"Life, Machines, and a Foundation for Synthbiosis"

Watch Episode Here

Listen to Episode Here

Show Notes

This is a ~21-minute talk titled "Life, Machines, and a Foundation for Synthbiosis" that I gave to the Santa Fe Institute (https://santafe.edu/) conference on "Symbiogenesis in the Age of Machine intelligence" (https://www.santafe.edu/events/symbiogenesis-in-the-age-of-machine-intelligence-2027", September 17 and 18, 2026)

CHAPTERS:

(00:00) Space of embodied minds

(03:03) Rewriting bioelectric set points

(06:08) Xenobots and novel beings

(12:04) Molecular networks and learning

(15:26) Mathematics as information source

(18:38) Exploring anatomical morphospace

PRODUCED BY:

https://aipodcast.ing

SOCIAL LINKS:

Podcast Website: https://thoughtforms-life.aipodcast.ing

YouTube: https://www.youtube.com/channel/UC3pVafx6EZqXVI2V_Efu2uw

Apple Podcasts: https://podcasts.apple.com/us/podcast/thoughtforms-life/id1805908099

Spotify: https://open.spotify.com/show/7JCmtoeH53neYyZeOZ6ym5

Twitter: https://x.com/drmichaellevin

Blog: https://thoughtforms.life

The Levin Lab: https://drmichaellevin.org

Transcript

This transcript is automatically generated; we strive for accuracy, but errors in wording or speaker identification may occur. Please verify key details when needed.

Main Episode

[00:00] What I'd like to do today in the twenty minutes that I have is tell you a brief overview of an empirical research program in our group that includes both biological and computational components, which connects to some very deep issues that I think we're all interested in. I'm gonna go fairly quickly. If anybody's interested in the stuff, catch me later. There are long-form versions of all of this.

[00:00] One of the most important things, I think, for us going into the future is to realize that this is not about humans versus AIs or language models or anything like that. I think we really need to keep an eye on the fact that the space of possible embodied minds is enormously vast, that almost any combination of evolved material, engineered material, software, and this other thing that I'll mention briefly at the end is a viable being. All of Darwin's “endless forms most beautiful” are basically a tiny blip here among all the cyborgs, hybrids, chimeras, all these different things.

[00:00] In order to understand how we're going to have an ethical synthbiosis with them—I like this word; it was proposed to me by GPT when I was brainstorming about this beneficial mutual thriving going forward with all sorts of beings—I think we need to ask ourselves what we are, first of all, and how we come to be as embodied beings in the physical world. I'm going to take a couple of minutes to say what we actually are, or at least what our bodies are.

[00:00] First of all, this is the sort of thing that we are composed of. This happens to be a free-living organism called Lacrymaria. This is one cell. It has no brain, no nervous system. It's extremely competent in its tiny little physiological and even morphological agendas. At every level of organization in a living organism, you find problem-solving competencies. You find the ability to reach specific states, to get there even when circumstances change, in all kinds of different spaces: transcriptional space, physiological state space, anatomical morphospace, and so on.

[00:00] So here's how we come to be. When we look at an early blastoderm—maybe a million cells, let's say—we look at this and say, “There's an embryo.” What is there one of? When we say there's one embryo, what are we actually counting? What is there a single thing of?

[00:00] I think what's happening here is that what there's one of is a single set point. It's a goal state in anatomical space to which all of the cells are committed. They share the same goal of a journey that each of us has taken from being a single fertilized egg cell to the species-specific target morphology. All of these cells are aligned physically, but also physiologically and in various other ways, towards a specific future model of what they're going to be.

[00:00] This is a very unpopular view in developmental and cell biology, especially molecular biology. They will tell you that we need to do this in terms of chemistry. Nothing at this level knows anything. Nothing has any goals. Its emergence and complexity are what's going to handle all of this. There are no goal states. These things are emergent. I will argue that, at least in some cases, that is simply factually incorrect because we've learned to read and write those set points.

[00:00] These set points are encoded, at least in part, bioelectrically. We've developed some of the first tools to read and write pattern memories in tissues outside of the brain—all the same things that the neuroscientists do—but we try to ask: What are the cells and tissues thinking about long before they have a brain?

[00:00] Once you can read those patterns, this, for example, is the prospective pattern of the face that's going to be made in the future frog embryo. This is what tells you that the cell collective has a particular set point. We now have a way to rewrite that set point. When you rewrite it using optogenetics and various other techniques that we have, then you can ask the cells to build something completely different.

[00:00] We have some nice two-headed flatworms here that you can see, and I could show you all kinds of novel creatures for hours and hours. But the point is that we now have the ability to start to see what goal states are encoded in the cognitive glue that binds individual cells, competent individual cells, towards larger-scale endpoints in morphospace.

[00:00] This is how we come together. It is the set points that are holding together our components. One of the interesting things about biological materials is that the boundaries of those goal states—in other words, the size of the goals that any system can actually pursue; you could call it a cognitive light cone—are plastic. They scale. So whereas individual cells, both in development and early on in evolution, have tiny little goals.

[05:04] Scalars over a very small region of spacetime: things like pH level inside the cell, hunger level, little tiny goals. But the collective, due to various computational processes of these networks that we've studied, is actually able to represent much larger, grandiose states, like the formation of a limb. This is an axolotl limb. No individual cell knows what a finger is or how many fingers you're supposed to have. But the collective absolutely does, because if you deviate it from that position, they will grow, implement it again, and then stop. Actually, if you do all kinds of things to try to prevent it from doing it, it has lots of clever ways to get around whatever you've done and try to implement the same goal by different means. William James's definition of intelligence.

[05:04] If we want to rewrite the goal state, the cells happily build to whatever. I don't know if they're truly universal, but they certainly can build more than we think they can. Here's a frog where we've asked it to make a fifth limb, and so on. So much for the scale of the goals of these systems.

[05:04] Now we want to ask: what is the content of these goals? The reason we want to ask this is because we are going to be living with all sorts of novel beings that are not easy to classify, who do not share our entire evolutionary history, whether that's your neighbor that has some percentage of their brain replaced with whatever technologies. I don't think we want to be in the position of checking to see if it's more than 50% or less so that we can try to maintain this real human versus machine distinction.

[05:04] This is from a talk I gave to psychologists and psychiatrists, trying to point out that their clientele is going to be extremely diverse soon, and trying to improve the quality of life for beings whose dream symbolism and all these other things that we're used to doing with standard humans are going to be completely different.

[05:04] Could we develop a science of asking: when we do have novel beings—even if they have some of the same components, they don't share our evolutionary history—how can we understand what their competencies are? What are their goals going to be? What is it like to relate to these beings? We typically think of what sets these goals. Either evolution does or, occasionally, design or training. These are the three ways we have of setting it.

[05:04] I'm going to introduce you to a few of these beings that are basically model systems in our lab for trying to develop some understanding of this. First of all, these are called xenobots, and this is what happens when you liberate some epithelial cells. These are cells that are going to be skin from an early frog embryo. You take them away. We do not add anything. We don't change the genome. This is not synthetic biology. There are no scaffolds, no genomic editing. There are no nanomaterials. All we've done is liberate them from the signals that basically force these cells into a boring life as the outer two-dimensional covering of a tadpole and then a frog.

[05:04] They form these self-motile little creatures. Here they are self-assembling. Among other things, if you provide them with loose epithelial cells, they implement von Neumann's dream of a system that runs around and makes a copy of itself from loose materials it finds in the environment. It takes the cells. It combines them into these little balls. The little balls mature and become the next generation of xenobots. They do the same thing. So you get the next generation and the next. We call this kinematic self-replication.

[05:04] They have many other features that normal embryos don't have. For example, they upregulate a bunch of genes associated with hearing and sound perception. Turns out, if you put a speaker under them, absolutely, they can respond to specific frequencies of sound. If you track their calcium signaling and apply various causal emergence metrics that neuroscientists like to use, you find out that the difference between them and the null models is really no smaller than the difference between human fMRI data and their null models. I'm not saying that these things are brain-like. I'm saying that both brains and various other phenomena across implementations have the same features that are important for the things we're interested in.

[05:04] They can form memories. Hopefully, it will be published shortly. They can remember at least two different things for at least 24 hours, and we can now read out those memories.

[05:04] In case you're thinking this is some sort of frog-specific weird amphibian capability, I could ask you what your cells would do if we liberated them from your body. Adult human patients—not embryonic—give tracheal epithelial samples. We buy the cells. We put them in a different environment. They become anthrobots. This is what they look like. They have little cilia on their surface. They have all kinds of cool behaviors. You could not guess from looking at this that, if you were to sequence it, all you would ever see is pure Homo sapiens. 100% Homo sapiens. We haven't done anything to these cells. Just looking at it, you wouldn't know that.

[10:10] Also, looking at the human genome, you would have absolutely no idea that this kind of creature would exist or what its properties are. These guys express 9,000 genes differently than their cells of origin, so half the human genome is differentially expressed. They are younger than the cells they come from, so they manage to roll their age back. They have all kinds of interesting capabilities. They can heal neural wounds if you put them next to a sheet with a big scratch through it, and so on.

[10:10] So one reason I'm showing you all of this is that it starts to bring up some interesting questions. We know when we paid the computational cost—or we think we know—for this to appear. Here's a frog embryo with very specific developmental stages. Eventually, you get tadpoles with specific behaviors. And the story that we tell is that we paid this computational cost in the eons that this genome was bashing against the environment. There were some spandrels, but fundamentally, there was a set of selection forces that supposedly explain the features, the competencies, the shape, the behaviors of the creature we have now. There's supposed to be specificity between what you see and the series of environments that got you here.

[10:10] When did we pay the computational cost to make good xenobots? They have very specific behaviors. They have a developmental sequence. This is an 83-hour-old xenobot that's turning into something. I have no idea what it's turning into. They have, of course, this kinematic self-replication and so on. You could say that they had to do something, but that's not what we're looking for here. As biologists, we would like to understand: where do the specific properties of this creature come from if you can't lean on behaviors? There have never been any xenobots. There's never been selection to be a good xenobot. It's very hard to tell a specific evolutionary story about this.

[10:10] And one of the interesting things that we start to realize is that, even below this level, there are interesting questions about where competency comes from. This is really, fundamentally, I think, this important question of where competency comes from—not just order, but actual competencies recognizable by behavior scientists.

[10:10] I'll show you another one that's cool. If you look at models of gene regulatory networks or molecular pathways, these are not large. There may be four, five, up to 20 nodes. They're represented by systems of coupled ordinary differential equations, and people study them using dynamical systems theory to understand what these gene regulatory pathways are going to do. If, instead of that, you take the lens of behavior science, then you can ask the following question: What type of learning could they do if we treated the individual nodes—the individual genes or molecules in this pathway—as potential targets for stimulation and readout?

[10:10] It turns out that they can form at least six different kinds of memories. They can do Pavlovian conditioning, which is really important biomedically. We're trying to do this for drug-conditioning applications and so on. They can count to small numbers and so on. This is already baked into the substrate of which cells are made. You don't need a whole cell to do Pavlovian conditioning. Molecular networks already do this, and we're, of course, doing this now in real cells and so on.

[10:10] But one of the interesting features is that when you train them—with repeated stimuli so that you can get associative conditioning, habituation, whatever you're looking for—at least for some of them, their causal emergence goes up. With each training stimulus, the causal emergence of the network—meaning the extent to which it's more than the sum of its parts, the way that a rat is more than the individual cells that had contact with a lever or the reward; the association is owned by this rat, by this larger-scale creature—goes up. But what's interesting is that higher causal emergence makes them better learners. The process of learning actually raises causal emergence.

[10:10] Most amazing of all, if you force them to forget—we've developed tools to force them to forget; that's, again, important for biomedical reasons—they do not lose the gains that they've made in causal emergence. So this ends up being an asymmetric ratchet that points upwards for a learning ability and for integrated emergence of some sort.

[10:10] And now, two interesting things. First of all, you might think that evolution has done an amazing job selecting for networks that have this awesome property. It turns out that even random networks are incredibly biased for optimality in this kind of feature. Random networks already do this. Evolution didn't have to do much at all. And that property is a free gift from math. It does not come from any specific feature of physics. It did not need evolution. There are no replicating units here.

[15:08] The material on which evolution works is already primed for an asymmetric ratchet for intelligence over learning capacity and causal emergence. So that's interesting. Some of this kind of amazing property comes from mathematics.

[15:08] Generally, we can start to think about mathematics as a rich source of information beyond physics and biology. Typically, biologists like to think that information either comes from the physical features of the environment or from selection. Here's a pattern. This is a Halley plot of this function in complex numbers. There is no story about physics that you could tell about why this particular pattern versus some other pattern. There is no selection that went into this. There's no design. There's another source of information, which mathematicians have been dealing with for a really long time, that has very rich order that does not belong to the categories that we use in biology as bioengineers and so on.

[15:08] In my group, we've been exploring this concept of a free lunch. Whether it's actually free, I'm not sure. But what's interesting about it is that you get patterns that you do not pay for by either rational design, evolutionary selection, or training. And so one of the things that we've been working on a lot is to figure out what you actually get. This is a static pattern. It's pretty and all, but it's just a static pattern. But what else do you get?

[15:08] We've been exploring with actual experiments this notion that not only do you get static patterns, not only do you get trajectories between patterns, but you get behavioral policies. Some of these behavioral policies look like goal states. When we ask, “Where do the goal states of novel beings come from that were not selected for at that level?” some of these goal states come from the same place that the truth of mathematics comes from. Once you start chasing down “Where did it come from?” you end up in the math department. And there may be other things that you get as well that we don't know yet.

[15:08] But the bottom line is that when you make even very simple systems, you don't just get complexity. You don't just get unpredictability and things like that. You get policies for navigating problem spaces that are well recognized by behavioral scientists. You get patterns at different levels of the cybernetic hierarchy.

[15:08] So at this point, you might think that biology has some kind of magic associated with it. Maybe it's the evolutionary process, although we've already shown that this kicks in long before you have replicators. Maybe it's just complexity. Maybe there's some quantum stuff going on here. So this is the typical view of bioinspired computing and bioinspired engineering. We take things we see in biology. We try to implement them in our technologies.

[15:08] Actually, it doesn't take biology. Even extremely minimal deterministic systems, like this—this is our study of sorting algorithms—have unexpected features that really need to be paid attention to when we ask what kinds of intelligence might exist in unfamiliar places and where it comes from. We have a research program on this. You can think about it as inspiration, as policies that you didn't pay for that allow you to do useful things in your environment. We've been giving robotic bodies to all sorts of static patterns to ask what kinds of minds are frozen in these patterns.

[15:08] And the last thing I wanna show you is this. This is somewhere between a tool and a colleague. We call this Mom Bot. This is, I think, the first system that actually does experiments in morphospace. So this is not synthetic biology. This thing makes xenobots. And what it does is it has an AI that tries to make hypotheses about how to talk to frog cells. It applies stimuli that are electrical, optical, vibrational—lots of different things—looks at the xenobots that are produced, characterizes their behavior, and goes back to revise its hypothesis. It does this loop, basically the loop of science.

[15:08] And I think what's important is that there are three ways to look at this. The obvious way is that this thing is an interface between us, human scientists, and the cellular collective intelligence. It's helping us communicate our goals to the cells. This is what our entire regenerative medicine program is based on. It's trying to communicate novel goals to cells. That is one thing.

[15:08] But another thing it is is an actual intelligence, a synthetic intelligence that, while it's not moving in three-dimensional space, explores anatomical morphospace using the living tissue of the frog cells as its interface. It is a brain in a vat, as are we, and it knows about the outside world—the world it lives in, which is the space of form and function. Whereas we use our retinas and various other things, it uses the frog cells to explore the space that it lives in.

[15:08] Somebody put their turtle on a skateboard, and you can see this simple passive prosthetic has unlocked a completely new behavioral mode for this slow, shy reptile.

[20:13] The whole mom bot system is basically a prosthetic for frog cells that allows them to do things that they otherwise would have never been able to do.

[20:13] What I think is actually going on is that biology and technology are basically inspired by exactly the same thing. I think these are forms from a structured, tractable latent space. This is a very unpopular position. If you wanna hear me talk about it, this is where it is. Fundamentally, I think that we have to take very seriously that this is not just the province of mathematicians. It has one layer that has low-agency patterns that sit still long enough to be amenable to mathematics and formal models. But there are other patterns in that same space that are patterns of physiology, patterns of morphology, of behavior that we would recognize as kinds of minds. I think this is a slightly different way of looking at some of these issues.

[20:13] I think the future Garden of Eden isn't gonna look like this. Here, Adam is naming the animals. It's actually gonna be very weird, and we're gonna have to reimagine what we are in order to have the symbiosis with all these beings.

[20:13] I thank all the people who did the work. There are three disclosures that have licensed some of this technology. Thank you so much.