Michael Levin is a biologist at Tufts University working on novel ways to understand and control complex pattern formation in biological systems. Please support this podcast by checking out our sponsors:
- Henson Shaving: https://hensonshaving.com/lex and use code LEX to get 100 free blades with your razor
- Eight Sleep: https://www.eightsleep.com/lex to get special savings
- LMNT: https://drinkLMNT.com/lex to get free sample pack
- InsideTracker: https://insidetracker.com/lex to get 20% off

EPISODE LINKS:
Michael's Twitter: https://twitter.com/drmichaellevin
Michael's Website: https://drmichaellevin.org
Michael's Papers:
Biological Robots: https://arxiv.org/abs/2207.00880
Synthetic Organisms: https://tandfonline.com/doi/full/10.1080/19420889.2021.2005863
Limb Regeneration: https://science.org/doi/10.1126/sciadv.abj2164

PODCAST INFO:
Podcast website: https://lexfridman.com/podcast
Apple Podcasts: https://apple.co/2lwqZIr
Spotify: https://spoti.fi/2nEwCF8
RSS: https://lexfridman.com/feed/podcast/
Full episodes playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4
Clips playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41

OUTLINE:
0:00 - Introduction
1:40 - Embryogenesis
9:08 - Xenobots: biological robots
22:55 - Sense of self
32:26 - Multi-scale competency architecture
43:57 - Free will
53:27 - Bioelectricity
1:06:44 - Planaria
1:18:33 - Building xenobots
1:42:08 - Unconventional cognition
2:06:39 - Origin of evolution
2:13:41 - Synthetic organisms
2:20:27 - Regenerative medicine
2:24:13 - Cancer suppression
2:28:15 - Viruses
2:33:28 - Cognitive light cones
2:38:03 - Advice for young people
2:42:47 - Death
2:52:17 - Meaning of life

SOCIAL:
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/lexfridman
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Reddit: https://reddit.com/r/lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

It turns out that if you train a planarian and then cut their heads off, the tail will regenerate a brand new brain that still remembers the original information.

I think planaria hold the answer to pretty much every deep question of life.

For 1 thing, they're similar to our ancestors.

So they have true symmetry, they have a true brain, they're not like earthworms, they're, you know, they're much more advanced life form.

They have lots of different internal organs, but they're these little, they're about, you know, maybe 2 centimeters in the centimeter to 2 in size.

And the first thing is planaria are immortal, so they do not age.

There's no such thing as an old planarian.

So that right there tells you that these theories of thermodynamic limitations on lifespan are wrong.

It's not that well over time of everything degrades.

No, planaria can keep it going for probably, how long have they been around?

So these are the actual, so the planaria in our lab are actually in physical continuity with planaria that were here 400 million years ago.

The following is a conversation with Michael Levin, 1 of the most fascinating and brilliant biologists I've ever talked to.

He and his lab at Tufts University works on novel ways to understand and control complex pattern formation in biological systems.

Andrej Karpathy, a world-class AI researcher, is the person who first introduced me to Michael Levin's work.

I bring this up because these 2 people make me realize that biology has a lot to teach us about AI, and AI might have a lot to teach us about biology.

To support it, please check out our sponsors in the description.

And now, dear friends, here's Michael Levin.

Embryogenesis is the process of building the human body from a single cell.

I think it's 1 of the most incredible things that exists on Earth from a single embryo.

I think it's maybe the most magical process there is.

And I think 1 of the most fundamentally interesting things about it is that it shows that each of us takes the journey from so-called just physics to mind, right?

Because we all start life as a single quiescent, unfertilized oocyte, and it's basically a bag of chemicals, and you look at that and you say, okay, this is chemistry and physics, and then 9 months and some years later you have an organism with high-level cognition and preferences and an inner life and so on.

And what embryogenesis tells us is that that transformation from physics to mind is gradual, it's smooth.

There is no special place where, you know, a lightning bolt says, boom, now you've gone from physics to true cognition.

And so we can see in this process that the whole mystery, the biggest mystery of the universe, basically, how you get mind from matter.

How do we get from information encoded in DNA and make physical reality out of that information?

So 1 of the things that I think is really important if we're gonna bring in DNA into this picture is to think about the fact that what DNA encodes is the hardware of life.

DNA contains the instructions for the kind of micro-level hardware that every cell gets to play with.

So all the proteins, all the signaling factors, the ion channels, all the cool little pieces of hardware that cells have, that's what's in the DNA.

The rest of it is in so-called generic laws, and these are laws of mathematics, these are laws of computation, these are laws of physics, of all kinds of interesting things that are not directly in the DNA.

And that process, you know, I think the reason I always put just physics in quotes is because I don't think there is such a thing as just physics.

I think that thinking about these things in binary categories, like this is physics, this is true cognition, this is as if, it's only faking, these kinds of things.

I think that we really have to understand that it's a continuum, and we have to work up the scaling, the laws of scaling, and we can certainly talk about that.

There's a lot of really interesting thoughts to be had there.

So the physics is deeply integrated with the information.

The DNA is integrated in some sense in response to the laws of physics at every scale.

The laws of the environment it exists in.

Yeah, the environment and also the laws of the universe.

I mean, the thing about the DNA is that once evolution discovers a certain kind of machine, that if the physical implementation is appropriate, it's sort of—and this is hard to talk about because we don't have a good vocabulary for this yet—but it's a very kind of platonic notion that if the machine is there, it pulls down interesting things that you do not have to evolve from scratch because the laws of physics give it to you for free.

So just as a really stupid example, If you're trying to evolve a particular triangle, you can evolve the first angle and you evolve the second angle, but you don't need to evolve the third.

That's a gift for free from geometry in a particular space.

And if you evolve an ion channel, which is ion channels are basically transistors, right?

They're voltage-gated current conductances.

If you evolve that ion channel, you immediately get to use things like truth tables.

You don't have to evolve the logic function.

And the fact that if you have NAND gates, you can build anything you want.

All you have to evolve is that first step, that first little machine that enables you to couple to those laws.

And there's laws of adhesion and many other things, and this is all that interplay between the hardware that's set up by the genetics and the software that's made, right?

The physiological software that basically does all the computation and the cognition and everything else is the real interplay between the information and the DNA and the laws of physics of computation and so on.

So Is it fair to say just like this idea that the laws of mathematics are discovered, they're latent within the fabric of the universe, in that same way the laws of biology are kind of discovered?

Yeah, I think that's absolutely, and it's probably not a popular view, but I think that's right on the money, yeah.

Then embryogenesis is the process of revealing, of embodying, of manifesting these laws.

You're just creating the capacity to reveal.

Yes, I think, again, not the standard view of molecular biology by any means, but I think that's right on the money.

Some of our latest work with these xenobots, right?

So what we've done is to take some skin cells off of an early frog embryo and basically ask about their plasticity.

If we give you a chance to sort of reboot your multicellularity in a different context, what would you do?

Because what you might assume by looking at the thing about embryogenesis is that it's super reliable, right?

And that really obscures some of its most interesting features.

We get used to the fact that acorns make oak trees and frog eggs make frogs and we say, well, what else is it going to make?

That's what it, you know, that's what it makes.

But the reality is, And so you look at these skin cells and you say, well, what do they know how to do?

Well, they know how to be a passive, boring, two-dimensional outer layer keeping the bacteria from getting into the embryo.

Well, it turns out that if you take these skin cells and you remove the rest of

embryo, So you remove all of the rest of the cells and you say, well, you're by yourself now, what do you want to do?

So what they do is they form this little, this multi little creature that runs around the dish.

They have all kinds of incredible capacities.

They navigate through mazes, they have various behaviors that they do both independently and together.

Basically they implement von Neumann's dream of self-replication, because if you sprinkle a bunch of loose cells into the dish, what they do is they run around, they collect those cells into little piles, they sort of mush them together until those little piles become the next generation of xenobots.

So you've got this machine that builds copies of itself from loose material in its environment.

None of this are things that you would have expected from the frog genome.

There's nothing wrong with their genetics.

Nothing has been added, no nanomaterials, no genomic editing, nothing.

And so what we have done there is engineer by subtraction.

What you've done is you've removed the other cells that normally basically bully these cells into being skin cells.

And you find out that what they really want to do is to be this, their default behavior is to be a xenobot, but in vivo, in the embryo, they get told to be skinned by these other cell types.

And so now here comes this really interesting question that you just posed.

When you ask where does the form of the tadpole and the frog come from, the standard answer is, well, it's selection.

So over millions of years, right, it's been shaped to produce the specific body width that's fit for froggy environments.

Where does the shape of the xenobot come from?

There's never been any xenobots, there's never been selection to be a good xenobot.

These cells find themselves in the new environment, in 48 hours they figure out how to be an entirely different proto-organism with new capacities like kinematic self-replication.

That's not how frogs or tadpoles replicate.

We've made it impossible for them to replicate their normal way.

Within a couple days, these guys find a new way of doing it that's not done anywhere else in the biosphere.

LBW Well, actually, let's step back and define what are xenobots?

JF So a xenobot is a self-assembling little proto-organism.

How much is it biology, how much is it robot?

At this point, most of it is biology because what we're doing is we're discovering natural behaviors of the cells and also of the cell collectives.

Now 1 of the really important parts of this was that we're working together with Josh Bongard's group at University of Vermont.

They're computer scientists, they do AI, and they've basically been able to use an evolutionary, a simulated evolution approach to ask, how can we manipulate these cells, give them signals, not rewire their DNA, so not hardware, but experiences signals.

So can we remove some cells, can we add some cells, can we poke them in different ways to get them to do other things?

So in the future, there's gonna be, you know, we're now, and this is future on published work, but we're doing all sorts of interesting ways to reprogram them to new behaviors.

But before you can start to reprogram these things, you have to understand what their innate capacities are.

Okay, so that means engineering, programming, you're engineering them in the future, and in some sense, the definition of a robot is something you in part engineer versus evolve.

I mean, It's such a fuzzy definition anyway, in some sense many of the organisms within our body are kinds of robots.

And I think robots is a weird line because we tend to see robots as the other.

I think there will be a time in the future when there's going to be something akin to the civil rights movements for robots, but we'll talk about that later perhaps.

Anyway, so how do you, can we just linger on it?

From when does it start, and how does it become the glorious xenobot?

Yeah, so just to take 1 step back, 1 of the things that a lot of people get stuck on is they say, well, You know, engineering requires new DNA circuits or it requires new nanomaterials.

You know, what the thing is, we are now moving from old school engineering, which used passive materials, right?

That things like, you know, wood, metal, things like this, that basically the only thing you could depend on is that they were gonna keep their shape, that's it.

It's on you as an engineer to make them do everything they're going to do.

And then there were active materials and now computation materials.

This is, you're now collaborating with your substrate because your material has an agenda.

These cells have billions of years of evolution, they have goals, they have preferences, they're not just going to sit where you put them.

That's hilarious that you have to talk your material into keeping its

It's like getting a bunch of cats or something and trying to organize the sheep out of them.

It's funny, we're on the same page here, because in the paper, this is currently just been accepted in Nature by Engineering.

1 of the figures I have is building a tower out of Legos versus dogs, right?

If you build out of Legos, you have full control over where it's going to go.

But if somebody knocks it over, it's game over.

With the dogs, you cannot just come and stack them.

But the good news is that if you train them, then somebody knocks it over, they'll get right back up.

So as an engineer, what you really want to know is what can I depend on this thing to do?

A lot of people have definitions of robots as far as what they're made of or how they got here, design versus evolve, whatever.

I think as an engineer, what you want to know is how much can I depend on this thing to do when I'm not around to micromanage it?

What level of dependency can I give this thing?

Which then tells you what techniques do you use.

How much, how intelligent is your substrate?

And so now we're moving into this area where you're working with agential materials.

So basically the material has agency, meaning that it has some level of, obviously not human level, but some level of preferences, goals, memories, ability to remember things, to compute into the future, meaning anticipate.

You know, When you're working with cells, they have all of that to various degrees.

Is that empowering or limiting, having material that has a mind of its own, literally?

So it raises difficulties because it means that if you're using the old mindset, which is a linear kind of extrapolation of what's going to happen.

You're going to be surprised and shocked all the time because biology does not do what we linearly expect materials to do.

On the other hand, it's massively liberating.

And so in the following way, I've argued that advances in regenerative medicine require us to take advantage of this, because what it means is that you can get the material to do things that you don't know how to micromanage.

If you had a rat and you wanted this rat to do a circus trick, put a ball in the little hoop.

You can do it the micromanagement way, which is try to control every neuron and try to play the thing like a puppet.

And maybe someday that'll be possible, maybe.

And this is why humanity, for thousands of years before we knew any neuroscience, We had no idea what's between the ears of any animal.

We were able to train these animals because once you recognize the level of agency of a certain system, you can use appropriate techniques.

If you know the currency of motivation, reward, and punishment, you know how smart it is, you know what kinds of things it likes to do, you are searching a much smoother, much nicer problem space than if you try to micromanage the thing.

And in regenerative medicine, when you're trying to get, let's say, an arm to grow back or an eye to repair a cell birth defect or something, do you really want to be controlling tens of thousands of genes at each point to try to micromanage it, or do you want to find the high-level modular controls that say, build an arm here?

You already know how to build an arm, you did it before, do it again.

It's both difficult and it challenges us to develop new ways of engineering, and it's hugely empowering.

Okay, so how do you do, I mean, maybe sticking with the metaphor of dogs and cats, I presume you have to figure out the, find the dogs and dispose of the cats.

Because, you know, it's like the old hurting cats is an issue, so you may be able to train dogs.

I suspect you will not be able to train cats.

Or if you do, you're never gonna be able to trust them.

So is there a way to figure out which material is amenable to hurting?

Is it in the lab work or is it in simulation?

Right now it's largely in the lab because we are our simulations do not capture yet the most interesting and powerful things about biology.

What we're pretty good at simulating are feed-forward emergent types of things, right?

So cellular automata, if you have simple rules and you sort of roll those forward for every agent or every cell in the simulation, then complex things happen, you know, ant colony algorithms, things like that.

The difficulty with all of that is that it's incredibly hard to reverse.

So this is a really hard inverse problem, right?

If you look at a bunch of termites and they make a thing with a single chimney and you say, well, I like it, but I'd like 2 chimneys.

How do you change the rules of behavior-free termites so they make 2 chimneys?

Or if you say, here are a bunch of cells that are creating this kind of organism.

How do you control all the individual low-level rules, right, all the protein interactions and everything else?

Rolling it back from the anatomy that you want to the low-level hardware rules is, in general, intractable.

It's an inverse problem that's generally not solvable.

So right now, it's mostly in the lab, because what we need to do is we need to understand how biology uses top-down controls.

So the idea is not bottom-up emergence, but the idea of things like goal-directed test-operate-exit kinds of loops, where it's basically an error minimization function over a new space.

It's not a space of gene expression, but for example, a space of anatomy.

So just as a simple example, if you have a salamander and it's got an arm, you can amputate that arm anywhere along the length.

It will grow exactly what's needed, and then it stops.

That's the most amazing thing about regeneration is that it stops.

It stops when a correct salamander arm has been completed.

So that tells you that's a means-ends kind of analysis where it has to know what the correct limb is supposed to look like.

So it has a way to ascertain the current shape.

It has a way to measure that delta from what shape it's supposed to be.

And then it will keep taking actions, meaning remodeling and growing and everything else until that's complete.

So once you know that, and we've taken advantage of this in the lab to do some really wild things with both planaria and frog embryos and so on.

Once you know that, you can start playing with that homeostatic cycle.

You can ask, for example, well, how does it remember what the correct shape is, and can we mess with that memory?

Can we give it a false memory of what the shape should be and let the cells build something else?

Or can we mess with the measurement apparatus, right?

So it gives you those kinds of, so the idea is to basically appropriate a lot of the approaches and concepts from cognitive neuroscience and behavioral science into things that previously were taken to be dumb materials and you'd get yelled at in class for being anthropomorphic if you said, well, my cells want to do this and my cells want to do that.

And I think that's a major mistake that leaves a ton of capabilities on the table.

So thinking about biologic systems as things that have memory, have almost something like cognitive ability.

But, I mean, how incredible is it you know, that the salamander arm is being rebuilt not with a dictator, it's kind of like the cellular automata system.

All the individual workers are doing their own thing.

So where's that top-down signal that does the control coming from?

And how does it tell everybody to be like, whoa, whoa, whoa, slow down, we're done?

So the first thing to think about, I think, is that there are no examples anywhere of a central dictator in this kind of science because everything is made of parts.

And so we, even though we feel as a unified central sort of intelligence and kind of point of cognition, we are a bag of neurons, right?

All intelligence is collective intelligence.

There's this this is important to to kind of think about because a lot of people think okay there's real intelligence like me and then there's collective intelligence which is ants and flocks of birds and you know termites and things like that and and you know and and And maybe it's appropriate to think of them as an individual, and maybe it's not.

And a lot of people are skeptical about that and so on.

But you've got to realize that we are not, there's no such thing as this indivisible diamond of intelligence that's like this 1 central thing that's not made of parts.

And so if you believe, which I think is hard to get around, that we in fact have a centralized set of goals and preferences and we plan and we do things and so on, you are already committed to the fact that a collection of cells is able to do this because we are a collection of cells.

In our case, what we do is we navigate the three-dimensional world and we have behavior.

This is blowing my mind right now because we are just a collection of cells.

So when I'm moving this arm, I feel like I'm the central dictator of that action.

Like all the cells here are collaborating in some interesting way.

They're getting signal from the central nervous system.

Well, even the central nervous system is misleadingly named because it isn't really central.

Again, it's what we— It's just a bunch of cells.

There are no singular indivisible intelligences anywhere.

Every example that we've ever seen is a collective of something.

We're used to that, you know, we're used to, okay, this thing is kind of a single thing, but it's really not.

I mean, we're jumping around, but that's something that you look at as the bioelectrical signal versus the biochemical, the chemistry, the electricity.

Maybe the life is in that versus the cells.

It's the, there's an orchestra playing and the resulting music is the dictator.

That's Dennis Noble's kind of view of things.

He has 2 really good books where he talks about this musical analogy, right?

See all Lex Fridman transcripts on Youtube

Michael Levin: Biology, Life, Aliens, Evolution, Embryogenesis & Xenobots | Lex Fridman Podcast #325