All right. Thank you very much. I'm very pleased to be part of this amazing meeting. So, thank you for having me here. Hopefully everyone can see my slides. Let me know if the... Yeah, that's all good. Thank you.
Perfect. All right. So, what I'm going to talk about today are some approaches to the emerging field of diverse intelligence. And if anybody wants to follow up, all of the primary papers, the software, the datasets, everything can be downloaded here. And then this is my personal blog of thoughts about what some of these things mean.
So one thing that I find really interesting is that, of course, Alan Turing, who was very focused on different kinds of substrates for intelligence and computation and the ability to recognize mind in other substrates, actually wrote this amazing paper which was one of the first attempts to formalize how order arises in embryogenesis. So the chemical basis of morphogenesis in embryonic development. And I think that's because he recognized this very deep symmetry between the way that bodies and minds come together.
In other words, this idea that the formation of the body and the self-assembly of the body and the autopoiesis of minds are really a symmetrical process that have a lot in common. And I think that's very true. And this is the kind of thing we've been working on. So what I'm going to do today is try to cover three areas.
First, I'm going to show you a bit of unconventional biology as a kind of inspiration for thinking about embodiment and intelligence. And so this is going beyond the traditional view of the brain and neural architectures as bio-inspired computing. And then I'm going to spend some time showing you some progress that we've made in communicating with an unconventional embodied intelligence, which is basically the agential material of life.
And we've uncovered this bioelectric interface as a modality to talk to this cellular collective intelligence, which has many practical applications. And so I'll show you that. And at the end, I'm going to say a few pretty speculative things about where the intelligence of novel beings might get its goals and how we can learn to relate to novel beings that have never been here before.
What I'm interested in is developing a framework that would allow us to recognize, create, and ethically relate to truly diverse intelligences. So, not just familiar creatures such as birds and mammals and so on, but colonial organisms, swarms, synthetic life forms, various engineered chimeras, AIs, whether they be purely software or robotic, and perhaps someday exobiological agents, and even really some very exotic kinds of things such as active patterns within media.
And we work on trying to develop formalisms and tools that help use these ideas to move experimental work forward. That is not just philosophy, but actually using some of these unusual ideas to drive applications in biomedicine, in synthetic morphology. And most importantly, I think ultimately this is going to impact a lot of other issues around ethics and so on.
And so of course we're not the first to try to do this. So here's Rosen, Lettvin, and Wiener in 1943 trying to show a spectrum that goes all the way from passive matter up through human-level metacognition and beyond. So this idea that we're really talking about a spectrum has been around for a long time. And I want to show you some of the ways that we've been thinking about this.
And so the first thing we'll talk about is some biology that you may or may not have seen before, which are giving us a clue as to what the unique features of the material of life is, and the ability to basically self-organize models of itself and the outside world as a form of sense-making all the way down to the most simple elements of living material.
And why this is fundamentally driven by the unreliable substrate, which will tell us how this is different from all the computational technologies that we're working on today. So I'll just show you some examples. So here's a tadpole of the frog Xenopus. Here are the nostrils, the mouth. Here's the brain, the gut, and the tail.
What you'll notice here is that what we've done is prevent the primary eyes from forming. But we did induce an eye on its tail. And I'll show you in a few minutes how that happens. We induced an eye on its tail. And this eye makes an optic nerve. Here's one optic nerve that comes out of this eye. That optic nerve does not connect to the brain. It might connect to the spinal cord, or sometimes to the gut, or sometimes nowhere at all.
But the most amazing thing is that when that happens, these animals can see. And so we test them in this machine that we built that basically automates the training of these tadpoles on visual cues. And we find out that they actually can see and they can learn visual behaviors.
Now, isn't it remarkable that this radical change in sensory-motor architecture does not require new rounds of evolution? You don't need additional generations of adaptation, of mutation, all of the things that we like to think are required for pulling together a system that has this kind of complex sensory-motor behavior. Actually, it works. It works right out of the box, on the fly. No evolutionary adaptation needed.
And I'm going to argue that this is basically because the whole process of setting up the system every single time, even under standard conditions, is really an example of problem-solving from scratch. This animal never assumed that the eyes were going to be in the correct spot anyway.
This plasticity is also morphological. So for example, tadpoles here that need to become frogs, they have to rearrange their face. And it turns out that when we make these so-called Picasso tadpoles where all the organs are scrambled, so the eyes on top of the head, the mouth is off to the side, everything is in a wrong configuration, these animals still make pretty normal frogs.
And all of these organs move in novel patterns to get to where they're going, even though they're starting off from an incorrect position. So the genetics does not specify a set of hardwired rearrangements. What it specifies is a system that can do error minimization. It can recognize unexpected states and it can take corrective action toward a specific goal.
So this is a kind of cybernetic system that implements anatomical homeostasis. So it has the ability to reach a specific goal despite being deviated from that goal by an unexpected perturbation. Now we know that morphogenesis is very reliable. So here's an example from the plant kingdom.
So you've got acorns. They very reliably make these oak leaves. And so we like to think that this is what an oak genome knows how to do. It knows how to make this particular structure. And so we wonder, because morphogenesis is so reliable, does that mean that the goals are hardwired?
In other words, in that frog example I just showed you, we know that the movements of the organs aren't hardwired, but surely at least the pattern of the final frog face is somehow determined by the genetics. Well, what's interesting is that like many highly plastic goal-directed systems, morphogenesis is not hardwired. It is extremely hackable.
And we wouldn't know that except that there's this non-human bio-engineer, this amazing wasp that will come and give some prompts to the cells. And what they will build is something that's very different from the standard leaf morphology. So instead of this flat green thing that we're used to, it will make this remarkable round, spiky, red and yellow thing, or here, or something like this. There are many forms of these galls.
This is not made by the wasp cells. This is made by the plant cells. Okay? When it's been prompted by signals from the wasp, they take advantage of their morphogenetic competencies. We would never know that this was even possible. And we're now developing some tools to try to take advantage of this plasticity.
It took the wasp millions of years to learn to do this. Hopefully with modern tools, we can do it much faster. So we now see that life, the material of life, has this amazing problem-solving competency where it can try to reach the same goal by different means, and that competency can be hijacked by rewriting the goals.
And so the other thing that I want to talk about is the role of information and interpretation in this architecture. One of the critical things about biology, which makes it a little bit different than most of the technology that we build, is that it optimizes for the active salience of information, not the fidelity or not the correctness of information.
And I want to start by showing you this example. So here's a caterpillar. This is a soft-bodied creature that crawls around in leaves. And it lives in basically a two-dimensional kind of world. And it has a brain that's suitable for driving a soft-bodied creature in that world.
It needs to become a butterfly, which is a hard-bodied system that flies through three-dimensional space. And in order to do that, during this process of metamorphosis, it basically liquefies its brain. So it dissolves most of the connections. Most of the cells are killed, and it puts together a brand new brain that's more suitable for this lifestyle.
Now the remarkable thing is that if you train a caterpillar, let's say, to find leaves at a particular color stimulus, that memory can be still found in the butterfly. So one question we could ask, of course, is how does the information survive this drastic refactoring of the substrate?
But a more interesting aspect of this is that the actual memories of the caterpillar, the exact details of finding a specific cue and then crawling and then eating leaves as a reward, those kind of memories are useless to the butterfly. The butterfly does not move the same way. It doesn't want the same thing as reward. Butterflies drink nectar, not eat leaves.
So what you can't do in the system is keep the memory exactly as it was. What the system is actually doing is remapping these memories onto a completely different architecture. So if you want to think about transfer learning or other kinds of robotic plasticity, what you've got here are memories that survive not by remaining intact but by being remapped onto a very different architecture.
And so in its new three-dimensional life, it brings with it some of the generalized lessons from the past, but not the specifics. And so all that is described here.
And so more generally, what we're looking at here is this conundrum in which all living agents find themselves, which is that at any given moment we don't have access to the past. What we have access to are the engrams, the memory traces that past events have left in our brain or body.
In other words, we have generalized past experiences into very compressed representations that are biochemical, bioelectrical, and so on. And these must be reinflated every moment dynamically to keep a continuously updated story or self-model of what we are, what our past has been, and what our world is like.
The computer science folks in the audience will immediately recognize this architecture. What's happening here is that very rich input data from the past at any now moment has been compressed into this bottleneck of storage of memories. And so this part is algorithmic and deductive. It's a compression and a kind of thing.
But from here on, at any given moment, interpreting what do your memories mean, given changes that have occurred in your body, in the environment, the continuous process of reconstructing your memories into a dynamic story is creative because you've lost information here. This is not algorithmic. You can't simply recover all the initial information.
And so you can think of memories as messages from your past self, and those memories have to be interpreted exactly as you interpret signals from lateral interactions with others at the same time. And so one could imagine that we are a continuous reconstruction of a dynamic model of ourselves.
And because they need to be reinterpreted, and because you don't have a ground truth of what these physical traces, whether they be chemical or biological, what they actually meant to your prior self, and you have to in real time construct the best possible story you can, this is fundamentally what biology commits to.
And that kind of thing, the inability to trust that your material still means what you think it means, because you know it's going to change on an evolutionary time scale. You know you're going to be mutated. The biological substrate is very unreliable. You never know how many copies of proteins you have, and you know things like that. So it gives rise to this amazing plasticity.
Here's one of my favorite examples. This is a cross-section through a kidney tubule in the newt. And so normally eight to ten cells used to build something like this. But we can make polyploid newts that have many more copies of their genetic material. This was done in 1945 by Fankhauser.
What you find is that the cells adapt to that by becoming bigger, but the overall newt stays the same size. And so what you see is that it's using fewer but larger cells to build exactly the same structure. The most amazing thing happens when you make truly gigantic cells. What the cells will do is bend around themselves and leave a hole in the middle, still making the same structure by a completely different molecular mechanism.
I mean, this is basically the definition of intelligence that's measured in IQ tests. Can you use the tools you have in new ways to solve problems you haven't seen before? And just think about what this means for an egg coming into the world. What can you count on?
Well, you can't really count on knowing what is going to be in your environment, but you also can't even count on your own parts. You don't know how many copies of your genetic material you're going to have. You don't know how big your cells are going to be. You don't know the number of cells that you're going to have. And you have to adapt and build a proper egg regardless of what's happening.
And I think this is why that frog tail eye example works that I showed you at the beginning, because all development, regeneration, all of those kinds of processes are continuous problem-solving, taking the parts you have and doing the best you can with the information that's given, because you can't be certain of what it is.
And of course, Josh Bongard and Hod Lipson's group back in, I think, 2006, had this amazing robotics paper that kind of illustrates this idea of not being sure of what you are and having to reconstruct that from scratch.
And so in biology, we see this on multiple scales. The genetic material you are given is a generative model that's open to interpretation. And that's what morphogenesis does, is interpret the information passed on from your ancestors to see what is the best story we can put together now. What kind of system can we build now given the information that we have?
This happens on an evolutionary time scale, on a developmental time scale, and as we talked about, on a cognitive time scale as well. And you can see some of those things here.
And I think what actually happens is because that unreliable material requires life to be actively interpreting and reinterpreting whatever it's handed by prior versions of itself, because of the need to be an active, creative interpreter of that information, it actually ends up making a very interesting positive feedback loop with intelligence.
Because that process of dynamically interpreting your material in a way that is not a direct mapping means that selection can't really see the genome. What selection sees is the results of that interpretation, of the physiological software running on cells. That's what creates the phenotype, and that's what selection actually sees.
And when you hide information about the genome from selection, what we find, and this is all simulated in other papers, what you find is that evolution ends up doing a lot of its work on the actual creative problem-solving layer. In other words, not on the prompts, but on the ability to interpret them.
And this then, of course, makes it even worse. It makes it even harder for selection to see what actually went in. And this results in an amazing spiral of intelligence and evolutionary processes.
So the next thing I want to do is show you a specific example of how we applied some of these tools to start to try to detect and then communicate with a novel embodied intelligence. And basically what I'm going to do is talk about the model system, which is the collective intelligence of cells in the body having behavior in anatomical space.
So here's the journey that each of us takes. We begin life as a single cell. So here is an unfertilized quiescent oocyte, which looks to be amenable to the techniques of physics and chemistry and so on. And then through the process of a smooth, gradual embryonic development, we become something like this, or even something like this.
In other words, we transition from a domain covered by physics and physiology into something that needs behavior science, psychology, and psychoanalysis and so on. And of course, the critical thing to remember is that there is no sharp transition here. There is no magical red line where suddenly cognition kicks in.
So what we actually need to do is develop stories of scaling, not discrete categories of beings, but actually as an explanation of how cognition scales from the very primitive kinds of simple things that cells can do into the kinds of intelligence that we all can recognize.
And there are some other paths that we'll talk about here in terms of cancer and anthropods and some other things that can happen after this step.
So one thing to remember is that we are all built out of a very competent set of tools. We are a collective intelligence, and we are made of an agential material. Not just the bees and the ant and termite colonies, but all of us are collective intelligences because this is the kind of thing we're made of.
This is a single cell called the lacrymaria. This is a free-living organism. But all of us are constructed from cells. And you can see the amazing competency it has in its own tiny little world that it's working to penetrate.
And even below that level, if we ask what's actually underneath this, why are these cells able to do this? Even below, what we find is that there are molecular networks that are themselves capable of six different kinds of learning. So not just the cell itself, but just the molecular networks, for example, gene regulatory networks, biochemical pathways.
They themselves, if you do a learning paradigm such as stimulate condition nodes and look at response nodes and so on, they have habituation, sensitization, associative conditioning. And we are now building tools to actually do this in real cells, for example, for drug conditioning and things like that.
Many, many applications in biomedicine to the fact that even the molecular components inside of our cells already have certain learning abilities. So you don't need brains for this. You don't need the synaptic machinery. This is baked into the math that governs molecular networks.
So what we have here then in biology is what I call a multiscale competency architecture. That is, there are agendas, memory capability, goal-directed problem-solving all the way through at all of these different layers.
So in order to think about embodied intelligence, we really have to remind ourselves that while we as humans are pretty good at noticing intelligence in three-dimensional space, medium-sized agents moving at medium speeds in three-dimensional space, biology exploits lots of different embodiment in other spaces that are very hard for us to see.
So we have to be really careful here because when we talk about, for example, software AI not being embodied, we have to remember that embodiment is not just for three-dimensional space. You don't have to have a robotic body that wheels around or walks around in 3D space to have an embodiment.
Biology exploits the same kind of perception-action loops, memory, predictions, problem-solving in other spaces. The spaces of possible gene expressions. This is a very high-dimensional space. Physiological state spaces, and what we're going to talk about now, which is anatomical morphospace, that is the space of possible anatomical configurations.
And I think as far as biology is concerned, these are no less real than the three-dimensional space on which we are often fixated, both sort of personally and in science when we talk about embodiment. We have to remember that biology does not really distinguish this. All of these spaces are very serious for these agents that traverse them.
So let's talk about anatomical morphospace. How do we as collective intelligences, meaning groups of cells, how do we navigate anatomical space? Well, we start here and very reliably we end up with something very complex.
And the reason I'm calling this intelligence is not because it's reliable. It's not because there's a lot of emergent complexity, that those things are easy to get. That's not the intelligence part. The intelligence is actually the problem-solving capabilities in this space. And I've shown you a couple of them already.
But fundamentally, for example, if we cut these early embryos into pieces, you don't get half bodies. You get perfectly normal monozygotic twins, triplets, and so on. Because this is a system that knows how to get to its ensemble of goal states from different starting positions and actually avoiding lots of local minima and so on.
So it can navigate from different starting positions to the normal target morphology of where it needs to go. Not just embryogenesis. Some animals like this axolotl can do this throughout its life. So you can cut the limb anywhere here, and these cells can immediately detect that they've been deviated from the correct target morphology, and then they will work really hard to get back to where they need to be, and then they stop.
So this is the most amazing aspect of regeneration, is that it knows when to stop. When does it stop? It stops when the correct salamander limb has been completed. And so clearly again, this is this kind of example of anatomical homeostasis.
And we have to ask then the question of how does it know, how does it remember what the correct state is? So we've been investigating this using inspiration from one uncontroversial example where we know that there are groups of cells that store goals and pursue them, and that's the brain and nervous system.
So in the brain and in the nervous system, you have this architecture where specific ion channels, so these are little proteins that enable ions to enter and exit the cell. These ion channels enable voltage gradients to form, and then these voltage states some do or don't propagate to their neighbors through these gated electrical synapses.
So you get this electrical network that's capable of storing and processing information, and the output of that network controls muscles to move us through three-dimensional space. That's how we conventionally think about this. Information is processed electrically and then moves us through three-dimensional space.
And then there's this process of, this project of neural decoding, where scientists try to read that electrophysiology and infer what are the memories, the goals, the preferences and so on of that agent, by reading out the electrophysiology.
And so what's happening here is that the electrical signaling is a kind of cognitive glue. It's what makes us more than just a pile of neurons. It's a set of processes and policies that bind individual neurons together into goals and preferences and memories that no individual cell will have.
So for a large-scale collective intelligence, you need a mechanism like this. You need a cognitive glue to do this. So it turns out that this cognitive glue mechanism is ancient. Evolution discovered this long before nerve and muscle came on the scene.
And basically around the time of bacterial biofilms is when it was discovered by nature that electricity is a fantastic way to bind active subunits into a larger collective that does things in other spaces.
And so it turns out that every cell in your body has these ion channels. Most of them have these electrical synapses. And all of these cells are forming an electrical network. And instead of moving you through three-dimensional space, what this electrical network does is control cell behaviors to move your body through anatomical space.
This is what the activity of this network is, what takes you from a single cell egg all the way through to the human morphology that we have.
Now, if we want to ask what are the non-neural electrical networks thinking about, we know what the neural ones think about primarily. They think about movement, and then in humans they can think about other things, but these ancient networks were thinking about one thing, and that is navigating anatomical space.
So the electrophysiological software of life is basically a navigational competency that moves you through three-dimensional space. And as we saw, it's not just a matter of retreading the same path every single time. It has to be really clever in how it does this, and so this is why we're interested in this.
Now, how can we detect this? Well, we've created some methods to actually read the mind of the body, the way that neuroscientists try to read minds that are implemented by brains. And so we use, for example, voltage-sensitive fluorescent dyes to track the bioelectric state of different cells of a whole embryo.
So this is an early frog embryo. This is not a model. This is a real frog embryo. You can see all the electrical patterns that are here. And then we can try to decode them. So we have simulators, some of them are down at the transcriptional level, telling us where the different channels come from, and then the tissue-level electrophysiology, which we use to try to explain the repatterning of regeneration.
And then of course merging this with thermodynamics and connectionist models in computer science to try to understand, for example, regeneration as pattern completion and memories as attractors in these networks and so on. So we're able to detect and model the information processing that goes on in the collective intelligence of the body.
And then of course, the most important thing is to be able to rewrite those patterns to show that they're functional. And we don't use any applied fields. So this is not about electrodes or electromagnetics, there are no frequencies, no magnets, nothing like that.
What we do instead is use electrophysiological techniques, basically borrowed and remapped from neuroscience, to control the interface that cells are usually using to talk to each other. So the only reason the things I'm about to show you work is not because we are so smart. It's because we are just hijacking the system that these cells normally use to store information.
So that means we can control the ion channels on their surface using drugs, using light, using genetic modifications. We can control these ion channels, which controls the voltage of each cell, and then we can control the topology of the network, that is these gap junctions. We can influence who gets to talk to whom.
So we have control of the plasticity of the connections and of the electric states of the cells themselves. So now it's time for me to show you what happens when we do hack this interface. What kinds of information can we put in? Can we communicate novel information and novel goals to this collective intelligence?
Because remember, the whole point of calling this a collective intelligence, it's not just a poetic metaphor. The whole point of this is to enable the use of behavior science and cognitive neuroscience to do new things in morphogenesis that you couldn't do before, because we recognize that this is actually a behaving system navigating morphospace, and we can interface with its cognitive processes.
which by the way are as I pointed out homologous to the ones that are used in the brain and this is the same system basically pivoted into a new problem space and just made faster in the brain. So one thing you can do is you can provide prompts that tell the cells to build specific organs. So in this case we inject a particular ion channel that establishes a voltage state that we now know is interpreted by cells to mean build an eye. And so when we inject this into cells that are going to be the gut of an embryo, what they do is they go ahead and they build an ectopic eye. Now this eye has all the same contents that it should, retina, optic nerve, lens, all that stuff. And what's remarkable about this is that we don't give it the information on how to build an eye. That's an extremely complicated process. Lots of different cell types. We don't talk to the stem cells. We don't specifically tell which genes to turn on. That's not it. Because this is an agential hierarchically organized material. All we have to do is give it a prompt. We say build an eye here, a very low information content signal, and the cells do the rest.
Not only do they do the entire structure with all of the downstream steps that have to be organized for this to happen. Many genes have to turn on and off and so on, but they can also self-scale to the problem. So for example, if we only inject — this is a cross-section taken through a lens sitting out in the tail somewhere. If we only inject a few cells here, they will automatically recruit their neighbors to complete the project because they know there's not enough of them to build the whole lens by themselves. So the blue ones are the cells that we injected. All these cells that are not blue that are part of this lens were instructed by the blue ones to help to participate in that process. Much like other collective intelligences like ants will recruit their nestmates to carry heavier objects and so on. So you can start to see the competence of the material as hierarchical organization. There's the ability to do very complex things off of simple prompt, self-scaling to the problem.
We can rewrite specific memories. So here's a flatworm. These are planaria and planaria regenerate and so you can cut them into pieces and they very reliably regenerate. And the reason that this middle piece right here knows to make one head not two is because it has this biological pattern memory here that we can see that says one head not two. Okay. And so what we can do is we can go in and we can rewrite that memory. And these methods are still being worked out so it's a little messy but you can rewrite it. And when you do this, then you cut it and then the piece will build to the set point. Very important to note this biological map is not a map of this two-headed animal. It is a map of an anatomically normal one-headed animal which has normal expression of genes. What is abnormal is the memory it has of what it would do if it gets injured in the future. In other words, a counterfactual memory, the beginnings of the kind of time travel that we see in brains where you can think about things that are not happening right now.
So the normal one-headed body of a planarian can store at least two different representations of what a correct planarian should look like. And if we change that representation, the cells will build to that representation. Again, we don't need to tell it how to build a head. This is very high level communication that resets the goals of this goal-directed process. If you do this, you find out that it has a memory. In other words, these two-headed animals continue to form two heads in future rounds of regeneration. You don't need to do anything else to them. It's basically forever now. Once you've changed that memory, they will keep it. We do know how to set it back. You can erase that memory and take them back to the one-headed. But this has all the properties of memory. It's long-term stable. It is rewritable. They have conditional recall, meaning that they can hold on to a latent memory until you cut them, and that's when it becomes obvious. And so on.
And here you can see this is a movie of one of these two-headed worms doing its thing. Not only can you control the number of heads that way but you can control the shape of the head and make them form heads belonging to different species. So for example this triangular shaped head, this animal can be, by manipulating the pattern memories here, by the bioelectrics of regeneration, you can have it make flatheads like a planina, round heads like a nest. Mediterranean, about 100 to 150 million years of evolutionary distance between these guys and these. And yet this hardware with no genetic modification. Remember in that two-headed form that I showed you again, no genetic modification. There's nothing wrong with the hardware. This is totally a software and behavior issue. They are perfectly happy to visit the attractors of other species in anatomical space. Right? So here are the different types of shapes that you can make. The same hardware is very happy to visit these other attractors where these species normally live. And it's not just the shape of the head, but it's also the shape of the brain and the distribution of the stem cells that becomes just like these other animals.
In fact, the latent morphospace is actually quite large. You can make all kinds of things that don't look like planaria at all. You can make round cylindrical things that are very different from their flat architecture. You can make these hybrid things. You can make these crazy spiky forms. There's enormous plasticity as the system navigates the anatomical space. Okay. And so the very last thing I want to talk about is this. We've now gone over a number of examples where biology is very good at reaching their expected goals in anatomical space. It uses a system of decision-making and sensing that can accommodate lots of different problems, chemical, surgical, lots of different weird scenarios. And they always try to reach that correct species-specific pattern. So, and then what I just showed you is that that pattern is encoded bioelectrically and that by changing that set point, by resetting that, now that we've started to crack the bioelectric code a little bit, we can now reset that set point and we have many applications in regenerative medicine, birth defects, cancer which I haven't even talked about and so on. That is really focused on resetting those set points.
But now I want to ask maybe a deeper question which is where do those set points come from in the first place? Okay. And this has many implications for us going into the future because you might think that well in all of the creatures that I've shown you today, the set points come from evolution. That is eons of natural selection that have established specific set points. And that works for natural creatures. But what we are going to be having to deal with soon is a wide variety of novel beings. So just like humans stand at a center of a continuum leading from cells into what we think of as today as the kind of standard modern normal human, we are also at AI in the middle of a continuum of change. So both technologically and biologically there are going to be all sorts of different cyborgs and hybrids and all kinds of augmented humans with sensory augmentation and sensory substitution and so on. And of course there will be creatures like this, beings who cannot be dealt with by these simple questions of is it a machine, is it a living being. I think those distinctions are going to go out the window because of its plasticity and because of the fact that biology is solving problems every single time that it does not assume the meaning of its information. It's incredibly interoperable, right? Every level of organization here can be interfaced with engineered materials. So there are going to be an incredible number and variety of beings for whom there is no standard evolutionary trace where you can say this is what was selected as its goals.
And so now we have to start thinking about where can patterns of shape and behavior come from if it's not just the genetics and their history. Where else do patterns come from? And one thing that is important to think about is mathematics. So in mathematics we have an amazing example of patterns that exist that cannot be explained by physics. They cannot be explained by a natural history of selection. None of those things that we like in biology work here. So this is what's called a Halley plot. And it's a map of a root finding method applied to a very simple formula like this. So this is the prompt. This is the seed. And the algorithm will show you what this actually encodes. What this encodes is this amazing structure. If you vary these numbers by a tiny bit, you can make videos. And so these are the Halley plot movies that I've generated. You can make other kinds of plots. It doesn't hurt that they all kind of look pretty organic. But the biggest thing here is that if you ask why this particular pattern and not some other pattern, we come face to face with the fact that in mathematics we have patterns that cannot be explained by either physics or history.
Okay. And so now we have a new tool that we can start to think about what do we do with novel beings to try to understand where their patterns of form and function come from because physics and history are not the only sources of information. So let's look at what could those kinds of beings be? Well, I'm going to introduce you to this thing which we call an anthrobot. And you can see it moving around in the field of cells. And you might think that this is, if I asked you what this was, you might say that this is something we got off the bottom of a pond somewhere. It's some sort of primitive organism. You might guess that if you sequence the genome, you might guess that it would have a genome similar to a primitive organism such as this. Actually, if you were to sequence it, you would find out that it's Homo sapiens. 100% unedited normal Homo sapiens genome. These are human adult human cells that have been basically allowed to reboot their multicellularity in a protocol that we have and they put together this self-motile active little protoorganism we call an anthrobot.
The form and function cannot be guessed from the human genome. It looks like no stage of human development. Actually, human morphology and capabilities also can't be inferred from the genome, but certainly you wouldn't know that it could do something like this. It has all kinds of interesting properties. First of all, this is a volcano plot of new genes that it expresses. About half the genome, almost 9,000 genes are differently expressed in these anthrobots than they are in the tissue that they came from. They come from tracheal epithelial tissue, by the way. They're donated by adult human donors. They come from lung epithelium. But because of their new lifestyle, they have a radically altered transcriptome. They have four different kinds of behaviors. And you can draw an ethogram of transitions between these behavior states. And they have really weird capabilities. Like for example, if you put them on a field of damaged neurons with a big scratch that we made here, they will actually sit in this gap and they will heal the neurons. You know, who would have thought that your tracheal epithelial cells have the ability to make a self-motile little biobot that runs around and has the ability to heal other tissues in your body. You know, we would have never known this. Where is this coming from?
Here is a different construct we call xenobots. So these are made from frog epithelial cells, from frog embryonic epithelial cells. And you can see here they have all kinds of interesting behaviors. Again, they form this kind of motile little thing that swims by virtue of cilia, these little hairs that push against the water and swim around and you can make them in different architectures and you can trace their collective behavior. They have all sorts of interesting group behaviors. Here's one navigating a maze. And so it takes this corner without having to bump into the opposite wall at this point. It spontaneously turns around. So they have novel behaviors. One thing that they do is called kinematic self-replication. That is if you give them a field of materials, so loose epithelial cells, they will actually combine those into a little ball. They will polish these little balls and because they're dealing with an agential material too, these little balls mature to become the next generation of xenobots. And guess what they do? They do exactly the same thing and they make the next generation and the next generation. As far as we know, no creature on Earth does this kind of kinematic self-replication.
So where did this come from? Right? There's never been any evolutionary selection to be a good xenobot. There's never been selection to be an anthrobot that can run around and heal neural wounds. One of the interesting things about this is that when we looked at the, again much like with the xenobots, we looked at the transcriptome and there are many many genes that these new creatures, again remember no changes in genome, no weird nanomaterials, no synthetic biology circuits. All we've done is given them a new environment and a new lifestyle and they express lots of new different genes. And one thing that we discovered, this is just one example, we discovered in this data set a cluster of genes related to hearing. Okay. So normal embryo tissues do not express this. This is purely a xenobot thing that they turn on in their new life. And what's happening here is that they express all these hearing genes and we wondered is it possible that by expressing all these hearing genes they're actually trying to have a new sensory modality in their environment. And it turns out that they can hear. And so the way that we detect this is we have a speaker underneath the dish and we provide sound stimulations or vibration. And what you see here is that whereas embryos are much less reactive to this, xenobots actually, here's what actually happens.
[clears throat] Sorry, my keynote just died. Let me restart again. Oops. Let's see. Sorry, are people seeing my slides again?
