The Collective Intelligence of Cells During Morphogenesis with Michael Levin - Incredible Minds

Thank you so much. Really appreciate the opportunity to share some ideas with you and to be part of this event and some of the other speakers you've had, which were tremendous.

So, let me see. All right, can everybody see the slides? Is it good? Okay. So, if you're interested in this after the fact, you can find at this website here, you can find all of the primary papers, the software, the datasets, everything is here, and you can also get in touch with me here.

So, today what I'd like to transmit to you are some examples of the following main points. I'm going to stress this idea that all intelligence is collective intelligence, and in particular, the notion of scaling from the competencies of simple subunits to larger kinds of cognitive systems. We're going to talk about problem solving in diverse and unconventional problem spaces, and we're going to talk about the navigation of these virtual spaces as intelligence.

We'll talk about a kind of goal-directedness. We'll talk about agency in unconventional embodiments, and towards the end, this idea of universal hacking and of components of biological systems trying to manipulate the behavior of other systems both above, below, and lateral to them.

I'm going to spend about a third of the talk using anatomical control as an example of a collective intelligence of cells navigating a particular kind of anatomical space. And I'm going to show you lots of data on bioelectric networks as a kind of cognitive glue that is the ancestor of brain function. And we're going to talk about the impact that this is going to have on biomedicine.

And towards the end, I'm going to talk about some synthetic bioengineering and ways to think about creating and then relating to novel types of beings, which has lots of implications for science, for engineering, and for ethics.

So, the talk is going to go in three parts, and what I'll do is at the end of each part, I'm going to pause. I'll take some questions. We'll take a break. People can get up and stretch and get water and whatnot. And then we'll continue.

So, the first part is going to be my efforts to break down some common categories and some terminology that I think is really limiting the field. And so for this reason, we're going to talk about stretching that into a kind of continuum of various types of agents, and I want to show you some examples.

So, this is, even though this is fairly a very old view, this is a famous painting of Adam naming the animals in the Garden of Eden. This is still actually quite prevalent in many, many relevant discussions. What you see here is that there's a very sharp binary distinction between Adam here and the rest of the animals. So, it's very clear that there are different kinds of creatures, fundamentally different kinds.

But what we know now, because of Darwin and because of Von Baer, is that actually this notion that there's some kind of a sharp binary category of being a human is actually a smooth continuum, both on an evolutionary time scale where if you just slide back in evolutionary time, you see a very gradual set of changes that go from that human all the way back to a microbe and below that.

And of course, even on the life scale of the individual itself, you see we all start life as a single cell, and then slowly and gradually we become what is pretty much, you know, what is basically considered to be a standard human. So, we are at the center of this kind of continuum that stretches in two different time scales.

But it's actually even more complex than that because now with bioengineering and various other technologies, there are additional axes in this space where we can stretch the notion of what a human actually is. And you can imagine slowly adding various prosthetics, various technologies within the body, as is already happening. You know, people who can control wheelchairs with their minds via implants and things like this. And of course, biological modifications that we can make, both genetic and non-genetic.

And so we see here that at every level of organization, cell, tissue, organism, and so on, we can replace native, naturally evolved materials with all sorts of novel, either natural or engineered materials. And so we start to see that this notion of a standard human can be quite continuous in a number of different dimensions.

And so I've been trying to develop a framework called TAME, and we'll talk about this more. It stands for Technological Approach to Mind Everywhere. And the idea is to be able to recognize, create, and relate to truly diverse intelligences.

Not just the familiar creatures like us and maybe birds and dolphins and things, but also really unusual creatures like colonial organisms and swarms and synthetic life forms that might be engineered, software artificial intelligences, possible alien life if that's found.

So, all of these things ought to be able to be handled by a mature science of diverse intelligence. And they have to be dealt with the same set of concepts regardless of what they're made of. So, regardless of their composition or origin story, meaning evolved versus engineered or anything in between.

And of course, I'm not the first person to try for something like this. So, Rosenblueth, Wiener, and Bigelow did something like this in the '40s where they gave us this ladder where you can see that all the way from basically passive matter all the way up through complex human metacognitive types of capacities, there are a set of great transitions, but nevertheless, these things are all on the same spectrum.

The idea was that we should be able to understand how various types of material implementations can begin to have these kinds of functionalities.

And so, one of the essential ways of thinking about these things in my framework has to do with the relation of an observer, and that observer might be an engineer, it might be another conspecific, it might be some sort of a parasite, or in fact, the evolutionary process itself. Some sort of observer that is able to modulate what the system is doing.

And so you can think about different ways of modulating a system depending on how sophisticated it is. So down here on this end of the spectrum, the only way you're going to make changes is by engineering directly at the hardware level. You're not going to convince a clock of anything, you're not going to reward and punish it. That's all there is, is mechanical engineering.

But cybernetics takes us to things like this, these thermostats which have recorded set points, and so there you can do some interesting things such as modify the set point without rewiring the hardware and get the system to take on a novel set of goals without having to rewire it.

And then of course, more complex systems avail us of an interface involving rewards and punishments and various other behavior shaping techniques where you can rely on the system to learn from experience. And then of course, we have even more complex beings which are able to follow logic and on.

And so the difference in all of these different systems is not what they're made of or how they got here, but it's what are the set of tools and techniques that you might use to relate to that system, all the way from physical rewiring to something that looks a lot more like a psychoanalysis where you're literally trying to affect the model that the system has of itself, of the outside world, and so on.

So this shows you the kind of inclusive framework we're trying to make. And the interesting thing is that all of us made this amazing journey across the Cartesian cut. We were all once so-called just physics and chemistry, a little blob of protoplasm, an unfertilized oocyte, quiescent, and then through the process of embryonic development we became one of these functional creatures, or even something like this.

And the crucial thing about this process of self-construction is that there is no magical sort of sharp red line at which point you go from being just physics to being a true cognitive being.

So many people would like to look at themselves and say we have true, you name it, sentience, cognition, intelligence, whatever term you want, and then they will look at something like this, which is a little blob of chemicals, and they say, well, that certainly doesn't have any of those things. And so then you owe a story of how you got from here to there, and that story cannot be any kind of a sharp change. It has to be a gradual process that is in concert with what we know about embryonic development and so on.

So that's interesting. So we go from matter to mind in this way that, you know, I think developmental biology is maybe the most magical of the sciences. But at least we tend to think of ourselves, once we get here, we tend to think of ourselves as some sort of a unified intelligence.

And in fact, often people will contrast the kinds of things that we are with something like ant colonies or bee swarms or something like that and say, well, we talk about that as a collective intelligence, but that's a metaphor. We are a true intelligence, you know, collective intelligence is something else.

But actually, this is not the case. And so Descartes, for example, was really interested in the pineal gland because he said that the pineal gland was unique, it was an unpaired organ in the brain, and therefore kind of a singular unified location for the mind to interact with the body.

But if he had had access to good microscopy, what he would have discovered is that there isn't one of anything in the brain. In fact, inside the pineal gland you see this, which is a huge number of individual cells. And if he had had this technology, he could have looked inside of any one of those cells and discovered that they in fact are made of a huge number of different components and various kinds of active matter.

So there are mitochondria, there are little motors that zip up and down various filaments, just an incredible number of parts. And so all of us are collective intelligences. I'm not aware of any intelligence that is really just some sort of one indivisible diamond of mind. We are all made of parts of different kinds.

And so this is the sort of thing we're made of. This is a single cell, it's called a lacrymaria. This is a unicellular organism. It has no brain, it has no nervous system, no stem cells. It has remarkable competencies over its physiological, metabolic, and anatomical goals.

So you can see here that it's basically, this is it hunting, it's in its environment. If you're into soft body robotics, this is an amazing design challenge. We don't have anything remotely this competent. And cells have all kinds of other interesting capabilities. So this is the sort of thing that's studied in the field of diverse intelligence or basal cognition.

I'll show you another example. This is a slime mold. So the slime mold, this is a unicellular slime mold called Physarum polycephalum, and we've put a little bit of it here on this petri dish. We put one glass disc up here, we put three glass discs back here. They're just glass, they're completely inert, they're very thin, they have no food or any kind of a chemical substance on them.

But what happens is that this, and by the way, the slime mold is a single cell, it can be very large, but the whole thing is a single cell. So what it starts to do, it starts to grow out basically in every direction at once. But during this time, during these 600 minutes or so, what it's actually doing, as we discovered, is gently tugging on the material around it and reading the strain angle that comes back.

And based on that strain angle, it builds up towards a decision of where the heavier mass is. And the way we know this is because reliably what it's going to do is grow towards the heavier mass. And so by reading that strain angle, it's able to gain some knowledge of its environment, and then for some reason that we don't know, it prefers moving towards the heavier masses.

And so we see that we have this biological material that's very, very sort of ancient and early in life has interesting capabilities. And then the real question becomes, how do these things scale up? What is it that allows combinations of these cells to do something like this?

So this is a very standard scenario, the rat learns to press the lever to get the reward, and that's very well understood. But the interesting thing is that the rat is really a collection of cells, and no individual cell has both experiences.

So there are cells on the foot of the rat that interact with the lever, there are cells in the intestine that collect the nutritious reward sometime later, but no individual cell has the experience of doing this and receiving that. And so who is it that owns this kind of associative memory between two events that no individual cell experienced?

And so clearly you need some sort of, and the standard answer of course is the nervous system provides this, you need some sort of way to integrate individual cellular activities towards a single being. Now this is kind of very common. I want to show you some uncommon examples that people don't hear about a lot to kind of stretch our idea of what's possible.

So this is a caterpillar, a soft-bodied kind of creature that crawls around and eats leaves, and it has a certain kind of brain. Remarkably, it transforms itself into a completely different kind of creature that now lives in a three-dimensional world, it flies, and it doesn't like leaves anymore, it wants nectar.

And so very different set of behaviors, very different set of preferences, very different brain. The brain is actually largely dissolved during this metamorphosis process. Most of the synapses are broken and most of the cells are killed, and a new brain is reassembled.

But there's actually really good evidence that the memories that you impart onto the caterpillar that you train, the moth or butterfly will actually retain those memories. And there's work on this in beetles and various other creatures.

So you can think about two things. The practical issue of, isn't it amazing to have a memory medium that can basically be taken apart and put back together and drastically rearranged and still retain its information content? We don't have anything like that in computer science or anything like that.

And of course it raises some interesting philosophical questions, which is, you know, never mind the older question of what's it like to be a butterfly, the question could be posed, what's it like to be a caterpillar slowly changing into a butterfly? Because this is a change that happens not on an evolutionary time scale, but actually during the lifetime of an active individual.

And so this ability to retain memory through very drastic deformations is really maximized by this other system known as planarian flatworms. We'll talk much more about that. And these planarian flatworms, they can be trained.

And if you train them, for example, to recognize these little bumpy patterns as the place where they get fed, what you can then do is amputate their head, which contains their brain. You are left with a tail. The tail sits there doing nothing for about eight days. It slowly regrows a brand new head.

And McConnell, who discovered this in the '60s and got a lot of flak for it, was actually completely correct in that these new animals with a brand new brain still remember that original information. And so what we have here is first evidence that information is not stored exclusively in their centralized brain, but actually apparently spread throughout the body.

And you see this ability of those tissues to imprint that information onto the new brain as it develops. So this idea of information actually moving through body tissue.

And so you can think about, you know, there's this malfunctioning transporter experiment from Philosophy 101 where they say, well, there's this transporter, and what it does is you step into it and it destroys the original body and recreates a molecule-for-molecule copy somewhere else, and so now you've transported. But what if it malfunctions and it doesn't erase the original copy? How many of you are there?

Well, in planaria you can literally do that because you can cut them into pieces, many pieces, and each piece becomes a perfect little worm. So we start to see that the common workhorses of behavioral science and cognitive neuroscience, rats and things like this, are just the tip of the iceberg.

There are many interesting phenomena that speak to this question of what are cells, what are integrated beings, what's information, how does it move through the body?

In vertebrates you see this kind of interesting plasticity. So this is a tadpole of the frog Xenopus laevis, and here are the nostrils, here's the brain, here's the gut, and what you'll notice is that there aren't any eyes. We prevented these eyes from forming, but what we did do was graft an eye primordium onto the tail.

And so these cells actually make a perfectly good eye, and not only do they make an eye, but they make an optic nerve. That optic nerve comes out, and you can see what it's doing here. It doesn't go to the brain, it stops here, it synapses onto the spinal cord. And we made this machine that automates the training and testing of these animals in visual cues.

So what we find is that they can see. They can see quite well, actually. Despite the fact that they have a radically different sensory-motor architecture, it does not require evolutionary adaptation time. This happens immediately in one lifetime of a single tadpole, and they can do perfectly well with a completely different kind of connection between their sense organs and the brain.

And this kind of thing has massive implications for evolution, too, because some little, for example, a mutation that causes their eye to be pushed off doesn't immediately get weeded out because the animals can't see and die, but rather these animals, the behavior is quite robust, and so the evolution is free to explore any other consequence of that same mutation.

It really buffers, this competency really buffers the negative consequences of mutations.

So what we see, the reason that all of this plasticity works is that biology is built on this kind of multi-scale competency architecture where all the way from molecular networks, cellular networks, and tissues, and organs, and so on, every level here is not merely a structural set of nested dolls, but is actually a functional problem solver.

In other words, they are competent, they have competencies in specific spaces. This might be the space of gene expression or transcriptional space. This might be physiological space. This might be familiar three-dimensional behavioral space, and so on. And so each of these levels solves various problems at various degrees of sophistication.

And so what I'd like to do is really expand from this notion that behavior takes place in three-dimensional space and understand that to biology, all kinds of other spaces that we are not good at perceiving and thinking about, those are equally valid problem spaces.

So this familiar space of behavior, we are very good at recognizing intelligence when medium-sized beings moving at medium speeds through three-dimensional space are doing clever things, and we recognize that right away. But there are these other spaces that we are not good at.

For example, gene expression space and physiological space. I think that if we had a built-in kind of innate ability to feel our blood chemistry, for example, if we had sensors pointing inwards, not just outwards, I think we would have no trouble recognizing our liver and our kidneys as intelligent beings navigating physiological problem spaces and solving various problems the way that we recognize other animals doing it.

And so I'm going to spend most of the time in the next section talking about this problem solving in this anatomical morphospace, but I just want to show you one interesting example of a different kind of space called physiological space.

So here's an experiment we showed a couple of years ago. So here's a flatworm. You put these flatworms in a solution of barium. Now, barium is a non-specific potassium channel blocker. So what it does is it blocks the ability of the cells to swap potassium in and out.

The cells are very unhappy, especially the cells in the head, because all these neurons like to flow potassium in order to keep their physiology correct. And so what happens is their heads explode. Literally overnight their heads degenerate.

But the amazing thing is that if you leave them in barium, within a couple of weeks they grow back a new head, and the new head doesn't care about barium at all. No problem. It's barium adapted. So we asked the simple question, what is different transcriptionally? What genes do these barium-adapted heads express that the standard heads do not express?

And what we found out is that there's a very small number, on the order of about a dozen genes that are changed here to become barium adapted. Now, here's the kicker to all of this. Planaria never see barium in the wild.

Barium doesn't exist in nature in the way that we used it here, and so there's never been evolutionary pressure for planaria to know what to do when they've been exposed to this incredible stressor.

So when you think about how they found these few correct genes out of a sea of tens of thousands of other genes, it's a little bit like, I visualize it like sitting in a nuclear reactor control room. Everything is melting down.

So you have this very strong physiological stressor. What you don't have time to do is to start randomly flipping buttons. And so, not only because there's no time for that, the combinations are astronomical, but also because many of the things you do this way are going to kill the cell long before you find the barium adaptation region of that space.

So it's a search of a roughly 20,000-dimensional space to find a new region of that space that solves your problem. This is absolutely a kind of navigating problem solving that we really do not understand how this works, but this is an example of what I mean.

And I think that what evolution does is pivot some of the same mechanisms and some of the same tricks and some of the same problem-solving strategies through various spaces, including very early life had to solve metabolic problems, and then some physiological problems, and then gene expression, and then anatomical morphospace, and of course 3D behavioral space once brains and muscles showed up, and then linguistic space, and who knows what else.

So what evolution does is pivot organisms through these various spaces. And of course, even the distinction between these spaces is completely arbitrary relative to us as an observer.

So as for example, Chris Fields points out, if you are a bacterium sitting in a concentration gradient of some nutrient, you could move in three-dimensional space to try to go up the gradient, or you could move in transcriptional space, turn on some other gene that would let you instead metabolize some other kind of molecule that's in that space, and solve your problem that way.

So to the bacterium, it will not see a clean separation between three-dimensional space and transcriptional space. All of these things are effectors. All of these things are moves that it can make in its various spaces to solve problems.

Okay, so let's take a break here before we go on to the next section. And I'll pause, and if anybody has questions, I'm happy to talk about it.

Let's take maybe five minutes.

Yeah, if you do have any question, you can use the raise your hand function or just come in if I don't see you.

Yeah. So there are... No, please keep going.

Yeah, yeah, great questions. Thank you.

To me there are two different dimensions to this question. One is overall the concept of panpsychism, and I think if you would wait till the very end, and we'll talk about this after the third part.

But the other interesting part you said is about the driving of evolution. So I'll just say one thing about it here, which is that if you think about the kind of pre-Darwinian notion of what people associate this with, with Lamarck most of all, but this idea that evolutionary change is not random with respect to outcome, that basically by stretching his neck, the giraffe has children with longer necks.

The problem with that, up until 2018, I would have said the following. The reason that strong Lamarckism is impossible has nothing to do with Weismann's barrier or setting aside the germline. All of that is completely porous. People have shown that that doesn't stop anything. That's easy.

The hard part, and prior to 2018 I would have said impossible, is that the real problem is that genes don't encode specific characters. In other words, there's no gene for neck length. And so if you wanted a longer neck, you have to solve this inverse problem of which genes do you mutate to have a longer neck? It's an impossible inverse problem. So I would have said that's impossible.

However, looking at this barium experiment, what we see here is that the cells have basically done the hardest part. They have identified which genetic loci are relevant to the new problem they're facing. Once you've identified which those are, mutating them by some kind of CRISPR-type mechanism or some kind of genomic editing mechanism, that's trivial.

There are lots of organisms that edit their genomes in various ways, once you know which genes. And so now I have a slightly different view on it to say that I don't know how they do it, but we do have some evidence that cells can infer which genes are responsible for specific adaptive outcomes.

And if that's the case, I'm no longer sure that it's impossible that to some extent they're driving the evolutionary change. Because once again, this is a very controversial area. I have not remotely enough evidence yet to prove that that has implications on a global evolutionary trajectory scale, but I no longer think it's impossible because we can see cells solving this inverse problem.

Yeah, yeah. I mean, so I haven't shown you anything yet. There are some things coming that may relate to this, but I haven't shown you anything yet that suggests that the solution has to be any kind of exotic physics or anything like that, right?

Because for the whole next section, which we should start in a few minutes, what you'll see is that there's a tremendous amount of intelligent problem-solving that can be squeezed out of a perfectly conventional set of mechanisms. Not to say that there aren't others, but we haven't yet come to the point where we need to invoke anything more exotic. But, you know, we'll see.

Let's do one more, and then at the end I'll stay as long as people want afterwards, but let's move on to the second section.

Yes. Yes, you've anticipated a chunk of the third part of the talk. Stay tuned. That's coming. You are 100% on the money there. Yeah. Yeah, it's coming.

Okay. So, let me move on to the second part, and then we can come back to any of this.

So the second thing that I want to talk about is to really give you an example, a very concrete example of an unconventional collective intelligence, a problem-solving system that doesn't look like what we're used to thinking about as intelligent systems. And what we're going to use for that example is the distributed collective intelligence of cells as they undergo morphogenesis. And then specifically we'll talk about bioelectricity.

So let's... I like to start by noticing this interesting thing. So Alan Turing obviously needs no introduction, kind of the father of computer science and some other things. And he was interested in intelligence and intelligence in novel embodiments that are not natural, and computation and things like this.

And so he was interested in problem-solving machines, and in particular the notion of reprogrammability or intelligence through a kind of plasticity. And so he wrote these foundational papers in computer science, but interestingly he also wrote a paper on the chemical basis of morphogenesis.

So the question he was addressing is how can biological order, or at least in this case biochemical order, arise spontaneously? And so you might wonder why would somebody who's interested in computing and mathematics and machine intelligence, why would he be thinking about chemical morphogenesis and embryos and things like this?

And I think that, like with many things, this is just an example of Turing's genius. I think he saw very clearly that these are fundamentally the same problem. And I'm going to argue that that's true. Although he didn't write much, if anything, about this that I know of, but I think he was onto this idea that the formation of the mind and the formation of the body are fundamentally the same problem.

And so let me take you on a little tour of some of that. So let's ask this question. This is a cross-section through a human torso. Look at the amazing order here, right? So all the organs, the cells, the tissues, everything is in the right size, the right shape, next to the right thing, oriented in the right direction.

And it all comes from a collection of embryonic blastomeres during development. It's self-assembly. Where does this particular pattern come from? Now, the kind of immediate answer that most people give, well, it's in the DNA. It's in the genome.

And we're going to come back to this with a vengeance in the third part of the talk, but I want to point out already here that that is a kind of an untenable answer because we can read genomes now and we know what genomes encode. They encode proteins. They don't specifically encode three-dimensional shapes. They encode the tiny hardware that every cell gets to have.

And so there's still this massive question of using that hardware, how does the physiology of this cellular machine make this particular pattern and not some other pattern? And the thing is that the structure of the human body is no more directly in the genome than the structure of a termite nest or the shape of a spider web is in the genomes of the termites or the spider.

It's an emergent construction project of the collective intelligence of these cells. And so we can ask interesting questions. How do these cells know what to make and when to stop? And then for the third part of the talk, we'll talk about what else is possible.

As a regenerative medicine worker, we'd like to know if a part is damaged or missing, how do we get the cells to rebuild it? And as an engineer, we'd like to go one step further and say, okay, but can we get these cells to build something completely different, something that's not like this at all? And to think about, I like to sort of think all the way forward so that we know where we're going. Let's think

About what what is the end game of this field? You know, when can we all basically say we're done and go home? And the way I envision this is as something called an anatomical compiler. So, just imagine at some point in the future, you're going to be able to sit down in front of a computer and draw the plant or animal that you want. It could be something that already exists. It could be some organ that you need for transplantation. It could be some completely different thing, you know, some sort of six-legged frog with a little propeller on its head. It could be anything. The point is it could be absolutely anything. And if we had a mature science of this, if we knew what we were doing, this anatomical compiler would output a set of stimuli that would have to be given to cells to get them to build whatever you drew, like in this case this three-headed flatworm.

And so, the idea this is not a 3D printer. This is not something that meticulously micromanages the position of cells to print what you want. This is a communications device. What it does is it's a translator, really. It translates from your anatomical goals to the anatomical goals of a collection of cells, which are encoded in a different medium, which I'll show you momentarily. And by doing that, it basically resets the target morphology of these cells and gets them to build whatever it is that you want them to build. And I'll show you some examples of that, but of course, we don't have anything remotely this sophisticated. We can only do this in a very small number of cases.

Now, why is this important? Well, aside from the very fundamental aspect that we need this to understand how evolution works, we need it to understand our origins and so on. This is the answer to most medical problems. With the exception of infectious disease, if you could communicate your goals to a collection of cells, you would immediately have the solution to birth defects. You could regenerate after traumatic injury. You could reprogram cancer. You could suppress aging. You could solve degenerative disease. All of these things boil down to one problem. How do we tell a collection of cells what to build? If you knew that, all of these problems would go away. So, this is a major unmet need.

Why do we not have something like this? I mean, genetics and molecular biology have been not really hard for many decades now. Why do we not still have this? I want to show just a simple example. This is a baby axolotl. And axolotl larvae have little four legs. And this is a tadpole of a frog. And these guys do not have four legs. And so in my lab we can make something called a frogolotl which basically just means you take a bunch of cells from an early axolotl, you take a bunch of cells from the early frog, you put them together and you got a chimeric embryo that has some cells from here and some cells from there.

Now, we have both genomes. We've sequenced the axolotl genome and we've sequenced the frog genome. Can anybody tell me looking at those genomes whether frogolotls are going to have legs or not? And the answer is no. No one can tell you that. And that is because while we do have some good frameworks for understanding what a single cell is going to do biochemically by looking at its genome, we have almost no ability to look at a genome and guess what the collective is going to do. Either a pure collective like this. I mean you can't even predict the shape of a single animal from its genome either. But certainly not in these chimera cases.

What we do not understand yet despite all the biochemistry and genomics and everything else, what we don't understand yet is collective decision making. We don't understand how cells with specific hardware make decisions about large-scale shape. And so here's where we are today. We're very good at manipulating molecules and cells and making networks and so on. But we're really a long way from form and function.

And I'm going to argue that from controlling large-scale form and function. I'm going to argue that the reason for this and the thing that holds back really radically transformative regenerative medicine is that all of molecular medicine today is roughly like this. This is what programming looked like in the 40s and 50s. I mean you can see what she's doing. She's really literally re-wiring the machine by pulling wires in and out. This is hardware control.

Now the reason that you don't get out your soldering iron when it's time to go from Microsoft Word to Photoshop on your laptop is that what computer science has figured out is that there are higher level ways to interact with the system that does not require you to rewire the hardware. Meaning that you're taking advantage of the inherent plasticity, competencies, and various other capacities in the software of the system, in the activity of the system that you can guide with stimuli with a signal such as hitting a keyboard in the case of the computer. And that's what gave us the information technology revolution.

Biology to a large extent is still here because all of the exciting technology, so genome editing, CRISPR, protein engineering, single molecule approaches, all of it is down at the level of the hardware. All the excitement is at the level of hardware. We haven't even begun to scratch the surface of what the actual biological software can do. Specifically the intelligence of the various biological components.

Now what do I mean when I say intelligence? There are many definitions. I like this one. This is William James who gave us a strikingly prescient definition of intelligence. It's very good because it doesn't say anything about brains or what you're made of or anything like that. It's very cybernetic. It says this. Intelligence is some degree of the ability to reach the same goal by different means. That's it.

So in order to estimate the intelligence of some system whether it's engineered or biological or hybrid, it's an empirical question. You have to name a space that it works in. You have to say what you think the goal is. You have to make a claim about what you think the competencies of the system to reach that goal despite perturbations. That's the key is that you need to do interventions. You can't say how intelligent something is by simply observing it. You have to do perturbative experiments to see what goes on and what it can do to achieve its goal.

And in particular James gives a very nice example which is the spectrum between two magnets trying to get together and Romeo and Juliet trying to get together. And the difference is very simple. And there's many things in between, right? If you've got two magnets, they have some competency to try and meet each other. But if you put a piece of wood in their way, that's it. They're going to stand there pressed up against it. They have no ability to go around or do anything else. So that's sort of the minimal competency.

Romeo and Juliet have massive competencies to go around various physical and social barriers to get what they want. And in between there are all sorts of self-driving vehicles, worms, and cells, and you know, different kinds of creatures that have different degrees of competency to reach various goals.

So let's talk about this then. If that's the definition of intelligence, let's ask what kind of collective intelligence do these cellular swarms deploy? I've made this argument that the cells of the body making an embryo are a kind of collective intelligence. I'm not simply saying that they have a complex outcome. Feedforward emergence of complexity from simple rules is commonplace. Many systems do that. That's not what I mean. They have a specific ability to adapt to perturbations. That's more than simple emergence. And let me show you some examples.

So development is quite reliable and has the ability to reach the specific region of anatomical morphospace meaning the different configurations that our body might have to go from here to here. So you go from a single cell and then two cells to a whole body. But the amazing thing is that if you cut this embryo in half, you don't get two half bodies. You get two perfectly normal monozygotic twins. So this system is reliable but it is not hardwired. In the face of this massive perturbation, it is able to still get to where it's going.

So if you boil everything down to two dimensions just for sake of illustration, the normal embryo does this. But the embryo cut in half which is in a different location of that space actually is able to get to where it needs to go. In this ensemble of states, this is the goal space. The goal states in that space. This is what all normal sort of what are considered standard human anatomies are. And there are various other ways to get there and there are local maxima that can be avoided and so on. So being able to navigate to the same region from different starting conditions is a kind of intelligence that we view as navigation.

Here is an example in an adult organism. This is an axolotl. So this is a salamander. This guy regenerates his limbs, his eyes, jaws, portions of the heart and brain, spinal cord, ovaries, amazingly regenerative. What happens is that when a limb is amputated and they bite each other's legs off all the time when they're housed together. So this is a natural experiment that happens. Whether it's here or here or up here, the cells will very rapidly build what's needed and they make a standard salamander leg and then they stop.

That's the most amazing part of regeneration. First of all that it happens exactly the same way no matter where you start from. And second that it knows when to stop. When does it stop? It stops when the correct salamander arm has been completed. So you can see that what this is is an error minimization scheme. This is a system that is going to keep working and changing until a large-scale set point, a target morphology that looks like this has been reached.

Now by the way this is not just for frogs and worms. There are some examples. So for example the human liver is highly regenerative. Even the ancient Greeks knew that. I have no idea how they knew that but they did. Human children can regenerate their fingertips. So below a certain age they will have a cosmetically very very nice regeneration of a fingertip amputation.

And deer are a large adult mammal that regenerates a centimeter and a half of new bone per day when they're growing this. So bone, vasculature, innervation, skin, all of it. And in fact in the correct pattern. So this ability to regenerate is sprinkled throughout the tree of life. It is really fundamental. And I want to show you one of my maybe all-time examples of creative problem solving in this kind of space.

So in that salamander that I showed you in the newt, if you take a section through a little tubule that goes to the kidney, you see that it's like eight or 10 cells typically that work together to make this lumen. One trick you can do is you can make the cells large. You can force the cells to be larger. And if you do that, what happens is it adjusts. Fewer of those cells are there to make roughly the same size of lumen. That's already amazing that it adjusts the number of cells to the size of each cell.

And that, you know, to being a newt born into this world, the embryo cannot count on having the right size of cells. It has to adapt on the fly. This is one of many things you can't count on as an embryo. You have to solve these problems from scratch every time. So what it'll do is it'll use the right number of cells. But if you make the cells truly gigantic, one single cell will bend around itself and leave a space in the middle to give you that tubule.

What's amazing about this is that this is a completely different molecular mechanism than this. In other words, this is some sort of cell-to-cell communication tubulogenesis kind of a thing. This is cytoskeletal bending. This is a single cell bending around itself. So look what's going on here. Different molecular mechanisms are being called up in the service of a high-level anatomical goal. This is what's known as top-down causation. In the service of a particular anatomical goal, different molecular mechanisms can be called up.

So this ability to adjust to circumstance has to do with external circumstances like being injured or being cut in half or losing a limb, but also internal circumstances. Not only can't you count on your body maintaining a particular shape, you can't even count on your own parts. You have no idea what size your cells are going to be when you come into the world and you still have to be able to solve this problem. So this is amazing and again not through eons of evolutionary change. This happens immediately in one generation.

So, the final example I want to show you is something that we discovered a few years ago. Here's a normal tadpole. Here are the eyes, the brain, the nostrils, and the mouth is here. So, these tadpoles have to become a frog. In order to become a frog, they have to rearrange their face. So, the jaws have to move forward, the eyes have to come in, various things like that. And it was thought that this is a hardwired process. Basically, every organ in the face moves in a certain direction a certain amount. And that way, if you just know which direction and how much you'll go from a standard tadpole to a standard frog.

We made the hypothesis that this was actually a more intelligent process than that. And so, how do you test it? With perturbation. So, what we did was we created what we call Picasso tadpoles. So, the eye is off to the side, the jaws are on top of the head, everything is kind of scrambled. And what we found is that even those animals generally make pretty normal frogs because all of these organs move in novel, unnatural paths to get to where they're going, and then they stop.

Actually, they sometimes overshoot. They sometimes go too far and then have to come back. But all of these things move around until they get to a pretty normal frog face. So, what this is is this idea that genetics isn't specifying some kind of hardwired rearrangements. What it actually specifies is a problem-solving machine, and we'll get to more of this in the third part of the talk, that can recognize unexpected states and take corrective action. It's an error minimization scheme.

So, how might that work? And here's what we have. The standard story of developmental biology looks like this. There are genes that interact with each other as gene regulatory networks. They make proteins which do things. Some of them are sticky, some of them diffuse, some of them exert force, and so on. And then there's this process of emergence. So, simple rules lead to complex outcomes, and you get this kind of beautiful anatomy. And this is all true. This all happens. But this leaves out a really critical component.

Which this is all feed-forward emergence. But there are all these feedback loops that if you step away from this target morphology, either by injury or by teratogens or by mutations or anything else, all kinds of processes at the level of physics and genetics get activated to try and get you back there. There's that error minimization scheme, which is really a kind of anatomical homeostasis or pattern homeostasis. And a couple of interesting things about this.

On the one hand, biologists know all about feedback loops. It's pretty obvious. But on the other hand, most of the feedback loops we know about are scalar. All of these homeostatic loops require a set point, right? They need to know like a thermostat needs to know what is the thing with respect to which we calculate error. Most of these examples of homeostasis in biology are scalar numbers. They're a pH or temperature level or hunger level or something like that. They're a simple quantity.

In this case, we're talking about the set point of this process being a complex anatomical descriptor. Doesn't have to be an exact cell-by-cell kind of thing, but some general notion of what the correct shape should be or the correct length of limbs and stuff like that. So, that's unusual. And of course, the whole notion of a process that's in some sense goal-directed is a kind of taboo. There's a real what I call teleophobia out there about people are really suspicious about models of processes with goals.

But I think that cybernetics since the '40s has given us a mature non-magical science of systems with goals. So, I think it's okay. But something interesting about this is that if you only know about the first part of this, then when you want to make changes up here, let's say for biomedical purposes, your only game plan is to alter the genome. And that means you have to solve this inverse problem.

You have to ask, "If I wanted a different shape up here, which genes would I have to change?" And as we talked about a few minutes ago, that is in general an unsolvable problem because this process is so complex and so convoluted. You can't just reverse it. You don't know which genes to change. But if there is kind of this homeostasis relative to a set point, then you've got a different strategy.

Maybe this isn't at the mechanical clock part of that continuum that I showed you. Maybe this is closer to that thermostat level where you can actually change the set point. You can revise the set point without having to change the hardware, without changing the genes at all. And this is what I'm going to show you now. So, what we wanted was, and this is years of work on this now, what we wanted was to, if this is all true, we should be able to do a couple things.

First of all, we should be able to identify where that set point is stored, and I'm going to show you that, decode it, and I'm going to show you a couple of examples of how we can actually read it and understand what it says, and we should be able to rewrite it and let the cells build to the new pattern without changing out the hardware. Let me show you that.

So, how might this work? At the very beginning of this, I started thinking about, "Okay, how can we even talk about cells and tissues storing a goal state, storing some kind of a target for all this activity?" And of course, the one really good, uncontroversial example of a biological system that stores goal states and acts towards them is the brain. This is what complex creatures do all the time.

So, the brain stores memories of target states, and it acts in three-dimensional space to achieve those goals. How does it do it? Well, the hardware looks like this. This is an electrical network where you've got cells. The cells have little ion channels in their membrane, which allows them to pass things like potassium, sodium, and so on. By doing that, they acquire a voltage gradient, a voltage difference between inside and out.

And that voltage gradient may or may not be communicated to the neighboring cells through these little electric synapses known as gap junctions. So, you have a network of cells that are electrically connected, so it can store different states electrically, and they can propagate information, and that information spreads through the network. And that is what runs the software of the brain.

And so, what you can see here is this group made this amazing video of real-time electrical signaling in a living zebrafish brain. So, you can see here's what's going on. As the fish is thinking about whatever fish think about, and it is a commitment of neuroscience that if we were able to read out these electrical signals, we could decode them and thus know what the fish is thinking.

We should be able to recover the memories, the goals, the various preferences, the behavioral repertoires, and so on by decoding this. So, to summarize, this is the project of neural decoding here in neuroscience where you try to image the brain, and you make a model, and you understand what that animal is thinking or seeing, and so on.

And so, the idea is that according to neuroscience, the bioelectrical signaling in the brain is literally the medium that contains the cognitive content of the mind. So, you as a person, all your memories and thoughts and everything else is in some way encoded in the electrophysiology of your brain. Well, the amazing thing is that this ancient... So, that's a really neat trick, right? That combination of hardware and software is amazing.

But what's really interesting is that that architecture, the architecture of having this electrical hardware and thus this very plastic, goal-seeking, clever in problem-solving software is absolutely ancient. So, every cell in your body has ion channels. Most of them have gap junctions to neighboring cells. This is not something that's just for neural cells, and it's something that's been around since the time of bacterial biofilms.

This is evolution figured out this architecture very early on. And so then, one can try to extend this. And basically, this is the project and this is what we've been doing for almost 30 years now is this idea of extending neuroscience beyond neurons. The real kind of magic of neuroscience is not in the neurons. It's in this understanding, this scaling up of cognition.

And so, what you might imagine is doing exactly the same kind of neural decoding thing, but not on neurons. Let's say on embryos or some other morphogenetic system to ask. Here, you're asking, "What is the collective intelligence of neurons thinking?" Here, you're asking for what is the collective intelligence of the rest of the body cells thinking. Quite literally.

And I can use... The thing is, you know, this sounds very strange, and a lot of people will say that this is some kind of weird misuse of a metaphor. But the reason that I can apply all this terminology from neuroscience to outside the brain is because the concepts, the tools, the reagents cannot tell the difference between neurons and non-neural tissues. All of the same tools and concepts work everywhere in the body. This is what we've been showing. And so, the concepts are actually very isomorphic.

And so, what you have here is that whereas the electrical network, just to kind of summarize what I'm saying here, the electrical network of the brain is giving commands to the muscles to move you through three-dimensional space. That's the familiar version of neuroscience. And then we can modify it with various techniques, and we can read it out, and so on.

So, the electrical network of the brain tells muscles what to do to move you in three-dimensional space. But the reason it can do that is because it evolved from a much more ancient system in other tissues in your body that long before brains and muscles evolved were telling all the cells how to rearrange your body shape and thus move you in anatomical morphospace. Same strategy, same hardware, same tricks, except it's a different space.

And so this is what I said before when I said that evolution was pivoting us through various problem spaces. So what we did was to develop the first molecular tools to read and write this information outside of the nervous system. So how do we do that? So the first thing is a voltage-sensitive fluorescent dye. So this is a chemical that emits different brightness or different wavelengths of light depending on the electrical properties of the cell that it's sitting in.

And so what you're seeing here is a time lapse of a frog embryo during the early stages. And so the colors represent different voltages. And so this is a kind of brain scan, if you will, of all the conversations that the cells are having with each other as they sort out who's going to be left, who's going to be right, who's going to be anterior, posterior, dorsal, ventral, and so on.

And so we now have the ability to read out this kind of information. We spent a lot of time doing computer simulations to figure out how these patterns map onto the patterns of ion channels that are present in the cell membranes. I want to show you two kinds of patterns. The first we call the electric face. And this is again a time lapse. Now here it's grayscale instead of color, but it's the same idea. The brightness represents voltage.

And so in this early frog embryo that's putting its face together, there's all kinds of changes, lots of exciting stuff happening. But look at this one frame taken from it. Long before the genes are turned on to make the different parts of the face, we can already see what this animal's going to do. Here's where the animal's right eye is going to be. Here's where the mouth is going to be. The left eye comes in slightly thereafter. You will see that momentarily.

Here are the placodes on the sides. The pattern of where everything is going to go is already there in the stable electrical states of the cells if you know how to read them out. So this is the electric face. I'm going to show you in a minute that if you make changes here, you will change gene expression, you will change anatomy. Therefore, this pattern is instructive.

It drives. This is not just some sort of epiphenomenon. This is actually the memory of what the cells are going to be doing to make the face. So that's a natural pattern. But here's a pathological pattern. We'll talk more about cancer in the third part of the talk. If I inject a human oncogene, there will be a tumor. And but before that tumor forms, I can already see where the cells are starting to defect.

They acquire this abnormal bioelectric state. And I can already see where the tumor is going to be. So this is a pathological pattern. So in addition to being able to observe all this stuff, which is nice, but it's just observation, what we developed were a set of tools patterned on exactly the same thing done in neuroscience to study synaptic plasticity and intrinsic plasticity.

And basically be able to now, so I showed you how to read the patterns. I'm showing you how to write the patterns. So you can rewrite those patterns by manipulating the gap junctions, these electrical synapses. So that controls the connectivity of this network. Or you can actually open channels, open them using drugs or close them using drugs, using light, the various kinds of stimuli to control the voltage of every cell.

So to be clear, there are no applied fields, there are no electrodes, there's no electromagnetic radiation, there are no magnets, nothing like that. This is molecular physiology using the natural interface that these cells are exposing to us. This is the same interface that they use to hack each other. This is the way the cells are communicating and we're just hijacking it. So there are no external fields, nothing like that.

Okay, so using this technology then, we know what happens when you use it in the brain. So somebody like Tonegawa at MIT can use this to put false memories into the brains of mice using exactly this technology in the brain. We are going to put false memories into the somatic networks of cells that are building the body. So when you do that, here's what happens.

You can convince cells to make ectopic hearts like this. Here are some extra hearts in a tadpole. You can make extra limbs. So you can ask the cells in the jaw to start generating a limb. And here are some six-legged frogs here. You can make ectopic forebrain. You can make ectopic inner ear organs right here. Or you can even make fins. Now that's especially weird because tadpoles aren't supposed to have fins. And I'll show you some of that momentarily.

But here's one of my favorite examples. I showed you, if you remember that electric face, one of the things about that electric face was that there was a pattern that said where the eye should go. It represented the location where the cells should become an eye. And so we simply said, okay, if we try to reproduce that electrical state elsewhere in the body, can we induce some other kinds of cells to be an eye?

And so we did. So we injected some mRNA encoding some potassium channels into a different type of cells that were not going to be that region. In particular they were slated to be gut. And if you do that, if you create that pattern in those cells, sure enough they will build an eye. And these eyes will have all the right cell layers. So retina, lens, optic nerve, they'll have all that stuff.

Notice a couple of interesting things. Number one, this bioelectric pattern is instructive. It tells cells what to do. Doesn't just ruin the old pattern, it actually can call up whole organs. Number two, it expands the notion of competency. In the developmental biology textbook, what you will read is that only these cells up here in the anterior neuroectoderm are actually competent to become eye.

And that's true if you use the traditional way of inducing eyes, which is this master eye gene called Pax6. But it's not true if you go up a level and use the bioelectric signals, you can make eyes anywhere. So it completely changes our idea. Learning to hack these cells in a different way changes our understanding of their competency.

And notice the third thing that's amazing is that we didn't provide very much information at all. I mean eyes are very complex. We didn't tell this animal how to make an eye. We basically just triggered the eye morphogenetic subroutine with a simple trigger that says build an eye here. That's it. We didn't say where the stem cells go or how to make a retina, none of that. It's very modular. It's very hierarchical.

And so the next kind of part of this that's my favorite is this. Here's a lens, an ectopic lens sitting out in the tail of a tadpole somewhere. So we've induced the lens. The way we've induced it is just like here we injected some ion channels. These ion channels are labeled with this blue beta-gal signal. So what you can see here is that the blue cells are the ones to which we said build an eye.

But there's not enough of them to build a proper lens. And so what did they do? They recruited a bunch of their nearby buddies, all of this stuff here that we never touched. Right? So it's a secondary instruction. We instruct these cells to make an eye. They instruct their neighbors to work with them to complete the task.

This is something you see in other collective intelligences like ants where when a few ants come across a project that's too big for them to handle, what do they do? They recruit others. And so this ability of the system to self-scale to the available problem is something that you see in other collective intelligences.

So we don't have to tell these cells how big an eye should be or how many cells there should be. All of this is part of the competency of this amazing agential medium in which we're working. We'll get to that notion. But the fact that these cells are solving problems collectively. They know how big this group is. They know how big it's supposed to be. The collective functionally knows all these things because you can see it.

So we use the same kind of strategy to address the issue of leg regeneration. So frogs normally do not regenerate their legs unlike salamanders. But what we can do is after amputation, we can treat with a cocktail of ion channel drugs. And in fact they turn on immediately these pro-regenerative genes. They begin to make a leg. You've got some toes. By 45 days you've already got a toenail and some toes. Eventually very nice leg.

The leg is touch sensitive and motile. And all of this happens from a brief modulation. In fact in the recent work on adult frogs, we show that just one day, 24 hours of application of our cocktail, you get a year and a half of leg growth. So this is not 3D printing. This is not scaffolds. This is not us telling stem cells where to go or what tissues go next to what tissues.

We have no idea how to build a frog leg. It'll be a really long time before anybody can micromanage that process. But you don't need to because the cells already know how to do it. What you do need to do is to convince them that that's what they should do instead of scarring. And that doesn't take very long. 24 hours is plenty.

And once they've decided that that's where they're going in morphogenetic space, they will keep going. And this is how regenerative medicine can take advantage of the competencies of these cells. And here I should do a disclosure because David Kaplan and I are co-founders of a company called Morphaceuticals Inc. And what we're trying to do is the same thing in rodents.

So David makes these amazing wearable bioreactors that apply to the wound site our drug payloads, which are various things that attempt to convince the cells that they should regenerate. So hopefully eventually we'll be using this in clinical practice. For now, we've done it in frog. We're currently trying to do it in mammals and eventually hopefully in patients.

Okay, for the next story I want to switch back to planaria. So this kind of animal. And I want to at this point make good on the promise that I made about 10 minutes ago, which is to show you how those goal states are actually encoded. How does the tissue remember what a correct animal should look like? And the planaria are a great example of this because if you cut them into pieces, and the record is something like 276 I believe, each of these pieces regrows a perfect little worm.

They make exactly what's needed, no more, no less. And in fact, this is a tough problem because even if you make just one cut here, these cells back here have to make a tail. These cells up here have to make a head, but they were exact neighbors right up until the time you cut them, they were exact neighbors, and yet they have radically different anatomical fates.

So, you cannot decide if you're going to be a head or a tail based on your position in the animal. They have the same position. You can cut anywhere. It's after the cut you have to now figure out what am I missing, what do I still have, and what does a correct planarian look like? So, oh, and by the way, these animals are also immortal in case you're interested in aging. Planaria, they do not age. There's no such thing as an old planarian because they continuously regenerate their tissues.

So, we wanted to ask this simple question. Okay, here's our planarian, and we amputate the head, and we amputate the tail, so you have this middle fragment. The middle fragment reliably regenerates to have one head, one tail. How does this piece know how many heads and tails it's supposed to have? So, we investigated and found out that there's actually an electrical circuit, which I don't have time to go into, but there's an electrical circuit which actually encodes the number of heads that they're supposed to have.

And so, what we were able to do then is this. Here is a visualization of the normal electrical pattern in these fragments, so one head, one tail. But what you can do is you can rewrite that pattern to say two heads. Now, this is kind of messy still, so the technology is still very much in development, but you can already see you can do that. And when you do that to this animal, they go ahead and they make a two-headed flatworm. This is not Photoshop, these are real two-headed animals.

Notice something interesting. So here's anatomically one head, one tail. The gene expression for the head is of course in the head, and so on. This animal, this bioelectrical map is not a map of this two-headed guy. This is a map of this perfectly normal, anatomically normal, one-headed animal. Not only is his anatomy still normal, his molecular biology is still normal.

The anterior marker is still only on in the head, not in the tail. And yet if you were to cut them, they go on to make two-headed flatworms. So, let me pause here and just explain to underline what's going on here. This bioelectrical pattern is what I promised before as a readout of where the information for the target morphology, the target, the correct pattern to build, this is where it's stored.

It is literally stored in this electrical map the way that memories are thought to be stored in the brain. Also, a single normal planarian body is able to store one of two different ideas about what a correct planarian should look like. In other words, this again sounds very much like the nervous system where the exact same hardware, your brain, can store many many different types of ideas.

And you do not have to genetically edit your brain to have it store different types of information. It's just built to have different targets that it can remember. And the third amazing thing about this is that this pattern memory is a kind of counterfactual, meaning that this is a memory of what I am going to do if I get injured in the future.

It's not what I'm doing now, it's what I will do if at some point in the future I get injured. So, that ability to... it's a very primitive version of what's sort of magical about brains,

which is this ability to mental time travel, to think about, to have memories of things that happened long ago and are not true now, and to have anticipation of things that might happen in the future that are also not happening now. People think that this is very much the kind of thing that brains do, and it is, but the evolutionary precursors of that are ancient, and even the planarian morphological intelligence has a degree of ability to do this kind of mental time travel.

Now, why do I keep calling it a memory? Because if you take these two-headed worms and you cut them again in plain water, no more manipulation of any kind, they continue to form two-headed animals. So, no genetic change, the hardware is untouched. These genomes here are exactly the same. We haven't done any genomic editing, nothing wrong with their genomes. And yet, this is a permanent line of animals that will continue to generate two-headed creatures even though they have the normal wild-type genome. So, the question of where is... Remember that slide I showed at the beginning, that human torso, and I asked where is that information stored? Where is the information stored for the number of heads a worm should have? It isn't genetic because you can take the exact same genetic hardware and have it be two-headed. But what the genome does do is to build an electrical circuit that by default stores the information for one head, but it can be pushed to other fates, for example, to store information of two heads. So, this is why I call it memory. It's long-term stable, it's rewriteable, and we can also, by the way, take two-headed and put them back to one-headed. And it has this latency, which I just showed you, this conditional recall. It doesn't kick in until you've injured the animal. And so, here you can see this is a video of these two-headed critters doing their thing.

Okay. And so what we're doing now is lots of computational modeling, and actually using lots of ideas from connectionist machine learning to try and understand specifically how the electrical circuits that are in these planaria can store different types of target patterns and reconstruct those patterns even when the information has been damaged. Now, I've shown you this story about two heads and so on. But it's not just about the number of heads. One other amazing thing you can do is push them to make heads of different species. So, here is a planarian with a nice triangular head appropriate to his species. We can cut off that head, perturb the electrical network topology, let them settle back down, and when they regenerate their heads, sometimes they make the normal head, but sometimes they make round heads like this S. mediterranea. Sometimes they make flat heads like this P. felina. Again, no genetic change here. This has nothing to do with rewiring the hardware. What it is is pushing that same machine into other attractors in that enormous space of morphogenetic configurations that they're able to traverse. They're normally very good at going exactly where they need to go, but they can in fact visit other regions of that state space that normally belong to other species. And not just the head shape, but also the shape of the brain, the distribution of stem cells, just like these other species 150 to 100 million years distance. So again, this notion of reprogrammable biological hardware. They can do other things beyond the default that they normally do. And by the way, they can go even further into the latent morphospace of a typical planarian. And D'Arcy Thompson, as early as 1917, he had this notion of living beings exploring various kinds of mathematical morphospace. You can make them go further than a typical planarian, and you can make these things that are sort of, they're not flat anymore, they're round and cylindrical, and they're sort of hybrid shapes, and these crazy spiky things. These wild-type cells are capable of building lots of different stuff.

And so, what we're doing now is really trying to make these full-stack computational models that go all the way from the lowest-level molecular details of which genes are turned on and off, and then build models of the bioelectric states, the cellular and tissue-wide bioelectric states that result from these ion channels, and eventually all the way up to body-wide decisions and algorithms, which is what you need for that anatomical compiler that I talked about at the beginning, which is to understand how the tissue is actually making decisions about large-scale shape, number of heads, type of heads, and so on.

Okay, so just a couple more minutes, and then we'll take another break. Just the last thing I want to show you here is a biomedical intervention that emerges from this kind of approach. And it has to do with fixing frog brains. So, here's a normal tadpole brain. Here's the forebrain, here's the midbrain, here's the hindbrain. This is the normal structure. And if you expose these guys to various teratogens like nicotine, what you'll see is alcohol, nicotine, lots of different things. What you'll see is a real defect in brain formation. You see the eyes are connected to the brain. The forebrain is basically missing. Lots of problems here. And so, we wanted to know, well, how would you fix something like this? It's very complex. We don't want to try to tell every cell what to do. How do we fix this? And so, Alexis Pietak, who's a member of our Allen Discovery Center, made this computational model. And with Vaibhav Pai, who's a staff scientist in the group, they simply asked the question, what underlies the bioelectric gradients that normally tell the cells how big and what shape a normal brain should be. And so, when you have a model like this, you can ask it questions, and you can say, well, if the pattern is incorrect, what would we have to do to make it correct? And so, the model gave us an answer. It said, well, there's this particular channel called HCN2, and if you were to open that up, the pattern would be restored and the cells would know what to do. And so here's an example, and this is actually a much more difficult case than that teratogen case I showed you. This is a notch mutation. So notch is a really critical neurogenesis gene. In the absence of notch, you can see the brain is, the forebrain is gone, the midbrain and hindbrain are a big bubble full of water. I mean, these animals are profoundly affected. But even though that notch mutation is there, if you force the cells to have the correct bioelectrical pattern by using the strategy given to us by the computational model, the brain becomes normal. So they get their shape of the brain back, they get gene expression back, and they get their IQs back. Their learning rates are indistinguishable from controls.

So this is an amazing example of not only instructively repairing complex defects using this bioelectrical approach, but specifically fixing a hardware defect in software. And I'm not saying that's going to work for all genetic defects. There are many things that bioelectrics are not going to work for. But it's again this idea of plasticity. The idea is that broken hardware is not the end of the line, that if the algorithm is good, which for example in planaria it's absolutely amazing, if the algorithm is good, you can often get around these defects and still get to where you need to go if you can remember what the correct pattern is. And so the biomedical roadmap looks roughly like this. We're making this sort of turn-key system where you can sort of play with parts of it that are already online, where you know what the incorrect pattern currently is, you should know what the correct pattern is, and this is something that we're still working on to acquire these physiomic data sets. And then the model can actually tell you which channels and pumps to trigger, so how to communicate your anatomical goals to that set of cells. And then of course, once you know which channels, then it's pretty easy to pick drugs that you already know. So looking forward, the impact on biomedicine is to complement all of these current technologies. So surgery, transplants, stem cells, genomic editing, rewriting gene regulatory networks, molecular drugs that target proteins and so on. All of these things are bottom-up. They're micromanagement. They're targeting the pathways to exactly do what you need to do. You can complement that by taking advantage of the tissue intelligence. And so we've talked about, in this part of the talk, we've talked about many examples of cells and tissues solving problems. They can find new ways to deal with physiological challenges, with anatomical challenges, and with gene expression challenges. So they have various competencies. They can recruit neighbors as needed. They can take different paths to the same outcome. And if we can...

sort of take advantage of all those subroutines through either cell trading or tissue trading or various kinds of other triggers, which we call morphoceuticals and a subset of electroceuticals, then I think real, and you can find all the details here, really transformative types of regenerative medicine are going to be available. And I think that overall, and possibly surprising nobody who works in these sorts of traditional alternative medicine modalities, the future of this approach is going to look a lot more like a kind of somatic psychiatry than it is like chemistry, because bioelectricity is that interface layer that enables top-down controls, in other words, memories of global patterns in various spaces, to filter down to the molecular machinery that actually activates them. And so that's something that's really quite unique to this bioelectrical layer. And it's not just some rare thing that works in the case of hypnosis, hypnodermatology and mind-body medicine and weird altered states and biofeedback. It works 24/7 in your body. When you decide in the morning you wake up and you decide that you're going to work, that high-level cognitive intent of what you want to do has to depolarize your muscle cells for you to be able to get out of bed. Right? So voluntary behavior is already showing you this mind-body connection, this idea that these high-level executive cognitive goals have to eventually trickle down to control molecular events in your muscles and in your glands. So this is the way our bodies work as a multi-scale system all the way from high-level cognitive structures down to the molecular pathways, and bioelectricity is the layer that does that. So I'm going to call a break now. Let's take a couple minutes. Everybody get some water. I'm going to be back in about 3 minutes, and then people can ask questions. But take a break for a few here. All right. I'm back. If anybody wants to start asking questions, yes, I can see Matthew's got his hand up already.

Yeah, so let's say this. I think that the right way to think about the relationship between the experiments and all of that and these different philosophical takes is as a set of... and my framework, you can in fact see it here. It's called the TAME, Technological Approach to Mind Everywhere. And the reason that I emphasize technological is that I want the philosophical concepts to be testable and to do specific work. In other words, what I am looking for is for different lenses or different ways of looking at these phenomena that facilitate progress. So a philosophical worldview is, to me, useful to the extent that it can be applied to the things we do to drive new discoveries, new capabilities, to drive a research program, basically. So the reality is I've borrowed in order to make all of these kind of new findings that my lab has found, these were done because we borrowed from all kinds of unusual ways of thinking about the world that are not standard in the field. And so to that extent, I do think we absolutely interface with interesting different philosophical views, be they Bergson or whoever else. But that's what I'm looking for. So for example on the specific example that you mentioned, if you say that memory is sort of everywhere, that's fine. I suspect that has a lot of value to it. The next question is, okay, how do we turn that into a specific intervention that shows us how to be able to do something that you could not do if you didn't have that worldview. So that's where the rubber has to meet the road in terms of facilitating experimental advances and finding new research. So I think that's a longer... we can have a whole other discussion about where I've borrowed a lot of this stuff from because it is interesting, but everything to me, everything has to eventually interface strongly with the research enterprise.

Let's keep going and then I'll do the last part and then we can talk about whatever people want at the end. Okay. So in this final part I want to come back to... so let's remember where we were. We started out acknowledging this kind of multi-scale architecture of the body and trying to widen this idea of intelligence towards a continuum of competency from very simple kinds of things to much more complex things in diverse problem spaces. And then we went through some very specific examples of bioelectricity as a kind of cognitive glue that holds together these competent components, for example cells, and allows the collective to solve problems in morphogenetic space and the implications for biomedicine. I want to circle back to a somewhat more philosophical aspect to really start to dig into what does it mean to be an agent or a self? And how does it work that components are able to make these emergent selves? And then I want to show you some novel organisms as a kind of tool with which to explore this latent space that I've been hinting at. So here we go. Remember back to the original kind of goal for this framework was to be able to think about very diverse kinds of agents all in the same space. So how do we think about synthetic beings and possible aliens and cells and tissues and humans and birds and everything else kind of using the same framework. And so what I came up with was this notion of a cognitive light cone. So cognitive light cone, which is shown here, and this is kind of a... if anybody's into physics you'll recognize this is a kind of upside down Minkowski cone where all three dimensions of space are collapsed onto this axis and time is on this axis. And what the cognitive light cone is, is simply the size of the largest goal that a given system can work towards. So not how far away can it sense or how far away can it act, but the size of the largest goal state that it can pursue. So if all you care about is the local concentration of sucrose, then maybe you're a bacterium and you've got a tiny little cognitive light cone that has a tiny bit of memory from what the concentration was 20 minutes ago, a little bit of predictive capacity, and that's it. If you are something like a dog, well, you have a lot more memory going backwards, more predictive power going forwards. You have the ability to have the goal of, let's say, protecting the whole sort of property. So you have some idea of a yard, the size of a yard, and what's going on. But you're never going to care about what happens three towns over a month from now. It's just that that's too far in the future and also in space for that cognitive system, as far as we know, to have any of that as a goal. If you're human, you could have an absolutely enormous light cone where you might be working towards world peace over the next 100 years and you could have these very large light cones. And by the way, you have the added psychological pressure of knowing for a fact that some of your goals are not achievable in your lifetime because your cognitive light cone is bigger than your actual survival, your possible survival. So if you're a goldfish, you don't have that problem. All your goals are achievable in your lifetime because they're all very kind of short path. And of course, as maybe it was Beth or somebody had pointed out, we are of course compound. We are made of components with their own little cognitive light cones in their own spaces. So we are this huge patchwork of compound intelligences, all of which see different worlds in different spaces and are all cooperating and competing and so on, both across and within levels. So with this notion, let's think about a few things. First of all, let's go back to this idea of a collective intelligence. So if somebody shows you a human brain, if you didn't already know what a human looks like, and maybe even if you did because it's not an obvious question, and somebody asks you how many individuals are in there? How many selves are in there? And it leads to the fact that we do not know actually how much brain matter it takes to support a single human self. I mean, we see the normal case, but this leads to people saying things like, "Wow, for a bee or something like that, with such a tiny brain it does all..." Well, we don't know how much it takes to be a single self. We don't have any way of guessing what that's going to be until you already know what a human is. So let's look at this example. This is a schematic of an early blastoderm of an amniote embryo, let's say it's a chicken, a bird, or a human. And what you see is a bunch of cells. There's 50,000 cells. And we look at that and we say, "Oh, there's an embryo." Now, what are we counting when we say there's one embryo there? What is there one of? There are tens of thousands of cells. What is there one of? What we're actually counting is alignment. We're counting the fact that all of these... and that alignment is literal because the cells physically have to align with each other and also alignment in morphogenetic space because all of these cells normally are going to work together to build one individual. They have one particular large-scale pattern that they're all going to build and that is what we're counting as one. And now, could you guess in advance how many embryos there are or how many individuals are in this brain? It's a very similar, as I was saying before with the allusion to Turing, it's a very similar problem. Can you tell how many are going to be in there? Is it genetically specified how many are in there? Well, the answer is it is not. So here's an experiment that Lutz first did in the 40s and I used to do as a grad student in the duck embryo. So what you can do is you can take a little needle and you can put some scratches in this blastoderm. And so what happens is that once you've made those little scratches, until they heal up, which they will, but until they heal up, every region which does not feel the presence of these other regions becomes an embryo. It self-organizes an embryo. And then it heals and then what you have is multiple areas that are conjoined twins or triplets. And then in between them you have these little weird disputed zones where these cells can't quite tell who they belong to, but in the end they have to make a decision and they become part of this or part of that. And this is really... there are a couple of interesting things here. First of all, because of this, because you can split this up into some number of islands, the number of selves, I just imagine this is a human embryo, the number of individuals in there is not fixed. It is not determined by the genetics. It is set in real time and by the physiology. It could be zero, it could be one, it could be two, it could be some number of embryos, probably up to half a dozen, let's say. And it's very important for the making of selves that they are self-organized, meaning that they don't know, each cell does not know where it... every cell is some other cell's external environment. So where do I end and my neighbor begins, or where do I end and the outside world begins, right? This is something that is not true for almost any current AI or robotic system because they don't have to put themselves together from scratch. They are told right from the beginning. They're programmed to say, this is your border. Here are the inputs, here are the outputs. You are here, the outside world is there. That is not the case for living beings. We have to self-construct and we do not know right away where our borders are, how many of us are in there, and so on. This also, taking it back to the brain of course, is this issue of individuation because in split brain patients and dissociative disorders and so on, you find out that there actually are more than one self within a standard human brain. How many could fit in there we do not know, but the record for these dissociative kind of cases, I mean, I think I've read in the 60s that there could be 60 different personalities in there. So this idea of an active medium, an excitable medium that differentiates into a number of coherent selves is really key. And so now you have to start asking yourself, okay, so as we are self-constructing ourselves out of this Freudian ocean of possibility, out of 20 or 50,000 cells, three or two or one embryo will form. What are the dynamics that actually allow individual cells with their own tiny little goals? What are the goals of cells? Well, every cell wants to become two cells and they have metabolic goals that they try to meet and so on. What allows them to scale up to work towards very large complex goals like building a whole embryo? I mean, that's huge. No individual cell can know what a whole embryo is, but the collective does. So what allows that? Well, here are some hypotheses. One thing that allows it, I think, is the kind of memory wiping that happens when cells are electrically connected because when two cells signal each other the old-fashioned way where they just sort of secrete a chemical molecule and it comes over here and this cell senses it, it's very easy for this cell to know whether that signal came from the outside or not. So these two cells can maintain their individuality and they can maintain different memories and the cell can ignore the signal or do something about it, whatever. But there's something magic about these electrical gap junctions. When the cells are joined by these gap junctions, this signal, let's say it's a calcium spike from being injured. This cell got poked with something and this calcium spike is a trace of that event, a kind of memory engram of that event. It will immediately propagate through these gap junctions and the interpretation of that signal for this cell is exactly the same as it would be for its own calcium spike. Calcium doesn't have any metadata on it that says, "Where did you come from?" So by moving through this, there's a kind of memory, kind of identity wipe that happens where this cell now has that same memory trace and as far as it's concerned, that's its own natural memory. It has no way of knowing that's a false memory. So to this cell, that's a false memory, but to the collective, it is an absolutely true memory that the collective had. So that kind of wiping property, it strips ownership information. It makes it very hard for each cell to maintain an individual identity with respect to what its history was and it makes collectives out of individual cells and of course it makes them bigger. So my memory becomes our memories, my stress state becomes our stress and my goal becomes our goal. And so when a single cell in this diagram, when a single cell has some size of a cognitive light cone and some ability to anticipate and remember, a network of these cells has a bigger computational capacity and a bigger spatial sensorium. And it can actually have larger scale, more complex states as set points. And now you can start to see how a collection of cells could actually work on something like building an arm or building a whole body. And so we can imagine, and so we're doing this now, is doing these very specific computational models of how little tiny homeostatic loops like this, for example, measure the pH and act, become very large homeostatic loops about something completely different. Single cells cannot remember anything like this, but a giant collection of cells could. And so this is a very active area of research is to figure out exactly how that scales up. The implications are these. Here's what evolution has done. Evolution has enabled single cells to work together to enlarge their cognitive light cones where the goals scale from the very humble goals of single cells to the more grandiose goals of collectives where they can work on something gigantic like making sure you have four fingers. This cell doesn't know anything about what a finger is, but this collective does. So evolution gave us the scale up, but there's a failure mode to this. The failure mode is cancer. So here's where when these cells are electrically isolated from each other, this is a human glioblastoma in culture. So these individual cells can roll back to their unicellular past. They do not become more selfish. A lot of game theory kind of approaches treat cancer cells as more selfish. They don't become more selfish. Their selves get smaller. They're just as selfish. It's just that the self here is huge. Their self is very tiny and they do the best they can. Biomedically, if you have that kind of weird perspective on cancer as a disorder of selfhood, then what you can do is something interesting. And we did this years ago. You can induce tumors using human oncogenes, but what you can do is use these ion channels to force the cells to be in the correct electrical state to maintain proper communication. And when you do that, even though the oncogene is blazingly expressed, it's labeled in red here. You can see it's all over the place. There's no tumor. This is the same animal. There's no tumor because the genetics might have a defect. The hardware might be a little bit broken, but as long as the cell is physiologically part of this network that contains large scale goals like building nice skin, nice muscle and so on, they're kept away from tumorigenesis and harnessed towards these large scale anatomical goals.

Okay. So the last piece I want to show you goes beyond this. So we've been talking about goals in different spaces. So the tiny metabolic goals of single cells, the morphogenetic goals of cellular collectives, and now we can ask the following question. Where do these goals come from and what other goals can novel beings have? And so this is a project that I also have to do a disclosure here because Josh Bongard and I have another company called Fauna Systems which is all about AI-designed bio-robotics and this is all work done in collaboration with the Bongard lab at the Institute of Computationally Designed Organisms. We asked the following question. How much plasticity do these cells have and when they're liberated from their normal boundary conditions of the animal, what will they do? How will they reboot multicellularity? How will they navigate anatomical and behavioral spaces? What can they do? So here's the experiment and all the biology here was done by Douglas Blackiston, a staff scientist in the group. What we did was we took an early frog embryo and at this stage where they have about a thousand cells or so, we cut off these cells up here which are the kind of skin. This is all fated to be skin cells and we dissociated them and we put them back in this little depression here. And so they're just skin. There's no nerve. There's no muscle, just skin. And what could they have done? Well, they could have died. They could have spread out and gotten away from each other. They could have formed a nice two-dimensional monolayer the way that cell culture does. What they did instead was something quite different. Oops. Here. What they did instead, after we put them in this little depression, all these little skin cells, they got together and they formed this round thing we call a xenobot. It's called a xenobot because Xenopus laevis is the name of the frog and we think it's a bio-robotics platform, so xenobot. What they do is these little guys achieve spontaneous motion by using their cilia. So frog skin has little hairs, little motile hairs called cilia which it uses to distribute mucus down the side of the animal. So instead what they do is they use it to row against the water. So they start to swim. So they can go in circles. They can sort of patrol back and forth like this. They can have these collective behaviors where they can go on a long journey or they can interact with each other or they can sort of sit there, you know, kind of doing nothing. They have all kinds of behaviors. Here's one moving in a maze in water. So you'll see what it does is it takes the corner without bumping into anything and then at this point it spontaneously decides to turn around and go back where it came from. We are not pacing it. We are not controlling it. There are no neurons here. This is a piece of skin living its new best life and you're watching some of its behaviors. It has all kinds of active calcium signaling. This is the same technology used to read out brain signaling. We are currently in the process of trying to decode all this to see what the cells are saying to each other. What are these two xenobots saying to each other? They can regenerate. So if you cut them almost in half, they will seal up. So look at this. So here's like a 180-degree hinge right there. Look at the power it takes to fold itself back into its new xenobot shape. And something else that's amazing is by modeling, and so Josh Bongard and Sam Kriegman in his lab did a bunch of computational modeling in the simulator of these bots running around and they noticed a kind of unusual predicted behavior that they tend to rearrange their environment in this weird way. So we went ahead and we built some of the ones, then again Doug sculpted some of these things in ways that the AI suggested and we discovered something remarkable which is called kinematic replication. Notice that we've made it impossible for these animals to reproduce in the normal froggy fashion. Okay, they don't have any of the things there. They're just skin. But within a couple days, they discover a novel way of making copies of themselves that's basically von Neumann's dream. It's able to construct copies of itself from material it finds in the environment. So this white stuff here, these are all loose skin cells that we sprinkled into their world. What they do is they run around and they both collectively and individually, they collect them into little balls and then they sort of polish these little balls and compact them and because they're not working with some kind of passive particle, they're working with an agential material just like we were. The engineers, you know, Doug Blackiston was working with an agential material, these cells that have an agenda. They like to assemble into the xenobot. These little blobs become the next generation of xenobots. And guess what they do? They run around and collect more cells and they become the next generation of xenobots and so on. And so this ability to make copies of itself using kinematic replication is just one of them. I mean, we're just scratching the surface here of what they can do. There's much, much more coming. And so this suggests an interesting set of questions. What did the frog genome actually learn during evolution? Well, it can make frogs, sure, and this is the normal tadpole path, but it turns out, and something we would have never known if we hadn't tried this experiment, it actually can do other things. It can make these exact same genome. We didn't edit the genome here. We didn't give them any weird new nanomaterials. The same genome can have a different developmental path. This is what an old xenobot looks like. I mean, it has this weird structure that who knows what it's changing into, and then it has different behaviors that's different from what tadpoles do. So here's some interesting things about it. First of all, for any other creature, when you ask why does it look the way it looks, why does it have certain behaviors and properties, the answer is always the same. Well, evolution. So selection, right? So for eons, the system was shaped for, was selected for specific behaviors. Everything else died out, and there you go. But here, you can't use that story here because there's never been any xenobots. There has never been any selection to be a good xenobot. Where did these capacities come from? Where did their ability to do kinematic self-replication, which to our knowledge doesn't exist anywhere in the biosphere, where did that come from? Now, we don't know what their cognitive capacities are. We're studying that right now. What can they sense? Do they have preferences? Can they learn? We don't know any of that yet.

But we do know that they have novel morphologies and novel behaviors. And so where does that come from? So one way to think about this, and that same genome by being probed in this way is showing us a bit of the latent space, a bit of what they are capable of. So this is kind of, you know, Jeremy Gay, who's an amazing graphic designer that does a lot of our artwork, did this homage to the famous Hofstadter book Gödel, Escher, Bach, where he was pointing out that how you look at certain objects reveals what you see and what they can do. And so through that frog genome, when you look through it in a particular way, yes, you can see the tadpoles, and yes, you can see xenobots, and maybe you can see other things. And so what we can say here is that these kind of synthetic organisms are really a tool. They're almost like a periscope or a telescope that lets you look into another space that we normally don't see, into the space of possibilities of, for example, anatomical structure. And so for the last couple of minutes, I just want to talk about this and the future of all this in bioengineering and so on.

And the concepts that we want to talk about are these. We've already talked about multiscale competency architecture, this idea of polycomputing, which Josh Bongard and I came up with to indicate the fact that any physical process can be treated as a computation, as a different computation by different observers, depending on how they interpret it. And that these observers are cells observing each other and various conspecifics and parasites and so on. This idea that really because of this competency, because cells and tissues are not a passive material, every part is able to hack all the other parts by providing signals that exploit their functionality. Just like we as engineers, and that whole middle part of the talk, I was trying to show you that triggering complex morph organ formation and so on, you don't have to try to micromanage the molecular steps if you can communicate with a larger system to tell it what you want it to do. You can take advantage of that. Well, biology, you know, we're late to that party there as engineers, and biology has been doing this all along. So I want to show you a couple of examples.

The first thing I want to just mention is building with a gentle materials is very different. So when you build with a passive material like making Legos, you need it's on you to put every Lego where it goes. All they do is hold shape. They have very little competencies besides holding their shape. So you have to control everything. This is molecular medicine. This is synthetic biology now. This is bioengineering. You build everything yourself. But so if you wanted to make a tower out of Legos, you have to put everything where it goes. It's easy. There are a few surprises, but the problem is if it falls down, that's it. You could try to make a tower out of dogs, for example. If you made a tower out of, if you tried it the old-fashioned way by stacking dogs on top of each other, that doesn't work. The dogs have their own agendas. They're going to walk off. Not going to work. But there's a new set of tools, basically from different parts of this spectrum that we talked about, different ways of interacting with systems. If you don't insist on treating them as this kind of system, but you're open to the idea that your material might actually be this kind of system. And in the case of dogs, it's very obvious. In the case of cells, it's only now becoming apparent that they are also, you know, somewhere here. What you can do is train the dogs. It's a very different strategy. You're now doing behavior shaping as opposed to old-school engineering. But the good news is that this tower repairs itself when you knock it down. They'll get right back up. And so the question about this competency is what can you expect the subunits to do when you're not there to manage them, and how do you construct complex things out of units like that?

And the way you do it is by using their natural interface, in this case learning, rewards, and punishments, to control their behavior. This is hacking. Hacking not in the sense of, you know, the people who steal your credit card numbers, but in the sense of using a system in not necessarily the way that it was intended to be used, but in some clever way that exploits the system's own properties for a complex outcome. So we see this all the time in behavioral science, right? So there are these crazy parasites, this Toxoplasma, that causes animals and humans to become less risk-averse. So captains of industry and people who die in motorcycle accidents going too fast tend to be infected with this stuff because it actually changes your risk-aversion level. And, you know, we've got a fungus that controls ant behavior and so on. So we know that creatures hack each other's behavioral policies for navigating three-dimensional space. But that's not all. What we're now starting to understand is that there are really interesting hacks that happen in anatomical and physiological spaces. In order to do that, the material, right? This agentive matter, the material on which evolution works, so cells and tissues, has to have these various competencies. It has to have memories. It has to be able to measure things, to take in inputs. You can't hack something that doesn't have any of that. But life does. And I want to show you a couple of examples of this plasticity. So here is, and this is not Photoshop, and these are real animals. These are flies, and on their wings, they are running a kind of stripped-down ant morphogenetic program, like a two-dimensional ant morphogenetic program. And the reason they do it is because when they flap their wings in a particular pattern, it looks like ants scurrying around, and their predators don't want to deal with the ants, and they scare them off. So it's kind of amazing that on their wings, they're able to have these very specific patterns. It's no more amazing than the fly itself. I mean, this is just a simplified version of what the fly itself is doing. But in computer science, it would be kind of virtualization, the fact that the same tissue can actually make different kinds of patterns. Now, I want to show you one of my favorite examples of that kind of hacking. This is the traditional thing that happens. So here's an acorn, and we know that what acorn genomes know how to do is make oak leaves. So they're green, they're flat, they have this particular structure. This is what that genome does. Turns out that with just a little bit of prompting from a wasp, these cells are also able to make something like this, or like this. These are galls. These are made by the plant cell that is infected by this parasite. The parasite is hacking these plant cells to get them to build something completely different. Would we have any clue that these flat green cells are able to build something like this if we didn't see this example? So this is another one of these cases where you need to do perturbation experiments. In this case, this is a natural one done by this wasp to see the latent space of what the system is competent to, what is it capable of? What else can these cells do? And can you imagine all the things that these plant cells could do if they can build this kind of stuff?

So just to finish up here, to really, I think all of this pushes in a particular direction, kind of again going back to this sort of the philosophical side of things. Is this an organism? You know, it's got a body, and it's got some sort of cognitive architecture, this frog. Is this an organism? We tend to think so because we as humans have really tiny little observational slits through which we see the world. We like, you know, all the things that happen not too slowly and not too quickly. We only can perceive in a certain range, and thus we only see the animal right here. But we now have tools like the xenobot and others that can collapse the bigger picture, which we normally are pretty blind to, into a scale we can actually appreciate. Because this is not the organism. This is the organism. It's the cells plus all of the things that can manipulate them to do various things. It's also the evolutionary cycle wrapped around it. It's complex beings like us engineers, simple beings like various microbes and the parasites, and of course the low-agency aspects of the environment like chemicals. Why do I say all of this is part of the organism? Because if you don't understand all this other stuff, you really do not understand what else these cells are capable of. This manifestation is the default. It is the most common thing, but we are completely blind to all of what it's capable of if we do not understand the other inputs into the system. And this is somebody else had mentioned that, you know, the environment and the context. I mean, they're not always there. In fact, we need to explore all kinds of ones that don't naturally exist to really understand what this organism is. And so the xenobot is really a kind of tool you can use to peer into other spaces, be it physiological or transcriptional. I mean, we're studying now all the transcriptome of these xenobots. It is not like the transcriptome of frog embryos, by the way.

This is what it is. The environment is full of affordances, some of them low agency, some of them rather high agency, and everything in between. And all of this is part of what actually defines what a complex organism is. So, having seen all this, just to end here, having seen all this, I will point out that all of the biological natural forms that Darwin was impressed by when he said endless forms most beautiful, all of that N of 1 run down the tree of life is a tiny region, just a tiny region of the state space of all possible beings. Because life is so interoperable and so hackable, pretty much every combination of evolved material, so DNA, cells, and so on, designed material, so active matter, the various smart appliances, and so on, and software, any combination of those three things is going to be some viable being operating in some kind of set of spaces with some cognitive system. So, cyborgs, hybrots, various kinds of chimeras, various kinds of robotics, all of these things are somewhere in this giant space of possible minds. And this has massive implications for us on a societal level.

Because going forward, none of the concepts that we grew up with, the sharp distinction between humans and non-humans, between machines and living beings, between intelligence and, you know, sort of faking intelligence, none of these categories are going to make it through the next couple of decades. We need to scrap all of that towards a continuous view and a new system of ethics that does not depend on the two things that we used to depend on. In the olden days, you could ask what are you made of and how did you get here, meaning evolved versus engineered, and based on that, you could make a decision of how to ethically relate to something. You know, sort of knock on it, and if you hear a metallic clangy sound, you would say that, ah, you come from a factory, you're going to be pretty boring, I can take you apart, whatever. None of those things are going to be good anymore because we need to develop novel ways to recognize and to relate to beings that are nowhere with us on the tree of life. We simply cannot judge them that way. And this is going to be an exciting and grand challenge for the species to get beyond a very simplistic view of agency towards a more expansive compassion towards other beings of radically different composition and provenance.

So, I'll stop here and just say that lots of papers that go deeply into all of the things that I sort of briefly went over in terms of, you know, machines and everything else. I'm going to thank the people who did all the work. So, these are the postdocs and graduate students who did all of the work that I showed you today. Many collaborators, our various funders, the disclosures again. Jeremy Gay did all the amazing graphics. And most of all the animal model systems because they do all the heavy lifting in teaching us about these things. So, thank you so much and I will take questions.

Yeah, yeah, it's, yeah, sure. It's my fault that we went over the 2 hours.

Well, thank you so much. I appreciate it. I must say I love your background. As you can see, I have a lot of affinity to that. So, thank you. Let's see. Here's what I want to start with, and I will answer all of that, but here's where I want to start with. There's a standard kind of view that even implicitly underlies a lot of the concern when people ask me about these things, which is basically this: everything is fine and you scientists better not do something to really screw it up, right? That is kind of the mainstream assumption. And that was sort of understandable back in the day when we really could think that things were created by someone for us to be the best that they could be and really we ought not to screw it up, right? But I think now, I think it's safe to say that we have a different vision, and this is my vision. Things are absolutely not fine. And I'm not talking about just the various environmental calamities, but I get, and I'm not a clinician, but I receive daily phone calls and emails from people in the most unbelievable biomedical suffering. I mean, you know, my kid has a birth defect, I got a tumor, my, you know, lost a limb. I mean, you can't imagine. So people call me every day with this stuff. And our ability to address that degree of suffering and to improve people's opportunity to reach their potential given the limitations of being born even in a healthy body, right? These limitations. I consider us to be so far off of fine that it's, you know, I can't even begin to describe that. I think we are not starting here and hoping that we don't drop. We are down here. And to me, so here's my ethics to get back to your question. When people ask me what keeps me up at night about all this stuff as an ethical concern, here's what keeps me up at night. It's the possibility of failure to live up to the promise of using this kind of information to improve the lived experience for all sentient beings. That's literally what it is. And I should say, this is not, you know, I'm definitely not the only person working on this. There are many, many scientists throughout the world who have the exact same issue, who are all working towards the same thing, various degrees of freedom of embodiment and regenerative medicine and all of these kinds of things. So this is, you know, this is not just, you know, so I don't see it that I myself am going to sort of fix all these problems. It's I'm speaking now for my view in the community that we have a moral responsibility and an absolute kind of imperative to use this information to improve the state of beings in this world now. I do not believe we were sort of both physically in the world and in terms of our IQ level, our susceptibility to various weird diseases, to brain degeneration. I do not believe that any of this is optimized. Evolution did not optimize this for our happiness, it did not optimize for intelligence. Evolution optimizes for biomass. That's it. And so it happens to have dropped us at this level, which is, you know, a little bit of IQ and a little bit of survival over 7, 8 decades tops and that's it. And I don't see this as optimal in any way. This is due to the vagaries of the evolutionary process and it is now on us, now that we can, to make improvements for everybody. That's how I see this.

Absolutely, and in two ways. Let me say this about you. So goal-directedness, and I'm saying never mind just us, like all the way down to you can even, you know, with least action kinds of things, you can even talk about particles and simple matter, you know, simple matter systems all the way down. And the reason people who don't like any of it say don't use the word. People who really like teleology say use teleology, don't say teleonomy. You know, teleonomy was apparent to the goals, right? And it's sort of this hedge that some people use to sort of like, you know, try to avoid being on the bad side of the modern paradigm and so on. The reason I actually like teleonomy is not to go soft on the teleology angle. I'm full in. The reason what's interesting about teleonomy is the fact of apparent means that there, it means that you're acknowledging that there is an observer from whose perspective you're describing this. And I think that's critical. Most of what, you know, if we sort of drill down into the philosophical underpinnings of the things I've been talking about, it is very critical that there is not one objective right answer to all of this, you know, how intelligent is something? Well, you're really taking an IQ test yourself when you make that claim, you know? You might have a human brain and you might think it makes a great paperweight because you weren't smart enough to know what else it knows how to do. So all of this from the perspective of some observer being that, and that being another cell, a human scientist, a conspecific, a parasite who seeks to hack the system, an engineer, everything is from somebody's perspective. But so that's why I like teleonomy, but, you know, I'm 100% into the teleology angle.

Well, yes, although I want to be really clear is that on this, it is not simply, I mean, those kind of hypotheses are often criticized in the following way. Well, you can just put the, you know, you can just add those labels to anything. You can sort of look outwards and say there's nature, I'm going to call it Gaia and here we go. So I'm not saying that. What I'm saying is that actually I'm insisting that all of those kinds of things are actually empirical hypotheses. So that means that if you have the Gaia hypothesis, what that should do for you is to be able to say which tools can you use to benefit from the agency you're hypothesizing. So for example, if you think that some sort of ecosystem is a higher level protocognitive agent, you should be able to train it, you should be able to find and decode the memory, you should be able to show some tools from behavioral science that work at that level. And to the extent that you can do that, your hypothesis is validated, right? All of this, and we do this all the time. We apply these kinds of, you know, to me all of these claims, what these claims really are are protocols for engineering protocols. So if somebody gives me something and says I tell you that the level of intelligence in the system is X, what I hear is, ah, so here are the things in my toolbox that I can use to interact with it. That's what you've told me. And you might be right and you might be wrong or I might be more or less clever and find some better way to interact with it. But that's what I hear. So the Gaia hypothesis, maybe, but it is a hypothesis that should be empirically tested and somebody should come out and say, look, I've done associative conditioning on a whole ecosystem and it works. And we've done that on molecular networks, by the way. We did that on molecular network models which look very deterministic and very simple and all that. Turns out they can do six different kinds of learning. Who knew? Nobody would have known until you do the experiment.