 Hello, everyone. Welcome to the Active Inference Lab. Today, it is August 20th, 2021, and we are here in guest stream number 9.1 with some awesome guests. So we'll just start with introductions and then head into a short presentation before a question and answer period. So if you're watching live, definitely write questions in the chat. So I'm Daniel. I'm a researcher in California and Blue. I'm Blue Knight. I'm an independent research consultant in New Mexico, and I'll pass it to Rafael. Hi, everyone. Thank you guys for having us over. I'm Rafael. I'm also an independent researcher. I have a lead job at Google that has some, but not a lot to do with this topic. And I've been working with Pranav and Jacob on this topic for about a year and a half now. It's been a wild ride. So Pranav, go ahead. Hello, everyone. Thank you, Daniel and Blue, for inviting us. I'm a PhD student at Carnegie Mellon Depper School of Business. My areas of research are collective intelligence. Where does it come from? Building a causal model in social technical architecture, looking at collectives as adaptive systems. I was very interested in the Active Inference and Work we met. Rafael, I, and Jacob met across different CI conferences and got this project going. So we're excited to talk more about and learn from all the discussions that we have before the US. Sounds cool. So let's go to Rafael's presentation, and then we'll just have a bunch of time for questions. So go for it. All right. Before I get started, can you all confirm this? You can see me all right? Yeah, we see the slides. Perfect. All right, cool. All right, so what are we doing here? So as we mentioned, the interest here is to try to understand better the phenomenon of collective intelligence. And so two examples here of what we're talking about, and they're pretty different examples. One of them is a pretty classic examples, the Marmoration Starlings top left. I know the bird that's solving a puzzle on the top right is a crow, not a starling, so I'm cheating a little bit, but that's fine. So on the bottom, the example is the discovery of the exposure, which is an amazing example of a large-scale, long-term, sensitive collaboration. So those are both legit examples of collective intelligence, and yet the way that we approach analyzing these things tends to be a little bit different. So if you're doing something around complex aerodynamic systems or computational biology, you tend to look primarily at the collective behavior and to be able to do that in a computable way. You assume that individual agents are following simple rules to react to the environment and to their peers. And if you're doing something like anthropology or social science, you're probably thinking of collectives that look more at the bottom, and so you're focused on these really sophisticated agents that have these complex social cognitive capabilities that are the thing that endows them with the ability to act as collectives. And you try to build a picture of the collective behavior kind of bottom up from the sum of those interactions. And both of our valid perspectives, of course, but they're speaking pretty different languages. And we think that there must be a missing link here, because if you think about us humans, I mean, I definitely feel like I am, at the same time, an autonomous individual and a member of not just one, but a bunch of collectives. So I fluidly go in and out of companies or sports team or a polity. And of course, I'm part of my family. So all these collectives can demonstrate intelligence behavior that cannot be explained exclusively by aggregation of the individual intelligence. And yet, the collective itself is a complex system that is composed of components that are in and of themselves highly autonomous complex adaptive systems as well. And that's how we frame the goal of our research. We're looking at what we call the missing link, which is a plausible theoretical description of the functional relationship between these two scales. And in particular, we think that the missing link will be found by understanding what kinds of social cognitive features enable the formation of an effective collective within these kinds of autonomous agents. All right. To do that, we've been using active inference as a framework, since the group here is probably way familiar with active inference already. I'm not going to go a lot into it. Just pointing out here that we're approaching this. You might see a language that's a little bit different from some of the canonical stuff that Friston and Team put out. We are using some different nomenclature, but it's really the same principle we're all looking at an agent as a mark of blanket and trying to model exactly what's the generative distribution that creates behavior in that agent and look at what happens when you put it against some both true system dynamics that are happening in the environment. And of course, you run the system as an agent model. All right. Except in our case, it's a multi-scale model because we're interested both in what happens at the scale of individual agents and the global system. For us, it's actually quite interesting to try to be a couple things and look at what's going on in a very specific way. Modeling many to many relationships is really hard. So we really focus on these pair-wise subsystems. So our model is really composed of a collection of pairs of agents. They only have to interact with each other within a single pair. And we basically compose them what we call the system scale model bottom up from copies of these pairs of agents. All right. So let's look first at the individual scale where we have each pair is operating on the same physical environment. They're sharing. You can think of them as sharing one of these circles. If you're familiar with the McGregor model that has been published in 2015, it's a very simple, intentionally stylized active inference model that was originally designed to be analytically soluble, actually. So it allows us to have a really clear view of what's going on on this system as a dynamical system. And we just added some minimal capabilities to the agents in that model. So each of the agents that are marked out as little can think of them as little parts within this circle environment. Each of them can only do a couple of things, which is their capabilities. They can only observe their own position. They can't directly observe the environment or that is, meaning they can't observe their own position. They can only get a clue, which is an on-off sensor generated by a chemical signal on each cell. I'm getting a ping here from a friendly viewer, also known as my wife, saying that my image may be blurred. Now it's good for others. Yes? It might look a little different, but I have a high resolution one saved locally no matter what the final processed YouTube one looks like. Thanks, YouTube. All right. Thanks, YouTube. All right, cool. So as I was saying, the main way that agents get information about their physical environment is by this sensory state, which is a binary on-off sensor, that you can think of it as being generated by a chemical signal or a kind of smell on each cell. And it's emanating from this source position. So that's pretty much it about the physical environment. And the agents can go left and right or a safe sensor on that environment. Now, talking about the relationship between agents and their peer on a pair, their peer on a pair is a funny thing to say, but the agent can observe their peer or their partner's position, relative position, meaning I'm one or two or however many hops to the left or to the right. They can be actually in the same position as well. And they can also observe their motion right there. They can observe whether the agent, the other agent, the partner has gone to the left rider or sits on the last iteration. So let's talk a little bit about reward functions. Agents have independent reward functions that are two peaks with one common peak and one private peak. That's the reason why that support's going to become clear a little bit later. Cool. So that's it. That's pretty much it about the individual agents. As I promised, it's a really simple model. And the global scale, as I said, is a number of agent pairs. You can think of 80 copies of these two pairs of agents, meaning 160 agents in total or whatever. And from that perspective of the higher scale, the Markov Blanket is actually composed of the collections of all of these pairs. So they are each interacting with a global environment. And the cool thing is that the way that the sensory inputs are related to each agent is as desires. So you can think of it as the top level as bringing in some nudges, some target states that are pretty much the sort of smells that the agent picks up. All right, the other important thing is that the optimal state for the system corresponds to the agent's shared target only. But so from the system's perspective, it, quote unquote, wants all the agents to be in that shared target. But the agents themselves, they don't know it. Why it's important is that it means that there's no exogenous incentive to achieve the global goal. So it's not like just we're nudging, we're asking all the agents to do the same thing and, lo and behold, they're going to do it. All right, and let's talk a little bit about how free energy minimization happens here. On the individual level, the individual agent, we just assume that there's some process, some neurological process within each agent that does gradient descent. That's the traditional technique. But at the global level, what we're actually doing is that the behavior that we're simulating at the collective of all these individual agents, the aggregate of that behavior is what we want or we would like to see approximate gradient descent. So that's the thing that we actually want to demonstrate here. We want to look at the behavior of the collective, calculate the free energy, the collective free energy, which is f sigma here in this equation at the bottom, and see if it's actually minimized. If it's minimized, then we can say that the system, this group of very simple agents, as a collective, it's performing active inference at the global scale or not. And shock review, what we find is that the extent to which the collective is actually able to do that free energy minimization depends on the social capabilities that we endow these simple agents with. So we focused on two cognitive capabilities or features that are stylized versions of things that we traditionally look at in social science. The first one is what we call theory of mind. And what we define that is that is the ability of an agent to put itself in the shoes of another. And we actually tried to model that in a pre-literal way. So if you think of shallow versus deep representations of the other, you can think of a shallow representation as a mechanistic, oh, if they're going this way, it means this rule-based kind of way which you could think of being an internalization of a rule that I think somebody else is following. But what we actually did here is something quite different. So it was based on this philosophy of active inference that means that every agent has a self-actualization loop. And what we said is, hey, the self-actualization loop, the internal part of it needs to be implemented by some kind of quote unquote neurological circuits. And what if that's exactly the same type of circuit, the same structure we're being used to model the partner so that you actually also had a partner actualization loop? So this is quite literally putting yourself in the shoes of your partner, right? So you have your own beliefs about your position and about your desires. And you also have literally the same kind of belief about the partner's position and their desires. And you can use that, plus, of course, whatever sensor you have or information you have from the environment to predict the partner's actions is exactly the same way that you're predicting your own actions. The only difference, of course, is that you don't actually control your partner, so you're predicting that that's the dotted line from A2 out or from A1 out in the right. So you're predicting how you're creating counterfactuals about what your partner is going to do. And then, of course, you can observe whether that prediction matches the outcome in the next round and use that. Each partner uses that to calculate their model of their partner. And the other thing is you want those two predictions to be consistent between themselves. You want the beliefs about yourself and the partner to have some consistency. And that actually helps you do what you call triangulation. So you know maybe very little about your physical environment, but you're seeing your partner do stuff and moving around. And you're using that information to inform your belief about your own position. So that's really the power of theory in mind for our model. And we're going to see how useful that is. Now, the other one is Goal Alignment. I now realize that I've been speaking for a long time, so I'm going to go a little bit faster. And so the other feature that we talked about is this thing called Goal Alignment. And it's really about the ability for an agent to adapt its goals to match another agent's goals. And this is a bottom-up incentivization mechanism. And remember what I mentioned that the system actually has an optimum but has no direct way to tell the agents about it. So the assumption here is that to the extent that you can't, that you have overlaps between the two agents' goals. And it turns out that the overlapping goals are aligned with the system optimum that provides a bottom-up way for collectively optimal behavior to happen. And we did a model dynamic, the dynamics of how the Goal Alignment happens. We just assumed that it can be flipped on and off just again for simplicity. And that's a thing that we'd like to explore further. Cool. So we did a virtual experiment. We ran simulations with all four possible combinations with or without the Remind, with or without Goal Alignment. And basically what we, and the baseline here, which is without the Remind, without Goal Alignment corresponds to a scenario where these agents are pretty much blind to each other. They're not interacting with them with each other at all. We can see each other, but they're not using that information. And we're going to see how those features actually drive the behavior of individual agents and also the system itself. Cool. So basically what we find is that I don't think I mentioned this before, but we define in each pair there is a strong, a perceptually strong and a perceptually weak agent. The weak agent is pretty much following almost random noise that they have very, very poor smell receptors. But they can still use the partner's position motion to triangulate. So the effect here that we see, agent-wise, are mostly what benefit these social capabilities endowed the weak agent with. And so we see that the model with both together actually is the one where the performance of the weak agents consistently improves really well to the point where it gets almost as good at finding an optimum as the strong agent. And why is that? So the basic intuition is that if you remind by itself in this model is not able to improve outcomes that much because remember each agent has two goals, which correspond to the two peaks here in the middle drawing on the left side. And because there's this great ambiguity, if I'm going to a certain direction, it's hard to interpret that in terms of going to a specific goal if there's two possible goals where you're going. So that limits how much useful information the agent can get from the strong. And goal alignment itself doesn't help by itself because now you have a single reward target, but you're still not super well-defined where you're going. But putting the two capabilities together actually does result in a grasp improvement. So I'll let that sink in for a second. Theory of mind, knowing, putting yourself in the shoes of another, plus having a goal alignment capacity to share a goal, it actually creates benefits for the agents that are compensated, helps a lot to compensate for deficiencies in the agent's ability to navigate its physical environment. And that also is true looking at the collective. So remember we were talking about whether these models can minimize the collective free energy, F sigma. So we're interpreting all these runs of optimization for the individual agents as part of one single Bayesian inference step for the collective. And what you see here is if we're doing this right, the graph should look like a gradient descent down the free energy gradient. And this only really happens consistently towards a symptom of zero in the fourth model, where you have both of these social capabilities. As we mentioned, you get some improvement. You have some abilities from these agents that are coming from their individual capabilities. But when you put the social cognitive together with the capabilities about the environment, that's when you get the ability to really minimize the system free energy. All right, so that's what we found in our paper. And we think it's a pretty interesting and compelling example of being able to link. So in a very stylized way, but very clearly linked social cognitive capabilities at the individual level that are not just simple rules. This is not about simple rule following, but it actually has to do with interpreting the other and using that information and linking that directly to active inference at the collective level. And where we'd like to take that, I think the four most exciting things that we've been discussing the last couple of months are, first of all, can you actually look at experiments with humans? And because these capabilities, these social capabilities are things that you can measure theory of mind and ability to align goals or things that you can test out with human subjects. And can you use our model to represent and simulate and contrast with what happens in groups of humans and see if differences in their abilities to have theory of mind and to align on goals matches up with what we're defining here as free energy immunization. That's one idea. Another one is, you know, collectives. One of the things that collectives can have is behaviors around shared resources and these resources endow them with capabilities in different ways. So that creates, we think that creates sort of a non-ramp to understanding the idea of economic value. That's, we've been joking and jokingly calling it for synomics. This is, the further down you go, the more speculative we get, but we think that active inference can actually help us understand this notion of where does value come from? Is it really just related to scarcity or is it actually also related to the capabilities that it's endowing with different agents? How these two things relate to each other? So I think there's a powerful idea there. This notion of a guy inomics, I wrote a little bit about that. I think it's been starting to pick up scene in some areas. And, you know, ultimately the all of us are embedded in one big collective, which is Gaia from Gaia Theory. It's the planetary scale of collective intelligence that sometimes works really well. Sometimes it doesn't work really well. Can we use these sorts of ideas of multi-scale active inference modeling to improve how we define and model the behavior of that guy assistant? Finally, going from the pretty big to the pretty small individuals, an individual humans brain, we've been looking at Jeff Hawkins' A Thousand Brain Theory from his recent book. And there's a lot of interesting ideas that look a lot like active inference, but as far as, you know, nobody's tried to model what he says is the fundamental unit of intelligence, which are these cortical columns in the new cortex. Nobody's tried to model those using active inference. And that's name implies the main thesis of this theory is that the intelligence in our brains comes from the collective behavior of all these cortical columns interacting by a certain algorithm. And so what would it look like to try to implement this here in the context of the brain and what inter-portable agent capabilities are required? So I'm gonna pause there. I've overstayed my welcome. I'm really curious to hear what kind of questions or thoughts we have in our homes. Yeah. Thanks. Awesome, thanks for the fascinating presentation. You're muted then. Oh, thanks, sorry. Great presentation. So maybe we could start with just Pravnav and then Blue. Just what is bringing you to being right here and what are your first thoughts either being involved on the work or on the collective behavior side more generally? Blue, are you going first or like? Go ahead, sir. Authors, Pravnav. I think so. Thank you. Thank you. No, I think so this work in particular, I'll start with this work as a context to set up the bigger idea is a part of the thing that we want. I'm interested and we hope you're interested together is figuring out one understanding causality of where does collective intelligence emerge from? Like what are the interactions between humans or humans and machines, those things. While you can have causal theory from a management social science perspective, it's very hard to know that it is optimal. Like is that the outcome? Like a statistical mechanics approach to looking at ideal gas behavior and saying that, oh, at the aggregate level, this is how the gas should behave, but you can look at the micro interactions underneath. So I think from my perspective, active inference is, I know this may not sound correct, but from a social scientist perspective, it is an a theoretical way of physics way of looking at the Markov Bank, like looking at the statistical mechanics without worrying about how exactly does theory of mind come about? So as Rafael explained, we had a partner actualization loop. We think the analog in social science is theory of mind. And what we start to demonstrate is that even if there are individual active inference agents, they will need some form of partner actualization. And in humans, you can observe that as the capability of theory of mind to be able to use each other as sources of information. So if I cannot work in my environment because my physical skills, like looking at smelling the different chemicals at different levels is low, but my social perceptiveness is high. So I can look at my partner and infer something about my environment, then I can still work in the environment. But that alone is not enough for us to work as a collective. The free energy of the collective will not go down just if everybody can understand each other's perspective or step in each other's shoes. So then we need something more. So the second thing that we introduce, and test out is some form of aligning goals. Like there has to be a reason to find common peaks. Again, the current model does not say how it comes about. That is a social science type of question. Like what's the causality? Like do we trust each other? Do we have a good communication processes or not? Like there will be accuracies and efficiency issues in communication themselves. All of this is not, AIF doesn't worry about that, but it tells us that these two capabilities at the individual agent level are important to get a collective outcome. Like I think from our perspective, that is really the core contribution that we are saying that there are some cognitive capabilities that are beyond an individual AIF model, the way McGregor had set it up, that are needed for it to show up as a collective. There's much more to be done here, but we are starting to draw these parallel. So the exciting thing that Rafael started to talk more about was can we map those? Which is experimentally, if we run human agents in a similar environment and ask them to do a similar task and we have AIF agents working together, can we create a mapping of what is the theory of mind index, like the score for this person on this scale? Because that would help us preemptively say that, all right, if you have a collection of people with this distribution of theory of mind, like they are ability to understand each other, like social perceptiveness, which is a standard scale in social sciences, I think it's called RME, reading the mind in the eye, that's a standard scale. So like if we can create a mapping from a real indicator of theory of mind onto these things, we can start having some predictions that this team is likely to succeed more than this combination of people. So that starts to open up the idea, all right, now we have some interesting understanding of how the whole system works and can we draw it to real humans. So this is very much the beginnings of a grander research statement and this is the latest direction that we are trying to pursue and using AIF in tandem with causal theories or management or social science theories to figure out more that how can we help teams better or design things better. Does that, is that enough? Like I didn't go much beyond that, but like very happy to talk more on other kinds of words. Blah, blah, blah, blah, blah, blah, what do you? Thanks for not, that was great. Yeah, I really enjoyed this paper from my background is like biological sciences. And so like I've done cellular and molecular neuroscience like all the way through my PhD until I kind of took the job off the diagram board. So for me, I'm not as well-versed in the social science aspect, but I think a lot about collectives and I'm super drawn to collective intelligence and even hierarchical systemic organization, right? Like so how, you know, bodies are built, right? From like molecules to cells to tissues to organs, et cetera. And even like to societies and ecosystems, like the Gaia theory that you were talking about earlier. So this paper was awesome for me. I do have like quite a bit of questions, especially at like the social science front, like in the beginning of the paper, you guys outlined like a whole lot of different, you know, potential ways that we operate like as multi-system, multi-agent systems. And so like a lot of them, you had theory of mind in there and also the Bola alignment, but you also had like shared norms and folk psychology. And so what made you choose just those two, as opposed to like testing all the other ones out? Or did you try others? Or did you not figure out a way to like mathematically get it together? Or what was your logic there? Experience? Pranab, do you want to take that? Yes, I think so we had about six months of discussion of what would be the scope of the paper. And that's where, because it was like so many things, like as you can see, you're already excited about looking at it at the level of, Christianomics at the level of bionomics. But before we could jump onto any of those things, I think it was important to boil down as to what is the key pieces that are needed before we can go to specifics. So the key piece was at this point, AIF does a very good job of explaining an individual interacting with an environment and how it operates. So yes, we can go to norms. Yes, we can go to these things, but these are details about how does norms emerge? And at the end of the day, okay, I could be misunderstanding some pieces of it because I'm coming from social sciences to active inference. So pardon the errors there and please correct me. My understanding is that AIF does not give you sort of a human level explanation. It is the same kind of collective outcome that you would see if of cells clumping together. We would not call it theory of mind in cells because I don't think we have a feature that says, oh, cells have theory of mind of other cells. It could be like it is still an analogy. So AIF is not telling you that this is the mechanism. It's saying that you need this as a functional piece to it. So that the energy minimization happens and that has demonstrated the math. So the reason we did not go into a specific implementation was because I think that would be the next set of steps. We first need to say that at the abstract level, at the baseline level, concept level, we need to have certain additions to an AIF framework when looking at collectives. As against, so one way of looking at collective that we do in the system-level free energy is that we look at the entire system as AIF thing and plots free energy. But that is very different from being able to say that, oh, how are these two members or these 15 members within the system interacting with each other? What are the capabilities that these people have that is different from this whole system as an active inference system? And that is where I think if I am not, I don't mean to complicate this or sound vague. I think we first wanted to establish, all right, what is the minimum mechanism that we need? If we do that in one paper at the simplest level, then let's start digging into how does the goal alignment or norm formation come about. So that can then be a parameter that certain teams or certain set of individuals or subgroups of individuals are better at forming norms while others are not. You would expect these people to minimize their free energy as a subgroup better than this other subgroup. So that would be an exciting extension in which you can go, but we didn't want to jump there because it would be doing too much. At least that's where we were at. Rafa, is that accurate or did I completely go out there? I think it's one thing to hold on. I just wanted to add that. From my side where I got to the beginnings of what we were discussing is I had this question in my mind of it takes a ton of cells to create collective behavior. It's usually a pretty big discrepancy between the different scales at which different levels of hierarchy are set up. If you think about a cell and then an organ now, sorry, an organ or clearly not a biologist, but even in these examples that we gave with starlings or other examples with other animals other than humans, it looks like the behaviors around colonies and so on. It's a collective behavior, but it's really driven by the statistical, mechanical implications of these relatively simple rule following behaviors. Whereas the sort of, I don't claim to have demonstrated as I'll continue calling it a hunch. The hunch that we had is because humans have this more sophisticated set of abilities that imply some ability to do things in a non rule following way and that they're not cause effect relationships at the individual level that allows collectives of humans to be formed in a more, with fewer humans and with more of a fluid aspect and meaning the collective can also have more flexible behaviors. Again, no time to have proven it, but that's where I feel like the two capabilities that we picked came from because theory of mind, true, what distinguishes true theory of mind from just having a heuristic about how your partner is going to behave is really this ability to put what you're seeing in the world through the same active inference loop as you're used for yourself, which is what you call putting yourself in the shoes of another and that's inherently not a simple cause effect thing. The other one with Google alignment, even though we didn't demonstrate it quite as deeply I think it follows the same principle of being this kind of mutual adjustment of expectations that eventually gets you to a shared role. And so we feel like that has the kernel of an idea about what makes human collective special. Awesome. One comment and then I'll ask a question from the chat. You kind of pointed out that there's a lot of modeling at the single agent level. And then sometimes when you have a massive number of agents interacting, you can almost take these mean field approaches or statistical averaging and then the kind of complexity is in the middle where there's small groups and reconfiguring. So that was really interesting. And also just the challenge of different words used for the same phenomena across scales or across systems. I understood it as sort of like you were saying this is the continent of theory of mind. These are some of the keywords or the questions and then that's just a first connection and now you could reconfigure what the collective could do or do other experiments. So here's a question from the chat. They wrote, a feature of the model is that order emerges through the bottom up self-organization rather than top-down priors like most active inference models on morphogenesis. However, the goals were pretty much hard coded into the model. What self-organization did is only to select the collectively optimal decision among a set of possible decisions. So I'd like to know more how the model improves non-existing literature and importantly what it says about intelligence, which is understood as the capability to understand the world rather than just optimize outcome given a set of possible decisions. All right, I can talk a little bit to that and then Brown I'll be sharing anything to add. I think this really points to another limitation presented by time when we decided to publish for Entropy, we had this notion of being able to investigate how goals emerge at the collective level as well. And that turned out to just not be feasible in our timeframe, but it's already agreed with the critique that ultimately what you do want to understand is also how goals emerge. We talk about that a little bit in the paper. I think the one possible approach is this notion of exploring states and as a collective and seeing what happens to individual outcomes. And if there is an evolutionary knowledge that would align incentives naturally from the perspective of the various combinations of individual incentives as well as if you think about social incentives, these ideas of keeping up with the Joneses or however, there's a bunch of ways to model that. And we think that given that kind of evolutionary knowledge towards goal alignment, we could see that emerge as a totally bottom up model as well. Given that having said that, I don't think that the model is completely top down given, it's not giving like such strong constraints. And the way we try to tease apart that thing is that this notion of everybody has two different goals but the system has a single optimum state. But it's not like directly telling, it's not directly giving incentives to the individuals to do it right. It's just like quote unquote by coincidence, the shared goal is also the optimum for the system. Of course, as we showed, without a way to align on that shared goal, it's just as likely that each individual will just pursue their own private goal. We did play around with some models where the shared goal was actually a lower peak in the reward function than the private goals. And of course that meant that absent goal alignings meant the agents were more likely to pursue the private goals than the shared goal. And so you have a little bit of that flavor of, okay, I think we can at least say something about that relationship that it's not just the traditional bottom up opposing goals and of course they are gonna strive towards those goals. Yeah, I think I'll jump in and add a couple of, mostly how we arrived at this vision. I think great question. This is the kind of, we started with a complex model and evolved it into what's the simplest instantiation for the goal of this vehicle. So let me walk you through a couple of different things which we started out with and now planning ahead upon. So we first started with the idea that, okay, if you're looking for a minimal mechanism, we need something through which, and I'm talking in analogy in like two agents, team's terms, not in AIF terms, we need the two agents to not only understand each other's goals or distribution of desires, but also compile them, right? So the simplest way we initially started with that, we gave each individual their own 20 or three feet or however peak distribution of goals in the environment and said, okay, you can just take them and combine them. But what we started encountering in that piece was that should we be thinking about having an additional value for approaching shared goals? That is when the peaks are combined, if my partner also has this similar peak, it should be given high value. So there is like, there is a synergy bonus or there's a collective goal bonus that should be given that minimizes my selfish desire to go after goals that is only mine, like my private goals. So like we've played around this sort of like decision space, but we found out that I think we'll have to create another parameter on giving them how much is their preference for personal goals versus preference for collective goals. So that's an additional parameter that can be added to the model to each individual that they have, I forget the name that you're getting like selfish bias or like team bias, something, we created something of that sort and that could determine. So you could also end up in a place where me as an agent is always going after my personal goals, but Raphael as a team member is always going after team goals or shared goals. Like that would be the distinction that we would be able to draw out. That was not part of the scope of this paper. That's why we removed the way that's a thing and simplified it as everybody has only two peaks and it is very clearly known that this is a shared peak and this is the unshared peak to simplify the model and the explanation. What you can do is make a general version of it that you can have a way of compiling them. You can add on a parameter, which can say that I have a team preference, like I have the social science equivalent would be my collectivism orientation, individualistic versus collectivistic orientations. If I am very collectivistic, I'm very likely to choose a goal that is good for the team. If I'm individualistic, I'm likely to choose a goal that is good for me, right? I'm not optimizing for the team. So like those kinds of things can be built in and hopefully we'll be able to explore that further. But we realize that that orientation by itself is not required to minimally explain the emergence of the address, right? So one interesting piece, again, not explored in the paper was over here, the idea was that both agents should end up at the shared peak. So they're trying to seek the common peak. What if, so that is how high system performance was defined? But what if it is what you would call a disjunctive class? That is, there are multiple goals in the environment. The team as a whole should hit all the goals as against compile on a single goal. What does that look like? That is cooperation to achieve all the goals. Our model currently cannot handle that. It does not know how to coordinate separate goals that are contributing to a team. Over here, the assumption there is that synchronizing or converging onto a position is what is good collective behavior. That is not true. Like as you try to do more analogies on which direction you want to grow the model, that would be an extension that one has to think about. That how are we defining what better collective outcome looks like? Like that depends on the environment you're working in, depends on how the goals are broken down. So those are the places that we would like to go. So you're absolutely right that it is essentially right now an optimal decision-making situation because the simplest feature, the goal was not to make the optimization, the complexity of the paper. The goal was to say that you need a puristic to find the common ground. That's it. And here is the simplest one that we presented. Does that make sense? I think it brings us to this question about what does make the higher order group in active inference agent versus really what does make something specifically in active inference agent at any scale? Is it just enough to have input of information and policy selection? Isn't that just control theory more broadly? So how do we figure out who isn't the active inference agent at the level we designed it at explicitly, which was the lower level formally defined and then where really your claim is kind of building the bridge to, which is statements about collective phenomenon, collective behavior. So maybe either the authors or then blue. Yeah, and I don't have a ton to add to that. I just wanted to say that I'm not, I don't have a strong background in complexity through theory or even in dynamical systems, but I was reflecting on those sorts of things and it struck me that this assumption of gradient descent, the assumption that optimization is even possible is such a kind of an ingrained thing when you think about this kind of modeling that exploding that assumption a little bit and just focusing on exactly this question of, is it even possible for a system that looks like this to actually like in practice minimize this free energy functional, right? Is it? And I think it, again, my day job has to do with some issues of organizational dynamics and so on. And I think that has a lot actually to do with system dysfunctions because the basic assumptions of what needs to happen in order for things to fall into place is or sometimes not trivial at all and the efficacy of a team of the collective at doing that kind of minimization which we really understand as achieving some kind of of sharing interpretation of the world is really what's at stake here. I also want to add really quickly that we cheated a little bit. Our system is not an active, the global system is not an active inference model, it's just an inference model. It again came down to expedience. So it's the collective as a whole is not acting back into the environment. So I think there are bigger questions there around how does that bigger loop evolve and how does that influence the capabilities that agents have, right? If you're really trying to just to get your system to act optimally on the world and the levers you have are the kinds of social cognitive capabilities that individual agents have, would that select for more agents with more theory of mind with more go alignment with the mix? So that's the kind of things that we would have expected to see with a broader dynamic lens at the system scale. Thanks, Luke. Yeah, so from the complexity perspective and piggybacking off of the bigger system and what is the bigger system? So I don't know if you're familiar with maybe integrated information theory of consciousness, but it really looks at to what degree the whole is more than the sum of just its interconnected parts. And then there's that angle, but then there's also a recent paper by David Crackauer that talks about the information theory of individuality. And they really talk in this paper about how in order for a unit to be a new individual, there has to be like, so, and this goes even back to like the origins of life theories, right? Like so you start off as like a collective of whatever molecules and then well, what makes you form a cell, right? And so what has to happen in the integrated information theory paper, or sorry, the information theory of individuality paper is that there's this bi-directional information flow from both the bottom to the top and then from the top to the bottom, which you also see like in human organizations when like you get a group of individuals that align on some kind of collective goal, like somehow a leader emerges, right? Like so everybody's kind of like, well, I don't know what to do and like how people are there like that. So then somebody will take charge and direct the group even in an emergent unstructured way that there becomes this bi-directional information flow that makes something a collective unit. And I just wonder, like I just find myself questioning whether you have that if you've built a new collective unit or not. So, and do you think you could and how? Can I jump in, Rafael? Okay, so I think, Lou, you bring up fantastic points. I think the way I have simplified and understood active influences that you have an agent, it is considered to have an active working in an active influence paradigm if it is interacting with its physical environment to reduce its surprise, right? Like it is trying to act on it, build a hypothesis and then act on it again, as simple as that. But when we try to move to collectives, what we are changing here. So like when, so for example, if you were to say is this single human and active inference agent, we ask does it interact with its physical environment to optimize an outcome? Then we ask, is this organization or a team acting as an active inference as a system? You can say that for this organization is interacting with its physical environment with the resource environment and doing things. The distinction that we're trying to build is how do you go from here to here? Which means when there are multiple individually IF agents, in addition to having a physical environment, you also have a social environment. And that is where you have to be interactive, right? So like now you're the idea of theory of mind or whole alignment, all of this resides in the, from my perspective in the social environment place, which leads to a system as a whole behaving such that it looks like it is an active inference agent. It appears to be an active inference agent because it is able to very dynamically work on its physical environment, but coordinate or combine its information that is coming from decentralized sources internally to get the minimum free energy type of action. So now, looking back from there, the idea of, so we don't do in our current paper, I think as Rafael said, we are not modeling the entire system as an IF agent. What you're saying is, this is simply a statistical look. So if you look at all the different runs and compile them together, it would seem as if in most cases, the behavior of the system is such that it aligns with reducing free energy. It is not actually actively doing things. This is just an compilation or aggregation of the different simulation runs that we've got. So because you're not trying to build an AIF model at all, because at the whole level, if you were trying to build an AIF model, the model would look different. You would not have separate individuals in the model. They will have to act as somewhere joint at the hip in which there is some form of information integration going on. So the idea that I think you brought up was that in an unstructured environment, there's a leader developing and they bring it down. So this fascinating idea, and I think the way we were thinking about this was, if we are going down the direction of when goals are merged, do we have a bonus for convergence? Where does that bonus come from? So I think one way of thinking about this would be that that is the leader's job. The leader's job is to change the weightage that each member gives to pursuing a common goal. So if left to my own devices, just carrying the previous example, I might be very strongly focused on my individual goal that is not shared with it. But if we have a leader amongst ourselves, the leader is changing the way I weight these different goals in my perspective. And if they're able to convince me that the common goal would be a better outcome for you. In my type of theorization, I would say our negotiation and goal alignment. So that would be that we're talking and we're saying like, all right, let me understand what are you after? Are you looking for status? Are you looking for exciting job opportunities? Are you looking for higher pay? Let me understand what your underlying desires and motivations are and let me try to get them closer to what the common goal is like. By pursuing the common goal, you will get X and that motivates you and from like a very modeling perspective, it changes the weightage that you assign to pursuing common goals. And that kind of individual interactions, I would not say that it doesn't have to be a single leader. It could all be decentralized that people are helping each other understand why following the common goal is better for each one of them. That is a discovery process. In that sense, it is very directly looped that once a collective starts forming or once consensus starts forming, it becomes easier and easier or more profitable for other individuals to join into the group goal because there is very high chance of success or some other resources that you get as a payout. Like those kinds of directions, we can think about doing again, none of this is done in this model. It's a fairly simplistic model to talk about, let's think about social environment. How do you bootstrap from AIF in a physical environment to what do the agents do such that it looks like at the high level it is AIF? It is not at this point, but minimum cognitive architectures. Can I ask you for follow up? Yeah. So in the global system, that's not an active inference agent, but each individual agent is an active inference agent. And then they form this dyad. So is the dyad at this MISO level, is the dyad an active inference agent also? Or not? I think Rafael, feel free to disagree with me. I don't think it is an active inference agent at the dyad level because there is no changes that are happening in how I infer about my goal states or how I, the only thing that is happening is now I have a secondary source of information. So in the dyad up until now, my only source of information was the physical environment. If I was low skilled, like I think when Rafael was presenting, like the orange line was the low skilled person, they had very little, they had a very low ability to understand where they are or are in that environment. And hence they were not very successful in achieving their goals most of the time. But the high skilled, the blue line which came down really quickly, had a very accurate understanding of where they are on the map. So they could quickly drive us to their goal. The thing that changed was by having high social perceptiveness, now I have a secondary source of information about the environment. My primary skill is gone, like I cannot infer from the chemicals, but I can infer from my partner's behavior. So in that sense, there is additional information that is helping the individual agent be a better AIF. Because now they have a source of information that we can trust and build accurately on. That said, that source of information is not changing their own weighting of their goals. And that's why there's no top down. Like there is no, as a collective we are not trying to influence each other. We are only taking independent decisions, now assisted by a social environment also. And that much is enough to see collective outcomes emerge. But if you're looking for more complex scenarios in which you are supposed to make a decision of should we break down across multiple tasks or how many of these people should do task A and do task B or understand who is good at what, those kinds of things will require more sophisticated machinery. So I would not say it is a proper AIF. To that I would add my counterpoint that if you define a proper AIF model as just any Markov blanket, then yes, the dyads are Markov blankets composed of the, where the internal states are union of the two agents. And yeah, even the bigger collective is just a very boring agent, right? The collective is an AIF agent. It has its inputs and it has its active state, which is the null state, right? But proper there mean, is it useful to think of that in those terms? I guess it depends, right? We wanted to learn some salient facts about the collective. We didn't quite ask that many questions about the dyad as its own unit, but we could. So in the dyad formation, did you guys play with, like I know that you did the strong and the weak and you had like, the strong one had strong environmental preferences and you had some parameters that you played with, like, and the weak agent didn't have any, like was really stupid about the environment but like could lock onto the partner and infer based on the partner. So did you ever play with parameters? Like what if you make one agent weak in both and one agent strong in both? Or what happens? Like, did you mess around with those at all? Is it like the scale of it, like as you were tuning to get, like what was just right for you? Yes, I think Bernab can go into a lot of detail on that. We did learn one thing very interesting, which is it could very easily replicate the phenomenon of the blind leading the blind. We turned out, it turned out that if the strong agent was also very socially perceptive, it was very easy for that agents to lock into the partner's behavior and try to learn, try to infer more than was actually legitimate from what the partner was doing, which was of course almost brown in motion and that actually, it can throw a poor agent into a tailspin. So it is when you start looking at these more, like as I said, less rule-following behaviors, they are pretty sensitive to kind of this balance of partners. Well, that is exactly it. I think a fantastic question, Blue, and Rafael has captured the essence of it, is there is a combination. So like we had the entire, like I was doing robustness things and like there was an entire space of what are the different combinations and how do they pan out? The most striking example was when the strong agent has a very strong social perceptiveness also they, even if they've achieved their target, they start to move away from that because they're trying to get something. They're getting confused about their own state because they think that the other member is still trying to move towards their goal state. So either there must be some other goal state or like this is me putting explanation on top of the behavior that we saw. So like we did not see a stabilization coming up. Now what this tells us, it tells us a very important thing that our model is not complete at all, right? It simply says that you need these two features but these two features are not enough to see a fairly sophisticated version of collective behavior in which there is error correction at that level. And from my perspective, again, and I know that Rafa and I have been going at understanding active inference and how we define it. And that's why I come out on the side of the collective is not completely an active inference model yet because it does not have error correction. And how do we get there? Like there's more ones to be built into that. How is it different from if you have both all the agents in your team are very good in their physical environment? So that's like having an all-star team. We know sometimes all-star teams fail. Can we bring up that kind of niche of behavior? Now, all of these questions are unanswered and extremely exciting, important areas to go into. But all of that, and that's what I come down and we started with norms and we thought that we should have this and we should have that, but we boiled down to, okay, what is the minimum thing that we need to say? Without this, nothing works. So the collective does not emerge even in minimal circumstances without these two teachers in play. If I could give one comment on that and it's such an interesting tension that your two answers brought out where if the system is designed as an active inference agent, it's a no-brainer. Yes, it is. And then this debate, which we've seen play out increasingly across so many different areas. Okay, it's a system that wasn't explicitly designed as an active inference agent, whether it directly computationally emerges from a designed agent or whether it's a performer on a stage or some other social system governance. And then there's the realism question, is it actually, quote, an active inference agent versus the instrumentalism question? Can we use this framework statistically to do inference? And just like the distribution doesn't have to be perfectly normal in nature for the use of a t-test, there's ways to use active inference very usefully. So it's so cool to see like a sandbox for exploring some of these fundamental questions about collective behavior, which oftentimes it feels like every single species or every system, they have their own programming language or they have their own norms or the edges are transient, so they go down the rabbit hole with the math that way instead of stable edges. So what's the framework that's gonna have the flexibility to actually cross these different bridges within the collective intelligence or collective behavior area? So it's an awesome good question. And I think that it goes to the heart of, like, look, I don't know if you guys have ever read the three-body problem where they have, you know, spoiler alerts, there's an alien species and they have this big computer that's made up of soldiers with flags. You know, the soldiers with flags make up a computer that's by definition, it's a computer because it was designed. You know, they're following rules that cause it to compute certain functions, right? And, you know, that's how we design computers now. There may be a certain, you know, Turing machines that emerge that are, that we can even sometimes even prove that they are universe Turing machines. And that's pretty awesome, right? And, you know, that's another way of saying that they are computers. And I think there's the same sort of question going on here with active inference where, you know, is it a Markov blanket? Yeah, I can draw the lines around it. It's a Markov blanket. Is it, to what extent it's performing free energy immunization depends on how you define the generative density, how you define the internal state so you can calculate the free energy functional and the areata and the proof is inputting, right? When you look at it, does it actually look like it's optimizing that free energy function? Sometimes the answer to that is really hard to get or only feasible to get if you make some pretty aggressive assumptions as we're saying about mean field and so on. So it's, at the heart, there's a, there's an epistemological question as well. I'm almost in an ontological. I'm gonna ask a question from the chat. This is from Dean who wrote, is there a difference between a self-actualization loop and what might be described as a sense of self-awareness? And what is the pursuit of shared goals dependent on if common interpretation and translation such as symbolic manipulation is not generating a common context? Thank you. That's an awesome question. I don't present any answer to it. I'll give you my sort of opinion. I'd also love to hear from the others. I think my mental model of this is, hey, these agent service, they only have, to the extent that you can say that they have beliefs or opinions about anything, they have beliefs or opinions about two things, right? They're their own position and their partner's position. So they, in particular, they don't have, and these are literally just numbers from one to 16, right? So there's two probabilities for descriptions. In particular, they don't have anything that we could associate with the self as a concept, right? Because they're only looking at two positions, right? Now, I think there's definitely a, it's not trivial to say, well, what is, does self-awareness mean a symbolic sense of self or is it just the epiphenomenon of having beliefs about some aspect of itself, right? I'm not qualified to judge that, but in my opinion, what the agent has is just information about a specific aspect of itself. And it feels like without the symbolic, without having a sort of more sophisticated causal model that has opinions about this, opinions about that, and opinions about how these other things connect. And one of those things is the self, which representing is like a pointer to self. I don't think that qualifies itself the worst. Awesome. These are fun discussions. Blue. Daniel, you're muted. Oh, it was on the stream, but not on Jetsy. So blue. I will follow up to that question actually. So you have like this self actualizing loop with the agent, and then you have this like other actualizing loop, right? And I kind of looked at it like your, maybe your self actualizing loop is like your IQ and your other actualizing loop is like your EQ, maybe? Like, or either way, but what I really thought about is like, you know, we are by nature, like such incredibly selfish creatures. And so like it seems that you like, maybe gave equal weight in the model. Like, can you maybe dig into like, because in reality, we don't give equal weight. Like we would give the self loop way more priority than the other loop. And was that a parameter that you played with? Yes. Like I have a lot more to say, but like Rafael go first and then I can come back. Yes. Rafael Blake, please. Yeah, no, no, I was just going to say yes. I think the primary question that we were looking at when we were thinking about these loops, if you think about the just the basic function of taking some input from the environment, forming some belief states that are just really like collection of numbers and then acting on it, that can represent pretty much anything. That's what we were, when we talk about a partner actualization loop, you can really just think of it as two copies of the same circuit. One of them is wired out into my own actions. And the other one is not, right? So the one that's wired out of my own actions is the self actualization loop, just by virtue of the fact that it's completing the, it's the thing that's completing the Markov blanket. The other one is wired out into nothing. So it's just generating predictions about my partners but it has the same formal structure internally to the agent. Now I think the part where it gets interesting is that consistency requirement where we were talking about. Like we have beliefs about the other and as the predictions that we're generating about their actions are related to those beliefs. But I think what you're asking is really about what matters most to me, right? Is it about am I paying more attention to my actions than my goals or to my partner's actions and to their goals, right? So I think the definition of matters is really about what's influencing my actions. That is primarily about my beliefs but they are informed about my beliefs. It's primarily about my beliefs of my position but they're also informed by my beliefs about my partner's position which is that process of triangulation that we talked about and that's the theory of my parameter that we were squeezing out. So it is, of course, it's very simplified but in principle that's a range that can go anywhere. I think we did play around even with having some anti-social behavior programmed in and limited amount of interpretability. Yeah, I think a very insightful question again and I think we were also thinking about as we self as being IQ because it is individual skill and the partner actualization group being more social intelligence because it is about understanding the other person's interaction. We sidestep that problem of how do you wait one over the other by choosing the space. And again, I think this maps on very well with your previous question that what happens if an agent is good at both? We see blind leading the blind which means that as soon as you see that diversity in agents that can be really good at both or really bad at both you'll need some form of waiting mechanism that to what extent do I value information that I have received from my primary skill versus I value information that I've received from my social environment and my social environment. Now, as you can imagine if you have more than one partner you will get, you will also get conflicting information because the partners might be moving differently because they are differently able to, right? So at that time we cannot sidestep the question of how do I wait the social information coming in from different partners. So there has to be a trust or a credibility index that I have to maintain internally. So this is very close to a concept like a transactive memory system in management literature, right? Where you're trying to set up that, I know that Raphael is an expert at skill A and that's a credible source. So even if I find a new team member who is equally skilled at the same skill A I would more likely to go to Raphael for help when I need it because he has credibility, he has some history. So that is an index that I've maintained internally a meta memory of mapping of who has what skills and how credible they are. So like you cannot sidestep that question as we start building and working towards bigger connectors, right? So you are spot on with that and that will start to happen even at the simplistic level of this discrepancy can be created between my primary information source and my social information source. We did not, like we played around with, oh, let's give it antisocial. Like they always think that their physical thing is better. This led us to having conversation like do they have meta memory? Which is like, do they have self-awareness that they are low skilled in the physical environment as compared to their partner, right? So this all started building more complex things. So like in my other stream of research I build those kinds of agent based models which are social cognitive architectures that you give them a memory, they'll give them a meta memory, you give them meta attention, you give them meta reasoning, a way to map who is working with what, how much do I trust them? So like, and that's where these ideas were flowing in from that all that complexity we cannot sidestep that complexity if we want to build, you know, that sophisticated behavior. That was not the goal of the paper and that's why we didn't, like we just removed all that. Like you just sidestep that complexity by saying, let's look at what is the minimum gains that can be had by doing X or Y. But the next step really if you are trying to build a more sophisticated behavior outcomes, right? Not just minimizing system level free energy would be to poke in that direction. So brilliant question. Like this is actively of how we've been thinking and you know, making the model more complex and then simplifying. This is the exact tension and interesting insights that we will get there. So very happy to, you know, if you have more ideas on, oh, this would be something that would be nice to pursue. I'm all yours. Yeah, one thought that that brought to mind was when there's just a differential equation model, like some kind of smooth line that's going to be generated, the model can't really fail. Like you could fail to read in your dataset but it's not like the line's ever gonna break. So then people are wondering about how well does it fit and they kind of focus on the performance. And then with agent-based modeling, it's always understood that there's actually only a few islands of parameters and of course the choice of which parameters to include from which effective collective behavior is going to arise. Like it's not just enough to be an ant forager, not every foraging strategy works and the one that works in the desert is not gonna work in the jungle. So it's actually just these islands of stability and then you bring up all these questions like well, what's the resilience of it to being perturbed? Those are the kinds of questions that we care about with the real world systems. And it's almost like the affordance from a scientific and a modeling communication perspective of having a modeling approach like linear regression that pretty much accepts any kind of dataset without really questioning. It leads to different questions that are being asked and it leads to different uses of the model. So it's just so cool that those kinds of questions are arising and I'm sure to come up in interesting ways in your own research show. We'll keep on doing that. Yeah, this is super exciting because that's the direction that, sorry, did I speak over you, Prabha? No, no, no, so this is clearly as it is communicated. This is super exciting to me because I think a big question that we've not asked yet is what is the goal of this kind of a modeling? Like there's one way to look at active inference can do XYZ. Like from my perspective, the goal is, okay, how can we have these socio-cognitive models and active inference models go hand in hand so that we understand something about real teams? And real teams have, as you really elegantly put, they have nonlinear behaviors. They have islands of places where they'll behave in a structured manner and suddenly it breaks because one parameter moved out of that island. To be able to get there, we need some minimal machinery. So I think that is the space where I live in when we, with Anita and with Chris Riedel, and I think I've not worked directly with David Engel, but like Anita has pre-solved with it. That is the direction that we are pushing in, in that space. It's like, oh, let's find this island of area where meta-tension and meta-memory is reasonably developed or there is diversity within the team. How does that, where are the stable coordination zones there, right? So that is one way. But in AIF, at this point, we did not even have the minimal machinery that I think we could have understood. So we started at like, what's the first building block that we need to build? And the confession is, I don't think I completely understand AIF rules. I think they are fluid. Like, how do you define an AIF? Like Rafael was bringing up, if it is only a Markov market, then yes it is. If it is this, then it is that. So this could be my shortcoming because I've not read enough or it could be the field is actually really updating itself. So that's why this is exciting. Like, let's build that. But for me, my eye is always on the, like how do we pull it out into, is this application? Like we don't want only equations that can never be wrong in a manner of thinking. You don't get nonlinear dynamics out of it directly. Sorry, yeah, Rafael. Yeah, awesome. Thank you. Yeah, and I 100% agree. And I think that's, for now if you alluded that and I want to pull it up, it's really about the scientific endeavor as a collective active inference behavior, right? To think about how scientific modeling in the kind of linear paradigm works, it's about establishing universal truths that are not contingent. Contingent only on a very specific, very explicit set of assumptions that are knowable. And within that, you know, we pull it out. You can explain it to an alien or you can pull it out and write it into stone tablets and the next civilization can read it and so on, right? And that's not what we're doing here, right? I think we're feeling our way as a collective towards a more useful, more interesting and applicable way to understand the real systems that are around us and then to maybe to learn something to help them function better. So that's probably working at it. I'll ask a quick question from the chat because I think it applies at the end blue. So Dean wrote, where would the authors see citizen learning organizational structures? Blind leading the blind or basic research at scale? No, it's not a BuzzFeed article from 2027. Might be, maybe we're seeing glimpse of the future. I don't know enough to answer, but I'll lose you guys want to take a step. But now you're muted. Yeah, then. Sorry, yeah. I was just responding to Dean's comment. It's such an insightful comment. I think one pope said it is basically sort of scale. It is like it brings up the thought experiment that will democracy thrive if you don't have independent thinkers, right? So if it is blind leading the blind that nobody is actually thinking, everybody is just based on secondary information, then it would look one way and it would be the same at the AIF level to the same behavior. Or it is independently thinking and basically such at scale and you get completely different outcomes in both cases. So like, I don't think I have good answers there, but I think very, very legitimate question, something that we can figure out. I think that's, that's modeling, that's modeling. And it relates to the discussion earlier about convergent collective behaviors versus divergent. So whether it's like honey bees deciding on where to relocate or some other type of consensus protocol, you want convergence, but then you talked about an example where there was preference to be on separate niches. And it's also important to have divergence like in discourse and in other areas in creativity. So how to design that space is so critical. I think your spot, but there is value in both of those. See convergence is important to get the outcome. Like the bees have to decide on where to go finally, but divergence is important to assess the accuracy of the information. So to the extent or rather to explore the space, right? Like accuracy and exploration of the space itself. So to the extent it is blind leading or blind, I would say that it's biased towards convergence, which means the system is opening itself up to not settling on suboptimal outcomes because everybody agrees very quickly on what the solution should be. Lack of creativity, that kind of thing. If you are too divergent and you're not able to converge, then the system is opening itself to not getting the rewards because nobody's agreeing on anything. So you need both. And that's what I think the independent research and the blind leading brand are fantastic things. I don't think can resolve the issue unless there is one agent that is always understood to be, they have the most accurate information always and the entire system understands that, right? Like it could be like a small family unit that says, oh, whatever mom says is correct. So all the children's are lined up because they are not trying to question it. They agree, they trust that that is true. And in that case, convergence is great. I don't know what you're saying, which is that, yeah. It's great to have a family than mine. And actually to connect that to some of the analytical components of active inference in the free energy minimization, there's one term that's often framed as pragmatic value. And then the other one is the exploratory, the epistemic component, an accuracy minus complexity, these different framings that we've seen. And so it's so cool that you describe the need to have a balance of the converging and the diverging collective algorithms, not worrying about whether it is a formal active inference at the higher scale, but just some of the same principles, which is that you have to have some target seeking or signal detecting aspect, and then some aspect that prevents at the very best premature convergence if not actually keeping options open to remain resilient. So those are some of the principles of collective systems that kind of go beyond even active inference alone. It doesn't matter at that higher level exactly what it is. And so it's really interesting stuff. Yeah, so if you look in the literature in the management, so you look at Megra and even collective and complex adapter systems from Holly Arrow, like they have a book on groups as complex adapter systems. And like one of the key things that in a separate line of research I build on that is, there are two functions that any group has to do. One is efficiency, the other is maintenance. The idea of efficiency is that the resource that you have to deploy very well and get the maximum out of it. The idea of maintenance is that, oh, you need to choose the right goals so that the system can survive, right? Because if you're very efficient, but you're going after the low rewarding things, then the system will not thrive. On the other hand, if you're choosing the right goals, but you're wasting away the resources and not applying them efficiently, the system will not survive. So those two things together are functions of any group or any collective. So I think again, like for me, it's like they're looking at the same phenomena from two different sides. One is coming from, what's the mechanics of it? How do we achieve that? So you look at all the social cognitive architecture from collective memory, collective attention, collective reasoning. The other is like, does this look like an AIF and can be modeled mathematically without worrying about what exactly is the mechanism, right? So it can prove as a ground truth to this team has way more to achieve than it is currently achieving because an ideal AIF model can be better performing, right? And you're missing something. So there is some losses happening in the processes, in the groups, right? It could be, you're not understanding, you're not choosing the right strategy, or we are biased towards convergence, whatever that is, right? Like that's something to be diagnosed about the team. And that therein lies the practical application. Like you have to figure out from year to year, what could be going wrong. The thing that I wanted to add to that is if you look at Gaia Theory and if you shake it hard enough and it looks like a dynamical system and you throw in the organicity assumption. Yeah, you know, maybe from the whole system throughout it's like billions of billions of years timescale perspective and you can back out and say, yeah, it's a single, it's a single active inference agent and therefore there is one, you can back it into like one big system goal for all time that is only ever visible in retrospect from the perspective of like some of the timekeepers in the multiverse or whatever, right? We hear simple agents trying to do our part, don't have access to that. And we have to do with more approximate goals, right? These are often gonna be at odds. These are often gonna turn out to be wrong and there's always this tension between how we construct the goals that are reconstructing them for exploitation or for exploration at what level? How can you can internalize the work of other systems around you to get towards more of a more bang for the buck for everybody and so on. But it's all, it's all really imperfect and I think that's what the question trying to finally tie it back to that question on citizen learning, it's from a long run perspective probably to some definition it works out okay but what okay means and in the long term, we're all dead the devil is definitely in the details on that one. Nice, that reminds me a lot of Bucky Fuller and humans role as sense maker locally and making it work for everybody. So it's really cool to see that kind of come back to active inference through design science and other areas, so Lou. So I have like a small technical question and I'm just gonna load it up followed by my big question. So I am curious about your like social science mapping and meta mapping and memory and meta memory just like how much compute power that takes like is it monstrous but then I'll add another a couple of couple more things. So I have done a little bit of agent based modeling if you like need somebody to interrogate you over your model parameters I'm super good at that so I will harass you anytime and then the real question that I have is maybe because of the nature of agent based modeling maybe it's you know you always have to define the parameters of the system and so we're looking to create a model where essentially the rules are broken, right? Like when you make a collective the collective has broken the rules it's done what the individual agents it's not collective behavior it's something new, right? And so we don't know what that newness is gonna be and so maybe agent based modeling is maybe the wrong tool like how do you allow in an agent based model where you define all the parameters how do you allow for this like newness to emerge? Again, blue, fantastic question. I am so happy to keep listening to your questions they're just very insightful. I'm not sure if this is, if I understood the spirit of your question but I think based on what I understand I would say the use of agent based modeling is predicated on the goal that you have. So if then in a social science or sort of like social theory perspective we are building a transactive systems model of collective intelligence. That is we are saying that collective intelligence emerges out of the formation and core regulation of processes that involve collective memory, collective attention and collective reasoning. So intelligence is an outcome of these three functions but these are not independent as in every person has that that is true but there is coordinates of collective memory which is called transactive memory in the literature emerges out of the skill based interactions that coordination that people do same thing with attention, same thing. So we're building that kind of a model the role of an agent based approach to this kind of theorizing is essentially to say that okay I have this fantastic narrative theory that people understand are aware of each other skills meta memory, right? And they use that information to make allocation choices that is if I have a certain task and it requires help with X skill who do I go to or who do I assign this to in the team? So like if that is the theory how do we know that this is enough? This is parsimony. So this is the idea of generator efficiency, right? Like are these minimal processes enough to explain the emergence of collective memory, right? Like will a clock emerge if each bird, each bird is given these three rules, right? And that kind of research question or building of that model is important to basically give some credence to these are the minimum sociocognitive processes that are important to give you this outcome. So we are not trying to code flocking. We are only giving individuals certain capacities and the flock is emerging. So this is minimal enough now. Is that flock adaptive to let's say a predator approaching them? Maybe not because you've not told them how to react to a predator, right? So in that sense, we now know there is a boundary condition that with only X rules, it can handle high levels of workload, but it cannot handle high levels of uncertainty, like those kinds of things. So computationally, how difficult or how complex is the thing depends on how many features you are trying to build into it. The model that I've developed for my dissertation and I'm presenting as I'm going on the market is fundamentally say that at the simplistic level give each individual agent three capacities of having skills and meta memory, understanding my team members, their skills, having attention or focus like ability to work on something and understand how other people are working on different things. So like what priority that has, how important it is for them, which leads to the meta reasoning, which is I have selfish motivations and I know that other people also have selfish motivations, but is the task aligned with each other motivation? So I should send this task to somebody for whom this is already aligned. So they'll get to it faster and that will benefit the team. Like none of this is a centralized, here's the leader who knows everything and is making the allocations, right? This is not top-down command control. This is very decentralized. How will an intelligence at the collective level emerge to the extent these members are able to figure these things out? So that's the building of the model that's headed towards that direction. So what it becomes useful is under what circumstances? So the useful question to us is under what circumstances is meta reasoning or collective memory important? So like this is the question that is very different from saying that, oh, do we need this parameter or not? And how computationally intense it is? The answer lies in, okay, what's the environment? What's the, so like as Daniel was pointing out there are islands of operation. Does this island in a phase space land in a place where these parameters are important or not? So like those are the kinds of questions you can answer with that. And I think that would drive the choice of parameters. I would not, I would not try to say that, oh, it should be a all-encompassing model. It should be the simplest model. If anything, I've tried to always say that we're trying to go to the simplest model that explains a part very clearly. Does that answer your question? I suspect it does not. Yeah, so definitely like with flocking and like I've seen similar models of like fish schooling and stuff like that. So we have that as a model of maybe collective behavior versus collective intelligence or like has something actually, has something new emerged in the flock or in the school of fish? And so my point is, is that when you just give these agents their individual parameters, how do you know when there's another level? Like we're looking for another level that does active inference. So not like a yes to collective intelligence, you can eat collective intelligence and collective behavior can be modeled with a group versus the individual agents, but when do you have another level? Like in the hierarchy, when do you, when do yourselves become a tissue? When does your group of people become, you know, a squad or a team or something like this? So that's what I don't know if agent-based modeling is suited to determine or not. Well, that's a good question. I think it depends on, so it's partly a definitional issue. What do you mean by another level, right? So intelligence, the way it has defined at least in that literature, collective intelligence is an ability to adapt to changing environments. So as the task changes in its complexity, that uncertainty is changing or the level of knowledge interdependence or the level of workload is changing over time, the team is able to seamlessly adapt its internal coordination processes to maintain performance. So it is not just because the task has changed in its complexity. So the environment is moving around just because the environment move, it should not be so specialized that it starts to pay. It is able to adapt to it. So if intelligence is defined that way, to what extent can the team regulate internally to maintain performance at the highest level possible? Then I would say, yes, it is emerging because the diagram that you show, right? Like essentially the plot that you show is a team that does not have these coordination mechanisms is failing, like it's dropping in performance. Anytime there is a perturbation, while a team that has these mechanisms is like seeing a inventory dip but is able to maintain high levels of performance. So it is a relative thing. Now as we change the composition, the internal coordination would be different because now it's a different set of people trying to coordinate. But it would not invalidate the idea that a different set of people still have collective intelligence or not because it's given by, given our internal resources, how are we adapting those changes? Like if you're able to track their, so in that sense, I think agent-based model does a fantastic job of showing you that we are able to see this emergent behavior of maintaining high level performance or not, right? And that's where you go into the mechanisms that this mechanism is essential for this outcome versus not. Does that help? It does, it does. But it's the question of at what scale is it an active inference agent, right? So like is the team the active inference agent or is it the each individuals? Is it like their collective active inference versus the team collective? I can see Daniel like, oh wait, I was just gonna say that. Yeah. To speak to the question of novelty, there's the question of generating novelty and then responding to novelty. And in the case of generating novelty, especially for humans, we haven't really talked about stigmergy, niche modification, cumulative culture. So there's many avenues from collective behavior to more niche modifying algorithms that are maybe explaining some of the phenomena that we are interested in here. But then this question about responding to novelty and like how do you deal with a threat that maybe hasn't been seen before or a failure mode that's never been observed or rare? It kind of just reminded me of homeostasis, which is generalizing across specific situations to help you keep target variables within a preferred range. And then instead of just reacting to changes, it makes one move towards anticipating, which is the whole area of variously like cybernetics and signal processing. And so those features of responding and anticipating and planning, Azure-based modeling is a great place to computationally and physically test that and clearly active inference at the very least can be sketched out over a lot of these systems. And then it will be a debate always like about the details and the mechanisms like you brought up, but it's just to have a framework that we can integrate scales and just nest them cleanly is at the very least a path towards answering some of these questions, if not philosophically, at least empirically. Yep. So, on you, Rafael? Yeah. Thank you. Yeah, and I think about the philosophical aspect also as backing down from unless you're looking at the guy theory thing and you don't believe such a thing as the selection multiverse thing. So if you back down to like subsystems of the big thing, whatever the big thing is, then you can back down into functions, right? And you can use that as your benchmark for what performance or success or intentions looks like and we construct down and match that up with your get from the bottoms up, from the bottom up view. And I guess the interesting question is when and where does it actually do the bottom up and the top down approach line up, right? And, you know, there's, I think in general, there's no answer there from a necological perspective. I think there's some pretty clear patterns of functional specialization and some of them are hierarchical, some of them are, you know, tend to be like almost hierarchical, you know, hierarchical with a little bit of sprinklings of horizontal conditions. And maybe that's a useful, you know, philosophical approach to look at things that from the instantiation, at least from the instantiation of all the kinds of systems that we're interested in. It's when does a group of cells become a tissue, but it's also, what are the conditions that a thing needs to meet to be meeting the function of a tissue and to be worthy of being called the tissue in that functional context, right? And then you can talk about, well, what if that tissue is not made of cells, but of like nanobots or whatever and so on. Yeah, I don't think it is either or type of question. I think it is more of the two approaches of like coming at it from a mechanistic perspective versus coming at it from an active inference perspective are complementary. They're trying to look at the same phenomena but estimate them differently. So the idea of, I can then started mentioning that morphogenesis and morphostasis, I think that's like when novelty happens, how do you respond to that novelty is part of one term in the AIF equation. And it is also a way of looking at saying that, oh, will the system adapt to it or not? So there's a mechanism of how it adapts. And then there is the equation version of it. If it adapts well, it would be reducing its free energy because that is what we are able to estimate. I think both of them are very complementary approaches. What, so one way I think I look at it is how do you know that the adaptation from a mechanistic standpoint that is happening is correct or wrong? Is it maladaptive? Is it going to lead for better outcome or is it going to lead to a worse outcome? One way of looking at it is, is that reducing free energy or is that increasing free energy of the system? Now that is where you can marry the two things and work as complements that or there are certain adaptations because as an agent is sitting inside the system, not knowing the bigger picture, there is no way you know whether these variations that you're trying are good or bad. Like you need a system level out. So I think that is how I think that they're complementary. Again, I don't have good answers yet because that's the area of research that you're trying to figure out. But I suspect that that is where it could go. You could combine the two approaches to get more insight on the both the mechanism and the overall outcome. This has been great. Any last thoughts or anytime you're always welcome to join again. And it's just super interesting to hear about. We hope that people watching it live and replay could think about some of these questions, but anyone's welcome to give a last thought. Otherwise, this was just a great first appearance with you both and we hope to see you again. Yeah, just thanks. This was super fun to read and to discuss. And I think that the theory underlying the model is hot and super on point and building it out and fleshing it out into a more thoroughly developed model will be awesome. Thank you. Now, I really appreciate it. Thanks to both of you guys for inviting us. Thanks of course to everybody who is watching this or reading transcripts or connected to the mind meld and getting an instant download of all of humanity's knowledge from 200 years from now. And yeah, I mean, as I said, I think the ultimate purpose of all of this is to be useful and it's an evolving conversation. So we're really curious to hear thoughts, questions, other ideas that might pop up. And yeah, I mean, I don't know if our emails are available, but if not, can make them available or can reach out to us on Twitter or to the active inference labs organizers and they can connect us. Yeah. Yeah. Same here. Thank you both of you. And thank you to Sir Will and Dean for the questions. I think both of us are available on Twitter or otherwise to have questions. Would love to follow up on certain things and see if we can make interesting projects out of it. Thank you, folks. This has been fun. Thank you all. Bye. Cheers, bye. Good times.