Transcription
What I'm going to do today is give a very different, uh, talk than what I usually do. Um, Nicholas asked me to, uh, kind of be, uh, you know, get on the edge more. And so I'm, I'm going to do that today. Typically, I'd spend about 45 minutes on the biology and either not mention consciousness at all or mention it at, you know, a little bit at the end. And so I've, I've sort of flipped that and a large part of what I normally discuss on the biology end, uh, I'm, I'm going to skip, and we can talk about it in Q&A, or you can find it in, um, at this, at this, at our lab website here, or many other places.
And so, uh, you know, it's, it's a little, um, intimidating to be talking about consciousness to people who are experts in, in the field, because typically my lab does not work on consciousness per se. We work on different aspects of unconventional intelligence. But I'm going to go out on a limb here and, uh, say a few things and, and see what, see what you think. So, uh, just to tell you up front, what I do not claim is to have a new theory of consciousness. Uh, nor do I have definitive data uniquely supporting a specific theory of consciousness. But here are some things that, that I do claim.
First of all, uh, we've expanded the use of tools, and both, both, both bench tools and conceptual frameworks from, uh, cognitive and, and neuroscience way outside of that domain. And so we've applied it them to all sorts of other things. And this has led to new discoveries and new capabilities. In other words, it's had empirical fertility, not, not in the, not question of philosophy or linguistics, but actually of discovering new things. And so I think, I think that's actually quite useful and helps us to, uh, to, to think about very diverse kinds of minds in, in other, other embodiments.
I also think that if you use any of the typical criteria that we use to ascribe consciousness to each other, be they, um, joint cellular mechanisms or problem-solving observable behaviors or evolutionary relationships, or maybe you like microtubial dynamics or specific kinds of causal architectures as evidence of consciousness in each other, um, then for the exact same reasons, we are going to have to take very, very seriously the possibility of consciousness in our other body structures. And that is basically a stepping stone to asking more broadly what other kinds of, of very diverse, um, embodiments could be relevant to the study of, of consciousness. So I think we can, we can, we can get there, uh, step by step by first considering the reasons that one might or might not, uh, assign, um, at least, at least some, some plausibility to consciousness outside of the brain in our own bodies.
But more broadly, I think that, uh, issues around AI and, and meth, many, many, uh, ethical problems are not going to be resolved if we maintain a focus on humans, a focus on 3D space for embodiment, uh, if we view the brain as necessarily privileged in some way, and if we try to maintain crisp binary categories of being conscious or not. I think all, all of these things are preventing progress in this field, and I will, uh, talk about why I think that is. And towards the end, I'm going to say some, uh, some, some really, I think, controversial things about, uh, what I think this all means for consciousness. And basically, the idea that I think physics and evolutionary history do not tell the whole story of, of what's going on here. And I'm going to, uh, show you a model that, in some ways, is a very ancient model, and in other ways, I think now is in a, a much different place.
So, uh, so, so first, I'm going to talk about the framework, the approach that we take to these kinds of problems. And I think it's, it's really, uh, analogous to what's happened in our understanding of electromagnetism. You know, prior to having a good theory of electromagnetism, uh, first of all, we had all kinds of diverse phenomena, um, static electricity, lightning, light, magnets, and so on, that people thought were fundamentally different things. And so having a good theory allowed us to see that, no, they're actually different, uh, aspects of the same underlying thing. And also, it enabled us to create technologies that allow us to operate at, uh, areas of this, of the spectrum to which we were previously completely blind and unable to participate. In other words, the only a tiny region was available to us because of our own evolutionary history. And I think this is exactly what's going on with, uh, our mind blindness and our, really, uh, inability to, to envision and to recognize all kinds of very diverse, uh, intelligences all around us. And I think we need to create these kinds of enabling technologies, much as we did here, to, uh, to enable us to recognize and interact with these other beings.
So, um, this is, you know, this is, this is, uh, the sort of, the sort of picture that, that underlies a lot of thought in the field. Um, it's an, an old piece of art called Adam Names the Animals in the Garden of Eden. And I think, um, there's, there's something deeply wrong and then also something deeply correct about this, this, uh, this, this, this view. What's wrong here is that it suggests that there are, uh, very clearly delineated natural kinds. So, so here are the different animals. These are familiar kinds of things that are easy for us to recognize as, as having some kind of mind. And Adam, of course, is different in a way from all of these. And it suggests that this is, this is the sort of thing we need to, we need to understand. And once we've got all of these, then, then, then we're done. And I think this, this is going to have to profoundly change.
But what's interesting about this is that in these, um, in this, in this ancient kind of biblical story, the idea was that, uh, Adam had to be the one to name the animals. In other words, God couldn't do it. The angels couldn't do it. It had to be Adam that named them. And in these, uh, old traditions, the idea of naming something means that you've discovered its true inner nature. Right? That's the, the kind of, the deep meaning of giving something a name. And I think this is, this is profound for us because I think we are going to have to name, in this sense of discovering the inner nature, a whole variety of novel beings with whom we are going to share our world. And so, so I think that's just, just as was true for Adam. And so I think we, we're trying to develop some tools for, uh, for making that easier to do.
And one important thing to, that, that I think is, is true is that because of our understanding of evolution and embryonic development, that where on different time scales, we all start as a single cell and then we become this, uh, typical human that features prominently in philosophical discussions of, of philosophy of mind. Uh, we now know that this is a slow and gradual process on both time scales. And so I would claim that the continuity thesis is the null hypothesis. That is, if, if we think there should be some sort of emergent phase transitions and specific categories that are just, uh, uh, different than other categories of beings here, that, that is what needs very strong evidence. The, the, the baseline, I think, in the null hypothesis should be a continuity and the drive to develop, uh, models of scaling and, and so that we can talk about what kind of consciousness and how much, not whether something is or isn't consciousness. And I'll sort of hammer this theme a couple of times in my talk.
Um, now, moreover, whoa, uh, sorry, I don't know what's going on with this, with this visual here, but, um, there is, uh, there is another, uh, there is another, uh, spectrum here, another continuum, which is the, uh, gradual, uh, technological and biological changes that can and, and will and are being applied to this kind of standard human with this, sort of, special agential glow. And it, it be, because biology is so highly interoperable, pretty much any, uh, you know, you can, you can introduce engineered material anywhere along the scale. And so, you know, questions of life versus machines or what computers, quote unquote, can and can't do. These, these are all very difficult now because of the existence of these, of these chimeras. And I think these are the difficult cases that we need to talk about, which again speak to, uh, models of transformation, not sharp categories.
So, uh, in my, uh, in my framework, what I try to do is, uh, create, uh, a way of thinking about all of these different, uh, agents, no matter what their, uh, composition or province is. In other words, we need to be able to think about familiar creatures like primates and birds and maybe an octopus and a whale and things like that. But also very weird creatures such as colonial organisms, engineered new life forms, of which I'm going to show you one today, uh, uh, AIs, whether purely software or robotic, maybe someday exobiological agents, and even something that I won't have time to talk about today, which is, which is even weirder still, which are, which are actually, um, patterns within media.
I would like to, I would like to have a way of understanding what all of these agents have in common, and in particular, in a way that moves experimental work forward. So it's, this is very important for us in our lab is that whatever we say about these things actually has to be, um, actionable in an experimental way that shows us that why thinking of, of, of these things is, is, is useful. But I also think it's critical for better ethical frameworks as we learn to recognize and relate to these, these unfamiliar kinds of minds. And obviously, I'm not the first person to try for this. So here's Rosen, Bluth, Weiner, and Bigelow with with their sort of cybernetic, um, kind of ladder leading up from, from passive matter all the way up to human metacognition. But, but I think, I think we have a new way of, of thinking about it. And the sort of 1.0 version of that is, uh, is all described here.
So, uh, the fact is that all of us made this journey. So we all started life as a quiescent blob of chemicals as an unfertilized ovum. Uh, people tend to think that that is well handled by the laws of physics and chemistry, and that pres, that you perhaps don't need, um, any concepts from, uh, from cognitive or behavioral science to deal with this. But slowly and surely, uh, gradually through the process of morphogenesis, we become something like this, or even perhaps something like this. And this, this, uh, journey from, from, from chemistry and physics to, uh, to the land of psychoanalysis and, and so on, is, is gradual. By the way, this is not the end. There are a number of interesting things that can happen after that, and we can, we can talk about, talk about that, but, but I think that developmental biology offers no special point when any of, any of the things that are of interest to, to, to, uh, this group, um, kick in. Okay. There, there is, there is no special stage at which, uh, some sort of, um, uh, uh, discrete, uh, discreet thing happens, at least, I'm, I'm not aware of any such.
And the other, the other important thing about the architecture of which, uh, of which we're made is that we are all collective intelligences. We are all made of parts, uh, that are themselves quite competent. So here's a single cell. Now, this is a free-living organism, but you get the idea. This is one cell, and it has no brain, no nervous system. It is highly competent at its own tiny little, um, cell-level, um, agendas. And you can, you can ask the question of, of whether you can reward and punish this. And I think most, most people will say you can. And in fact, of course, you can do experiments where you provide, uh, positive and negative reinforcement and, and so on. But, um, then you have to remind, remind yourself that, uh, this thing here. So here's another, here's another unicellularial organism. This is kind of a sad video. What you're watching is, uh, is the death of a minimal, uh, minimal being. So here it is. So, so this poor guy is falling apart into its constitutive parts. And what you see now is just, uh, some chemicals. You see some, some chemicals that are quite different from what it was, uh, just, just minutes ago. And so now we can ask the question, can you, in fact, reward and punish a collection of chemicals? Because that, in a certain sense, is what we have here. And with some time, if this was an egg, with some time, you would have a being that you certainly can reward and punish.
And what these things are all made of are chemical networks. And by the way, these chemical networks themselves can learn. So, just by virtue of the mathematics that, that, um, governs collections of, of chemicals that turn each other on or off, uh, you find that these kind of systems have at least six different kinds of learning, including Pavlovian conditioning. So, and, and we're using all this in the lab for all kinds of interesting biomedical applications, such as drug conditioning. But you can, you can actually train the material that is inside the cells themselves. So what I'm painting here is a picture of a, of a collective intelligence, um, stack, where the material, I've called it an agential material. It goes all the way to the bottom of, um, of collections of chemical networks that have behaviors that would be recognized by people who work on, in behavioral science. And in fact, this is, this is very recent, the recent data. In fact, it turns out that as you, as you train these things, their causal emergence goes up. Sometimes it goes up gradually, sometimes it goes up in, in spurts where they, in fact, um, fall asleep or sort of disband into their collective subunits in between, uh, the interactions you have with them. The activity doesn't stop. They're, they're, they're highly active the whole time. But actually, the causal emergence falls when you're not talking to them. So they can't keep themselves awake, so to speak. But as you, as you, as you train them, in particular, in an associative conditioning paradigm, the, um, the, the, the integration goes up and up and up. Then you can find all that here. So the material of which we are made is itself has has significant relevance to this problem of, of minds and unconventional, um, is to, uh, understand how limited we are in terms of recognizing embodiment.
So we, as humans, are obsessed with three-dimensional space. We are pretty good at recognizing medium-sized objects moving at medium speeds in that space. And so these kinds of things are, are pretty, pretty easy for us to understand. Uh, however, biology has been navigating other spaces long before we had nerve and muscle that allowed us to explore the three-dimensional world. So living things navigate transcriptional space, they navigate physiological state spaces, and they navigate anatomical morphospace, which is basically the journey that, that we took during embryogenesis from being a single cell to, um, you know, whatever body shape we have now. And so my point is that, uh, there are unconventional beings that are doing this, uh, perception-action loop. They have competency at problem-solving. They, they strive for specific goals. They, um, they, I, I think they, they, they, they, they suffer and they do various, various things that we would find very relevant to this problem in spaces that we can't see. And that is, that is critical, just like in the, in the case of the electromagnetism. That means that we need to, the, it's on us, actually, the responsibility is on us to develop tools to be able to recognize and, uh, and think about, uh, the lives and the, the inner lives, in particular, of beings that live in spaces that are evolutionarily, we are not evolutionarily prepped for.
So the next thing I'm going to very briefly talk about is one model system that we study, which is the collective intelligence of morphogenesis. And this is, this is typically what I'd spend 45 minutes just, just on this part and, and not do any of these other things. But I think, I think it's important for us. And, um, the, the reason that it's important is that it allows us, it allows us to, to start to think about, uh, how it is that we might, first of all, uh, extend, uh, ideas from a cognitive and behavioral, uh, science to other substrates, and then how we can actually, in a very practical way, communicate with these beings, and, uh, and, and discover new science.
So, uh, the, the kind of basic foundational idea here is that there's very deep symmetry between the self-creation of bodies and minds. And I think Turing understood this. It's, it's quite amazing that somebody who is really, um, so, so interested in understanding intelligence in different guises and, and, uh, you know, machine intelligence and, and computation and things like that. He also wrote, wrote this amazing paper, The Chemical Basis of Morphogenesis, where he thought about the question of, how do embryos, uh, self-assemble from, from, from a, a soup of unorganized chemicals? And I think the reason he was interested in this, in this question, which at first doesn't seem like it has a lot to do with this, I think he understood that there's a deep, um, symmetry here. And in fact, it turns out that, uh, what evolution and developmental biology are telling us is that, uh, all of the amazing things that, that brains do were actually, uh, an evolutionary pivot of much more ancient mechanisms by which, uh, multicellular beings navigated anatomical space. And we developed a tool. I used to have my students do this by hand, but now, now we made a tool that, that, that you can all try. You can take an, uh, abstract of a paper in neuroscience, and what it will do is, every time it says neuron, it just says cell. Every time it says millisecond, it will say minutes or hours. Uh, and with very few, uh, similar shifts, you have yourself a developmental biology paper. The symmetries are really, um, are really quite, um, quite striking.
And I don't have time today to show you all the examples in which the collective intelligence of cells during morphogenesis solves problems, navigates the space in a context-sensitive way, contains memories of goals that it pursues. I'll show you a little bit of that, uh, just a couple of things to, to, uh, point out some interesting ways in which the collective intelligence of cells, as it navigates morphospace, is doing all of the things that we usually take to be, uh, pretty, pretty relevant evidence for, um, for consciousness.
First, first of all, uh, it has this really interesting top-down, um, control structure where, uh, higher-level goal states eventually trickle down and control the molecular biology of what happens in the tissue. So, so here's an example. So regenerative, uh, amphibians, such as this axolotl, what happens is, if you take a tail like this and you graft it to the flank here, what happens is, so here it is grafted. Over time, this thing will turn into a limb. So, so this whole thing gets remodeled into a limb. Okay, now notice something very interesting. These cells up here are tail tip cells, sitting at the end of a tail. Why do they turn into fingers? There's nothing wrong with them. Uh, this is, this is, uh, locally, there's, there's no issue, but the collective, not the individual cells, but the, the group actually, uh, uh, knows very well that that having a, a tail in the middle of the body is not the correct goal state. So that attempt to reduce error trickles down to the molecular biology that, that is needed to turn tail cells into finger cells. So this idea of, of, of, uh, control of, of low-level phenomena, uh, in the service of higher-level, perhaps very abstract goals that they know nothing about, is something that is not just, not unique to voluntary motion and things like that in, in our conventional brains. It actually is baked in all the way down into, uh, the way our bodies are, uh, are assembled. And so, so this is an example where the local order ends up obeying a very high-level plan.
There's another, there's another profound thing that unifies cognition and morphogenesis, which is that at any given moment, uh, we cognitively don't have access to the past. What we have access to are engrams, memory traces that, that we have formed, uh, in past, uh, instances of ourselves as a compression of our past experience. So you end up with this, with this kind of architecture, this, this, this kind of bow tie architecture, where, uh, there's an algorithmic encoder that takes past experiences and sort of squeezes it down into some sort of memory, that is in our brain or body. But then, but then there's has to be a process of, if you want to use those memories, of interpretation, and this has to be creative. I, I don't think it's algorithmic, as this end is, because you've lost information. You don't exactly know what your memories mean. And I think this is a critical part of being a conscious agent is being in charge of constantly, uh, reinterpreting your own memories. Maybe the same way they were interpreted before, but maybe not. And, um, all of this is described in detail here.
But notice that that living bodies face the exact same problem. We get squeezed down into a bottleneck, which is the, uh, the egg, and then we have to decode. And under normal circumstances, the decoding goes exactly the way the encoding went. But it doesn't have to. And this is why anthropots and xenobots and various chimeras and all kinds of weird things that we can make work, because this is a creative interpretation process. There is no allegiance to what the information meant before. You tell the best, most adaptive story you can tell now. So this, this process of creative confabulation, forward-looking, not, not just a recall of static information, but forward-looking reinterpretation of whatever you have. I think in our, in our brain, our ability to do this comes from this ancient need to re-expand your, um, the genetic information that, that you've been given, as opposed to, to sort of obey it in some sort of rote fashion. And this is why living things can do amazing, they, they have incredible plasticity.
I, I won't have time to go into this in detail, but basically, if you want to, if you want to make a kidney lumen, and you happen to be given normal-sized cells, this is what you do. Cell-to-cell communication, eight to 10 cells build a thing. If you're given these enormous cells that you've never seen before, this is experimental perturbation that makes the cells gigantic, no problem. It will just take one of these gigantic cells, bend it around itself using a completely different molecular mechanism than this, right? So, this is, this is a kind of intelligence that you see on IQ tests. You, here are some, some, some objects you have. Use them in a new way to solve a new problem that you've not seen before. Yeah, no problem. These cells figure out how to use cytoskeletal bending to, uh, to create the same large-scale outcome using very different molecular mechanisms. And so, and so this kind of, uh, problem-solving, uh, requires us to ask, how does the collective intelligence hold together, and how does it know what it's supposed to do?
Now, in, in the one sort of uncontroversial example that we have of, of, of systems that, uh, try to implement goals that, that their individual parts don't have. Right? So the, the cognitive glue in the brain is electrophysiology. So we know that this is, this remarkable, um, architecture where, uh, cells generate, um, electrical properties, which they do or do not share with their neighbors in a dynamic fashion. And that allows the network to perform computations, to pursue goals, and, and so on. And that's what, uh, that's what underlies the fact that we are more than the sum of our neurons. We know things and we have goals that the, that individual neurons in our brain do not. And it also means that you could start to do this kind of neural decoding project where, if we could, if we could just read out the electrophysiological signals and, uh, decode them, then, then we could, uh, we could basically understand the goals, the preferences, the memories of, of the, the cognitive content of that being. It's, it's supposed to be literally in the electrophysiology.
Well, as I said before, um, evolution, figured out this architecture long before we had neurons or brains. Every cell in your body has these ion channels. Most cells in your body have these electrical synapses. This is highly conserved homologous machinery, not merely, um, analogous. They are actually evolutionarily homologous. And so for, for years now, we've been, uh, operating on this hypothesis that we could also learn to do decoding, non-neural decoding, to understand the goals, the preferences, the problem-solving capacities of morphogenesis, of navigating not three-dimensional spaces, which is what the nervous system does, but actually navigating anatomical space. And so we have, we've developed tools to read the biological states here of individual cells. So the colors represent voltage. These are voltage-sensitive dyes. These are not models. These are the, are simulations. These are real, um, real biological data. This is a frog embryo where you can see all the electrical activity that's, uh, that's going on. So you can, we're, we're scanning this thing the way that, um, you might, um, you might scan a brain. We do lots of, uh, quantitative simulation to try to understand at multiple levels of description, uh, how these dynamics are, are changing, all the way from the molecular biology to tissue-level dynamics, to, trying to link them with ideas in connection is the connectionist, um, machine learning, and also, uh, uh, you know, dynamical systems theory and things like that to try to understand pattern completion. So regeneration as pattern completion and so on. So, so, so we, we use all the same computational tools.
And then more importantly than just reading out those patterns, you want to be able to rewrite them. And the way we do this is not by applied electric fields. There are no magnets, no frequencies, no waves. We basically hack the same interface that the cells are already using to talk to each other. In other words, we can control, you know, at a molecular level, the topology of the network by controlling these electrical synapses. And of course, we can turn individual channels on and on, whether that be optogenetics or drugs or whatever. So, we have the ability to, to both read and write information into this electrical interface as you would in the brain. And when you do this, you, and I could show you, you know, dozens of dozens of different examples. I'll just show you one. When you do this, you can communicate new goals. So, for example, we can take RNA encoding a particular potassium channel. We can use that to establish a particular bioelectrical state somewhere in the early embryo. And we have one here that encodes the message, "Build an eye." And when we do that in cells that are going to become gut, sure enough, they get the message and they build an eye.
Now, the way this works is, uh, not because we are turning specific genes on and off, although certainly genes are, are changing their expression levels. We're not talking to the gene level. We're not talking to the stem cells. We have no idea how to micromanage an eye anymore than when I'm talking to you, I don't know how, and I don't have to know how to rearrange all the, uh, proteins in your synapses so that you remember what I'm saying to you. You will handle that on your own. We're using a very thin interface. This linguistic interface, where I give a very, kind of, high-level, uh, message, and the receiving system, because it is a multiscale cognitive system, takes care of making the molecular biology fit. And the same thing is happening here. All we say is, "Build an eye," and then all this lens, retina, optic nerve in the right position, the right orientations, the system takes care of all of that. We are, we are giving a very minimal, high-level prompt.
Now, there's something interesting that happens here. If we inject only a few cells. So the blue cells are the ones that we injected. They will do the job of, uh, trying to recruit their neighbors to participate. We don't have to tell them to do that. The material is already doing it. There's not enough of them to build an eye. So they try to recruit these guys. But these guys actually resist. They have a cancer suppression mechanism that says, if your neighbor has a weird voltage that's telling you to do something strange, try to resist. In fact, try to normalize them and make sure they end up like you. So there's this kind of battle of, of, of worldviews going on here between this, the, these cells that have bought into our message and these cells that may or may not. And I want to show you one kind of cute example of, of what that looks like. So, uh, again, the colors are, are voltage here. These are two frames. First, no, notice, notice that this, uh, this cell has a particular voltage before it touches these, uh, and then this is right after. Look at that. The tiniest touch. And so here's the video. So, so it's coming along, coming along, minding its own business, having its own voltage. This guy reaches out. Boom. That's it. As soon as it touched it, just a little bit, the voltage changes, and now it becomes part of this mass, and it's going to do things that the, that the collective wants to do. This is a very convincing, uh, piece of biological data that takes over, basically takes over this, and that is that is exactly what we did when we created that ectopic eye that I just showed you.
You can also do interesting things like this. Um, this is a planarian. These flatworms regenerate. You can cut them into pieces. And this middle fragment knows exactly how many heads or tails it should have. And so very reliably, it makes one head, one tail. And it does so because there is a, a memory of how many heads you're supposed to have, which we can read out exactly as, as you might do in a neural decoding context. Here it is. And what we can do is we can rewrite it. So, so we've made an animal that now has, and this, of course, a little messy. We're still working out the technology, but now it has, it has a pattern that says, "Two heads." And when we cut this guy, sure enough, it makes a two-headed animal. This is not Photoshop or AI or anything. This is exactly what, what they look like. But notice something very interesting. This biological map is not a map of this two-headed animal. It is a map of this anatomically one-headed animal. In fact, the molecular biology is normal. So, uh, anterior markers in the head, not in the tail. So what you're looking at here is the fact that a single body can hold at least one of two different representations of what a correct planarian will look like. In fact, this is a counterfactual representation because this is not true right now. This is what you will build if you get injured at a future time. And if you don't get injured, then this stays a latent memory. They, the tissues ignore it until they get injured, and then they construct that, then they obey and construct what, what this is. So, so you can think about these very, very minimal, and by the way, as any good memory, it's permanent. So, if you take these two-headed animals and keep recutting them, they will continuously re-regenerate as two-headed even though we haven't touched the genetics. So, this ability to, to change your, your memory without changing the hardware. So, normal, normal genome, but now a completely different, uh, body plan, architecture, and the ability to store patterns that are not reflective of what's true right now, that is past or future, but not, not right now. I think these are basal model systems for some of the amazing things our brains do in terms of the time travel that they can, the mental time travel they can do, and so on.
Okay. So I'm, uh, so, so what I've shown you, uh, so, so far is that we can start to think about, uh, somewhat alien minds by looking at a situation where the system is navigating a space that is very hard for us to visualize. So this kind of anatomical morphospace, but it has the advantage that all of the mechanisms that it's using to do this are homologous to what our brains use to do the things that we do. So I think this is a nice stepping stone. It's a nice way before we start thinking about aliens and, you know, all kinds of weird, uh, intelligences, we can start, uh, kind of sharpening our tools here. Here's something that's using the exact same machinery, uh, to do this in a space that's hard for us, right? So that's, that's why I think it's a good, it's a good model system.
So I want to move on from this to a synthetic morphology model, because what I've shown you so far are systems whose, uh, goals in anatomical space are determined by evolution. So I want to show you, I want to show you, uh, some xenobots, and the way that they're formed is that we take a frog embryo. We take some of these, um, some of these prospective epithelial cells up here. We put them in a dish. We dissociate them here. We put them in the dish. Now, there's many things they could do. They could die. They could walk away from each other. They could form a two-dimensional cell culture layer. Instead, what they do is they coalesce into this little unit. And here, this is, this is kind of neat. Um, this is a close-up. Each one of these little circles is a single cell. So, you see this little group of cells. Doesn't hurt that it looks like a little horse. I, I just think this is, this is cute. It, um, moves over and it sort of has some, some interactions with this. You see a little calcium flash here. I'll get to the calcium momentarily.
So, the individual cells do these kind of things. Then they, then they sort of coalesce into this, and then you get this, uh, little, little motile creature. It has cilia on its outer surface. It uses them to propel through the water. It can go in circles. It can patrol back and forth like this. It, they have, kind of, collective dynamics that they do. Here it is, um, uh, traversing a maze. So this is still water. There's no, there's no flow. There are no gradients. It takes the corner without bumping into the wall. And then at this point, it spontaneously turns around to go back where it came from. These are not like the biobots people make out of muscle cells where you have to pace them and you control where they go. This thing is completely autonomous. Uh, I'm not making any claims yet about their level of intelligence. We're testing all that. We have some interesting data on their memory formation and so on. Uh, or what goals they have, or their consciousness. Not making any claims about that, because you, I, I don't think you can tell any of that from just from observations of behavior.
However, uh, we can start doing some interesting things. So this is, this is calcium signaling. Remember, there are no neurons here. There's just basically prospective skin cells. Um, but what you can do is look at the calcium signaling in these guys. And just, just to focus on one of them, it's also interesting if they're talking to each other. We don't know that yet, but, um, but just, just to look at one, we can, we can look at the connectivity that we see and use the kind of tools that, uh, that are used to analyze brain connectivity and in particular to analyze things like integrated information. So to compare, uh, this, uh, this FIR of our xenobots with the null models, as people do in fMRI data, right? And this is work of Thomas Barley and Josh Bongard's lab and PI in my group. Uh, start to look at what, what, what exactly is the same and what's different, as, as what these kinds of metrics tell you about brain signaling in the brain and, and beyond. And, and if you do see things like this, you get an interest, you know, you have to answer an interesting question. Do we, do we learn from what the, uh, from what the method is telling us, or do we somehow sort of by fiat assume that it's just income, um, doesn't apply here for whatever reason that we've, we've stretched these things beyond where they need to be applied, and if so, why? Uh, so, so clearly, I don't, I don't think that, I think these, applying these methods here is, is, is quite informative.
Now, they have all sorts of interesting behavioral capacities. For example, if you give them loose, um, epithelial cells, what they do is they run around and they collect into little, little balls. And because they themselves are working with an agential material, these little balls mature to be the next generation of xenobots. Guess what they do? They run around and they make the next generation, which makes the next, and so on. So, um, we call this kinematic replication. As far as we know, no other creature on Earth reproduces this way. It's a little bit like Von Neumann's dream of a robot that makes copies of itself from stuff it finds in the environment. We certainly didn't teach them to do that. This is, this is spontaneous. I don't think any other creature, um, does this or, or has in the lineage.
And we can do one other, one other interesting thing. Um, we can ask, what their transcriptomes look like? What do, what does this novel, novel being, what, what genes do they express? Well, it turns out that they express hundreds of new genes that their, uh, original age-matched embryos do not express. And lots of interesting stuff here. Among these genes is a cluster of genes related to hearing or sensitive sensory perception of sound. So, so, um, basically, we decided to test this, and PI put a speaker underneath these guys and and caused some, um, some sound and vibration. And what we basically see is that they, they respond. They actually, you can, you can tell from their, from their behavior that they actually react to these stimuli in ways that actually embryos do not do. So this is novel. These are novel gene expressions, novel behaviors such as kinematic self-replication and responding to sound stimuli that, that the original material does not do.
And so this raises an interesting question, and I'll explain why, why I think this is relevant to consciousness work. Uh, what you would typically say is that the frog genome, over the years of, of interactions with the environment, has produced this kind of a things, the developmental stages, tadpoles, you know, that have certain behaviors, but it also produces xenobots with their own weird developmental sequence and behaviors that have have never existed before. There's never been any xenobots. There's never been any selection to be a good xenobot. And so one wants to ask the question, we know when the computations were done to derive a frog, a frog embryo. It was done for the eons that, that this genome has been budding up against the environment. When were the computations done to figure out how to be a good xenobot? Because that's never existed. And you might say that, well, it has sort of happened at the same time as, as, as you evolved a frog, but that, of course, undermines the whole specificity of environment to outcome. That's kind of the point of evolutionary theory is that there should be very tight specificity between what you're selected for in your environment and what actually happens in the end.
So, so now this, this raises an interesting question in anthropots and xenobots. Where do their goals come from? We could say that the goals of these creatures, anatomical goals, physiological set points, transcriptional set points, and so on, come from a history of selection. You can't say that for these. Where do they come from? Where do the cognitive, and, and other capacities of, so, so the forms of, of body and of behavior, where do they come from, if it's not going to be a history of selection?
And so now here's the kind of the weirdest part of the talk, where I'm going to say some, uh, some kind of, uh, out there things, and, and we can all, we can all decide on what we think. So the, the first thing to realize is that there are patterns that are not set by either physics or history. So what you're looking at here is something called a Halley plot, and then I've made, I've made kind of videos of them that basically comes from, uh, plotting this formula in complex numbers. It's very short, very simple, but this is the kind of richness that this encodes, if you know how to look at it. And the particular shape of this thing. There is no fact of physics that sets this. There is no, uh, evolutionary history that tells you why this is the case. The explanation for why, or for why this is the way it is, comes from the properties of mathematical objects. It is, and, and so this is, I'm going to, I'm going to kind of push on this, on this concept for, for a few minutes. This idea that there are specific patterns in mathematical objects that are neither determined by physics nor by any kind of evolutionary history. There is another source. Biologists don't like to think about that. Biologists would like everything in biology to be explained either by some necessity of, of physics or by an evolutionary history of, of selection. But that is clearly not the only, um, game in town.
And so, um, what I would like to suggest the following, uh, the following hypothesis, that basically what's happening is that, uh, is, is what Platonist mathematicians think is happening, which is that there is another space, an ordered, structured space of truths that are not determined by any facts of the, of the physical world. They are sort of independent in that sense, but of course, they affect the things that do happen in the physical world. And that what they're doing is exploring this space, in a systematic, um, structured way. And so these, these patterns are, they have a, they have a causal influence on the physical world. They, uh, in physics, we see them as constraints. In biology, I think we actually see them as enablements. I think we call biology systems that are actually making use of some of these as, as free lunches. And so my hypothesis is, is this, basically what I'm taking is extending this, which is, you know, there are many mathematicians that already believe this. I'm extending this in the following way. I think this, this world of forms, or the Platonic space, or the latent space, or whatever you want to call it, contains low-agency patterns, like facts about numbers and logic and the, you know, the truths about the prime numbers and things like that. But also, it contains more complex, active patterns that behavior science, behavioral scientists would recognize as kinds of minds. And so what, what I think is happening is that the relationship between, uh, mind and body is basically the same relationship as between truths of mathematics and physical objects, because that space contains not just mathematical facts, but actually behavioral propensities or, or, or kinds of minds. And so there are two implications to this. First of all, that I think that whatever we build, be it, uh, machines, robots, biobots, embryos, bodies, organs, whatever, whatever we build, these are basically interfaces. They're front ends. They're like thin clients for, uh, for these patterns that ingress into the physical world and have a huge impact on what actually happens. And I think, well, the first implication is that biologicals are not the only interfaces for this. I think this stuff is pervasive. I think it, it goes all the way down to the bottom. It is not just the, the biological bodies that have the advantage of being able to host these kinds of, these kind of patterns. And I also think that we are really bad, as the scientific community is really bad, at predicting what is going to ingress when you make these kinds of things. And I, I'll briefly mention an example in a minute, but I think this is, this is very important for, uh, the status of AIs and all this kind of stuff, because this, I, I do not believe that we make.
Intelligence, either biological or by or or technological, I think that what we make are interfaces that um that basically enable some of these some of these patterns to uh to to come into the the physical world.
Now, at this point, most of you are probably saying, well, interaction is is dead. This was like killed off at the time of Decart, that that that you know, there's this interaction problem. There's a mental world and there's a physical world. It just doesn't work to say that there's some kind of uh external uh external patterns that that are part of the consciousness of of of beings. You know, most people think that this has been put to bed a long time ago.
I think that uh because of because of mathematics, I think that physicalism was already dead in Newton's universe in the classical boring deterministic universe. Uh you don't need quantum mechanics for this, as many people have tried to sort of resurrect this with a quantum mechanical interface. I don't think you need any of that. Although maybe that's true, too. I don't know.
But but because Newton's universe was already sort of haunted by truths that are not determined by anything that happens in the physical world but in turn affect uh what does happen uh I think I think this was this was already dead. And I think the relationship here, so this is this is kind of a hypothesis that I'm floating, is that the way to think about mind-body interactions is a is is the same way that people think about what how the truths of mathematics determine physics. And there are there are lots of aspects of biology, for example, cicada data has come out at particular at 13 and and uh in 17 years because of the way prime numbers work, not because of anything in in in in physics or biology. It it rests on the the properties of prime numbers. And Darcy Thompson's book on growth and form in the 20s shows many amazing examples of how biology makes makes use of this stuff. And and so I think epiphenomenalism is as hopeless for mathematics as it is for for models of of mind. I think I I actually argue for an interactionist model.
And so the final thing I I I want to just bring up is this, and I don't have time to go into it in detail, but you can you can grab this paper and take a look. What what we found long and short by looking at very minimal models, in particular here sorting algorithms. So like bubble sort, okay? These are six lines of code, completely transparent, completely deterministic. People have been working on them for for many decades. If you look at them the right way, you see that yeah, they do the thing they're supposed to do, which is to sort numbers, but they also have these weird side quests. uh that that we measured as as clustering and some other properties um that are basically nowhere in the algorithm. In other words, uh the the they have they have behaviors that would be what again, quite recognizable by behavioral scientists, that are not what the algorithm is forcing them to do. Okay.
And so I I think what this means is that uh there is something else besides chance and necessity. And this is a way to to start thinking about um uh what what significant minds do. This is obviously very minimal, that even something as simplistic and dumb as a sorting algorithm has is is doing things that you neither prescribe for it to do nor for nor for for forbaded from from doing. And I think there's a lot of humility necessary here because if we don't understand bubble sort, which I think we haven't, then I don't think we understand language models or anything like that. And and by the way, just to mention, if you think the situation is better in the in the neural sciences where people think they know what parts of the brain do and things like that, we've gathered some some really interesting examples of human clinical data that really also suggests that in some ways the brain is basically a thin client uh because the mapping between brain architectures and uh and and the cognitive performance really breaks down in certain cases.
So my point my point is this. Um I I think cognitive uh uh ways of handling cognitive beings go all the way down because neither digital nor biochemical machines are what our formal models say they are. You know, McGreet was reminding us, this is not a pipe. This is a um a depiction of a pipe. And I think we need to remember here, this is not a Turing machine. You know, Turing machines have limitations. Um our formal models of computers and computing have limitations. But I don't think really I don't think anything matches fully matches these models. I don't think they describe anything in the real world. Uh they're their limitations and their competencies are not fully um describing the limitations and competencies of uh systems, either either technological or biological. And I think we have to we have to remember that. Um
And so so here's kind of a con a summary of what I'm saying. I'm not sure there are any dumb machines or dead matter at all. Uh I think our ancient these these kind of ancient categories that just feed the mind blindness that we already have. I think that in in making machines and uh whether digital or analog, there are things we force them to do, such as language, which we force language models to do. But unlike biology, where the thing they're doing and the communications they're expressing are very tightly linked, what these things are doing may have very little to do with their intrinsic motivations. In other words, just like in that sorting algorithm, we make it sort, but the intrinsic motivation that it has to do other things has nothing to do with the sorting itself. And so we could really be fooled by the fact that these things speak. Uh I think language may be maybe sort of a red herring as far as what any of these things are actually doing.
And so these AIs, I I don't think they have anything resembling a human mind, but that doesn't matter. Neither does your liver. They these things have other kinds of minds. And by creating these sorts of interfaces, we may be fishing in a in a new area of that space that has never had embodiment before. And and that's a that's a different way of thinking about it than this idea that you are creating a new intelligence whose u cognition is described by the language that it that it speaks. I think these two things might be very decoupled. And I think that ultimately uh we might end up talking about freedom, not about predictability or determinism, but really about the quality of the ingressions that are operating uh at this at this uh other level between um chance and and necessity.
And I think in the future we need to uh uh focus less on emergence as merely a catalog of surprises that we didn't expect, but a systematic study of the structure of that latent space of patterns of form and behavior, and to really try to work out that relationship between the structure of the physical pointer, meaning the embodiment, and the uh the patterns that are coming through it. And I think xenobots and anthrobots and various other tools, including language models and so on, are exploration vehicles because uh pretty much any combination of evolved material, engineered material, software, and these patterns which affect all of them are viable beings and cyborgs and hybrids. And th this these are the kinds of things that are really going to uh challenge our our notions of of of consciousness. And if we're going to have an ethical synth biosis with these things, we have to get better at understanding uh minds outside of our normal embodiment. And this is what I think the future of Garden of Eden is going to look like. It's going to be a very weird place. But this is all this is all coming. I mean, it's not going to take long before um all we're surrounded by beings that are that are radically different.
Um, yeah, I'm going to I'm going to skip I'm going to skip all this and just to uh just to say that if you're interested in in any of these things, there are papers where where we go into this in detail. And the most important thing is to thank the uh the postdocs and the students who did the work, our um our amazing collaborators, uh our funders here that have supported this work um over the years. And there are three disclosures. Uh these are these are companies that have licensed some of our um stuff for various purposes. So I'll stop here. Thank you for listening.