Athena Aktipis: From Multicellularity to Multi-Agent AI | Intelligence at the Frontier 2026

Transcript

Great, thank you so much. And it's such a pleasure to be here and to have the chance to bring some of my research that I do looking at cooperation across many different systems from looking at our bodies as multicellular cooperative systems to looking at kombucha as a multicellular cooperative system to looking at human societies and how human societies cooperate and then how do we take those insights and use them to understand what's going on with AI? Now, since I only have 15 minutes for you today, I'm only going to be able to give you a small selection of some of the work that we've done and how it's related, but we'll go ahead and kind of jump right in. So, the one of the most

important things to realize is that this question or this sort of problem of cooperation, it is a universal problem that applies whenever you have agents, whether it's cells in a body or humans in societies or AI agents, the question of how do you have cooperation be viable, especially as you scale up without exploiters taking advantage of that cooperative system? And so, when we think about the evolution of multicellularity, this actually was a problem of cooperation.

How do you go from having cells that are operating as individual entities and have them scale up to what we have now just with our own bodies, which is 30 trillion cells that are cooperating and coordinating every millisecond to make us functional. It's quite astounding when we actually think about it. So, we can think of multicellularity actually is this unbelievable cooperative achievement. We have division of labor, we have complex communication, we have all these policing mechanisms, and a lot of this works because there's a sort of shared fate that all of our cells have. They really don't have an opportunity to survive and thrive outside of our bodies.

Now, you might already be thinking, well, there's an exception to that, right? Which is um certain kinds of cancers. And in fact, cancer is a problem of multicellular cooperation. So, when we look at what cooperative foundations are required for multicellularity to be viable, we see that it's important that cells are able to limit their own proliferation, that they have controlled cell death, that they can allocate resources effectively among each other, that they can divide labor, and that they can create and maintain an extracellular environment. These are all foundations of multicellular cooperation that we can think about as we are working on building and scaling multi-agent

systems. For example, this idea of controlled cell death, that a cell is able to notice that something has gone awry and do something called apoptosis, which is controlled cell death. That's something that we can think about in terms of kill switches, for example, for agents. Like having them have the ability to maybe tell themselves if something is going wrong. Now, what happens when these foundations of multicellular cooperation break down? Well, what we have in the middle here are those five foundations of multicellular cooperation I had on the last slide.

On this next round here, or as um pentagon outside, we have what happens when those break down. And those are the phenotypes of cancer. That's what you see when cancer emerges. And on the very outside, we have what are called the hallmarks of cancer. So, this comes from a sort of long-standing characterization of cancer's characteristics. And as you can see, this very much maps on to this framework that my colleagues and I built for understanding cancer as a breakdown of these foundations of multicellular cooperation.

Now, as we look at all of these features, we can see that there are parallels when we're thinking about building technological systems. Um for example, systems that can evade immune destruction, right? Ones that maybe are more resilient to uh being shut down if they're found to be problematic. So, my first book is called The Cheating Cell and it's about this very dilemma, this very issue that we can look at cancer as this breakdown of cellular cooperation and that there is this tension where inside a system, cancer cells inside our body, for example, are are under selective pressure,

evolutionarily speaking, to do things that benefit themselves. To replicate more than they should. To not commit apoptosis, to not take themselves out of the system, to monopolize resources, to actually destroy the environment around them if that lets them benefit. And because of that inherent tension, that means that in order for bodies like ours and other multicellular bodies to be viable, had to evolve many, many mechanisms that suppress that cellular cheating. So, we actually have systems on many levels.

One level is the intrinsic level. So, cells have within them the ability to monitor their own behavior and then change what they're doing based on if they're sort of detecting that there's a problem with how they're behaving. There's also neighborhood systems. So, cells monitor each others that in in a neighborhood and are able to then send each other signals like, "Oh, something seems a little wrong with the proteins you're expressing, for example, or maybe you're infected by a virus. You might want to check that before you replicate.

So, you you have these communication systems that are also happening locally inside the body. And then if all that fails, you have the immune system, which is the sort of system level, you could say more top-down centralized system. But what the immune system does is basically kind of detects that there regions where things aren't going quite right, and it seems like those decentralized mechanisms for regulation aren't working. So, we can learn a lot from looking at the multicellular body as we are scaling these systems up to levels that are completely unprecedented when it comes to technology.

So, does anybody know what this is? You can sort of see it. This is kombucha from my kitchen from about 10 years ago. Um so, the reason that I now have a thriving kombucha lab at Arizona State University where we've taken over sort of whole wing with kombucha is because I started brewing kombucha myself, and then I would come home from work and sit down and Oh, that was not what I hoped to do.

Um and I would look at it and I'd be like, "Huh, what are you?" And, you know, being an academic, I decided I would look online and just see. Well, yeah, what what is this? And at that point there was very very little research on it. So, of course, what did I do? I packed up some of it and brought it into the lab and we started brewing it and uh then looking at various um aspects of, you know, what happens if we challenge it by introducing pathogens from human hands, what happens if we uh actually look at the changes over time in the genetic composition of it.

And it turned out that there's actually a huge amount of cooperation going on metabolically inside kombucha. So, you have yeast and um what the yeast are largely doing is converting these sugars into a very small amount of alcohol and CO2. And you have bacteria that are um metabolizing some of that ethanol and they're contributing to the flavor that we have um that comes from it, but they're also making this sort of cellulose. And a really interesting thing happens when you have the production of CO2 and you have the production of cellulose. So, that's the biofilm that you see, the scoby. And that biofilm is sort of sticky, so it

provides a way actually for the yeast and bacteria that are cooperating to stick together in space to make these little clumps. And then the CO2 that gets produced inside those sort of sticky clumps stays with them until there's enough of it that it makes the whole association of yeast and bacteria and their biofilms rise to the surface. And then what ends up happening is they meet up with all the other little cooperative micro biofilms filled with yeast and bacteria on the top. So, you have the system that's actually selecting for cooperators just by virtue of the spatial dynamics and the physical dynamics of the system. And I This is a

really, really interesting model because it is it's it's entirely decentralized. And it has another really cool feature, which is it seems to be able to protect itself from invaders using something that seems like a multi-species immune system. So, let me show you what some of this looks like. So, on on top, you know, it's a little bit hard to see, um but we have some boat scobies that encapsulated some invaders. You could sort of see those little dark mold-like things. Um and then on the bottom we have some normal scobies that had no invaders.

So, essentially what we found is when there's more of a challenge to the system because there's an invasion, you get a very, very large scoby. And that scoby seems to encapsulate those invaders. If you don't have that perturbation, then you end up with a very small scoby. And we actually found this accidentally as we were doing other research. And um one of the previous students in my lab observed that, "Huh, it seems like these scobies look really healthy when you put an invader in." And it's like, well, they look healthy, but really they're just big because it seems like, you know, really what they're doing is overproducing this biofilm in order to

encapsulate these invaders. So, then, you know, that keeps those invaders from being able to use the resources and take over the system. Now, if we switch gears a little bit to kind of think about the big picture of cooperation, and we can go back in time a little bit. This was my dissertation work on something called the walkaway model, which is a very, very simple rule that allows um to get cooperation in decentralized systems. So, basically, you have individuals, and they could be in partnerships, for example, in a prisoner's dilemma game, or they could be in groups playing public goods games.

And they have a very simple rule, which is always cooperate and leave if your partner or group is insufficiently cooperative. And it turns out that is sufficient to allow for cooperation to be to evolve and be maintained in the these agent-based models. And, you know, as now we're thinking about agents interacting autonomously, these agent-based models of cooperation that me and many of my colleagues have been doing for several decades, I think are extremely important because they tell us about some fundamental principles that we can utilize to help cooperation thrive, including something called positive assortment, which is basically just any way that cooperators

can preferentially interact with other cooperators. So, as we embark now on the sort of new adventure in cooperation with technology, there are a number of problems that we have to think about and try to solve. So, one is how agents interact with each other and how we can use these principles to enhance cooperation. For example, figuring out how they can represent trust and interdependence. There's human-agent cooperation. For example, how do we divide labor effectively with agents? And then there's the human-human cooperation part of it, which is how could we actually leverage what we can do now with AI to help humans cooperate with each other

better? And these are all things that me and my team at the Cooperative Futures Institute are working on. And I'm writing about these questions very regularly on my Substack called Not for Peer Review. If you're interested in kind of hearing what we're doing at Cooperative Futures Institute, this is a good place to keep up to date. These are all extremely short. I write them so you can read them in 3 minutes. Um, and then big picture, many design principles that we can take from evolution, this idea of positive assortment I already talked about, how to police defectors, and do that ideally in a decentralized way, but sometimes you do need to sort of

activate a central system. That's at least a, you know, message from looking at multicellular bodies. Dividing labor, having um communication signals that are reliable, and then having mechanisms for aligning fate. So, having that interdependence be something that the agents understand and know that they're in these situations where you have that. Uh at ASU, I'm the director of the Cooperative Futures Institute. We're working on a lot of these projects, including um building benchmarks for looking at cooperation. If you're interested in this, um me and my team, we have a table set up if you sort of go back through those doors and around the corner, um where you can learn more about all of this. We're really excited

about it and happy to talk about it. We have a bunch of preliminary data. Um and finally, I'm going to leave you um with this, actually. Um and I'm going to turn my sound back on. So, we have a show of a pine The Science of Cooperation. What? I am so sorry, you guys. That was That was the worst possible thing that could have happened right there. Uh you are so awake. So, uh have you guys recovered? Are your nervous systems just completely jarred?

Um so, this is a a show that we have developed um Cooperative Futures Institute in partnership with Zombified Media and some other groups. We actually had a uh um early version of this show um here at Frontier Tower a few weeks ago. And the idea of this show is that it's a variety show about the question of how we survive these transitions with AI. And in particular, how do we build a positive future between humans and AI? All right, are we ready?

A year ago or so, I realized that we really need a scientific approach to AI. And that nobody was applying the signs of cooperation to it. What does consciousness mean in the context of machines? I I think I'm kind of like a tech optimist. I would I worry a little bit about the hellscape. You either go sex robot or you got friends who come to your house.

Because if we want to program them, we have to be careful that they don't program us. How can we actually, using science, get to a positive future with AI? The party of the future is optimistic, and being optimistic is a hell of a lot more entertaining than being pessimistic. All right, that's it for my talk. Thank you all so much for listening, and love to chat with you. I'll I'll be in the back there after my talk. Thank you.