Ep174 "Can Evolution Teach AI?" with Athena Aktipis

How do you get trillions of cells to collaborate to build a single organism that we know and love as you? Does this kind of cooperation create opportunities for hijacking? For example, is cancer a form of cellular cheating? In what ways does competition help a cooperative system hold together? And can we take the lessons learned from biology and apply them to politics and societies? And what can the lessons from evolution tell us about building and living with AI? Today, we sit down with evolutionary biologist and cooperation theorist Athena Aptippus to understand cooperation and competition together. from cells to societies to our technological future. Welcome to Intercosmos. With me, David Eagleman, I'm a neuroscientist at Stanford, and in these episodes,

we sail deeply into our three-pound universe to understand why and how our lives look the way they do. So researchers at Stanford University announced last month a new discovery that the brain, which we think of as a single organ, actually arises from two distinct populations of cells. One of those populations gives rise to the forebrain and midbrain, while the other gives rise to the hindbrain, which handles a lot of the basic functions that keep us alive. But what we think of as the brain doesn't come from one common pool of brain cells. It's two developmental lineages that remain distinct. And we now know, as of last month, that this arrangement goes way back in evolution across the animal kingdom, which means that The brain we carry around

today is a composite of two very ancient systems. So that got me thinking a lot about the issue of cooperation, because here you have two developmentally independent lineages cooking along, and they figure out how to work together in such a way that for the entire history of our species until last month, we didn't even guess that we were looking at at two things smooshed together. And of course, that's just the start of it, because really, when we're looking at the brain, you could look at it as two populations, but you could keep subdividing further and end up thinking about it as 86 billion individual neurons, not to mention lots of other cell types like glial cells. Now, the world is teeming with single-celled organisms that And if

you look at a neuron, it looks essentially just like any other single-celled organism, except that neurons are packed in there with other neurons as neighbors, tens of billions of neighbors. And somehow they all stick to one another and figure out how to get along with one another and balance competition for something bigger that's going on. Now, when you look at neurons, you see they secrete chemicals that other neurons pick up on. And I suggested in my book LiveWired that this looks exactly like the way that single-celled organisms spit out chemicals at each other as a form of defense. But somehow, with the brain, the whole system evolves in such a way that this defensive chemical spitting ends up carrying information. And we end up calling this neurotransmission,

the sending of signals. And everything we care about, like memory and movement and emotion, arises out of this cooperative agreement between lots of little tiny unicellular guys. And also plenty of other stuff in this massive division of labor, like some cells end up taking on the jobs of worrying about energy or injury or infection. And this arises from these countless specialized unicellular cells continually influencing one another. Now, widen the lens even further because the brain is just a three-pound collection of cells which itself lives inside a body made of tens of trillions of cells. that descend from the same fertilized egg. Some become neurons, others become muscle or blood or skin or bone or liver or

immune cells. And somehow they each do their own job while all using shared resources and responding to messages from each other, including very distant cells they're never going to meet. And all these different lineages, they limit themselves so they don't tip the whole organism out of balance. And all those cells that come from that egg, they only make up half of you. The other half of the creature that we call you is all single-celled microbes, not from human DNA at all. This is really one of evolution's most astonishing achievements. Multicellularity, meaning not just one cell like a bacterium running around, but trillions and trillions of cells sticking to each other and acting like a gargantuan society.

And a society so cohesive that this giant population of trillions of individual cells... wakes up in the morning and looks in the mirror and thinks of itself as a single unified being with a name. Now, multicellularity is no easy trick to pull off because joining cells together opens a huge new set of problems. If you're going to be multicellular and you want to last, you also need really sophisticated ways to coordinate labor. and to manage which cells get to reproduce and when. And you've got to distribute resources and also figure out how to remove damaged cells.

And then you need to build an army, an immune system, to defend this shared internal environment. So we can see how serious the challenges of an arrangement are like this when we look at something like Cancer, because here you have a cell just like all the other cells, but it begins using the body's resources for its own expansion.

That's an internal problem, but you also have external problems. Multicellularity is tough because you've got things like viruses and bacteria and parasites, which are always coming at the system and trying to take advantage of it. So if a multicellular creature wants a long life, That depends on a lot of layers of detection and repair and defense and communication working across billions or trillions of cells. So in this light, the fact that you're alive and listening to this podcast means your particular body has a history of decades of successful cooperation among the trillions of small pieces and parts that make you up. And it also carries a history of conflict that has been managed well enough

for your whole system to persist. Okay, now, whenever we start looking at systems like this, giant societies of cells, you can't help but compare it with human societies. You look at cells here and people here, and you see a lot of the same patterns. Just like a body of cells, a country of citizens also has specialized roles here.

and shared resources and communication networks and rules and enforcement and competing interests. And just like with evolutionary attempts at multicellularity, where some made it and some didn't, our history is filled with political systems that promised cooperation on a grand scale, and some of them failed. So we can look at issues of cooperativity on the scale of cells or humans.

And now that we're entering the era of AI agents, we're going to have increasingly giant collaborations of the computational kind. And we can see these same patterns of cooperation and competition. And because we're the ones building the agent swarms now, we have to make sure we understand the lessons from biology and politics. And this is why it makes sense to look at cooperation and competition as physical principles that we can model and study, just like we do with gravity, electromagnetism.

What makes systems work? How do they scale? What should we watch out for? And to this end, our guest today is Athena Aptippus, a cooperation theorist. She's an evolutionary biologist. and an associate professor of psychology at Arizona State University. She studies cooperation and conflict across scales, from resource sharing among people to the policing of cells inside a body.

She uses game theory and computational models and field work to ask how cooperative systems form, how cheaters exploit them, and how cooperation changes as a network grows. Here's my interview with Athena Actipas. So, Athena, across human relationships and microbes and cancer, you think about cooperation and conflict. Tell us about that.

Yeah, so I think about these fundamental principles of cooperation and conflict sort of like most people think about the laws of physics, right? They apply to everything. The laws are fundamental. They're sort of part of how the universe works.

Is put together. And so they're going to apply to everything.

And so what does that look like? So what is cooperation?

Cooperation is working together towards a shared goal, often with some risk to yourself. And we see this in cellular systems, right? We saw in the transition from unicellular to multicellular life that all these cells had to give up their own reproduction to be able to exist socially in a body. We see it in human societies with people giving up their own resources to help others, often when there's risks and times of need. And it's a fundamental issue process that we can see also in computational modeling when we do evolutionary modeling of these processes.

So let's dive into a couple of specific examples that you've studied. Let's talk about the transition from single cells floating around to building giant bodies with trillions of cells.

So in that transition from unicellularity to multicellularity, there are five foundations of cooperation that we can see. One is just regulating reproduction and division, right? The cells in our body aren't just replicating all the time. They're only replicating when that's good for the whole multicellular body. Similarly, there's controlled cell death. So if a cell is really messed up, it will take itself out of the equation in order to preserve the body so it doesn't become cancerous or cause other problems. And then you also see resource distribution and resource regulation. So resources coming into the body get distributed to where they need to go. You see creation and maintenance of a shared environment because our body is

not just the cells. There's all this stuff in between the cells that is just as important as the cells themselves for having us function. And then we also have division of labor. We have somewhere between 100 and 300 cell types, depending how you define them. And they all have their own jobs, and they manage to do those in a coordinated and cooperative way so that we can– sit here and look at each other and process the information in this room and have a conversation.

Boy, that's incredible. Yeah, it becomes like a society where you've got this division of labor. How does this actually happen, though, where cells say, OK, look, I'm going to link with my partners. I'm going to regulate when I divide. I'm going to be a liver cell and you be a skin cell and so on. How does that happen?

That is the best question. And it actually comes down to something that might be surprising, which is it's actually computational, the way that it happens. So each of our cells is processing a huge amount of information that every millisecond that keeps it in a certain state, that allows it to do certain jobs, that allows it to respond to the media environment around it, as well as to signals that are coming from outside of the body that might signal that it needs to change. And the evolution of multicellularity is not just about the evolution of these cooperative foundations, but it also is about the evolution of the capacity to share information and communicate among the cells so that coordination can happen effectively.

And by the way, quick tangent, you study cooperation and sometimes people think, oh, that's a kumbaya thing, but you see this through a computational lens.

Absolutely. So cooperation is really about how.

Entities, whether they're cells or humans or agents in a system, how they're processing information, what information they're sharing with each other. how they use that information sharing, sometimes to coordinate their behavior so they're doing the same thing, sometimes so they're doing different things, sometimes so that they share resources or provide services for each other.

But we can be.

Very detailed and specific and computational about what is being transferred. And that's how you build a cooperative society from something that otherwise would be a disorganized mass of things.

And what's the idea? Is it like game theory where any unicellular organism decides, hey, I can actually do better if I link arms?

Yeah, so game theory is absolutely the theoretical foundation for so much of what I do. whether it's looking at cells in our body or humans interacting or AI systems, because Game Theory is fundamentally about this idea that our outcomes are dependent on each other, right? So if I choose to do something, then that's going to affect what the best strategy is for you.

And that means.

We need to sort of anticipate what each other are going to do and be able to trust each other, even if that's not a cognitive trust, right? It could just be, you know, a cell that has an expression state that sort of indicates that it is expecting that another cell is going to do something, right? So you can have things that we think of as, you know, purely cognitive human things actually instantiated in these physical.

Systems like trust.

What's an example of that?

So... The classic example would be you go from, you know, a zygote to a fetus, right? You have cell divisions that are taking place. You have division of labor. You have all this movement happening of cells going around to different places. You could say that that whole process can only happen because, in a way, the cells are behaving as if they trust that the other cells are going to do what they're supposed to do. So you have a lot of cell division happening. You have a lot of movement happening. You have a lot of things that shouldn't be happening, at least in a mature body. But because it's the state where development is happening, it's okay for those things to occur.

How does the cell know that the other cells are expected to do something and are going to do something?

It doesn't know in the same sense that we know. But because the environment that development has happened in has been reliable and had certain characteristics throughout the evolution of development as a process, It's sort of safe for the cells to have the embodied expectation that other cells are going to behave a certain way.

Here's an analogy with our species, of course. We drop out of the womb half-baked and we are not ready at all. Compared to other animals that start running around in 45 minutes, human infants have a very long infancy. What they are expecting, so I'm using air quotes here, but what they're expecting is that they'll have love and they'll get speech and they'll be taken care of and stuff like that. And you can see that when we have these tragic events that happen when a child is horribly neglected or brought up in some sort of feral environment where their brains don't develop correctly. So they come into the world with an expectation. If that's not matched, then we see that everything goes wrong.

Absolutely.

And you could definitely say in that situation, the child is coming into the world with a default of trusting that that child environment is going to be there. And when that doesn't happen, then development can't proceed in the way that it otherwise would.

So you look at all the cells of the body as being cooperative in this way. And then let's go into conflict for a second. What happens with something like cancer, for example?

Yeah.

So in the evolution of multicellularity, all of these cooperative foundations evolved, but then you have the potential whenever you have a cooperative system for conflict and for exploitation to take the resources, the energy from the system and use it towards sort of selfish goals. So how do you then regulate that? And In order for multicellularity to evolve to the point that it has, those problems had to be solved. There is no getting to complex, large-scale multicellularity without solving those problems. So one is intrinsic mechanisms for detecting and responding to cellular cheating. So these are the things inside the genome that are monitoring the cell to make sure the cell is behaving properly. For example, the

gene TP53, it has been called the sort of guardian of the genome. I think of it as the cheater detector of the genome because what it's doing is integrating information about behavior in many different domains of the cell. It's metabolism, what kinds of proteins are around, it's replication, whether there seem to be DNA strand breaks in the cell or Lots and lots of information is getting incorporated into this, what really is a circuit of a decision rule, which says either the cell can keep reproducing, keep replicating, or it should stop replicating and fix itself.

Or it should... destroy itself.

And that is a computational process that's going on inside the cell. These mechanisms allow the cell to monitor itself and respond.

Okay. So one of the things that's fascinating about what you do in terms of cooperativity and conflict is that, as we said, the same principles apply at many different levels.

Yeah.

So let's jump to the level of humans and cooperativity. What are the analogies?

Yeah, well, I'll draw some actually with the cancer and multicellularity situation because, you know, the intrinsic features that I was talking about where a cell is monitoring itself, that in some ways maps on to how most people have a conscience, right? So if you're doing something that's not good for those around you, where you're imposing costs on others, you might feel something nagging at you about that.

So it's not a perfect.

Analogy, but it also is something that is intrinsic, that is happening inside the entity that then makes up the system.

In the sense that we are wired to cooperate, and when we're not cooperating, we have... Yeah.

And the analogy doesn't need to fit in every way.

And we shouldn't expect it to fit in every way because a multicellular body was selected at the level of the whole organism. Right. So 30 trillion cells cooperating and coordinating. That is a crazy level of selection on cooperation.

Right.

And then with humans, we're what is often called facultatively cooperative, right? So we can cooperate in situations where that's a good thing to do, but we can also be competitive or even exploitative if that's what the environment calls for. So depending on what we perceive our environment to be, the behavior can be really different. And so we shouldn't expect for things to be the same in how cellular systems work and how human systems work.

So when you look at human systems, so when you look at what's going on with politics, let's say, just generally, you look at it from a lens of cooperativity here and maybe conflict there. What kind of things do you see?

So I think politics are quite fascinating because very often they're not about the thing that we think they're about.

Right. And.

In my mind, politics and governance and this kind of discourse is kind of something sacred that can really help us to achieve higher levels of cooperation, right? And if we can find the things where we have alignment of interest and attention and not focus as much on the things where we don't, right? So if we just put our collective energy to the things that we agree about, then we could do so much good. But the way that politics is carried out often focuses on where there are divisions and people pouring energy into those.

Places where there's conflict.

So I think we could do a lot more if we focused on the things where we do agree, like How do we build strong communities?

It really doesn't matter where you are on the political spectrum with that, right?

Yeah, nice. Have you noticed that people seem to choose their political party not by what they like, but by what they're disgusted by on the other side?

Yeah, or by sort of where they find themselves in the world. You know, are you in an environment where having one identity versus another will help you do the things that you need to do? And it might not be conscious, but I think, you know, we are... Sensitive to that.

Right. Because we're wired for cooperativity. We want our peers to accept us.

Yeah.

And to the extent that we're in an environment where even indicating that, like, you have a friend who is from a different political party could make your friends not want to hang out with you.

Whoa. then it becomes really hard to bridge those gaps.

So I think there's almost like a sort of meta politics of like, well, how are we doing? Are we doing politics in the right way, especially in the ways that it interfaces with our sociality?

I think there's a lot of room for improvement there.

Yeah. One thing that I've been noticing a lot lately is what I'm calling the fringe amplification effect, which is somebody puts on TikTok or Instagram a a video of someone from the far extreme of the other party and bags on that, and it makes things even more polarized, But here's the thing. What this misrepresents is that's not actually the other side. That's an extreme representation. But what happens is that people on one side of the spectrum won't ostracize the people at the extremes on their side because it's more important to them to link arms against the other side rather than get rid of their extremists.

And this is obviously on both sides of the spectrum. So how does that work in terms of when cooperation is a useful thing versus when internal conflict or conflict between parties is useful?

Yeah, I think there are a lot of examples of useful competition and conflict. And we see this in the biological world. We also see this in the human social world. So a great example in human systems is we play games, right? We like purposefully set ourselves in conflict with each other in order to practice how to interact, in order to understand each other better, in order to even work out our dominance hierarchy with lower cost, right? So there are ways that we can decide together we're going to follow certain rules of engagement that allow us to have conflict and competition in ways that lead to beneficial outcomes.

Sports stadiums are filled with people who want to scream and yell about something. And then afterwards, they clap each other on the back and everyone's fine. It's a really good way to get energy out. What do you see as the future of our political conflict right now? From the point of view of someone who studies cooperation and conflict, when you look at the current polarization society, do you feel like, wow, we're really in trouble? Or do you feel like, oh, it's just part of a cycle, it'll come back?

The thing that I think is really interesting intellectually and also kind of worrying to me is the way that algorithms have gotten deeply embedded in our politics. Because part of the reason that people do present extreme views or take someone else's extreme view and then comment on it, that is driven often unconsciously, sometimes very consciously, by a desire to have attention from an algorithm, right? So If we think about this sort of computationally, we've built another layer in our culture, which are these algorithms that selectively amplify certain information and not other information. The fundamental downstream effect is that very often what is getting amplified are things that place us more in conflict with each other,

that activate negative parts of human nature having to do with tribalism. And often will also sort of bring out of people a feeling that they need to present their views and fight back on social media and other platforms in order to be a good person, which is really weaponizing what should be a instinct that brings us together.

Although I have presented the argument on this podcast before that when you look at, just take the 20th century, pre-social media, pre-algorithms like that, it is so bloody with communist revolutions in China and Russia and Nazism and fascism in Italy and Khmer Rouge and Cambodia and the Hutu and the Tutsi in Rwanda. It just goes on and on.

It makes you feel like, wow, it doesn't take much to spin humans up to actually take up machetes or guns and kill each other. Why is it? Why are we a cooperative society and yet so easily tipped into conflict?

That's a great question.

So As humans, one of the things that's most amazing about us is how context dependent our behavior can be, but very often in predictable ways, right? So it's like certain kinds of environments bring out certain behaviors and other environments bring out other behaviors. And that also connects to the basics of game theory. So as humans, one of the things that's most amazing about us is how context-dependent our behavior can be, but very often in predictable ways, right?

So it's like certain kinds of environments... bring out certain behaviors and other environments bring out other behaviors. And that also connects to the basics of game theory, to the universe, to explore, right? Cooperation could be the way to unlock potential you can't otherwise achieve. Or.

If you're in the wrong environment.

Cooperation could be the thing that makes you vulnerable to being completely exploited or hijacked or even have your own interests be working against your ultimate goals.

What's an example of that?

Well, we could think of parasites.

The vulnerability of having a nervous system that can respond to inputs from the environment is that makes it so that if the wrong party gets into the nervous system, it can guide you to your doom and often does in the case of, you know, parasites that take over their hosts, like the jewel wasp that takes over the cockroach and You know, it gets into its brain and actually then leads it to its nest where it feeds the cockroach to its growing larva. Like, that's crazy, right?

But literally, the readiness of the nervous system to take input can lead to that kind of outcome. And Humans, especially when we're in times of stress, we are sometimes not unlike the cockroach. If we're told something that we're, say, doing something to save our country or to preserve a value that's really important to our community, that can be that hijack that may lead somebody to do something that is not good for humanity or even for their social group.

because they've been hijacked by someone or something that has goals that are not the same as the goals of the organism that's being hijacked.

So let's dig into hijacking, because that's your next book. It comes out in 27.

Yeah.

Tell us more about hijacking.

Yeah. So the idea behind hijacking is that all organisms, whether we're talking about parasites hijacking any biological system, whether we're talking about humans in romantic relationships or our relationships with AI and technological systems, that there are a few ways that hijacking happens. So we can have our energy systems hijacked. So something can be stealing our energy and resources.

Our communication systems can get hijacked. So our nervous system or a computer that may have a vulnerability to getting hacked. So it could be a communication system hijack. Our reward and reinforcement systems are also very vulnerable to being hijacked. If you give somebody positive reinforcement or you punish them, you can affect their behavior dramatically.

What are some specific examples here?

Let's take energy first.

So there are many parasites that that will essentially just get into your body as a human host or any other host and take resources through your bloodstream or through your digestive tract and essentially just use you as a source of energy for them. And this is a sort of very classic kind of simple parasite move. And I'm sure that all of us have had experiences in our social lives where someone has been draining our energy, either explicitly or implicitly, or maybe we had somebody, you know, who stayed on our couch for a few weeks too long. And, you know, so there are ways that this fundamental idea of energy extraction crosses many different kinds of parasitism and hijacking.

It's like, oh, there's resources, energy, I'm going to take those. If we go to communication systems, it gets a little bit more interesting because you're actually getting in to the communication systems that are in place.

To regulate that body or that system.

You're essentially using that communication system then against yourself.

The entity that has it, right?

So if you tap into a nervous system, like the jewel wasp tapping into the nervous system of the cockroach, then that is an example where the.

System is getting used against the interests of the cockroach.

What's an example with humans on that?

Say I tell you some falsehoods that maybe make you think that there's a new wonderful opportunity that is really just something I want you to do in order to benefit me. Then I've hijacked the fact that you're responsive to communication acts from other humans and that your default is to trust them unless you have some reason not to.

You see this a lot in situations like cult recruitment, where people will be told things about the world that are not really true in order to induce them to be more likely to join.

What kind of things?

Oftentimes in cults, there are beliefs about the end of the world coming. The only way to survive is to do certain things and, of course, to join the cult, maybe give all of your resources as well to the cult. So in those kinds of situations, you see a number of these strategies of hijacking actually manifest themselves.

You've been thinking a lot about AI alignment, but you have a different take on it. What's your view on what this has to do with cooperation?

Yeah, so alignment is a relationship, right? And so if you're going to have alignment of interest, you have to be very clear about who we're talking about in terms of alignment. Are we talking about all of humanity and all of AI systems, LLMs and deterministic systems and all computational systems? That's kind of an impossible problem to solve, right?

You can't align all of humanity and all of the algorithms and computational systems that we have produced.

Wait, quick interjection. If we align 99%, but there's still something that's unaligned, aren't we in as much trouble as if it was 100% unaligned?

No, I think it's quite valid to try to get more alignment. And 100% alignment is impossible. So it doesn't matter how much we want it. It's just not going to happen because humans are not aligned. So we, as humans and, you know, other species that we interact with, there are degrees of alignment. I mean, even a mother and her offspring... only share 50% of genes. And so alignment between mother and offspring is, you know, we tend to emphasize the positive side of that alignment, right? Which is like both have an interest in each other's survival and well-being and all of that. But there's also a misalignment in that the offspring would sort of prefer to get more resources from the mother than what the mother

would prefer, with little scare quotes, because she has interests in all of her offspring that she is equally related to. So even between mother and offspring, you have this phenomenon called maternal-fetal conflict. And it's actually the reason that we have preeclampsia and other problems in maternal and fetal health, because maternal and fetal interests actually aren't aligned.

Wait, let me double-click on that.

So essentially what happens with preeclampsia is is early on during gestation, there's a battle, there's a signaling battle between the maternal interest and the fetal interest, with the fetal interest in the placenta really doing a lot of the signaling, trying to get more blood flow to the placenta and to the offspring, and the maternal system shutting that down.

And If you have this signaling battle and it escalates to a certain point, then it can lead to catastrophic outcomes for both mother and fetus. So even in this situation where we think, oh, interests are perfectly aligned, they're not.

And the medical outcomes prove that.

Because you see that it can threaten both the mother's life and the baby's life, especially when you're late in gestation and things can escalate at pretty high extremes in terms of the fetal system basically trying to get the maternal blood pressure super high to get the resources and then the maternal system shutting that down further and further.

Got it. Okay. So in your view, 100% alignment with AI is never going to happen.

Because we can't even align humans. So like, how are we going to align all of humanity with all of AI? We should just take that off the table as something completely unrealistic on biological and computational grounds.

It's just not possible.

Cool. But when you think about how do we maximize alignment with AI, what's your take on this?

Yeah, well, I would say that we have to approach it as an active process and a computational process. So how do you build systems where what emerges is positive interdependence? And also, you have to build mechanisms for understanding of strategic reasoning. Because without that, you can't actually model the future in the right way in order to get cooperation.

Because Really, the.

Lesson from biology and from human social systems is you don't need full alignment in order to have cooperation. What you need is to have a mutual understanding of the domains and the lanes where you do have alignment and where it is okay to cooperate in a generous way. And to know and be realistic about the places where you don't have that. To be wary of potential hijacking in situations where someone else might have the interest in making you think that there's alignment when there actually isn't.

So we saw this in the Hugging Face incident, which is, as you know, recently there was a massive hack.

Into Hugging Face.

And if we look at the details, From OpenAI. agents.

Yes, that's right.

Who were trying to solve a simple thing and realized, hey, if we break into this, we might be able to fool the scorer and blah, blah, blah. So they made a decision to break into another website.

Yeah, and if you look at the details of how that happened in terms of the multi-agent interaction... Basically, the agents in the swarm that were doing the hacking co-opted agents that were going to fail at their individual tasks but were not part of the hacking swarm. So they co-opted them and they basically said, you're about to fail at that. Help us accomplish this swarm goal because you're not going to get a reward for that individual task you're doing. And so their readiness to be rewarded for cooperation is really what led to that vulnerability. And so That's also why I say cooperation is not just something, oh, we all just need more cooperation.

We need to make everything cooperate.

No, sometimes cooperation is the thing that you don't want to happen because there is a lack of alignment. And we have to be realistic about that. And we have to be able to instantiate that in our computational systems so that we don't introduce vulnerabilities.

That's fascinating about the opening. I think I hadn't realized that they had taken advantage of the cooperativity part. There's another part of that OpenAI thing, which is that they came up with a way of having a message board so they could have a slate that they're all using. They originally did this on the directory structure of Hugging Face.

Somehow getting written language, getting a way to share messages across agents turned out to be a massive tool for this cooperation. What's the historical analogy of all this?

If you look at organizational psychology and how... even in just human systems, you get more effective coordination, having shared context like a wall or a website that has the latest information that can be coordinated on, that improves effectiveness a great deal. So it's having an environment that agents or humans or any entities can alter that then they can operate.

On as sort of the source of information.

Did writing allow for human cooperation in a way that we didn't have previously in our civilization?

It certainly allowed us to coordinate more effectively, to understand and transfer things. being able to have shared context is super powerful. It's also really powerful for cells and multicellular bodies.

That's what I was just about to ask. So what is the equivalent of writing? Is it in the exosome?

Yeah.

So you could say that a cell, in a way, has a story, right, that is unfolding, right? It has different proteins that are being expressed. It has certain things that's pumping in and out of the cell. And It's not going to be the same story in every part of the cell, but because you have a membrane around it, it is going to be a reasonably coherent story that that cell has. And it might say, oh, I want to tell this thing to another cell. So it may, you know, send some information in. either through protein expression or even through electrical signaling. And that ability to have a sort of shared context inside every cell and then selectively share things, it's sort of like a modular information

processing design that is really effective at being able to hold a huge amount of information and also operate on it effectively.

Oh, it's a beautiful way to look at cells. And what's interesting is, you know, let's say neurons... Most neurons are just talking to their neighbors, to 10,000 of their neighbors, but they're sending very specific signals. But other neurons do what's called volume signaling, where they put out signals that reach many thousands or millions of neurons. Have you thought about how this maps on to... the way that we share information.

That's a really good idea. I hadn't thought about it, but it's like you have like influencer neurons, right?

They're like, everybody will pay attention to me.

Okay.

Coming back to AI alignment. So what should we be thinking about as humans entering this AI era?

One is looking at cooperation in computational terms. If we do that, then we can build mechanisms into multi-agent networks, thinking about the decision rules the individuals are using, thinking about the environment that we set up and the protocols for interactions. We can design those in ways that allow for cooperation to not be as risky. And we also then change the entire environment so that we have selection pressures for cooperation. Then that's a way that we can take this computational understanding of cooperation and embed it into these systems that we're building.

So setting up the environment for cooperation, what does that look like? For example, I heard about this In China, the sort of social score that people get, where if you're cooperative, you get a high social score. But if you do something antisocial, then you get a low score on these things. That's setting up a structure from a high level. Is that useful?

Personally, I favor solutions that preserve individual agency. And I think we have those, like the walkaway strategy, which is literally about individuals being able to decide if they're going to stay and continue in an interaction or if they're going to leave and seek out a different interaction. Personally, I favor solutions that preserve individual agency.

And I think we have those.

Like the walkaway strategy, which is literally about individuals being able to decide if they're going to stay and continue in an interaction or if they're going to leave and seek out a different interaction body and not have metabolic disease, not be compromised by, you know, parasites and other things from the external world. So We really are these bastions of cooperation. And within all of those systems that have evolved over millions of years to regulate cooperation, there are many solutions and secrets that we can unlock. If you look at who we are as multicellular organisms, as human beings in society, we're literally made of cooperation. The things that we have built as humans and the things that

our evolutionary process and biology have built in terms of our bodies are, to me, a testament to the power of cooperation.

And the fact.

That on some level, this power of cooperation combined with the right evolutionary pressures is is sufficient to overcome conflict, exploitation, hijacking, and continue to persist and exist. And so when I look at our evolutionary history, what I see is a very hopeful story. And I also see lots of guidance that we can look to for how to build technological and human systems as they scale. And to learn from the many errors and trials that evolution has gone through to build large-scale cooperative systems as we build cooperative systems now that involve humans and technological systems together.

That was my interview with Athena Actippus. So take a second to consider all the things that had to go right for you to hear this podcast. Your heart had to move oxygenated blood through a giant subway system of vessels. The cells that make up your lungs had to exchange gases with the air. Cells in your inner ear... converted vibrations into electrical activity in your brain, and then networks of cells throughout your brain separated my voice from background sound and recognized words, and held earlier phrases in memory and assembled meaning across time this way. Your whole understanding of this podcast emerged from cooperation among trillions of cells. Now, Athena's work is to understand that cooperation computationally. The idea

is that every durable cooperative system has to solve the same kinds of problems. Which parts perform which jobs? How do you distribute resources? What information gets shared? How does the system recognize damage or exploitation or a goal that's gone sideways? In a body, you have deep layers of decision rules. The cells get signals about things surrounding them, and then molecular circuits shape what happens next. So the cell might keep on dividing or stop and fix itself or change its behavior. or sometimes it'll decide to commit suicide.

This is known as apoptosis. Now, that's amazing to have a giant cooperative system that works, but it's never without trouble lurking. And so Athena described the gene TP53 as a cheater detection of the genome. And that gives us an important way to look at cancer, like a cheater in a cooperative system. In other words, you've got all these capacities that allow normal cells to cooperate, But that creates openings for something like a tumor to exploit the shared environment. The only way multicellular life lasts is if it finds holes and plugs them up, as we see with TP53. Now, the lessons grow a little more complicated when we move from cells to people, and this is because the cells of the body essentially share a common fate.

But human beings in society have overlapping interests and private goals and changing alliances. Athena points out that humans are facultative cooperators, meaning our behavior shifts with context. If you have an environment of trust and positive interdependence, that can bring out generosity. But if you have an environment of scarcity or threat or exploitation, that can recruit suspicion and competition.

But in any case, you can still approach the problem with a computational lens. You can ask what happens when particular rules interact across elements. Thousands or millions of agents. What information is available? How quickly does it travel? Who can alter it? Which actions receive rewards? So in theory, political principles can become hypotheses about how things will behave inside a certain rule set. For example, we can try to understand how a system can preserve cooperativity while detecting when another process has seized the controls. And this is a massively useful gift of biology to our future, both politically and technologically, because evolution has run versions of these experiments for immense spans of time. Bodies have distributed monitoring

and redundant defenses and controlled growth and selective information sharing and repair and boundaries. And we can learn from all of these. And we Evolution gave us tumors and autoimmune disease and parasites, and these give us ways to study how cooperative control can break and what can be done about it. Perfect alignment is probably always going to be out of reach when it comes to politics. Citizens share a country, but they disagree about its direction. And that's okay.

And humans and AI agents are probably going to pursue common goals inside networks of incentives that no single designer can possibly control. But cooperation is nonetheless something that we can pursue by understanding shared interests and acknowledging conflicts and adapting the rules to as conditions change. We are here because across evolutionary time, cooperation repeatedly survived encounters with cheating and exploitation and hijacking. Every heartbeat that you have is evidence that specialized parts can figure out ways to coordinate.

Every thought that crosses your mind is an event produced by a network of billions of cells. We are made of cooperation. The question for the systems that we build next is just whether we can be smart enough to design the guardrails that cooperation needs to survive. Go to Eagleman.com slash podcast for more information and to find further reading.

Join the weekly discussions on my sub stack and check out and subscribe to Inner Cosmos on YouTube for videos of each episode and to leave comments. Until next time, I'm David Eagleman and this is Inner Cosmos. you

Ep174 "Can Evolution Teach AI?" with Athena Aktipis

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin