Today what we will talk about are a number of questions across computer science, biology, and some philosophy. And if anybody is interested in chasing down more details afterwards, everything can be found here — all the papers, the code, everything else is here. And of course you can also contact me here.
So the main point that I would like to get across today is this: I'm going to talk about my framework, which I call TAME — it stands for Technological Approach to Mind Everywhere. And the most important part of this thing is that I will argue that the philosophy actually matters; it drives new discoveries. And I'm going to show you some things that we've done that are directly emerging from a specific way we think about these things that otherwise was not found.
I'm going to talk about goal-directedness as a kind of invariant for recognizing, building, and communicating — or in some cases controlling — agents in unconventional embodiment. And specifically we'll talk about this cognitive boundary model and some other things like polycomputing. I'm going to use heavily the example of anatomical control as the kind of collective intelligence that navigates a problem space. That problem space is anatomical morphospace.
And I'm going to talk about how electrical networks are the kind of ancient evolutionary medium, an ancestor of what happens in the brain, that functions as a kind of cognitive glue to bind individual cells into intelligent collective behavior. And this has many implications for biomedicine and so on. And at the end we'll talk about synthetic bioengineering and the huge option space of new bodies and new minds. And so we'll talk about that.
Okay. So the first thing that I want to talk about are just some basic fundamentals. So this is a picture that is very old. This is a classic painting. Adam is naming the animals in the Garden of Eden. And although this is kind of a pre-scientific worldview, I think in many ways it still shapes a lot of discussion both in science and philosophy because it really emphasizes this idea of discrete natural kinds.
So here is the human, and people often talk about, let's say in the context of artificial intelligence or whatever, they talk about the human mind. And so what they mean is a modern, mature human like this. And then there are some other things, and everybody knows what the difference is between these different things. Okay. But we now know, both on the evolutionary time scale and the developmental time scale, that actually there are no discrete natural kinds here, in the sense that all of this is a very smooth, gradual continuum.
So we stand at one particular point of a very long set of very smooth, gradual changes that eventually make significant alterations to our capabilities. But it's really — you have to understand the whole continuum to understand what this thing is capable of. And this sort of magical agential glow, where people say, well, humans do this and then machines or animals do something else — it becomes very clear that you can't just draw a nice, crisp boundary around a modern human. You have to understand where do these things come from.
And in fact, it's even worse than that, because not only are we part of this natural continuum, but there's actually now — we see that we are part of an engineering continuum, both in terms of biological changes and in terms of technological hybridization, where at every level of the organism we can mix in different components. Some of them are evolved, some of them are engineered. And so again, there's this gradual, slow set of changes to where one might ask: where do these various properties begin and end, and how do they change?
So my framework really is focused on this goal of being able to simultaneously, together, consider very unconventional agents. So I want to think about all possible beings. So this is familiar creatures like us, birds, apes, octopus, things like that. But also very strange creatures like colonial organisms and swarms and engineered synthetic biology that might be made, artificial intelligence, whether in software or hardware, and possible exobiological agents. We should be able to think about all of these things using the same tools.
And that forces us to ask: what do they all have in common? If it isn't origin story and it isn't the material from which they are made, then the question is: what do they all have in common? And of course, I'm not the first person to try for something like this. Here's a Wigner and colleagues scale that goes all the way from passive matter all the way up to kind of human-level second-order metacognition and so on. And so my framework needs to be able to say something about this.
What do they all have in common? And to move experimental work forward — this is critical. I'm not interested in just conceptual philosophy. I'm interested in things that interface with the real world and drive new discoveries. And so kind of a central component of this is the idea that the kinds of systems that we're interested in exist on a spectrum, and it's a spectrum of persuadability.
What that means is that there is an observer, and it might be the system itself, but it doesn't have to be an external observer. So systems are also observers of themselves and their own parts. But there's an external observer that must ask the question: what is the most efficient way to functionally interface with that system? So for example, for certain systems like mechanical clocks, you're not going to convince them of anything. You're not going to reward or punish them. The only thing you can do is modify them via hardware or hack them in that simple sense.
There are other systems that are these kind of cybernetic homeostatic systems where you might be able to take advantage of the fact that they pursue simple kinds of goals, and you might rewrite or edit some of those goals and let the system take care of the rest. And then more complex agents, you can harness rewards and punishments and let the system make all of the necessary internal changes in itself to behave differently in the future according with the experiences that it's had. So you rely on its learning interface.
And then of course at some level you get to systems that are able to pursue reasons, not just causes, and there you can interact with this system in a very different way, and of course everything in between. These are just four waypoints. There are many more waypoints, but the point here is that we cannot assume where things stand. You cannot sit back in a sort of philosophical armchair and say, well, this is just a machine or it's just a piece of physics or whatever, and therefore I know it has to be down here. You actually don't know that.
You have to do experiments and you have to try these various stances and see what affords — it's an empirical question which affords the richest and the best interaction, you know, hardware modification or maybe even something like this. So the key to starting to think like this is to realize that all of us journeyed across this, what used to be called the Cartesian cut. This idea that we all start life as a single cell.
It's a quiescent oocyte, and people look at it and say, well, this is a little blob of chemistry and physics. It is not cognitive. It does not have — you name it, whatever you're interested in — it's just chemistry and physics. Slowly and gradually we become one of these things, or maybe even something like this, a deeply thinking human that's going to make claims about being more than a machine and so on.
And so the thing about developmental biology — I think it's the most magical of all sciences because there you literally see in front of your eyes, at a benchtop, if you look at for example a frog embryo, you literally see the journey from physics and chemistry to a scale-up of mind. And the key is that developmental biology offers no special moment at which some sort of lightning flash gives you true cognition, whereas before you just had chemistry and physics. There's no such thing.
So it's just a slow, slow transformational process. But we can think — at least we're a unified intelligence, right? We all feel ourselves as one unified being. And we don't think we're the same as, let's say, a colony of ants, which is like a traditional collective intelligence. We're not a collective intelligence, right? We're unified in some sense.
In fact, Descartes was really into the pineal gland because there was only one of them in the brain, and he felt that the unified character of human experience required a singular structure in the brain to mediate it. But if he had had access to good microscopy, what he would have seen is that actually there's not one of anything. Inside that pineal gland is this. This is what it looks like. There are many, many cells, and inside each of those cells there's all of this stuff. So just a tremendous — and this is by no means all.
So we are very much a collective. In fact, I think all intelligences are collective intelligences in the sense that they're made of parts, and we really need to understand how this scale-up happens. This is the kind of thing we're made of. This is a single cell. This happens to be a free-living organism known as a Lacrymaria. You can see what it's doing — it's hunting for food in its environment. It has huge competencies for single-cell-level agendas.
So it handles its physiology, its anatomical needs, its metabolic needs and so on, all at the scale of one cell. There is no brain, there's no nervous system. This is the agential material out of which we are all made. Individual cells have lots of competencies, and we need to start thinking about this because we're used to thinking about coarse-grained, higher-level beings.
For example, here's a rat. The rat presses a lever, learns to get the reward, and now the rat owns the associative memory between pressing the lever and getting the reward. But in fact, there are no individual cells in this creature that had both experiences. There were cells at the bottom of the foot that interacted with the lever. There are cells in the gut that reap the sugar of the reward, but no individual cell has both experiences. So who is it that actually owns the association? What can make the association between these two experiences?
And what is this rat that we say — this whole process requires a kind of cognitive glue. It requires a way to synthesize from the experiences of individual cells to something much larger that can have goals and preferences and competencies, memories, agendas and so on that no individual cell has. And so we need to understand that. And this seems very common — this happens all the time. So we're kind of used to this, but there are some very interesting cases that tell us that the story is actually much more complex.
So this is a caterpillar, and it lives in a two-dimensional world of crawling on flat surfaces. It has a certain kind of brain appropriate with it. It eats leaves. This caterpillar is a soft-bodied kind of construction. It needs to turn into something quite different, which is this hard-bodied machine that flies around and drinks nectar and has all sorts of different behaviors. It lives in a three-dimensional world.
In order to get from here to here, this creature basically deconstructs its brain. This brain is taken apart. Most of the cells are killed off. The connections are broken. It creates a new brain that's suitable for driving this kind of body. Something very, very critical here is that it's been found — and you can read about it here, this is all the work of Doug Blackiston — that caterpillars that are trained to associate feeding with a specific stimulus, that memory persists.
So the butterfly still remembers and can be tested behaviorally to show that they could still remember. So here are a few remarkable things. First, that memory survives basically the deconstruction of the brain. We don't have any computer architectures that work like that. So how does the memory even survive? That's the first thing, and why I don't think we understand very much about how memory is stored at all.
The second interesting thing is that the memory is generalized and transformed, because butterflies and caterpillars don't eat the same things. So butterflies drink nectar, caterpillars eat leaves. So you cannot simply learn that that stimulus associates with the presence of leaves, because that will not serve the butterfly at all. That memory has to be converted and generalized into the idea of food — not leaves, but the generic concept of food — so that the butterfly can now make use of that memory and acquire food that's completely different from anything the caterpillar ever saw.
And what's happening here is that these memories are being remapped onto a new architecture, not just in the sense of driving a different body, but in the sense of having actually novel, generalized meanings. And then finally, you can ask yourself this philosophical question. Thomas Nagel asked, what's it like to be a bat? Well, the second-order question is, what's it like to be a caterpillar slowly changing into a butterfly? Not on an evolutionary time scale, but on the time scale of a single organism. What is this transformation like? What's the experience like?
There are even more radical scenarios such as these with planaria, where these are flatworms. We'll hear much more about them later in the talk, but these guys regenerate their heads. And so what you can do is you can train them to recognize these little bumpy areas as where they get fed. So that's place conditioning, basically. And then you can cut off their heads. You can read about this. McConnell first did these kinds of things in the '60s. We repeated it with modern equipment in 2013.
You can cut off their heads. The tail will sit there doing nothing, and then eventually it regenerates a new brain, and the animal once again remembers the original training. So the memory is not strictly in the centralized brain that they have. It's somewhere, maybe throughout the body. But it also imprints — that tail is able to imprint that information onto the new brain that appears. So the movement and the transformation of information through the bodies is what we're actually interested in.
And of course there are many, many philosophical kinds of questions, like the malfunctioning transporter experiment, where you make a copy of yourself and which one is you? In planaria you can actually do that. You can cut them into pieces, and all of them have a good claim to being the original animal. So this idea of moving and converting information through tissues is something we're very interested in.
That plasticity is not only for these kind of invertebrates. Here's an example. Here's a tadpole. You'll notice that there are no eyes where the eye should be. But what we've done is we've induced an eye to form on the tail. So these cells form an eye on the tail. The thing makes an optic nerve. It synapses on the spinal cord here. It does not go all the way up to the brain. And these animals can see perfectly well out of that eye. We know because we've built this machine that tests them in behavioral learning assays, and they can see quite well.
And so it doesn't take novel evolutionary adaptation when you make a tadpole where the sensory-motor architecture is completely wrong. So instead of the eyes connected to the brain, you've got this eye on its tail connected to the spinal cord. Everything still works. You don't need multiple generations of adaptation to this. So what we're also very interested in is this kind of plasticity. Why does evolution make these kinds of flexible problem-solving machines that can operate in novel configurations and do these kinds of things that are not standard?
So the reason that all of this works is that we, as biological beings, are built on a multi-scale competency architecture. Not only structurally — you know, swarms are made of animals, which are made of organs and so on — but functionally, each level of biology is its own kind of agent. It has its own competencies to solve problems in specific spaces.
And it is the flexibility and the goal-directedness of these different mechanisms at different levels in different spaces that allow for this kind of remarkable plasticity that I just showed you some examples of, and we'll show you more momentarily. Now, with respect to intelligence and competency as problem-solving in these different spaces, we have to understand that we are pretty good — not great, but pretty good — at noticing intelligence in three-dimensional space.
So medium-sized objects moving at medium speeds in three-dimensional space, like crows and dogs and things like that, primates of course — we're pretty good at noticing intelligent behavior. But there are other spaces, such as spaces of gene expression, spaces of physiological states, and the space of possible anatomical states, that we really have a hard time recognizing.
And you can imagine that if you had a primary perception of your own blood chemistry, and you could innately feel all the things that were going on in your body physiology, I think you would have no trouble realizing that you live in this very high-dimensional space, and that your liver, your kidneys are navigational agents that are moving you through these spaces to solve the various problems and stressors that come up during the day.
But it's hard for us to think that way because all our sense organs are pointing outwards, and we don't really feel what all these amazing tissues are doing. So we have to understand that when we make a claim — any kind of an intelligence claim for a system, let's say a conventional or animal or an unconventional system — we're basically taking an IQ test ourselves.
It's we as observers who are saying we've identified or failed to identify some problem space, some set of competencies, and therefore we can make an estimate of what the intelligence of that system is to solve these problems. But that's us taking an IQ test, because we are probably missing quite a bit. I'll give you a couple of very simple examples of these kind of unconventional problem-solving capacities.
So here's an example with planaria. We take these flatworms and we expose them to barium chloride. Barium blocks all potassium channels. It's a non-specific potassium channel blocker, as a result of which their heads explode. The cells are very unhappy at not being able to transfer potassium. Their heads explode. If you leave them in the barium, a couple weeks later they grow a new head, and the new head is barium-adapted. No problem. The new head has no problem with barium.
So we asked — this is really crazy — we asked, how is this possible? What's different about this new head that it can function in the barium where this one cannot? And so what we did was we looked at all the genes that were expressed here and all the genes that were expressed here. And of course it doesn't have to be in the genes; it's just an easy technique that's available. And we simply asked, what's the difference? And the difference is only about a dozen genes are different.
So out of the space of tens of thousands of possible genes, these cells identified exactly what they need to solve this physiological stressor. Now the key point here is that planaria never see barium in the wild. There has never been an evolutionary selection pressure to know what to do when you encounter barium. This is an example of solving a novel problem. You're hit with a physiological stressor. The cells are dying. What do you do?
I sort of visualize this as being in this nuclear reactor control room. The thing's melting down. There's a million buttons. What do you do? How do you know which thing to press? Because you don't have time to just randomly turn genes on and off. You don't have time for selection the way you might in a bacterial colony. You don't have time for gradient descent. How does the system know?
So this is a kind of problem-solving of navigating transcriptional space, meaning the space of possible gene expression levels, to solve problems in physiological space. There are many other competencies. For example, if you study gene regulatory networks, you can find that in fact they have six different types of learning, where turning specific nodes on and off can give them memories that change the way they act in the future, including, for example, associative conditioning.
So you can do Pavlovian conditioning on gene regulatory networks. So not only do our tissues and organs and cells have these competencies, but in fact the pathways, the molecular networks inside of them, also have cognitive kind of capacities such as associative learning, habituation, sensitization and so on. And so one thing that we think about a lot is this idea that evolution might actually be pivoting over time some of the same strategies, some of the same navigational policies for traversing these kind of spaces into novel spaces.
So you keep the policy, you just change the sensors and effectors, and you can reuse the same types of intelligence that you might have for solving problems, and you can reuse them in novel spaces and of course scale them up. And so this goes all the way up through, for example, navigating linguistic space, where keeping a constant narrative through telling a story is actually a navigational task that's quite hard. Artificial intelligences today are quite limited in how long they can maintain the train of a story.
So all of these things share some important properties. Okay. So that's the first part of the talk. The next thing I want to do is really talk about a particular unconventional agent, which is a cellular collective of morphogenesis. It lives in an anatomical space and some other spaces as well. And we want to talk about bioelectricity as the informational medium that allows this to happen.
So this is Alan Turing. Alan Turing was interested in intelligence, in cognition in unconventional embodiment, of mind in thinking computation machinery and so on. And so he was interested in problem-solving machines through reprogrammability, a kind of plasticity. But weirdly, he also wrote this paper on the chemical basis of morphogenesis, which talked about how chemical order can come to be in embryonic development.
Now you might think that's kind of strange. Why would somebody who was interested in cognition and intelligence outside of the standard substrates, why would he be thinking about the chemical basis of development? And I think that although I don't think he wrote anything about this, my gut feeling is that actually his genius saw the deep invariant between these two areas: the building of the body and the building of a mind are actually the same question. And we're going to talk about that now.
So this is a cross-section through a human torso. So look at all the incredible order. All the tissues, organ structures, everything is in the right place, the right shape, next to each other. Where does this pattern come from? Where is this encoded? Why do all normal humans have this invariant pattern? Well, you might be tempted to say, well, it's in the DNA. Of course, we have a human genome. We can read genomes now, and we know that none of this kind of geometric information is directly in the genome.
What the genome does specify is proteins, the micro-level hardware that every cell gets to have. And so this process, going from a set of embryonic blasts to building something exactly like this, is a very important problem. How do these cells know what to make and when to stop? They have the hardware, but what does the software of life look like? If a piece of this is missing, how do we convince these cells to rebuild and repair?
And as engineers — and we'll talk about this even more at the end of the talk — what else is possible given a standard genome? What else can you build with it? And it's important to understand that this is a very similar problem to understanding the structure of ant and termite colonies, the structure of spider webs. The genomes of these animals don't specify these structures any more than the human genome specifies this structure.
All of these are the products of various collective competencies of cells and tissues that are not directly specified in the genome. So in order to understand this problem, it's good to think about the endgame of our field. Where are we going? What do we want to do? And to me, the endgame of this — meaning that here's where we can say that we've actually understood the problem, solved it and so on — is this notion of an anatomical compiler.
Someday you should be able to sit down in front of a computer and draw the animal or plant that you want, not at the level of molecular components, but the actual anatomy that you want. Such as, for example, here I want this three-headed flatworm like that. And if this system existed and if we knew what we were doing, it would be able to compile your anatomical description into a set of stimuli that would have to be given to these cells to make them build whatever you want them to build.
Now this is of huge practical importance because if we were able to do this, if we were able to convince a group of cells to build whatever we wanted them to build, most problems of biomedicine would go away. So birth defects, traumatic injury would be regenerated, reprogramming, cancer, aging, degenerative disease — all these things would go away if we understood how to communicate anatomical goals to collections of cells.
And that's why the actual solution to this kind of thing is not a 3D printer. The point is not to try to micromanage where the individual cells go. It's a communications device. It's a translator. It's a way to convert your anatomical goals into the goals of the cellular collective that's going to try to build something. And I want to just say very quickly why it is that we don't have anything like that today.
I mean, you might be wondering — we've had decades of genetics and biochemistry and molecular biology. Why don't we have an anatomical compiler? Why aren't we building organs and biobots and so on to arbitrary specifications? Well, so here's an axolotl larvae, and these guys have little — their babies have little legs. Frog larvae or tadpoles do not have legs at these stages. In our lab, we make something known as a frogolotl, which has a bunch of cells from frog, a bunch of cells from axolotl, and you get a frogolotl.
Now I can ask a simple question. We have the genomes. We have the axolotl genome. We have the frog genome. Now, here's my question. Knowing those genomes, can you tell me if frogolotls are going to have legs or not? And the answer is no, you can't. And there's that question. There's the question of if they do have legs, will those legs just be made of axolotl cells, or will they in fact convince some frog cells to come along and build? What will the shape of those legs be?
In fact, could we even have predicted the primary shape from the pure genome? No, we can't even do that. So there are huge areas, areas of knowledge gap despite the understanding of the details. So where we are today in this field is this: we're very good at manipulating cells and molecules. In fact, all the exciting stuff in molecular medicine is about single-molecule approaches, CRISPR, genomic editing, that kind of thing. So we're very good at this.
What we actually want is this: we actually want to be able to control large-scale form and function. And we're a very long way away from that. And I argue that the reason we're far away from it is that biology is still where computer science was in the '40s and '50s. This is how we interact with life. What she's doing is she's reprogramming this computer by physically rewiring it. That's how you used to have to program.
And of course since then, the reason we have this information technology revolution is because we've realized that you don't always need to interact with things at the lowest level. Some things are reprogrammable. And that means you can interact with them via inputs, stimuli, taking advantage of the native computational properties of the machine. You don't always have to rewire the hardware. Biomedicine is still here.
So I think the next frontier of regenerative medicine and biology is to understand the native intelligence of the material that we're working with, because it's not a simple machine. Now when I say intelligence, what do I mean? I like William James's definition of intelligence, which is the ability to reach the same goal by different means.
What I like about it, it's very cybernetic. It doesn't talk about brains or being evolved or designed or anything like that. It talks about what is universal and common to all examples of intelligence, which is the ability to — a degree of competency to solve a particular problem in novel circumstances. So he actually had this beautiful example of the difference between two magnets trying to get together and Romeo and Juliet trying to get together.
And the reason that all of these are on the same — I mean they sound like very different things. Of course we're not used to thinking about them in the same way, but there's a continuum here, which is this: the ability to deal with barriers in your navigational space. So these two magnets — neither of these magnets is competent to go around the barrier and get together, because in order to do that you would have to temporarily get further from your goal.
And these things are so simple. All they're doing is trying to minimize the free energy. And so they're never going to temporarily step away in order to get where they need to go. Of course here you have huge competencies and planning and all kinds of things. But in between you have all of these other things. You have autonomous vehicles and cells and worms and animals that have different degrees of competency to navigate around barriers in their space that stand between them and their goals.
And we can try to quantify that in different types of unconventional systems. So let's ask: given this kind of definition — the ability to get to the same goal by different means — what kind of competencies do cellular swarms have? So let's look. This is normal development. Normal development is very robust. So you start with an egg and you end up with a person almost always. It's quite robust.
However, what we find out is that it's actually not hardwired, because if you cut this embryo into pieces — and you can cut them into a number of pieces — you don't get half bodies, you get perfectly normal monozygotic twins. So this system, if you deviate it from its normal starting position, it doesn't just go to the incorrect location. It's actually from different starting positions, it's able to navigate to that same ensemble of goal states that corresponds to a normal human organism.
It has the ability to avoid local minima and so on. And so this is another example of this. These animals, these are axolotls. They regenerate their limbs. So if you cut the leg, you can cut it here, here, here, here, anywhere. In fact, they regenerate most of their body organs. They will grow exactly what's necessary, no more, no less, until it's fixed and then they stop.
So this idea of being able to get to that same location from different starting positions. So we can start here and we need to regenerate all of this, or we can start here and we only need to regenerate that. That's the ability to get to the same goal from different paths. By the way, these worms and salamanders are not the only ones. Mammals have some regenerability. So humans can regenerate their liver. Even the ancient Greeks knew that. I have no idea how they knew that.
Deer regenerate antlers. So every year, huge amounts of bone, vascular innervation in the right pattern. The bone grows at a rate of a centimeter per day. So really incredible rates of bone growth. And then even human children can regenerate their fingertips. So below a certain age, if you lose a fingertip and you just sort of keep it clean and not try to sew the skin, you will actually regenerate your fingertips.
So here's one of my maybe all-time favorite examples of morphogenetic problem solving. What you're looking at here is a cross-section through a kidney tubule, and it has a space in the middle. And normally what happens is that eight or nine of these little cells get together and they work together and they build this tubule. But what you can do is you can make newts with extra genetic material.
And the first amazing thing is that you still get a normal newt even though you have multiple copies of the chromosomes. That's already amazing. But what happens is the cells get bigger to accommodate the extra copies of the genomes. And as the cells get bigger, it turns out that fewer and fewer of them will work together to make exactly the same size lumen. So that's amazing thing number two: they scale their collective behavior to the size of the cells.
But here's the most amazing part. If you make the cells truly gigantic — and I believe this is either six or eight times the normal DNA that they should have — these gigantic cells will now just wrap around themselves, not work with anybody else, but wrap around themselves and produce and leave a hole in the middle that's the same lumen. So now, here's why this is incredible. It's because this is a completely different mechanism.
than this. This is cell-to-cell communication and standard tubularogenesis. This is cytoskeletal bending. And what you're seeing here is a kind of top-down causation. It's the ability of the system to call up different molecular mechanisms to achieve the same large-scale anatomical goal.
So this is, remember James' definition now, the ability to get to the same outcome, the same region of anatomical morphospace by different means, by using different tools in your toolkit to get this accomplished. Now, just think about what this means for biology.
If you're a salamander coming into this world, it means that you can't rely on, you don't know how many copies of your genetic material you're going to have. You can't rely on the size or the number of your cells. You don't know any of that. You have to figure out all of these things from scratch. And you have to be able to make a proper salamander of the exact same size, even if you can't rely on your own parts, not only weird perturbations from outside, like somebody cuts an embryo in half or some sort of damage. This is much more fundamental. Even your parts are unreliable and you still have to be able to get the job done.
So there are huge evolutionary implications of this, which I don't have really time to go into now, except to say that this kind of plasticity, the fact that evolution gives rise to problem-solving machines, not specific solutions to specific problems, has major implications for evolution itself.
So there's this constant feedback loop between the intelligence that's due to evolved mechanisms and the feedback on the evolutionary process of the fact that your parts can make up for problems, the fact that even damage, various kinds of damage, you still get a workable animal. And this results in a really interesting intelligence ratchet that we can talk about, where the evolutionary pressure ends up ramping up this competency mechanism at the expense of structural genomes.
So in order to get into this notion of bioelectricity, I just want to look at one simple example of this kind of navigational competency, which is the frog face. So these tadpoles, this is a tadpole of Xenopus laevis. They have to become this adult. In order to do that, they have to rearrange their face. Their eyes have to move, their jaws have to move, the nostrils, the whole face has to be rearranged.
And it used to be thought that this was a hardwired process, that basically every organ just moved the right distance in the right direction and then you get your normal frog. We decided to test that process and ask how much intelligence there is. So you see, these are all experimental kinds of questions. You cannot just assume these things.
So what we did was, much like James' definition, we changed the starting position. So we made a scrambled, we call these Picasso tadpoles, scramble the face. So the eyes off the side of the head, the jaws are on top of the head, everything is mixed up.
What we found is that these animals make largely quite normal frogs. All of this stuff moves in abnormal paths and abnormal configurations to get to the same final goal. In fact, sometimes they go too far and they actually have to double back. They have to come back a little bit.
So you see, what evolution gives you here is not a set of movements for tissue. It gives you an error minimization scheme. It gives you the ability to minimize error from different starting positions to get to a particular state. So that raises a very obvious question: how does it know what the correct state is?
And so what we've been doing is to this standard picture of developmental biology, where you have some gene regulatory networks and then some physics happens with these proteins and then there's emergence, this magical process whereby you get this complex shape. Yeah, it's not all feed-forward. There's this amazing ability both at the level of physics and genetics to return you to a correct state if you're deviated by injury, mutation, teratogens, whatever.
And so this is a model of a kind of anatomical homeostasis. And so in these kinds of models, what's very key is that, like with a thermostat in your house, any kind of homeostatic process has to have a set point. There has to be an explicit representation of what the goal is so that you can calculate error. How far off are we at any given moment?
So this has been our research program for years now, is to test this hypothesis by looking for that set point, learning to read it and rewrite it. And the idea is that you don't have to make all your interventions down here at the genetic level. You can actually rewrite the set point and let the system do what it does best, which is to solve to that set point.
So what could possibly be a mechanism that stores a set point as complex as an anatomical descriptor, as a layout? And we looked at the brain as a kind of inspiration. And the way goals and navigation works in the brain is using this system.
So you've got ion channels in neurons. They let individual ions go in and out. That creates a biological state which may or may not be propagated through these electrical synapses known as gap junctions to their neighbors. And the ability to control how signaling flows through this network is a very basic computational medium that gives rise to this kind of software.
Here you can see the physiology of a zebrafish brain as the zebrafish thinks about whatever it is that zebrafish think about. And it's the commitment of neuroscience that all of the memories, the preferences, the goals, the competencies of these creatures are encoded in that electrophysiology of the brain.
And there's this notion of neural decoding. If we understood how to read it, we could read memories out of the brain. It turns out that this kind of system is evolutionarily ancient. It goes back to bacterial biofilms. Every cell in your body has ion channels. Most cells have these electrical connections.
And we can do the same kind of decoding, this kind of non-neural decoding, by looking at electrical activity. This happens to be an early frog embryo. We're visualizing the electrical patterns and again try to decode where is it trying to get to in anatomical morphospace.
What you're seeing is a clear isomorphism between neuroscience and developmental biology, where typically the electrical activity of this network tells your muscles what to do. It controls your muscles to move you through three-dimensional space. The rest of the body would long before that, the rest of the body was using the exact same thing, electrical networks using the same components to move the configuration of your body through morphospace. Same trick, different space.
And so we developed the first molecular tools to read and write these things. So here's our use of voltage-sensitive dyes to control the bioelectric patterns within embryos, to read out what all the electric communications are. Of course, we do a lot of computational modeling and simulation to understand how they connect to the channels and pumps that produce them.
Here is just an example of two patterns. This is a time-lapse movie of voltage patterns as a frog embryo is originally putting its face together here. And you can see one frame out of this movie. I'm showing this to you because this is one of the easiest patterns to recognize. This pattern memory literally looks like a face.
You can see where the animal's eye is going to be, where the placodes, the mouth, you can see where everything's going to be. And I'm going to show you in a minute what happens when you perturb this representation of what a correct frog face is. So this is a normal, this is an endogenous normal pattern. This is what controls the downstream gene expression and anatomy of the face.
If you mess with this, then all of those downstream steps go awry. This, in contrast, is a pathological pattern. This is a human oncogene, and you can see that these cells, first thing that happens when you express an oncogene, the cells electrically disconnect from their neighbors and they're basically, they're starting, they're going to treat the rest of the tissue as external environment. They're no longer connected. They're just unicellular life forms at this point.
So beyond reading these, the kind of basically reading the mind of this morphological collective intelligence of cells, we have the ability now to write information into that as well, as people like Tonegawa do in the brain when they incept false memories into mice.
So what we can do is we can take all of the tools from neuroscience. And so this is very important. The tools of neuroscience really don't distinguish between neural and non-neural cells. They all work, all the same concepts, the laboratory techniques, the neurotransmitter machinery, like everything works.
So what we can do is we can control the gap junctions which cells talk to each other, and we can control the actual electrical states by opening and closing these ion channels. So remember, there are no magnets here. There are no fields, no waves, no radiations, no electrodes. We are using the native interface that cells use to program each other.
There's this electrical interface on their surface that controls the information processing, and this is how cells hack each other's behavior during normal development. And also, by the way, how some parasites hack their hosts. But we are just exploiting that same interface. The reason all of this works is not because we're so clever. It's because the cells have a beautiful, high-efficiency interface that they offer to us.
So when you do this, what can you do this way? What kind of false memories can you imprint onto a collective agent? And let's just take stock of what we've said so far. Groups of cells are a collective intelligence which navigates anatomical space to solve various problems and reach the correct region.
So what happens when we incept false memories into this collective agent that is storing and processing those memories in an electrical network? Well, one thing you can easily do is inject a particular ion channel RNA to trigger a voltage state that corresponds to that eye spot that I just showed you in that electric face model.
And when you do that, let's say we do that to a bunch of cells that are going to become gut, those cells become convinced to make an eye. And so here's a tadpole. Here's the gut. Here's the eye sitting on their gut. You can make eyes anywhere, in fact. And these eyes, if you section them, they have all the right lens, right? And the optic nerve, all that stuff.
So there's a couple things you learn from this. One is that the bioelectric pattern is instructive. It's not just about making errors or causing monsters. It's about telling the cells you should build an eye, and they actually build a complete complex structure. Number two, this is hugely modular. We have no idea how to micromanage the formation of an eye. It's very complex.
What we do know is how to put in a memory that's basically a subroutine call or a trigger that says build an eye here. All the competency is in the rest of the cells. We just found the language by which they encode which structure should be built, right? And downstream of that is all of the correct gene expression, all the size control, all of that stuff is downstream.
There's also another really cool aspect to this, which is this here is a cross-section of a lens sitting in the tail of a tadpole somewhere. And the blue cells are the ones that we actually injected with our ion channel. So the blue cells have had their voltage modified. All of these other cells are completely natural cells, but they all participate in this lens formation. Why?
Because these cells can ascertain that there's not enough of them to build a whole lens. And so they recruit their neighbors. And so much like other collective intelligences like ants and termites, which recruit their buddies when something needs to be done and there needs to be more of them, this collective intelligence scales to the task at hand. Again, natively. We didn't have to teach them how to implement this or try to micromanage it in any way.
And in fact, there's actually a really interesting battle going on here because what you can do is you can make any number of these blue spots. This is also frog, and these blue spots are other areas where we've injected our potassium channel. And what they're doing is they're expressing eye precursor markers, but in the end they might only make one eye, not the three that you might expect from looking at this.
It's because the cells are fighting it out. The injected cells are saying to their neighbors, hey, let's all make an eye. There should be an eye here, and you guys need to help. The surrounding skin cells are saying, no, you should be skin. There doesn't need to be an eye here. You should be skin.
And depending on how that conversation goes, you might make one, two, well, zero, one, two, or more eyes. So this is very much a collective intelligence kind of problem where they're actually communicating with each other to decide what's going to happen.
And so you can make other interesting things. You can make fins, and you can make ectopic hearts. I mean, fins are kind of wild because tadpoles aren't supposed to have fins. That's more of a fish thing, but we'll get to that in a minute. You can make ectopic hearts, you can make ectopic legs, ectopic forebrain. So you can very readily convince these cells to build new kinds of structures.
Why? Because without micromanaging all the gene expression, we're not talking to stem cells. It's very modular. It's because that is how the collective is processing information about what to build and where to build it.
And that means that we can do regenerative medicine applications like this, where we take this, the frogs, unlike salamanders, normally don't regenerate their legs. And so you can amputate here, and normally nothing happens by 45 days later. We make a bioelectric cocktail that we apply here that tells the cells rebuild the leg.
And immediately, within a couple days, they have a pro-regenerative blastema, and then by 45 days, they've got some toes, they've got a toenail, and the leg is eventually touch-sensitive and motile. So the whole thing becomes a pretty respectable leg after that.
So the idea is that a very low-content input, very simple trigger, gives a very high information content result, not because we know how to make a leg. We have no idea how to make a frog leg, but because the system is already competent to do this and needs to be convinced to do so.
This is where I have to do a disclosure because Dave Kaplan and I are co-founders of a company called Morphaceuticals, and we are working to take some of that technology now into mammals, eventually hopefully into humans.
Okay, the second model system that I want to show you is to come back to these planaria. These planaria are a fascinating system. They are so highly regenerative. In fact, they're immortal. So they regenerate everything, and they just sort of live forever.
And so one might ask this very important question: if you cut them in half, the bottom end grows a tail, the top end grows a head. But these cells that you just separated were next-door neighbors before you cut them. How come they have completely different anatomical fates?
That question cannot be answered locally. You can't tell from where you are what you should build. You have to communicate with the rest of the body and figure out what's missing. So this is non-local decision-making.
So we asked this question: when you amputate this planarian, how do you know how many heads you're supposed to have? And it turns out that there's an electrical gradient that we characterized that tells you how many heads you're supposed to have.
And what you can do is you cut off the head and the tail, and this middle fragment will regrow a nice perfect worm 100% of the time. But because we understood this gradient, we were actually able to make these two-headed worms by taking this electrical. So here it says one head, but now we can say, no, actually, you should have two heads.
And sure enough, this isn't Photoshop. These are real animals. They make nice two-headed forms. Now, here's the really important part. This bioelectric map is not a map of this two-headed creature. This bioelectric map is a map of this one-headed creature.
So what we did was we rewrote the stable biological pattern memory in this one-headed animal, and it stays one-headed with a normal gene expression. So here, anterior genes only in the head, only in the head, not in the tail, and it sits there perfectly happy like this until it's injured.
So a single planarian body is able to store at least two different representations, two different ideas of what a correct planarian looks like, and sit there, and it's a latent memory. It doesn't do anything until you injure the animal. The same body, the same one-headed body, can store two different representations.
So at the very beginning, I told you that this collective intelligence works towards, it's a homeostatic loop that works towards a specific set point, and the question was how do you store a set point for anatomy. This is it. You are literally looking at an example of how pattern memories are stored. They're stored as a stable biological state in tissue.
And this in particular is kind of cool. It's a counterfactual memory because this is not a representation of what's happening now. It's a representation of what you're going to do in the future if you get injured.
So you can think about this as a very basic precursor of this amazing ability to mentally time travel that we have, where you have memories of things that are no longer happening, you have predictions of things that haven't happened yet. You decouple the current state from back in time. And this is a very simple way of doing that that I think brains kind of appropriated this trick.
Now, why do I keep calling it a memory? Because if you take these two-headed animals and you cut off the primary head, you cut off the ectopic secondary head, you might expect that, well, the genome is normal. We haven't edited the genome, there's nothing wrong with the genetics. And once you get rid of this crazy secondary head, this middle fragment should just go back to normal.
And in fact, it doesn't do that. Once you've convinced these cells that a planarian should have two heads, it will do that in perpetuity. No matter how many times you cut, we actually have a way of setting it back to one-headed by again by targeting this bioelectric circuit. We can sort of flip them back and forth.
Here you can see what these two-headed guys are doing. This is memory. It's long-term stable. It's rewritable. It has conditional recall, as I just showed you, or latency. And it has discrete behaviors.
So what we're doing is we're reading and writing patterns into the collective intelligence of these cells that must navigate anatomical space. And so now what that means is that we can really do the kind of computational work of fleshing out the connections between the electrical state space of the circuit that we understand with all kinds of ideas from connectionist computational neuroscience to understand how electrical networks can store patterns, do pattern completion, and navigate spaces intelligently.
It's actually not just about head number, it's also about head shape. If you take this planarian with a nice triangular head, cut off that head and perturb the electrical signaling, you can get flatheads like this, you can get round heads like an S. Mediterranean. And no genetic change.
These animals are 100 to 150 million years distant from these. They get all the right brain shape changes, the stem cell distribution changes, just like these other species. So in the morphospace, we talked about this anatomical morphospace of possible configurations of structures.
There are attractors belonging to other species of planaria which they naturally go to. But this hardware is perfectly willing to go to these attractors if correctly prompted. It's just not what they normally do, but you can prompt them.
So there's real this notion of navigating that space using the electrical connections between cells to store the policies of where are you going, and we have the ability to control that. Not only, of course, in the standard morphospace of standard planarian shapes, but you can make some really weird-looking things.
You can make these spiky forms, you can make these three-dimensional things that look like a little hat, a little ski cap, or these kind of hybrid forms. The latent morphospace, which Darcy Thompson already in 1917 already foresaw, the different kinds of possibilities for this same genetically wild-type hardware is immense.
And so now what we want to understand is we want to build a full stack model where that goes from the genetics of which ion channels are present, and so that's the genome, that's your hardware, to the software that's implemented by the tissue-wide electrical dynamics that we can simulate using our various simulators, all the way to large-scale representations that eventually get cashed out as basically algorithms or navigation policies that make it much easier to control the system than if you try to do it bottom-up this way.
So for kind of, and I didn't show you some of our biomedical kinds of stories, but the idea is that what we want to do is complement the traditional bottom-up approaches that are all focused on the hardware with top-down strategies that seek to take advantage of the software.
And this has to do with literally behaviorally training cells and tissues, including the pathways within them, and really understanding the decision-making of the material to have much more efficient control. And for this reason, I've argued that future medicine is going to look a lot more like a kind of somatic psychiatry than it's going to look like chemistry, because it's not about changing the physical landscape. It's about changing the tissue agents' perception of that landscape.
And you can see more details on all this stuff is in these papers. Okay, so for the last, and I think I'm running somewhat behind, so I'm going to kind of go quickly here, but the last section I just want to show you a few new things related to this framework.
So the interesting thing about trying to put all of these very diverse agents, for example, a collective tissue navigating morphospace, together with conventional agents on the same scale, is that you can come up with a central invariant, which I have picked as the spatiotemporal size of the largest goal you can pursue.
So all agents have in common one fundamental thing, which is that they pursue goals. And so now the question is, are you only interested in the local sugar concentration and you have a memory of a couple minutes back and a predictive capacity of a couple minutes forward, and then you might be a bacterium?
Or do you have a bigger cognitive light cone, and you might have a memory going backwards and some predictive capacity going forwards, but you're never going to care about what happens in three months, two towns over. So you might be a dog if that's the case, or you might be a human which has a huge cognitive light cone and you're actively working towards world peace and what the financial markets are going to do 100 years from now. You could have a huge cognitive light cone.
So this idea of, and we of course are a compound intelligence where our cells have little tiny light cones and our organs do, and so do we, and possibly larger structures in which we participate. So this idea of a cognitive light cone is based around the goals that any particular system can follow.
And so that allows you to start thinking about what an individual actually is. What are cells? Given that we're all collective intelligences, we're made of parts. We're basically a bunch of cells in a trench coat. How does the collectivity and the unification happen?
Well, let's look at the very early steps of embryogenesis. This is an embryonic blastula. There might be 50,000 cells here. And we look at that and we say, ah, there's one embryo. What do we mean by one embryo? What are we counting? What is there one of?
Well, what there's one of is alignment, not just physical alignment, although that's true too. Cells have to physically align with each other in planar polarity. But we are counting alignment of purpose. What we're saying is that under normal circumstances, all of these cells will work together to build exactly this thing. This whole blastula, all of them together, have this one goal.
And we know that because if we try to deviate, they will try to compensate. They are all together trying to get to this one goal. We're counting goals in alignment, and they have a particular size of this goal state that they're trying to fit.
It's actually quite interesting to ask how many cells are in here. It's the same question with brains. If you didn't already know what a human was and somebody showed you a human brain and said, in this volume of substrate, how many cells can fit there? We actually really have, you would have no idea what the density of cells per unit material is.
Because what you can do is, and I used to do this as a grad student in duck embryos, you can take a little needle and you can make some scratches in this blastula like this. And when you do that, for the next four or five hours before these things heal back up, each of these little islands is not going to be able to feel the rest, and it will decide that it is the embryo and everything outside is external world.
And they will make individual embryos, and you get conjoined twins, and you can get any number of them. And so the question of how many embryos are here is actually not clear. It's not set by the genetics. It's the outcome of a dynamic process of autopoiesis or self-construction.
Every cell is some other cell's neighbor. They have to figure out who is part of, is this particular cell part of me as the embryo? Is it part of the outside world? And this kind of generative medium, this excitable medium, can give rise to zero, one, two, up to probably a half a dozen or so different selves.
And the same thing is true in the brain, actually. And I think this goes back to that point about Turing, is we typically feel like unified individuals, but we know from studies of split-brain patients and dissociative identity disorders that there's actually not clearly just one self inside the medium of our nervous system.
And organs have to decide this too. Instead of making one giant eye, this one's decided to make three slightly smaller ones. Why? We have to understand how these different collectives decide who is in and who's out.
And this, of course, has medical implications because while evolution has given us some nice ways to scale up our goals, so from the goals of individual cells to being able to, from pursuing little tiny goals like metabolic states and proliferation and things like that, to working on these very large construction projects where no individual cell knows how many fingers you're supposed to have, but the collective does, because every time you amputate one, it'll grow exactly the right number.
But that has a failure mode, and that failure mode is cancer. So in cancer, this is human glioblastoma. These cells have electrically disconnected, and as far as they're concerned, the rest of the body is just environment to them.
And so that idea that shifting boundary of the self, the shifting scale of the goals you care about from local single-cell goals to large anatomical goals to once you're a neural human system, much bigger goals than that, is a way to start addressing some biomedical problems. For example, cancer.
What we've done here is we've induced a human oncogene, and so they're going to make this tumor. But what we've also done is co-inject an ion channel that forces these cells to stay in electrical communication with their neighbors. So we don't kill the cancer cells. We don't fix the oncogene.
In fact, you can see here the red is the oncogene protein. It's quite blazingly expressed. It's all over the place. But there's no tumor. This is the same animal. There's no tumor because what? Because the hardware is not what drives. It's not the genetics that drives. It's the physiology and the decision-making of the collective.
And when you maintain the collective, the collective remembers that it has to make nice skin and muscle. It doesn't go off to tumorigenesis. So I'm now going to skip the rest of this because I'm already an hour in. So I'm going to skip all the Xenobot stuff, but we can come back to it if people have questions about it.
And just to point out a couple of quick and simple things. The basic development is so reliable that we get lured into a false sense of robustness, which is that we know that oaks make oak trees, and we think, well, here's what the oak genome can do. This acorn is going to make things like this, a nice flat green structure.
But what we don't realize, and we don't know this until some parasite like a wasp prompts these cells with chemical signals to build something completely different. This is a gall formed from the plant cells under some simple prompting from this other creature. Here's another one. We would have no idea that these cells with their standard genome could build something completely different.
That morphospace, that latent space of possibilities, that behavioral space is unknown to us until we start to experiment by probing it with various signals. And this is very profound, this idea that living things hack each other constantly.
And it's not just parasites that hack the host, but all of the cells of an embryo are constantly hacking each other in the sense of putting out signals to get them to do what the collective wants them to do.
So we can now say that synthetic constructs like these Xenobots that we have made are a tool. Together with the Xenobot, it's actually not this little tiny thing that we make. Actually, this whole structure, including the environment, including high-agency aspects of the environment like the bioengineers and lower-agency aspects of the environment like chemical signals and so on, is a way to look into the various spaces that a collective might be traversing, and these might be behavioral, physiological, transcriptional, and many others that we may not know.
So I'm just going to summarize my points here, which is to say that in this framework we look at a continuum of agency. I don't like binary categories. I think they're completely artificial. I think these questions are empirical. They are not philosophical. We have to do the experiments and see which substrates have what capacities.
We can define selves as a boundary of goals that the collective system is capable of pursuing. But it's on us as observers to recognize those goals in an optimal way. And we're not very good at it. We need a better science of doing it.
Developmental bioelectricity as the ancient precursor of what nervous systems do is one kind of cognitive glue that binds competent subunits towards larger-scale selves. And evolution, of course, interacts with this whole process, but in a bidirectional way.
And so the last thing I want to point out is simply this, that when Darwin looked at the richness of variety in the biological world, he coined this phrase, endless forms most beautiful. All of the natural forms are here. They're a tiny little corner of the possible state space of every combination of evolved material, engineered material, and software.
All of these things, hybrids and cyborgs and chimeras of various kinds. Many of these already exist, and many more are going to exist because all of this is possible because life is so interoperable. And the reason it's so interoperable is that evolution makes problem-solving machines.
All of these systems have to, they don't take their past experiences too seriously. They don't overtrain on their priors. They have to solve problems from scratch every single time. And thus novelty is not surprising to them. They're dealing with a very unreliable medium.
Basically, the whole architecture assumes that things are going to be broken, things are not going to work the way you expect, and you have to solve problems from scratch. And for this reason, all of these things are viable.
And that means that going forward in the next couple of decades, we are going to be surrounded by creatures that are nowhere on the tree of life with us. We cannot use the old familiar strategies of asking where on the tree of life is it? Is it more like a snake or a dog or a dolphin? These things are going to be nowhere on that scale.
And the old categories of is it designed, is it engineered versus natural? Is it a machine versus an organism? All of these crisp categories are going to become completely useless. I mean, they were never correct, but at least in the olden days they were a rough heuristic that you could use.
And so we're going to need to develop a new form of ethics for relating to minds that are not like ours. And that's really critical. So I'm going to stop here. If anybody's interested in these kind of things, here are some papers that are all on my website. I'm happy to send these out.
I want to thank the postdocs and the students who did all this work. Lots of people that contributed to the stuff I showed you today. All of our collaborators, all of our funders, and most importantly, the model systems, the animal model systems, because they do all the heavy lifting.
