crazy developmental thing. This is a roughly 80-day-old Xenobot. It's turning into something. I have no idea what it's turning into. And then it has different behaviors. We are now studying their learning and so on. Stay tuned for that.
But I think we're starting to see here that what you can't do is the thing that you would normally do, which is to say if you're wondering about the properties and the behaviors of a particular animal or plant, just to say, "Well, eons of selection. That's where it comes from," right? These things are baked in by selection for specific things. There's never been any Xenobots.
This is... in case you think this is some weird embryonic frog thing, I can show you this. This might look like something you got out of the bottom of a pond somewhere. If you were to sequence the genome here, what you would see is 100% homo sapiens. These are adult human patient cells. They are tracheal epithelial cells. In this environment, they assemble themselves into a motile bot. It has all sorts of fascinating... It has a different transcriptome that has thousands of new genes that are differently expressed than in the native tissue from the patient. It has new behaviors. I don't have time. I've run out of time, but it has all kinds of other things it can do.
What I think is going on here is that, much like simple physical devices, triangles and the Archimedean machines and so on are embodying aspects of what mathematicians call the contents of a platonic space. Basically, what we are doing when we make anthrobots and xenobots and those kinds of things is we are making vehicles to explore exactly that latent space. I think where they come from is basically exactly where all these other patterns in nature come from that are not specifically baked in or evolved or anything else. I think that by making these kinds of synthetic constructions, we actually start getting... we can start to map it out, what are some of the affordances that exist in that latent space?
I think it's important because everything that Darwin meant by endless forms most beautiful is like a tiny corner of this option space. Cyborgs and hybrids and chimeras, any combination of evolved material, natural material, and software is some kind of agent. These things, many of them already exist. There are going to be way more of them, and we need to understand how we are going to enter some kind of ethical symbiosis with other embodied minds that are very different from what we're used to.
I just want to mention one thing, that I think a lot of humility is warranted because, and I won't go into detail unless somebody wants to ask, but we found interesting not just complexity, but primitive problem-solving and emergent goal-directedness in very simple things, in this case sorting algorithms. The kinds of things people have been studying for many decades, bubble sort and so on, that we found. You can see that here. I think we've really underestimated what matter can do, and I think for the same reason we underestimate what very simple algorithms can do and what machines can do. Just because we've made it and just because we think we understand some of the parts, I think we really do not understand well other capabilities, not just emergent complexity and unpredictability, but actually emergent cognition. I think it begins very low on the spectrum of the kinds of things that we associate with minds.
That's it. I'm going to stop here and just say that I think intelligence is widespread. I think we have to learn to rise above our innate baseline limitations of how to recognize it.
We can now have principled frameworks that avoid either assuming there is no intelligence or conversely assuming high level minds under every rock. I think we can do better than this now. And we have all kinds of interesting opportunities in the future about using AI and other tools.
So here's... If you want to dig into any of this, there are some papers here, and I want to thank all the people who did the work. The biobot work done by Doug Blakiston and Gizem Gumuskaya; the ectopic eye work by Sherry Au and Vaibhav Pai; the algorithms' unexpected competencies with Te-Ning Zhang; Falon Durant for the planarian work; Niroshi Murugan for the physarum and the leg regeneration work; Juanita Matthews who's leading our cancer efforts; and Sarama Biswas who did the memory in networks in gene regulatory networks. And we have lots of collaborators that contributed to all of this.
And our funders. Here are the disclosures. These are three companies that have supported some of this work. And again, all the heavy lifting is done by the Model Systems.
So I'll stop here and thank you for listening.
