Watch Episode Here
Listen to Episode Here
Show Notes
This is a very fast (~33min) flyover of some ideas relevant to the relationship between biology, computation, cognition, consciousness, and related subjects. This was given at the amazing Progress and Visions in Consciousness Science series (https://amcs-community.org/events/progress-visions-series/) as the prologue to a discussion.
CHAPTERS:
(00:00) Talk Overview and Structure
(01:30) Expanding Our View of Mind
(04:30) Biology's Multi-Scale Architecture
(09:00) Bioelectricity: Ancient Cognitive Glue
(14:00) Biology Solves Novel Problems
(22:00) Creative Interpretation, Intelligence Ratchet
(29:30) Novel Goals, Platonic Space
(36:00) Conclusion and Future Research
CONNECT WITH ME:
Transcript
And what I thought I would take you through is some ideas that, at the beginning of the talk, are things I feel pretty strongly about, and then some things at the end which are kind of way at the edge of what we know and basically speculation that I think is useful nevertheless.
So, here are four things I'm going to try to talk about very quickly. First of all, I'm going to address this issue of biology solving problems at all levels in all kinds of problem spaces, and argue that genomes are not drivers. If anything, they are more like libraries of affordances in that evolution makes problem-solving agents, not solutions to specific problems.
I will not have time to talk about this. Normally, if I give a long talk, I actually show how we go about communicating with the agential material that underlies life because this is where we have most of our data.
What I will talk about is giving some thoughts about basically the fact that the architecture of life is fundamentally creative because it is dealing with an unreliable substrate, and that this gives rise to a really powerful intelligence ratchet. Towards the end, if we have time, I'll talk about some recent ideas about shifting the perspective of agents versus data and how the notion of platonic space can be actually useful for research.
So what I'm interested in is to address something that I think we have, which is a degree of mind blindness. It's an analogy you can think about this as the electromagnetic spectrum. It used to be back in the day that things that happen when you wave magnets around, light, X-rays, all of those things were thought to be very different things.
Actually, could we mute somebody? Thank you.
Something really interesting happened when we ended up with a proper theory of electromagnetism, which is that first of all, phenomena that were thought to be extremely different turned out to be manifestations of the same thing. And so we have an idea for how they relate to each other, and it allowed technology to be made that enables us to detect and interact with and make use of all kinds of things that were completely inaccessible to us with our natural set of senses.
And so I think something like that is happening with cognition, where evolution has given us some firmware that enables us to recognize a certain extremely narrow range of kinds of minds. And I think we need to develop theory that will lead to practical tools to enable us to recognize, create, and ethically relate to a very wide, diverse set of intelligences, regardless of how they got here or what they're made of.
And so we're interested, in my group, in understanding how to relate to all of these kinds of weird things, and in particular, it's very important to move experimental work forward. So I like philosophical ideas that end up being computational models and then end up being applications and eventually therapeutics. And this is why part of the lab does birth defects, regenerative medicine, cancer, and so on, because then we know we're on the right track.
And in particular, again, what I don't have time to do today, but what I have just sort of made a recent talk about all of this, is similar to what happened to different kinds of numbers, where we slowly, with certain kinds of schisms and people sort of freaking out when they discover certain new kinds of numbers, we've kind of enlarged, even though some of these sets are infinite, we've discovered novel degrees of infinity in new kinds of numbers. I think the same thing is happening and going to happen more with the sets of cognitive agents that are out there.
And so we all can recognize this. Some of us are okay with this or even this, but actually, it gets really weird. And this, I'm not prepared to talk much about today, but we can certainly talk about this. And the infinity of sentient forms, I think, once we understand what we're looking at, is getting bigger and bigger and bigger. And we can think about for each of these things, what is needed to be able to recognize these things, and then what does it break for our assumptions? And then how do we do something useful?
Okay. So the first section here is I'm just going to show you a little bit of biology and this idea that what we call life are systems that self-assemble into a multi-scale competency architecture where every level has agendas and certain capabilities which explores problem spaces by aligning its parts towards larger and more diverse goals.
And so this is the journey that each of us takes from physics and chemistry, which describe this quiescent oocyte. That's how we all start life. Eventually, through the slow, gradual, continuous process, no magic lightning flashes that suddenly push us into the realm of psychology and so on, eventually we end up being something like this. This is what we're made of. These are single cells. They're extremely... I mean, this is a free-living organism, but you can see in a single cell, no brain, no nervous system, it has all kinds of capacities to pursue its own tiny little goals.
And then life normally goes this way. It can take some twists and turns. It can have a breakdown of multicellularity, a kind of dissociative identity disorder of the collective intelligence that gives rise to cancer. It can also, with our help, go into new directions such as these anthrobots, which can continue life even after a patient has died, and we can talk about all that.
And what fundamentally happens to bring us into the world is alignment and a kind of dynamic storytelling. And what I mean by that is that when we look at an embryo, and there may be, let's say, at an early stage, it's a sheet of 100,000 cells, why do we look at that and we say there's one embryo? What is there one of if there are 100,000 cells? What's there one of? Well, what there's one of is a particular model, a commitment to travel in anatomical space to a particular location that describes whatever it is that's being built: a human, a giraffe, a snake, whatever. And all of the cells are committed to this one story of what it is that they're building, okay? And the fact that they're all committed to the same story and reliably make that journey together is what allows us to call it one embryo.
But if you were to... And I used to do this in grad school with duck embryos. You take a little needle and you put some scratches into this blastoderm. Each one of these little islands will align within itself, because it can't feel the others, and self-organize into a new embryo, and you get multiples, twins, triplets, quadruplets, and so on, which may or may not then heal up and reattach. But in the meantime, you have multiple individuals.
So the question is, how many selves are present in an embryo? It's not at all obvious. It's not nailed down by the genetics. It's a self-organizing process that leads to the manifestation of anywhere from zero to half a dozen or more individuals. And the capacity of this excitable medium is not really clear, much like it's not really clear in the brain at all, how many actual individuals fit into a particular amount of real estate in the brain. And we can talk about that.
So what you have here is alignment, both physical and functional, to a particular model of navigation of that space. Like any good collective intelligence, or any intelligence really, it can be... We can think about how reliably it does its job. So here are acorns. They make these oak leaves. We think that, okay, this is what the oak genome knows how to do is direct the formation of exactly this pattern. It's incredibly reliable. They're very good at doing their job, these cells.
But of course, it's hackable. For example, this is an example of a non-human bioengineer, this little wasp, that comes around and deposits some signals, a kind of bio-prompting, that causes these cells to build something completely different, right? So you'd have no idea that these flat green things were actually capable of building something like this or something like this. And there are lots of examples that I could show you about how life hacks itself and each other, and how we are learning to do it.
One of the most interesting interfaces to this collective intelligence, much like in the brain, right? So what is it in the brain? It's bioelectricity, or AKA electrophysiology, but that's actually extremely ancient. It was used as a cognitive glue by evolution long before brains and muscles appeared. And so we now have ways to manipulate, to observe it and manipulate it.
And just to show you a couple things because the visuals are cool, this is an early frog embryo putting its face together, and you can see the colors are voltage. The colors are the different voltages of the cells. This is one frame. What you're seeing here is you're reading out the memories inside the collective intelligence that's navigating that anatomical space. It is this is the memory. This is the face. So here's where the mouth is going to go. Here's where the right eye is going to be. Here are the placodes. This is what guides the downstream gene expression and the anatomy to form a proper face. This is how this... When you ask how do the cells know what to do, this is how they know what to do. This is their memory.
And that bioelectric pattern is not just for bundling together cells to form a single embryo. It actually connects multiple embryos together. So it's a multi-scale kind of phenomenon. If I poke this guy here, all of his friends... Each one of these things is a separate embryo. It's a separate frog embryo. Each one of them finds out about it, and you can poke this one here. Within some minutes, they all find out about it. And here you can see some cells doing their thing.
So once you start to read these, it's basically think of neural decoding. Think of trying to read the electrophysiology of the brain and extract memories, and preferences, and so on. We can do the same thing in morphogenesis, and we can start to manipulate it.
So while the normal flatworm has a very clear memory that it needs one head and one tail, so that when you cut it into pieces, they, every piece makes one head and one tail, we can change that. We can incept false memories into these things using optogenetics and pharmacology and various other means. And so now you can say, "No, actually..." And of course, this is still messy. We're still working this out, but you can say, "No, actually you should have two heads." And it's a perfectly normal one-headed animal, right? There he is, with the wrong internal representation of what a correct planarian should look like. It's a latent memory. In fact, it's a counterfactual because it's not true right now. Right now, he has one head, but what happens is if you cut him, if you cut the tail, the head and the tail, that fragment is actually going to give rise to a two-headed animal. And I now have to tell people, "This is not AI. This is not Photoshop. These are real results."
And what happens is, because this is what enables the collective to know what to build and when to stop, so you can read out these pattern memories. If you continue to cut these two-headed animals, the memory holds. It's a bioelectric memory. It is not controlled by the genetics. The genetics are totally wild type. Nothing has changed within the genome. Sequencing the genome will not tell you that these things have two heads. It will also not tell you that the whole lineage will have two heads, because you can keep cutting them, and in fact, when they reproduce, they cut themselves in half on their own, and then they regenerate. They will continue to be two-headed. Here's a video. You can see what these guys are doing.
So what we're seeing here is that there are a variety of computational systems at all levels of organization. And I don't have time to talk about all the cool things that the genetic networks are doing. They can learn too. You don't even need cells. Just the molecular networks alone have six different kinds of memory, including Pavlovian conditioning. But all of these things work together, and they solve problems in various spaces.
And as I mentioned before, we are tuned to recognize medium-sized objects moving at medium speeds in three-dimensional space. So crows and primates and maybe a whale or an octopus, we can sort of get that. But what biology is teaching us is that embodiment is really not what we think. So when people say that various kinds of AIs are not embodied because they don't roll around the room on wheels and they don't walk around, I think that's actually not the right way to think about it at all, because there's lots of perception-action loops and goal-directedness and intelligence that take place in spaces that we cannot see.
So your cells and your tissues are navigating a high-dimensional space of possible gene expressions, they're navigating a space of physiological states, and of course, what we study is morphospace, the space of possible anatomical configurations. And they don't just roll down hills and automatically sort of do the same thing every single time. They are really navigating with some amazing capabilities.
The kind of problem-solving that they do is astounding. I'm just going to show you two examples. We could talk about this for hours. And I think it's really important for us to understand in order to make a real artificial intelligence and things like that.
So here's just one example. You take these planaria, these flatworms. You put them in a solution of barium chloride. Barium is a non-specific potassium channel blocker. It makes their cells very unhappy. Their heads explode literally over the next 24 hours, their heads explode. But if you leave them in the barium, what happens is that a little while later, they will make a new head, and the new head is completely fine with barium, no problem. It's barium-adapted.
When you check what is different about the original head and the barium-adapted head, you find that it's less than a dozen genes that are different. And the trick is that these guys, planaria, have never seen barium, either these worms nor in their evolutionary history. There's never been selection for knowing what to do when you get hit with barium.
So just think about this problem. You've got an extreme physiological stressor. You have on the order of 20,000 control knobs that you can twist. So like a Moravec bush, with an incredible number of effectors, which of these genes do you turn on and off to solve a novel stressor? It's kind of like being in a nuclear reactor control room. The thing's melting down. Unless you know exactly how it works, unless you have a model of your own apparatus, what you don't have time to do is to start flipping switches randomly to see if life gets better, because you'll be dead and there's no time for that. So we need to understand how these living systems actually solve novel problems that they haven't seen before.
This is maybe my favorite example of all time. This is a cross-section through the kidney tubule of a salamander, so normally eight to 10 cells that make this thing up. What you can do is you can make these newts with more copies of their genetic material. So instead of 2n chromosomes, you can have 4n, 5n, 6n, and so on. When you do this, the cells get bigger, but the newt stays the same size. Well, how can that be? Well, it's because there are actually fewer cells that make up the same structure. The cells are bigger, but there are less of them.
And then when you make the cells truly gigantic, what you find is that one cell will bend around itself like this and leave a space in the middle. Same structure. Now here are a couple of amazing parts. First of all, this is a completely different molecular mechanism. So instead of cell-to-cell communication and tubulogenesis here, this is cytoskeletal bending. So what the system is doing is what you would do on an IQ test. You use various tools at your disposal to get to solve the problem, to get to the same goal, in novel ways, beyond what you would normally do, right? It's creative use of the tools that you already have.
And think about what this means from the perspective of the system. You come into the world, what can you count on? Well, you can't count on your environment. We all know that. But you also apparently can't... you don't know how many copies of your genetic material you're going to have. You don't know how big your cells are going to be. You don't know how many cells you're going to have, but you still have to make that newt, and in a minute I'm going to show you can actually make other things. But you try really hard to make that newt despite the fact that you can't even rely on your own parts, never mind the environment. You can't even assume that your parts are what they are.
So for the next couple minutes, I just want to talk about this, that unlike most of the things we program, life really operates with a polycomputing architecture where you are free to reinterpret all the messages that you have, and those are genetic messages from the past. Everything you have is up for interpretation because your substrate is unreliable. You cannot afford to care too much about what your ancestors thought that your genes mean or what your past self thinks that your memories meant. You have to reinterpret all this stuff from scratch.
And I'm just going to show you a couple of examples of how this looks on three scales: on the evolutionary scale, the developmental scale, and the cognitive scale.
So one of the things that I like to think about is caterpillars and butterflies. So caterpillars are this kind of soft-bodied creature with a controller suitable for soft-bodied motion, and they crawl around and they eat leaves. They have to turn into this thing that lives in the three-dimensional world, and it flies around, and it's a hard-bodied creature, and it needs a completely different controller for that. And so in the meantime, what it does is it basically liquefies its brain, kills most of the cells, breaks most of the connections, and then eventually it rebuilds a new butterfly brain.
Okay. So what is known is that if you train the caterpillar to, let's say, eat leaves when it sees a particular color disc, it's known that butterflies actually retain this information. So on the surface of it, the crazy thing here to think about is, well, how do you maintain information while you're drastically refactoring the substrate? Okay? And so that's a cool question, but there's an even more amazing piece to this, which is that the actual memories of the caterpillar are of absolutely no use to the butterfly. The butterfly doesn't crawl. It doesn't eat leaves. It wants nectar. The memories of the caterpillar, if they were trying to stay intact, would never survive. But if they are able to be remapped onto a new architecture, not just carried over, not just persist, but actually change, remap, then they survive into this new embodiment as the creature gains a new life in a higher dimensional space.
So you have here something like the paradox of change, where if you, as a biological organism or a lineage, if you stay the same, you will surely die when the circumstances change. But if you change to fit those circumstances, in a sense, you, meaning the old you, is still gone. How do you persist? You cannot persist in place, right? That seems to be the paradox.
So what I like to think about is this idea that your memories, and these are memories, these are both cognitive memories and genetic memories, are just messages from your past self. You don't have access to the past in any given moment in time. You have no idea what happened in the past. What you have are memory engrams, traces left in your brain and body that were constructed by your past selves. They were compressed. They were generalized. They went through this bottleneck.
So here's all this stuff that, I mean, all the computer science people will recognize this immediately. They're... all the inputs and the different diverse experiences came through. They were compressed somehow into this very thin substrate, which are the engrams that people still fight about what that is. But future you has to decompress them, and information was lost here. You don't actually know what these memories mean when you find a particular memory molecule or a set of electrical states in your brain. You don't actually know what they mean. You have to creatively interpret them.
And I suspect that while this part can be algorithmic and deductive, this part I don't think can. I suspect this is... and I don't exactly know what this is. I'll speculate a little bit at the end. But there's something else going on here which allows those memories to be remapped in a novel context. Like every message from others at the current time, messages from past you also need to be interpreted. And you don't need to really care about what they meant before. What you do need to care about is, how do I make them adaptive now? So what the biology is optimizing for is not fidelity of information. It's saliency. So you're not trying to keep the caterpillar memories as they are. You're trying to remap them to a new meaning and new functions.
So the same thing is happening in... so that's the cognitive side. The same thing is happening in the developmental side. When you have a genome, and I know I didn't have time to go through all these examples probably, and you can find some of this stuff in this paper. But the idea basically is this, that you come into the world with a certain genome, but now comes the creative process. So this is all of past history. This is your genetics. Now comes the fun part of, if you're an embryo or a regenerating organ or anything like that, is to say, what do these things mean? And in my current environment and given what's going on now, how do I use this information to do something useful? And we have lots of examples in biology of this happening. I'll show you a couple more now.
So basically, the idea is that you can't overtrain on what the past was. You assume that the hardware's unreliable, meaning you know you're going to get mutated. You know that proteins come and go. You have no idea exactly how many of anything is in your cell. All of that stuff is very fluid. And for this reason, you have to commit to interpreting on the fly. And what evolution does is produce systems that are very good at creative interpretation of data. First, that was the case physiologically, then genetically and developmentally, and then eventually conventional cognition. And again, all of this is here.
One of the cool things about it is that this gives rise to an interesting intelligence ratchet. So the more creative interpretation of your genome you do during development, the harder it is for selection to actually know whether your genome was any good or not. And we've done simulations of this to show that what happens when you evolve over a competent material like this is that the pressure on the structural genome comes off, and all of the work of evolution starts being done on the actual algorithm of the creative interpretation of the tools that you have. Because with each run through that loop, it becomes harder and harder to select better genomes because you can't see them. You can just see the fact that useful phenotypes had appeared. And so I think that ratchet started very early on and it contributes to the amazing conventional intelligence that we see in the animal world.
Okay. And so the last part, just maybe two minutes more. The last part is this. When we have these systems that I showed you, when I talk about problem-solving, what I've been showing you are systems that can get to their goal by different means, so that's William James' definition of intelligence. And there's sort of... they have these competencies to... the newt will be the same size even if you mess with the genetics or the size of cells or whatever.
But what happens when you make novel systems? Where do their goals come from? Because for all of the natural systems, you could say, "Well, evolution did that. Evolution gave you the goal." And that's what we normally say. Why does this animal look the way it looks? Why does it have behaviors that it has? Because of selection, eons of selection gave you that goal. Okay. What happens in novel systems? I'm very interested in where goals come from. I don't really think they come from evolution per se.
And so we wanted a model system that tries to study that aspect of it. And so in biology, we like genetics and we like environment, and those are the two things, the two sources of information that we think and so on. So I'm very interested in this kind of stuff. When we make novel beings, whether technological hybrids or biological, and cyborgs and augmented humans and all kinds of weird stuff that is already partially in existence, what else... I mean, these things here don't really necessarily have a history of selection for specific properties. Where else are they getting these goals if it's not... where do these patterns come from if it's not selection?
And by the way, one story I didn't show you is that there is an anatomical space. There are these attractors that correspond to different species' shapes, and we can actually show that the same genetic hardware can actually visit multiple attractors. It doesn't have to go to its species-specific attractor, but they're there. And then people will say, "Well, okay, fine. So these are shaped by evolution."
So what I'm really interested in are patterns that don't have a history and they don't have an explanation in terms of physics, and what could that possibly be? So here's one example. This thing is called a Halley plot, and there's a million examples. I just like these. It's called a Halley plot. It's a fractal that you get from plotting this very simple formula. This tiny little seed in complex number Z, if you unpack it, gives you this. And if you slowly change one of these numbers, what you'll get are videos and movies of all this kind of stuff.
So what's interesting here, and then you can make all sorts of other things. Clearly, clear complexity. It doesn't hurt that it's kind of organic-looking as well. But what's interesting about it is that it does not have, at least to my knowledge, I'm not a mathematician, but to my knowledge, it does not have an explanation in physics. It does not have an explanation of history. It was not evolved. It was not selected. It is, in some way, indexed into by this tiny little seed or pointer. I don't think you can call it a compression for reasons we can talk about. But clearly there are patterns that come from at least mathematics that do not have the conventional kind of explanation that biological patterns have.
So how much of this is actually relevant to biology? So biologists hate this idea because they want a nice, kind of sparse ontology where you have the physical world and that's it, and they don't want any weird stuff. But I think a lot of mathematicians are actually pretty comfortable that what they're doing is exploring a completely different world that is not, in fact, explained by any fact of physics. It matters for physics, so it has influence in the physical world, but it isn't something that you're going to find by studying physics.
And so we've started to ask, what other kinds of patterns might be interesting for biology? And so I'll show you just a couple of quick things. This little thing, when you're looking at it, you might think that this is some sort of primitive organism that we found from a pond somewhere. If you were to sequence the genome, you would find out that you're very mistaken. The genome is Homo sapiens, 100% Homo sapiens, normal Homo sapiens. These are, we call these anthrobots. They're self-constructing little bio-bots that are made from adult human tracheal epithelial cells. They are not embryonic. There's no embryonic tissue here. They don't resemble any stage of normal human development. They are what your cells do when liberated from the rest of your body and given the chance to have a new life. The patient that donated these may or may not be alive, but these guys have rebooted their multicellularity. They do some cool stuff.
They have novel behaviors which we can arrange in ethograms, like we do for animals. Half their genome is expressed differently than what they were doing when they were part of the tissue inside the patient. They have cool properties, like they will heal your neural wounds. So if you make a big scratch through this neural tissue in vitro, they'll sit here in this bio-bot cluster, and they'll heal it. And you can see here they'll... And who would have known, right? Who would have known that your tracheal cells can do this?
And we've done this with frog embryos too. These are xenobots, which are made from frog skin cells, from embryonic epithelial cells. They don't have the ability to reproduce the way normal frogs do, but they've figured out something different. This is basically von Neumann's dream of a robot that finds material and makes copies of itself. If you provide them with loose epithelial cells, they will push them together into little piles. They then sort of polish the little piles. The little piles mature into the next generation of xenobots, and guess what that does? It does the same thing. It makes the next generation and the next.
None of this existed before in evolution. There's been no evolutionary pressure to be a good xenobot or to be a good anthrobot. And so now we're confronted with two options in order to try to deal with these novel forms, behaviors, and other patterns that are not directly selected for by evolution. We can say they're emergent, which basically just means that they're surprising to us. We can catalog them, and when we come across them, well, there they are. We find emergent things. I don't really like it. It allows you to have a sparse ontology, but it's kind of a mysterian position. I don't like just saying that, well, it's a fact that holds in the physical world and then that's that.
I prefer what I think is a more optimistic version, which is to say, much like the mathematicians, there's probably an ordered space of these things, and we should have a research agenda to systematically investigate them, and that the things that we're making, such as xenobots, anthrobots, various chimeras, and some other crazy stuff such as minimal algorithms, are actually vehicles. They're exploration vehicles that you can use to look for patterns in the latent space of possibilities, and understand how those patterns come into the physical world and under what circumstances.
So this is my current hypothesis, and certainly other people have put out versions of this in the past. But the idea is that I think we now have actual tools that we can begin to study this and to figure out why it is that the frog genome doesn't just point to frog morphologies, but actually it knows how to make many different things that it pulls down.
And so what I think is that the platonic space doesn't just contain things that mathematicians like, such as facts about prime numbers, but also things that we would recognize as kinds of minds. In other words, behavioral patterns and capabilities.
And what we make when we produce embryos, xenobots, computers, pretty much anything, you're not really making these kinds of things. What you're doing is making pointers that allow the ingression of specific patterns from that space into the physical world. And our job is to understand the relationship between the pointers and the thing they bring down.
Okay, I'm going to stop because I've already run way out of time. But just to point out that natural living things are a tiny corner of the possible space of beings because any combination of evolved material, engineered material, and software is some kind of a system that's going to pull down interesting patterns from the space that we are very bad at predicting.
We cannot... We are really bad at predicting. We do not know when these things come. They come in extremely minimal forms. It doesn't take much. We found crazy capabilities in classical sorting algorithms, things like bubble sort, which have been studied for decades. We can talk about that.
And I think basically what Magritte was trying to tell us here when he said, "This is not a pipe," is that when people say, "I know what it is. I've built it. I built it with my own hands. I chose the materials, so I know what this is and I know what it can do," I think we have to be extremely careful because I don't think anything is exactly what the algorithm or the material say it is.
I don't think that's true for humans in terms of the laws of biochemistry, and I don't think it's true for machines because of the formal limitations of the algorithms that they're supposedly running.
So here's kind of just the summary of these ideas. We now have a research agenda for all of this stuff, and I'd just like to thank the people who did all the work. These are the post-docs and the students in my group, and we have lots of collaborators. And I have to do a disclosure. There are three companies that I'm involved with. So, thank you.
