Where Does Intelligence Live? | Michael Levin on Evolution, AI, and Platonic Space

You take a planarian and you put it in a solution of barium. Barium is a non-specific potassium channel blocker. The cells, especially the neurons, are incredibly unhappy. It's a toxin. Roughly overnight, their heads explode, and so they lose their heads. If you leave the planarian in the barium, within a couple of weeks it grows a new head, and the new head is barium-insensitive — couldn't care less. So what we did was we looked at the gene expression of a standard head compared to a barium-adapted head, and there were really less than a dozen genes that were altered. That's a crazy combinatorial problem. You don't have time to test all the possibilities. How do you identify the actions that you're going to take that solve your problem? If we had the answer to this — and I have some ideas, but we certainly haven't cracked it — the answer to this, A, I think solves a lot of modern medicine, because you could ask the cells, 'How would you solve this problem?' for problems that we don't know how to solve. And B, I think it would be massive inspiration for machine learning.

The minds everywhere thing has been a through-line through our work for a really long time. Even back, you know, when I wasn't specifically talking about that aspect of it, I was always interested in how many of the approaches from behavior science and things like that can we import into regenerative medicine, into cell biology and developmental biology, and so on. What we've been working on more recently, that I've only sort of started talking about in the last year, I'd say, although this has been on my mind for a long time, is this notion of Platonic space. And I don't — did we talk about this last time? I don't even remember.

Okay, so if you want, we can sort of dive into that. So now that with the advent of increasingly sort of alien kinds of agents — so both the biologicals that we make, xenobots, anthrobots, neurobots, the different kinds of hybrid and chimeric things that are made that we and others are making, that are partly evolved, partly engineered, and of course AI and all those things — it becomes really important to ask where the properties, especially the cognitive properties, of new beings come from, because typically people assume, well, it's selection, of course. Eons of selection determine what your goals are, what your preferences are, what your cognitive competencies are. But now we're dealing with beings that have never been here before, and so now it becomes especially interesting to ask: where do those things come from if it isn't selection? And I don't think the standard story of emergence even begins to scratch the surface of what I think is going on. So that's one area that we've been working on.

Another area is the implications of that, which means creating practical, useful tools to help us detect and communicate with all kinds of unconventional intelligences around us. So we've been thinking about ways to make interfaces that both facilitate the ingression of novel patterns from that space and have a little bit more control over what shows up and when, and then to communicate with the various kinds of minds in ways that are useful for engineering. And of course, you know, the ethics of it is critical. So those are some big areas.

So let's talk about two aspects of that. First of all, just in the purely biological cases. So the claim that the properties of natural things are primarily driven by selection — that is not a result, it's not a conclusion, it's a kind of an axiom. And it's an assumption that, in many cases, a story of selective forces is useful, and that's a good story. I'm not at all sure that's a complete story, even for standard biological things, but that's actually not the argument that I'm making now. That's a much longer thing, and you may want to even talk to Richard Watson, somebody that I collaborate with quite a lot, to hear more about that. I'm talking about something slightly different. So the thing is that even with traditional biologicals, my suspicion is that some of these other factors that I'm about to talk about have a much bigger role than we ever anticipated. But at least for the biologicals, it seems like you can tell a story of selection, because they've been around for a long time and subjected to various selection forces. It seems plausible that that will be an explanation for everything that you see. Plus, you know, there's spandrel effects and things like that. But it sounds like that might be a good story.

But let's talk about what evolutionary theory is supposed to do for us. Selection, as an explanation or a history of evolution for a given being, is often used as a kind of catch-all, because it seems like it should work. And the thing is, it's not enough to say, well, let's look at xenobots. Okay, so xenobots clearly come from living cells. It's not enough to just say, well, the cells they came from existed for a long time, so whatever they're doing it must be some sort of effect of prior selection forces. That's kind of an unfalsifiable thing. You can always say that. It doesn't actually do you any good. But let's remember what the theory of evolution was supposed to do. What it was supposed to do is provide a nice, tight connection, with specificity, between the properties you have now and the history of environments you've had. So the whole point is to be able to say, you look and act the way you look and act because very specific aspects of your past environments got you to be where you are today. That's the standard story of evolution: your interaction with specific environments got you to one outcome versus a different outcome, right? That's what a good evolutionary story is supposed to do.

So when we look at, let's say, xenobots or anthrobots, or even weirder things that we and others are making, we have to say, what is it about the past of frogs that actually would have allowed us to predict all the things that these things are doing now? We can look at a frog and we say, okay, it definitely doesn't come from the desert, it had this and that and features. But you have to be able to paint a tight picture. It's not enough to say, yeah, well, at the time you were selecting for a good frog, you also selected for kinematic self-replication, and in the case of anthrobots for something that can heal neural wounds. You know, that's just hand-waving. You have to be able to make a specific connection.

And in fact, one of the fun things I did recently — and this is unpublished, so grain of salt here — but what I did was, I made a pipeline where you look at the novel genes that the xenobots and anthrobots express that their normal tissue doesn't, right? So differentially expressed genes, new transcriptional profiles that these things have that they normally wouldn't. And you can sort of run that backwards and say, 'Okay, if these had been evolved, if they had been selected for, what would that world look like?' So if the properties of xenobots, anthrobots, and neurobots had been selected for, what would that world look like? Because that's actually quite different. If you try to roll it backwards that way, you will come up with a bunch of features: the level of oxygen, the degree of light, you know, other aspects of the ecology. And so you can think of it as a kind of virtual ecology. Like, what would the xenobot home world look like? What would the anthrobot home world look like? And it's quite different, actually, than the world that we have here. So that's part of it.

And then when you start making completely novel things with synthetic biology, with bioengineering, by including totally synthetic components — engineered components — you end up with things that really, in no reasonable way, have an evolutionary history of forces here on Earth. So then you have to ask yourself, well, where do their properties come from? And this gets us to some of the issues, if you want to talk about them, that I think have implications for computer science as well, where right now we have basically three ways of paying for competency. If you're going to have a competent controller of some sort — let's say it's to solve a problem, or it's an organism that's good at doing certain things — there are three ways that that comes to be. One is engineering or design: somebody knew how to solve the problem, they wrote a controller or they created a machine or an algorithm, and therefore that's what you paid for. You paid for it by creating the thing outright. Or selection: you ran through huge numbers of variants, killed off everything that didn't work, kept what's left. Okay, so you paid for effort there. Or third is learning: you have a system, it had lots of contact with the problem, it took its lumps, learned, and now it's good. Okay, those are the three things. So the idea is that specific kinds of competencies require you to put in effort along those three lines. Yeah, it's either design, evolution, training or learning. And the most interesting thing happens — and this is partly why we do this work — is when you create novel beings that were not specifically engineered for anything, they were not selected for those specific properties. There's never been pressure to be a good xenobot, to be good at kinematic replication and all of that, nor learning. There's never been many attempts to make xenobots, and then finally the cells learned. So that suggests to me that we're missing something. There's an input into this process that we're not accounting for. And that to me is extremely interesting. And that's what we need to understand, because I think as engineers, we both could exploit it, but also we have to be aware of it, because in some cases it isn't what we want, and in other cases it provides a massive amount of potential for us to cooperate with and collaborate with in our creations that we're not using.

Absolutely. So what I'm not saying is that selection never happens or that it's irrelevant or whatever. No, no, absolutely these things work together, for sure. But I want to emphasize two things. First of all, the problem with the bonsai tree analogy — and a lot of people like that kind of analogy, people talk about constraints a lot — is that that only tells you half the story, and I suspect it's not even half the story. Constraint is one thing, but enablement, meaning these free lunches that we can talk about, is massively important, and I think evolution exploits the hell out of it. I think it's actually really important, and it's completely different than constraint. People try to sort of lump it all under constraint, but I actually think that's not a helpful framing.

The other thing I'll point out is that we have to know that the material that evolution is working on, even in cases where selection is having a huge role, the material itself has important properties that don't come from biology and they don't even necessarily come from physics. So I'll give you a simple example: gene regulatory networks. So models of pathways of any kind, chemical pathways. When you have a pathway of a small number of chemicals that turn each other on and off or regulate them up and down, it turns out that these networks have some really interesting properties well beyond — so Stu Kauffman is kind of like the original person who discovered some of this stuff, right? Order for free and so on. But it goes way beyond that, because not only is there criticality and things like this, but these kinds of networks, A, they can show learning dynamics to stimuli. So they can do six different kinds of learning: habituation, sensitization, Pavlovian conditioning. But also when you train them, their causal emergence goes up, meaning the degree to which the system is a whole that's more than the sum of its parts actually goes up. We can dig into that if you want. But the bottom line is that it becomes a positive feedback loop that points upwards in terms of intelligence and agency, it becomes a ratchet. And that aspect of it — the fact that just any kind of molecular networks tend to have those properties — and this is as yet unpublished work, but it turns out they're actually optimized; like random networks are already optimized for those properties. You don't need to select them, you don't need to design them. They're already optimized for those properties.

So those amazing properties, which evolution absolutely uses, and which contribute to what happens in selection and all of those things, they're a free gift from math. There's nothing that physics gives you there. You don't need selection. It's already there. So the material that evolution gets to work with has these amazing properties, and that's not a constraint. I mean, you can call it... You know, the thing about constraints is you can say, 'Yes, it's a constraint because all of the things that don't do that aren't there.' That's not much of a constraint.

That's a popular way of thinking about it, and there may be new things that happen with complexity, but I want to emphasize a different aspect of it, because that aspect of it has been emphasized by many people. I want to go a different direction. I think a lot of this has really nothing necessarily to do with complexity. The simplest network that does Pavlovian conditioning is just four nodes.

That's it. You don't need... this is not a trillion-parameter crazy thing with 900 different kinds of bonds. This is not... at its root... complexity may give you interesting things. Some people bring up quantum, you know, some of the quantum magic. That may give you extra bells and whistles, maybe, but the fundamental thing that I'm harping on here is not about complexity. It's about the fact that math — the world of mathematics, which I think extends far beyond mathematics, and we can talk about that, too.

Well, I'll tell you. I'll tell you what I mean. It offers you things that range from simple truths, meaning static patterns like the truths of number theory or, you know, the distribution of primes, just like static patterns. Beyond that, it offers you other kinds of patterns that basically are recognizable as competencies. They're ways to navigate problem spaces and solve problems. Treating those as constraints is really, I think, shoehorning this whole thing into a conventional framing that doesn't hold it. They're not constraints. They're enablements and free lunches, and evolution absolutely exploits them.

But evolution is not the source of them. So evolution and selection were not the source of the fact that these gene regulatory kinds of... And by the way, it's not just gene regulatory networks. It's a property of many different kinds of networks. So we recently showed that models of ecological systems, so like predator-prey kind of dynamics, same thing: many different kinds of learning. This is brand-new work again, not published yet. Random electrical circuits — random, not selected, not engineered — random electrical circuits, all kinds of interesting properties. And all of these: they don't come from physics, they come from math.

And what I mean by it goes beyond math is the following. Most — not all, I know it's not all, but many, and I think most mathematicians, when you ask them: first of all, the things that you study, do they originate from the physical world? Are there aspects of physics that explain why the complex numbers and the quaternions don't act the same, or why, you know, you can't have like a quadratic formula kind of thing for degree six and above? Like those facts, do they come from physics? They'll say, 'No, this is an independent set of truths that we're studying.' And then you say to them, 'Okay, so are these just like random things?' And no, no, of course not. They're a structured space that we can systematically investigate and find connections between them. And are they things you can just make up by definition? Are they invented or discovered? And they say, 'Well, no, because once you commit to the idea of an empty set and a successor of empty set, eventually you find out that e equals 2.78 whatever — you don't get to pick, you don't get to say it's 17, you just sort of discover what it is, you're forced into it.' So that space from which these truths are drawn and that we can investigate — I'm sticking with the terminology of Platonic space, just because this is what mathematicians recognize immediately; this is what Pythagoras and Plato, you know, said years ago, okay.

So we already know that there are facts and patterns that are not part of physics, that you can't just fire the math department and hope the physicists will tell you why e equals 2.78 whatever — that's not going to happen. There's nothing you can do in physics to change it. At the beginning of the Big Bang, if you tweaked all the constants, you'll get a different universe; you're not going to get a different value for the natural logarithm. So that's fine, and we can say that there is a space of truths that are fundamental, and patterns that are fundamentally not physical, but they do impact the physical world, because they often serve as explanations for things that happen in the physical world. If you want to know why there's this kind of particle but not that kind of particle, eventually the answer is, oh, because this mathematical object has a certain symmetry group. Or, you know, why these come out at 7 and 11 years — it's because those are prime numbers. And if you want to know why those specific numbers are prime, you go to the math department; you don't keep asking the biologist.

And then typically that's where things stop. So people assume that the contents of that space are the things that we study in mathematics: low-agency static patterns that sit still long enough for you to apply the formal rules of doing mathematics to them. So I have a strange hypothesis on this. I suspect that that space contains not only the truths of mathematics as kind of the lowest level of things, but the entire range of high-agency types of patterns that we would otherwise recognize as behavioral competencies, aka types of minds. I think that space contains kinds of minds. And I think that minds are fundamentally not physical. I mean, this is crazy for a biologist to say nowadays, but that's what I think. I think the things we see in biology, in developmental biology, in behavioral science, and cognitive science, and the things we see in computer science — we can talk about that — are fundamentally drawn from that space. And what we provide physically are interfaces for it. When we make robots, embryos, biobots, AIs, whatever — when you write an algorithm, when you make a machine, when you produce a living being, what you're doing is basically providing an interface for a whole ecology of patterns of different degrees of agency. And they all fundamentally draw from that space.

So when I say it goes beyond math, I mean that I would define math as — I'm not a mathematician, so I don't know where I get off defining that — but for me, math would be the behavioral science of the lowest level of that space. And I don't actually know how low that level is, because no one's really looked at it. We may even be off there. But I think math is the behavioral science of one layer of that space. And other layers of that space are what people study when they are doing behavioral science, cognitive science, computer science, developmental biology. They're actually studying the behavior of other patterns in that space. So that's what I mean. I think they're all on one spectrum. And I think that the relationship between minds and bodies is exactly like the relationship between mathematical truths and physics. That's fundamentally what I think is going on.

Yeah.

Yes.

Sure. Yeah. So the first and most important thing I'll say is that some of these patterns are, as I said, simple and low-agency things that the kind of engineering exploitation approach is appropriate for.

But let's just keep in mind that some of the rest of these patterns are things that we recognize as kinds of minds. In other words, yes, there's a technology-engineering practicality aspect here, but we very quickly get into the ethics aspect, because for some of these patterns, exploiting them gets into the same issue as dealing with what people see as physical living beings. So it is not all the same. Anytime you do this kind of work, you have to very clearly...

Yeah, yeah. Yeah, I mean, I don't use that word because I don't know exactly what it means, but some of these patterns matter in a moral sense, like, to various degrees. And so that means that when you do this work, you have to think very carefully about what kinds of patterns you're actually dealing with, and what kind of lived experience are you providing for them? Because that's what happens. When you build these interfaces, you're providing a first-person lived experience for some level of pattern. And you know, sometimes these are minimal. I mean, we know we have to have some sort of a graded response. You know, we don't sort of sit still so that we don't crush bacteria and things like that. We kind of go about our life. Nevertheless, you have to think carefully about some of these things, because they become significant. Having said that, the fundamental premise of what you just said, I think, is absolutely right, and we're doing this. This is what we're doing in our lab: we're trying to figure out ways to detect and then benefit from specific free lunches that you don't pay for in a conventional way. I think that's massively powerful. I think if we can actually figure out how to do this, and how various interfaces facilitate novel patterns that you didn't have to either engineer, select, or train, the implications for technology and for society and for improving life, from biomedicine all the way through, you know, environmental kinds of things — the potential is absolutely enormous. And that's exactly right. I think these patterns are all around us, and we're mostly blind to them, not having the tools to work with them, but also the conceptual frameworks where we say, well, it had to be selection, you know, which... it's emergent.

Like, I think one of the worst things that slows down progress in this area is the concept of emergence, because people use it to be a kind of stopper for further research. So you do something, and then you get this unexpected result, and then people say, well, it's emergent. So what does that mean? They say, well, there are certain truths that just hold in our world for some reason. And so what happens then? If you think these are just random truths that hold, you're not motivated to study the structure of the space from which these truths are drawn, and why did you get this one as opposed to some other one? Because if you knew why — if you had a story about why that happened — it wouldn't be emergent anymore. It would just be a conventional story. So in many ways, emergence just means surprise. It just means you got surprised, you didn't see it coming. And it's only emergent until somebody can explain the specificity of why this as opposed to that. When they do, then it's not emergent anymore; then it's just conventional. So I think we have to get over this idea of magic emergence and really ask ourselves: what are the free lunches? How do we quantify them? And that's what we've been working on: quantifying the effort. So quantifying the effort you put in and what you got out. And that delta between what you worked on and what came out is a quantification of that free lunch. And then the specificity: how do we get the ones we want? How do we suppress the ones that we don't want?

Well, so we do in parallel — we do the biology and the computer science in parallel.

The reason is this. In biology, the ingressions that you get — so ingression is sort of a borrow of Whitehead's term for when these patterns from that space actually start to manifest in the physical world. So I call it ingression. So these ingressions in biology are extremely impressive. They're super cool. And then there are many examples. I mean, I have a talk online where I go through, you know — what's like the best one? What's the best example?

Well, I mean, okay, one example — let me just finish the previous thing, so I'm going to forget, and then we can talk about it. The thing with biology is they tend to be very impressive.

But you can never really quantify or prove anything, because in biology you never know all the mechanisms. So if I tell you, 'Look at this, you know, it wasn't engineered, evolved, or trained,' somebody says, 'Well, there's a quantum microtubule or something in there somewhere that you didn't know about.' So in biology we never quite know all this stuff, right? But you can make minimal computational models where you have all the pieces. They're right down to the bare metal — you know everything that's there. You can see the algorithm, and I like very simple, minimal things. It's not some enormous parameter model where you don't know what's going on. It's like the first thing we started studying was sorting algorithms — just a few lines of code, bubble sort. Now we have even simpler things: 1D cellular automata and things like that. So you know what all the hardware looks like, and yet it's doing things that you didn't know it was going to do. And that's not the same as complexity, it's not the same as unpredictability, it's not perverse instantiation. It is something that our standard ways of thinking about these things don't yet handle. It's not magic. I fully believe this is not an example of saying that there's some sort of mystical thing that we're never going to understand. I'm going in the exact opposite direction. I'm saying that's how it's been treated up until now — it's been treated as this emergence thing. No, we have to actually get a handle on how this works and how to work with it. So the reason we do it on the computational side is that even though the kinds of things you get — the kinds of competencies that ingress — are not as impressive as in the biology world, which is, I think one thing that evolution is really good at is giving you better interfaces that pull down better, higher-level patterns, which in turn — right, there's like a feedback loop that biology is amazing at. They're not as impressive, but they can be quantified. They can be rigorously characterized, and so we go kind of back and forth. Yeah.

Yeah.

Yeah, there are many, depending on which you think would be more impressive. I mean, I'll just give you a couple of examples. So one example — and this is true for all of these things — is that once you see them, you sort of adjust your expectations and you go, 'Well, yeah, I guess that's what happens.' And I think it's really important to think ahead of time what you would have predicted would happen, right? So...

Imagine you're an alien venture capitalist and somebody comes to you with a proposal, and here's the proposal. I want you to give me a billion space dollars. Here's what I'm going to do. There's this planet Earth, and there are these like mammals running around. And all I'm going to do is select for survival. That's it. Just survival. Nothing else. And sooner or later, they're going to put up a Webb telescope and walk on the moon and do quantum chemistry. All I'm going to do is select for survival. Would you give me any money for that? I wouldn't. If somebody said to me, 'I'm just...

Going to select for survival, and eventually we'll get quantum chemistry.

It's nuts, and nobody really thinks twice about it because, well, I guess, what else could have happened, right? But really, if you were in the position ahead of time to put any money on whether you think that was going to happen, you wouldn't think that would ever happen. You would think that you would have to select for specific things. So that's one example.

I show another video that I found on YouTube that's kind of wild. This guy, he has a turtle, and if I asked you ahead of time, okay, here's this slow, shy reptile. I want it to act like a cat. I want it to move at the speed of a cat. I want it to be playful like a cat. What would you need to do? Well, millions of years of evolutionary selection, or maybe some crazy neuroengineering that we don't know how to do. So he puts a little skateboard under the turtle. And luckily he has a cat which is super fun. So as soon as he puts the turtle on the skateboard, the turtle immediately figures out that you can push off of the ground, and it takes off, and it's chasing the cat, and the cat's like, 'What the — you know?' and trying to get away from it, and it wants to play. So it's looking at him, it's looking at the cat, it's chasing the cat... that's all it took to unlock this amazing thing.

So, you know, what is that? So for millions of years these turtles — what, they were sort of ready for this? Or why does it not take...? Then you have things like the slipping goat, you know, which was in the 30s and 40s. Some guy had a goat that was born without four legs, and so it learned to walk upright, and eventually it died, and he dissected it, and you found out that basically a lot of the adjustments that you need to do bipedal locomotion — so the spine, the hips, all the musculature, all that stuff that took a really long time during evolution to walk upright — yeah, it was all there, right? So that's just other people's examples.

We have tons of stuff on the xenobots, on the anthrobots, these weird kinds of things where there is no conventional story of where these came from, and I just think it's not enough to wave your hands and say, well, it must have been an emergent feature of evolution. No, we have to be able to predict them. We have to be able to facilitate the ones we want and suppress the ones we don't want. And if we can have novel competencies in Bubble Sort — five or six lines of code that have been around for 80 years, that every CS 101 student has had to play with — and we didn't know they were there until we looked last year, then it's got to be all around us. And then these complex things that we're building must have massive... and this is where it gets to some of the machine learning stuff that we can talk about. I think we're just not very good at noticing these things, and we tend to brush them off as one of these three things — these three paths of paying for effort — and I don't think they tell the whole story.

Yeah, I think that's 100% true. So let me just tell you really briefly about this ratchet. So this is the work of Federico Pegorini in my group. What we were looking at was the ability of small gene regulatory network models — so basically just coupled ordinary differential equations — to learn from experience. And so here are the things we found. If you give them stimuli, you will see phenomena like habituation, sensitization, Pavlovian conditioning. They can count to small numbers, things like that. When you do that, at least in some of these networks, their causal emergence goes up — meaning that, like phi kinds of metrics — meaning the degree to which the system is more than the sum of its parts rises.

So that's number two. Number three, as their phi goes up, they get better at learning. So to the extent to which you are a coherent entity, you are a better learner, right? And we know that — just to give you a simple example from biology. So you have a rat. The rat learns to push a lever, get a pellet reward. No individual cell had both experiences. The cells of the foot touched the lever. So who has this conditioning information? Well, the rat. What's the rat? The rat is the collective. So the extent to which you are an integrated collective allows you to learn things that your parts don't know. So these networks can learn. More learning gives you higher causal emergence. Higher causal emergence makes you a better learner. And this is the new stuff that's in review right now. If you force them to forget — because we found stimuli that force these networks to forget things they've already learned, that's very important for biomedical reasons. There's lots of memories your tissues have that you would like them to forget. When you force them to forget, you don't lose the gains you made in causal emergence.

That is insane. So first of all, what that means is that you have a positive feedback ratchet that points in one direction. It points upwards, just as you said, whereas entropy always points downwards and always points towards disorder. This thing always points upwards in intelligence and agency. It means that as you learn things and forget things, you just keep climbing up in terms of your ability to learn and the degree to which you are a new causal entity over and above your parts. So that points upwards. It is not a piece of physics in this story, it is not a piece of biology in this story, it is not selection. This is an extremely generic thing that applies to many things besides molecular pathways. And it's right there for you to use in the world whenever you make an interface that is describable by ordinary differential equations. That is crazy. And I think we've barely scratched the surface of this kind of stuff. I think that kind of thing is exactly the opposite of things like entropy, which kind of point in the other direction. Yeah.

Driving in Boston is a hell, and I think that definitely helps with finding the right way. It's mind-bending and life-saving technology. I highly recommend.

So, okay. So there's a few interesting things. There's a lot we don't know, but I'll tell you what we do know. First of all, we have — at least in some cases, we have now learned to see the target state directly. So in some cases, the set point with respect to which the loss function is calculated is stored bioelectrically as a bioelectric pre-pattern, and we can see it. We have tools now using voltage-sensitive fluorescent dyes and genetically encoded voltage reporters, where you can actually see this pattern. And we know that's the goal pattern because we can also change the pattern. When you change the pattern, the cells will build something else. So that is the reference towards which they are building. Now, what we also know is that the larger collective is really good at deforming the action space for their parts to make them get in line. The individual cells don't need to know this, and I—

It's a little more — let me describe what happens, and maybe that's a way to describe it, but so this is the reason why the individual cells don't need to know what they're building, and I don't think they do. I can't prove that. I don't think any individual cell knows what a finger is or how many you're supposed to have, but the collective absolutely does.

And what I mean by top-down transduction is this. First of all, the behavioral version. When you wake up in the morning, you may have extremely abstract, high-level goals — financial goals, social goals, research goals. And then your cells don't know what any of those things are, but in order for you to execute on those goals, you have to get up out of bed and go do them, which means that the sodium, potassium, and other ions have to cross your muscle cell membranes for you to walk and do those things. So your body — this is the everyday miracle of voluntary motion. People talk about mind-body medicine and yoga and placebo effects and hypnosis and whatever — sure, but this is an everyday thing. It's an incredible thing that our bodies are constructed in a way that the chemistry is made to dance in a way that is in tune with large-scale goals that nothing else in your body knows anything about. That transduction — making the lower levels do what they need to do so that higher-level goals are met — is massively important.

And the way that brains got that was to exact a much more ancient trick that has to do with body shape. So here's an old experiment. This isn't my work. This is old, old data. You take a newt or something like that, an amphibian, and you take the tail and you graft it onto the flank, onto the side. What will happen after that is that that thing remodels into a limb. And the most interesting thing to think about are the cells at the tip of that tail. So they're tail tip cells sitting at the end of a tail, they are locally fine. There's no damage, there's no injury, there's no problem. Why are they turning into fingers? Because as far as they're concerned everything's cool. But the collective knows perfectly well that you don't have a tail there, you need to have a limb there. And it issues a bunch of commands that eventually filter down to the molecular biology needed to change gene expression and make cells move and turn those cells. So you get the idea. So you get this top-down control going on.

And in that case, what every level has to know is how to hack its own parts. It has to know how to behavior-shape its parts. It has to know how to give stimuli, rewards, and punishments, and signals, and prompts, and all this kind of stuff to its parts to get the parts to do what they need to do. The parts don't need to know — in fact, I doubt they know what they're working on — which raises the interesting question, which people email me every day, which is: what if we are part of... what is a higher level that is deforming our shape, right? And we know that happens — when you get a letter from the town telling you to do your garbage differently or something. Well, who sent me that? It wasn't the lady behind the glass at the town hall, it was the town. Well, what's the town? Well, we all are. And so we know there's stuff like that, and I'm sure there are things like this that we don't even know that we're part of. But the point is, all of these higher levels — the degree to which they have good causal emergence is the degree to which they can deform the action space for their parts to do things the parts don't know anything about.

So, here's my current hypothesis, and this is where hopefully we'll have this paper out early next year. I think it has to do with stress. And what I mean by that is stress is spatially distributed. I think what evolution has done — and we've done a lot of simulations on this, and now the biology of it will hopefully be published soon. What biology has done is to capitalize on what happens when the error measurement is non-local. So I'll give you a very simple example. Imagine a group of cells here. And there's a cell down here that actually belongs up there. Let's say you're trying to remodel the tail into a limb, and you got a cell here that needs to be up there.

So this cell calculates an error with respect to what the bioelectrics say you should be doing. It's wrong. The cell is unhappy. Normally it would try to reduce its own stress and get where it needs to go. But the cells that are between here and there, they're in the correct position. They're not going to let you through because they're happy. They're not going to move. And so it's like the whole thing has now frozen down — it's like in an Ising model, you freeze the system and there are parts that still can't align the way they're supposed to go. But here's what the biology does. The cell is unhappy, everybody else is fine. But the stress molecule is leaky, meaning that the cell doesn't just sit there being unhappy. It sort of lets the stress leak out. And what happens then is because they're all using the same molecules to indicate stress state, they get stressed out. It's not their stress, but they don't know that. It doesn't matter. The stress molecule doesn't have metadata on it that says, 'Oh, don't worry, this isn't you.' They're like, 'No, we're all stressed out.' And what happens then is that it raises the temperature of the system in the simulated annealing sense — not physical temperature, but the plasticity of the system goes up. So everybody's now a little more plastic, a little more willing to try new things, and then eventually it can get where it's going. Then the temperature of the whole thing can come down. So stress sharing that way, letting it be leaky, is a way to make my problem become your problem. It's a cheap way to do cooperation without new mechanisms for cooperativity, for altruism. The cells don't need to know or care that this other cell is in trouble. They just want their own stress to go down. So one way you can think about this whole loss function thing is that everybody can look at the pattern, which is available to them all, see where they are, but the stress is...