 Hello and welcome everyone, welcome back. This is the second interval of the third Applied Active Inference Symposium at the Active Inference Institute on August 22nd, 2023. This is gonna be another packed and exciting interval. And we're kicking it off with Jean-Francois Cloutier, a collective of dearizers, first steps. So, JF, thank you for joining and to you for this presentation. And if anybody has questions for this presentation or any of the others, please just put it in the live chat and I'll do it, I can't. So thanks, JF, to you. Well, thank you, Daniel. I'm a software engineer at a company called SmartRent where I write software for smart home systems. I also work on a research project at the Active Inference Institute called the Robotics and Embodied Project. My current focus is unsupervised learning in Active Inference Agents. In my presentation today, which is titled A Collective of Dearizers, First Steps, I'll first start with a brief recap of the project, then dive into recent progress and then I'll conclude with what I see as the next steps of the project. Well, since 2017, I've been experimenting with Lego robots and models of cognition. I do this for my own education because like I think most of us, I understand something better when I build it. What you see here is my latest robot model. It's a rover. It has a whole bunch of sensors and a number of effectors, actuators. As a matter of fact, those sensors and actuators can be understood as forming the macro blanket of my robot. Last year, I presented at the symposium about the history of the project, its then status and its ambitions. I presented a cognitive model that implemented predictive processing from an active inference perspective. So there were generative models, there was predictions, there was prediction errors and these generative models animated the robots so that they would roam, avoid obstacles, observe their companion and also in a very simplistic way, build a theory of mind of the other robot. They could, from observations, infer that the other robot had detected food, food being presented here as a sheet of paper on the floor and then would kind of try to track the other robot to also get to the food and fun stuff like this. The implementation combined multi-processing and functional computing. There was nothing probabilistic in the implementation. I thought I'd maybe would see them in action. That's what's presented last year, but just to get a sense of what they do. I have two robots here, one named Carl and Andy. I named them before I knew I was gonna present at the conference. So they benefited from a number of training sessions. They learned how to find which policies would achieve their goals better than others and one here on the left found the food right away but got closer to the pedestal where the beacon is that simulates the scent of food. The fear was getting into a collision and then backs off in a hurry, as we'll see. The other robot observes all this, sees the robot backing off as being in a panic and decides to share in this emotion and backs up as well. So, you know, fun stuff. Since from the beginning I've implemented different collision models, but they all have income and the fact that they implement a society of mind. So what is a society of mind? A society of mind is a concept by which the mind is not a monolithic structure, but a composition of simple actors, independent actors that interact with each other in simple ways. That view of the mind was put forth by Marvin Minsky 50 years ago, so this is not recent. Again, what I presented last year was a society of mind containing a hierarchy of I call them cognition actors. Each cognition actor is an independent process. Each one has a scope, a level of abstraction as well. So for example, you would have a cognition actor that's concerned with the location of food and it would have beliefs and perceptions about food location. And these would feed into a higher level cognition actor, let's say food approach, which is concerned about getting closer to the food and so forth and so on. These cognition actors communicate again in simple ways. This is a society of mind and they communicate by emitting predictions, predictions about the beliefs of other cognition actors and they communicate by emitting prediction errors when predictions made about their beliefs by others are inaccurate. That leads to cognition actors processing these prediction errors and combining them with their own predictions to create an updated set of perceptions that are synthesized into beliefs and these beliefs lead to actions in order to eliminate negative beliefs and validate positive beliefs. Now, it's from the interactions of all these cognition actors that seemingly purposeful behaviors emerge. So that was successful, but learning was very limited. This hierarchy of cognition actors was given, was predefined and the only learning that a robot would do was to discover which action policies tended to be more effective. So clearly my robots are not monolithic active inference agents. And so the question is, why would I use a society of mind architecture? Well, because I subscribe to the notion that all intelligence is collective intelligence and this paper makes the argument quite cogently and I'm gonna cite a number of papers which were important to the evolution of my thinking. This is one of them. This paper sees intelligence as a process and not a property. It's a process enacted by the interacting parts as opposed to, again, a property of individuals. So a society of mind is basically that. It's actually a number of interacting processes together from which emerge apparently intelligent behaviors. Now, my own personal current definition of intelligence is self-sustaining and active sense-making. Sense-making is really important by autonomous agent as system in a dissipative and dynamic environment. And this guides all the work I do in this project. Now, if you want a description, a detailed description of where my project stood at the, a year ago, there's a paper that was published, that I published on Xenodon, which you can look up. Now, over the last year, I wanted to move away from a pre-built, a given society of mind and toward a learned society of mind. I wanted to see if I could program an autonomous robot to evolve it's society of mind through its interactions with its environment. Now, one might think that programming autonomy is an oxymoron. Would be a good point, but I don't think it is. If the program I write and install my robots imparts a constitutive autonomy, which is enable the robot to constitute its own identity. And if it imparts adaptivity, which enable the robot to modify itself from its interactions with the environment, then I think that the robot will be able truly autonomous. Now, for autonomy to exist inherently in the robot, something in the robot must be at stake. And in this case, it's the survival of the robot's society of mind. Now, the physical structure of the robot is not at stake, its survival is not at stake, unless of course it falls off a shelf. But as I intend to have my robot develop its society of mind from experiences, I also expect that if it fails that the society of mind will perish in its attempt to grow and sustain itself. So that's what's at stake in this current, and going forward in this project is the robot's survival, the survival of the robot's society of mind. And this survival would be an expression of being by doing, the robot will need to act and interact in its environment in order to survive. Now, whatever sense it's gonna make of its environment will then be grounded in this survival imperative. In essence, the robot's society of mind will have skin in the game. If this weren't the case, then sense making would actually reside somewhere else. It would reside in the mind of the programmer, in my mind as I observed the robots. But I would like the sense making to be grounded in the survival of the robot's mind itself. So that's key. Now, what do I mean by an evolving and growing society of mind? So instead of the given society of mind which I showed earlier, I want the cognition actors to be created to connect with each other dynamically through interactions with the environment. So I want an active self-organizing, self-optimizing collective of cognition actors. Now, is this feasible? Well, that's a big question in order to, in trying to answer this question, I'll be also answering questions like, what must be given a priori and what can be discovered? Well, I already have elements of an answer. I know that the sensors and the effectors and the primitive cognition actors that will wrap them will be given. This will be like what you're born with, basically. And there might be a metacognition actor which role will be to oversee and guide the evolution of all other cognition actors. Well, that's my hypothesis. I can imagine a test environment for my robot that will test its ability to survive. I can imagine, for example, that as the robot moves around or computes, it uses a limited store of energy, it's simulated. And that store of energy is replenished when the robot consumes food and that would be by being on top of pieces of paper, colored pieces of paper on the floor. And I'm gonna make sure that it needs two sources of food in order to survive. Like that would be represented by a yellow paper and a green paper, for example. So that to prevent the robot from just simply finding one source of food and just stationing itself over it. So the environment will also contain obstacles. So the robot will need to learn how to navigate, avoid obstacles, locate different sources of food, get to them and make sure that it alternate between various sources of food in order to survive. And the society of mind that will evolve will hopefully evolve to successfully do this. Else, if it doesn't, then it will shrink as resources disappear and essentially die. So that's what it's at stake for this robot. Now this effort employs a number of frameworks. And by framework, I mean a useful system of concept and constraints that guide the implementation. Well, there's obviously the free energy principle and the active inference framework. However, I see this, the active inference as an as if framework, it describes the what, what must be achieved. In this case, a reduction of the agent's variational free energy. But it doesn't guide me as to the implementation, how to build the robot. For this, I need as is frameworks. And I'll be using two frameworks. One, which I've been using since the beginning of the project, which is the actor model. The actor model views computation as a diversity of processes, processes that are independent of their own private internal state and who communicate with one another strictly through messages. Yesterday in interval one, Keith Duggar presented on the actor model and made the case that we should use the actor model to implement active inference agents. Well, I wholeheartedly agree with him. The other framework that I'm gonna be using is a special case of symbolic AI called the app perception engine. And much of the presentation will be about the app position engine and its implementation. So here's where we are, as the project is located at the intersection of active inference as a domain, the what? And society of mind as an architecture and symbolic AI as a form of computing. So that's where the project is at this intersection. Now, so where to begin? So I want to dive into a more extensive form of learning. And the first step logically is to learn how to predict. So I want to enable a single cognition actor, we'll start with a single cognition actor to learn how to make sense of its local environment. It's so-called Umveld. And making sense implies at a minimum to be able to predict incoming sensation. So it needs to learn to predict. So what will be given to a cognition actor? Well, there'll be a history of sensations broken into discrete units of time. So time n minus three and n minus two and minus one time n, which is the present moment. So these will be remembered observations. And then what we want to get out of this is the ability to predict the next incoming set of sensations at time t equals n plus one and plus two. And for this to be able to predict future observations, we need some kind of predictor function that is learned. From the remembered observation. Now this predictive capability can be built in two very general ways. One is from statistics. So doing pattern analysis and being able to predict what's most probable, which is the standard current machine learning approach or we could predict from an understanding of the observations by developing a causal model of what produced these sensations. And from this understanding, predict what rationally should be observed next. So this is all about sense-making. And now what is sense-making? How do I understand sense-making? Well, to rationally predict incoming sensory inputs, one must make sense of them. That's what making sense means to me. And to make sense of sensory inputs means to derive meaningful experiences from them. It's not just data, it's not just pieces of data. It's there must be meaningful experiences. And by experience, I mean a conceptualization of the sensations and a unification of them in time and space. So making sense of these inputs will mean to produce meaningful experiences that are conceptualizations and unifications of the sensations. Now an experience is meaningful if it is underwritten by a causal model. So the experience is perceived as the consequences of a latent generative model, a generative process that we have modeled. And I want meaning to be inherent to the agent. And that only happens if the agent is truly autonomous and if this meaning is grounded in the survival imperative as discussed earlier. So how does experiencing work? How can that be put into computer code? Surprisingly, for this we refer to the philosophy of Immanuel Kant. Immanuel Kant took a reverse engineering approach asking himself, what must entities do to achieve experience? This is akin to the free energy principles of high road which can be paraphrased as what most organisms do to maintain their existence. So Immanuel Kant tried to reverse engineer cognition asking himself, what's the minimal cognitive apparatus needed by an entity to have experiences? And that he documented in his critique of pure reason. Just a little parenthesis, the meaning of the title, Critic of Pure Reason is not what I thought it was. It's actually translates more closely to the case for a priori cognition. Critique is a legal term, is where you make your case. And pure reason we translate nowadays as a priori cognition. So his work wanted to establish what must happen to create an experience that's coherent, that is unified in time and space and to reverse engineer cognition as a system that is both complete and essential, that is minimal. Okay, so let's, I'm gonna try to give you the postage stamp version of Immanuel Kant's theorem focusing on his synthetic unity of our perception. Well, first of all, there's the real world which is outside of our direct experience. It's the numina. It's forever hidden from us as an as is reality. But it impinges on our sensorium. And then so we have a number of intuitions, sight, sound, touch, smell that are initially separate. And then we need to network them, connect them, both in time and in space. In space is the sound and the sight, describing a single thing is one thing behind or inside another and in time is this happening before. After something else. And then at a higher level, meaning is given to these networked intuitions, sensations via concepts and judgments, rules, which are generalizations as to what can and cannot be. And this is what we experience. Now, it turns out that a synthetic unity of our perception is a blueprint for automating a sense-making. And it's kind of interesting, I think, that 18th century philosophy would be relevant to 21st century technology development. And this is what happened and was published by Richard Evans and all in the paper Making Sense of Sensory Input, where they developed the app perception engine. They took synthetic unity of our perception as software requirements and successfully implemented them into a piece of software, the app perception engine, and applied it to a number of exercises where they got a very good result. So the app session engine is in an instance of machine learning, it's unsupervised machine learning, and it operates on very small data sets and generates human readable generative models. And when I read the paper, I realized, well, that's exactly what my robot needed. So what does an app perception engine do? Well, given a sequence of observed states, it finds a generative model that can recreate past states, but most importantly predict future states. And the state is defined as a set of simultaneous observations, sensations, intuitions. So an app perception engine searches for a causal theory that can recreate observations. I say searches because this causal theory is not determined by the observations. It has to be found, it has to be discovered, but once it is found, then it can be validated against the observations and see if it can recreate them and augment them into the future, as well as into the past. So what is a causal theory? A causal theory is a logic program that has a number of components. There in the causal theory, there will be the objects and the predicates from the observed relation. So from the observations, we can extract what objects were observed and what properties of these objects were observed, maybe what relationships of these objects were observed. That's the start. Then we have latent object types, objects and predicates. So we may want to imagine, the causal theory imagines hidden objects, maybe in types of objects and maybe hidden properties and relationships between objects, latent meaning unobserved. And given both of the observed and unobserved objects and predicates, it derives rules, first of all, constraints on those predicates, what's permissible. So for example, being in front of cannot, A cannot be in front of B and at the same time behind B. So an object cannot be in front of another and behind it. So there are constraints on predicates. And then there are rules that apply to any simultaneous sets of observations, what they must conform to. Then there are rules that given a state will infer the next state and then maybe some initial state from which we can run the causal theory. So what makes a causal theory unified? Well, first of all, it needs to be unified in order to make sense of the observation. There are various dimensions. So if a causal theory involves a number of objects, all these objects must be directly or indirectly related. There's no object that just floats in space, totally independent of the other objects. So they must all be related. So they're spatially unified. All predicates that make up the causal theory, like on, off, behind and in front, they must be constrained so that, for example, in front cannot be at the same time as behind or that a light cannot be both turned on and turned off. So there's some restrictions on the predicates and that creates conceptual unity. And then there's static unity where all simultaneous relations must satisfy the static rules and temporal unity where all the states must be sequenced by causal rules. We'll see examples. So let's start with an example here of a set of observations. What we are observing are two lights. And the lights can either be at any discrete moment in time, either on or off. So here we have a sequence of observations. And one moment in time, the first light was off and the second light was on. Then the first light was on, the second light was off. Then both were on, et cetera. And I put the gray bars there to show that maybe the observations aren't complete. So at one stage we can only see the second light or there may be other lights or other objects that we do not see, but that's what we observe. So you see these observations in discrete time and to the upper section engine. And the upper section engine searches for a causal theory that when applied to an initial condition, let's say that light A is off and light B is on, it will create a trace of recreated observations that cover is a superset matches the initial observations. And if this happens, then our causal theory is a good one. Now, the causal theory may infer the existence of hidden objects, hidden relations and whatnot. It may actually need to. Oh, so here's an example of a causal theory that is generated by my own implementation of the upper section engine, as I re-implemented the upper section engine as described in the paper by Richard Evans and all. And I ran it on this set of observations about lights on two lights, one on one off at any point in time. And it came up with, it found a result. It found the result in 64 seconds. It was a perfect match. And it actually invented a relationship which it called tread one, which we can, let's imagine that it's actually means connects two. And it found a static rule and a causal rule. It said that a light is on at any moment in time. A light is on if a light that connects to it is off. And it found a causal rule that said a light turns off if it connects to another light that was also off. So that's how the lights change over time. The status of on and off. And it came up with initial conditions that said that, well, A connects to, first of all, that there's an object one, a light called object one that we don't see but is there. We imagine is there, that A connects to it, the light object one connects to B and the light B connects to object one. So that's the causal rule that it discovered. Now, if we run this causal rule, produce a trace and as you can see, the trace matches the observation. It adds a new object. So the coverage being excellent and being perfect in this case and our causal theory is a good one. It's actually a perfect one. It's not necessarily the only one though. So is this causal theory unified? So going back to Dr. Kant's requirement of synthetic unity of that perception, not every causal theory will do, though it may predict correctly, it may not be meaningful unless it is unified. Well, is it, we've saw the four dimensions of unification, is it spatially unified? Well, all our objects are connected directly or indirectly to each other. That's good. So we have spatial unification. Do we have conceptual unification? So we have this new predicate. We have two predicates, right? Pred one, which we translate to connects to and then the predicate that says whether the light is on or off. Well, we have a constraint that says that a light can only be connected to one other light. So pred one has a constraint on it that says it's exclusive. So an object cannot pred one to two objects, cannot connect to two objects. That's a constraint that was discovered in part of the causal theory. And also implicitly the on relation of the predicate has the value on or off and it cannot be both at the same time. So it's conceptually unified. Is it statically unified or are the rules, static rules obeyed? Well, for example, the static rule would say that given that B connects to A, if B is off, then A must be on. So if you look at any place where B is off, A is gonna be on. And you could do that for every other light and relationships between lights. So they all obey the static rule. And the causal rule says in, for example here that if B connects to A, then if A was off, then B must turn off. So if you look at B, let's say E or B was off. Yeah, if, I'm sorry, if B connects to A and yes, and if A was off, B must be off. So if A was off, B becomes off in the next step. So that's correct as well. So statistically we are true and temporarily we are true. We are unified. And of course that we get a thumbs up from Dr. Kant. Our causal theory is unified. Thus it makes sense of the observations of the sensory inputs. Now it's no accident that Kant would be, would figure in an active inference project, there is a link between active inference and Kant. And it runs through the celebrated 19th century German engineer, Hermann von Hemm-Holz. He was a disciple of Kant and he developed the theory of visual perception that operationalized Kant's epistemology. And in fact, it anticipates predictive processing. In 1995, Peter Dayan and Jeff Hinton developed the Helm-Holz machine. Name is in this honor. It is a type of architectural neural network that's trained to create the generative model from an original set of data. And it can account for the hidden structure of the data. So there's a, as you see, there's a link which is discussed and elaborated in this paper, which is very interesting paper. All right, so close parentheses. So we've looked at the perception engine from the perspective of Kant's philosophy. So now let's look at it through the lens of machine learning. We, the observations that constitute a training set, it's a very small one. And the opposition engine is the learning algorithm. And what is learned, the output is a causal theory. So the learning process is unsupervised logical inferencing. And the output is a human readable logic program. So we see here that there's some profound differences with the more popular form of machine learning in that the training set is really small. That the product of the learning beyond is actually a human readable artifact, in this case, a logic program. So this is the training set as inputted. Lights, LEDA turned off at this time. Time one, B turned on at time one, A turned on at time two, B turned off at time two, et cetera, et cetera. So that's the training set. And you feed this into the perception engine's algorithm and outcomes a causal theory. So in a little bit more details, what the algorithm is and does, first it extracts the observed object, extent objects, the object types and predicates from the observations. So we have on, we have object A, object B, we have LED as an object type. So that's all part of the observations and that becomes part of the extent vocabulary. Then the opposition engine imagines unobserved objects types and predicates for the relationships and properties and that becomes a latent vocabulary. So there's a step of imagination. Then using the combined vocabulary, the both the extent and latent vocabulary combined, it looks for a unified causal theory and set of constraints, rules and initial conditions that obey the rules of the constraints of a synthetic unity of our perception. Once it has this causal theory and with initial conditions, it applies the causal theory to these conditions and produces a trace. It recreates observations if you want and augments them and extends them to the future. Now it looks at this trace and compares it with the initial observations for coverage and decides if this is a good causal theory or not. Then it also looks at the causal theory complexity, how many rules, how complex are the rules, et cetera and measures for complexity. So if we have a choice between two causal theories of equivalent coverage, the perception engine will select the least complex one using Occam's razor. Now, if you look at this algorithm, you'll see that the boxes in green are not deterministic. That's it, this is where search happens. We can posit different kinds of objects. We can find different kinds of rules. So this is where search happens. Now, our perception is implemented using logic inference. Actually, it uses three forms of logic inference. There's the one that we're more familiar with, which is deduction, where given rules and causes, we infer the effects. Then there's induction where given causes and effect, we look for the rules. This is what science does, right? We're looking for rules that would account for effects given the causes. Then there's an abduction where given the rules and given what we observe the effects, we're looking for the causes. In this case, we're looking for the latent objects, the latent relationships between these objects. And then you can combine both abduction and induction where you're given effects, essentially observations, and you're looking for both causes and rules, which is what the app perception engine does. And this is where in the algorithm, these kinds of inferences are at play. So positing latent objects, that's a form of abduction, imagining causes. Finding the rules, well, it's clearly a form of induction. And then applying the rules of a causal theory to some initial conditions to create a trace. Well, that's deduction, when we have the causes, the initial conditions, we have the rules, causal theory, and then we produce a trace, the effects. And so that's deduction. So the app perception engine uses all forms of logical inference. And now just a reminder that the output of the app perception engine, that is what is learned, is actually human readable. You may wanna compare that to a large array of folding points produced by traditional, the more popular form of machine learning nowadays. So here, this is what's actually produced by the app perception engine as it runs on a set of observations. It produces a logic program that is human readable. When you look at it, the only thing you need to kind of guess is what is meant by PRED-1. And if you think, well, maybe it means connects to, maybe the lights are connected underneath a board out of sight of the observer. But finding a unified causal theory is hard. So we have to guess what the latent objects and predicates are, and what are the hidden lights? What are the hidden relationships between lights? And we have to discover what constraints might apply on the predicates. And what are the initial conditions from which we wanna recreate a trace? What are the static rules that apply to simultaneous observations? And then what are the causal rules that given observations at time T will predict observations at time T plus one. This is hard. As a matter of fact, it's a non-polynomially hard. The search space grows exponentially with the size of the input, which is the size of the extent and latent vocabularies. And so just like in chess, you can't predict to the end the consequence of a move because of combinatorial explosion with the apperception engine, you cannot systematically traverse the entire space of possible causal theories to find a good one because it's impossibly large. So the job of the apperception engine is to find a causal theory in a ridiculously large haystack. How to do this? In my implementation, I follow the recommendations and I followed also the implementation in Richard Evans paper by breaking the search space into chunks. First, there's a region and the region says, so how many latent object types, objects and predicates will we allow? So what is the limit of imagination of the cognition actors that is trying to apperceive a causal theory? What are the limits of its imagination? And within that region of bounded imagination, we carve it into templates where we say, okay, we're gonna use these latent objects, types, these latent objects. And so basically with vocabulary, specific vocabulary we're gonna be using, we're gonna use object one, we're gonna use object pred one on top of the observed on predicate and observed A and B lights. And we're gonna set the maximum complexity on the rules and see if within, we can find causal theories that fit this template. So this is a carving up of the search space. And having broken the search space into regions and templates, we have scopes in which to apply heuristics. Now, why heuristics? Because the systematic traversal and it cannot be done in reasonable time. There's just too many candidate causal theories to look at to find a good one. So we use heuristics. We find ways of maybe getting to a good solution faster at the risk of missing it, but at least we'll have an answer or no answer in a reasonable amount of time. And there's a number of heuristics that I have implemented in my implementation of the app-assetron engine. While there's time boxing, at some point you'd spend no more than this amount of time looking into a region or into a template. There's multitasking. Well, the problem is actually, as they say, embarrassingly parallel. You can explore multiple reasons and multiple templates in parallel. And so make good use of a multi-core computer. You wanna make sure you don't repeat yourself. So you don't want to traverse the same region twice or look at the same causal theory twice. You want to satisfy. You may be a good enough theory is just fine. We don't want to look for the perfect one. necessarily, we may not have time. You want to fail early. If you're in a region where nothing good is found, you may wanna leave it quite quickly at the risk of maybe not finding the good one that is just over the horizon, but you wanna be impatient. You wanna throw the dice. You may want to kind of mix it up so that every time you run the app-assetron engine on the same problem, you may find a solution, a different solution first. You want to go for the simpler solution first. You may not want to try everything, just sample some. You wanna start with the easiest part of the search base first, be judicious, and so forth and so on. And most importantly, be selective. So reject any causal theory that would fail the constraints of unity of app-reception. With all these in place, the running, my interpretation of the app-reception engine gives pretty good results. So here I did a run. This is not cherry-picked. I decided to do one series of seven runs and collect the data and show it. And in this run, I set up the app-reception engine to only accept a perfect causal theory, one that would produce a trace that totally covers the observation. So, and I did seven runs. The first one succeeded, found it in four seconds. The second one, it took 102 seconds. So there's some randomization in the order in which things are searched, is luck is involved, as I said. The third one, one second, that was pretty cool. The fourth one, well, took 204 seconds. Then 96, 12, 99. So quite a good distribution here. Now I said, okay, I'm gonna run the app-reception engine again on the same training set that I showed earlier, those two lights. But this time I said, I'm gonna accept the theory that has at least 85% or more coverage. So it predicts, it recreates the observations well enough, but not perfectly. And I time-boxed it to 30 seconds. So yeah, 30 seconds to find it go. The first run, it find a causal theory with 75% accuracy immediately, then the same accuracy, same coverage, 10 seconds. 11 seconds, it hit 29 seconds. It found a perfect one. Then 11, it found 87% coverage and stopped right there. That's good enough, 75. Again, 100% is the first one it found above 85 in 18 seconds and 75% zero seconds. So a good distribution again. So we're getting into reasonable times. We're not talking about hours here, we're talking about seconds. And I'm hoping to do further optimizations and bring it down to something even smaller so that a cognition actor can say, I wanna make sense of these observations, query the app perception engine and get an answer, a causal theory within maybe a couple of seconds, that's my hope. Now, something interesting here, it so happens that what makes it hard for the app perception engine to find a good causal theory is formally equivalent to what makes cognitive science as a whole hard. And this paper here makes the case and proves the case quite cogently. So cognitive science wants to find models, functions or algorithms that explain account for situated behaviors. So you feed into the cognitive science machine, the cognitive science machine, pairs of situations and behaviors, and you wanna come out of it, a model, an explanation, a function or an algorithm that accounts for it. Well, the paper makes the case that if the explanation is to be bounded in size, then the problem is computable but it's not tractable in the sense meaning that it's combinatorially explosive. But once you have a solution, it is computable and tractable to verify that the solution is good, that it accounts for the data that you're trying to understand. Well, this is equivalent, formally equivalent to what the app perception engine is doing. My implementation was done in Prolog. I will not go into the details, it's about a thousand lines of Prolog. I'll just say that Prolog is a programming language that use deductive inference as its model of computation with backtracking. So essentially it searches for a solution and will backtrack if it took the wrong branch if you want it and will look for a different way of satisfying a line of the program. So let's just say that it makes traversing a search space. It gives, we get traversing a search space for free when we program in Prolog. I won't go into any more details, but you can see some Prolog code here. And it's, the fun thing is that a Prolog program is akin to a logical description of the problems trying to solve. I think it's very cool. And Prolog environment was augmented by something called constraint handling rules, which is an extension to Prolog that adds deductive reasoning. So basically in the program, you can say, assume this is true until proven otherwise. And the rules, the CHR rules are there to verify if it can be proven otherwise. So again, I'm not asking you to understand this quote at all, but I want you to realize that this code is the code that actually executes a causal theory, both the static and causal rules to build traces. It is that small. It's very powerful. So combining Prolog and CHR, I found extraordinarily powerful and very excited about it. I'm a programmer. Next steps. Well, next steps, now that we've solved individual learning by cognitive actors, well, I want to move to beliefs from sensations and to policies to validate or eliminate beliefs. A lot of these beliefs actually fall out of our perception. Layton objects and latent relationships and properties are can be considered as beliefs. Then there are other kinds of beliefs that can be obtained from what's been perceived. There'll be introspective beliefs that communicate how the cognitive cognition actor is doing in terms of competence, predictionary rates, how well its apperception is doing and whether it is engaging with other cognition actors. Is it relevant? I will have feelings which will provide normativity to these beliefs. So if feelings are signals of risks to homeostasis, loss of resources, physical damage, too many prediction errors. So that's anger, pain and fear. And feelings will take beliefs over time. And tainted beliefs and good beliefs, bad beliefs will want to be eliminated or validated through policies that will be synthesized by the cognition actor. And each cognition actor will emit, make available to others its API, what predictions can be made about the beliefs of this cognition actor, what actions are available to others to be asked of the cognition actor. And then as cognition actors connect to one another, as the cognition actors form the unbelt of other cognition actors, then the cognition actor will be able to predict the beliefs of others, will be able to compose policies made out of actions that are implemented by other cognition actors. And eventually we'll have a society of mine, which is a bunch of intersecting boom belts. So that's it. So I see what the society of mine is a complex system of collective theorizers. And I'm going to try going further with this project to answer the question if a collective of theorizers can self-organize to actively sustain itself. So thank you to the Active Inference Institute for inviting me to present and for providing a home for this project and for the constant support and encouragement. I'll see you later on Discord. Thank you. Awesome. Thank you, JF. Just to conclude the session, I'll read two questions and let's maybe address them in an upcoming robotics and embodied meeting. So if you're excited about this project, certainly we all are and about symbolic Active Inference. Join the Discord and participate in the robotics and embodied, but I'll drop these two questions from David Williams in the chat who wrote. One, how important is conducting this work in real world versus simulation? And two, what tools or components are missing in the robotics toolkit to make this research easier and better? I know those are things you have a lot of thoughts on so I'll look forward to discussing with you more. Thank you, JF. Thank you. Peace. All right, see you. And it's a great segue from collective behavior in surprise minimizing agents to collective behavior in surprise minimizing agents, I guess. Welcome, Connor. Hey, sorry, I just turned on my audio so I just heard the last thing you said. Yeah, hey, welcome to hear you. Yes. Thanks. We're looking forward to your presentation, multi-agents, Active Inference and multi-scale alignment, current developments and challenges. So feel free to share your screen or proceed however you prefer. Great, thanks. And before I start, can I just, because I've had issues with my voice, like saturating, like my microphone saturating, how's my audio? Is it clipping at all or is this on okay? This is good and I'm watching it. Okay, perfect. Thank you. I will share my screen. Whether share this one. The one thing I'll note, oh yes, okay. It's just JF symbolic implementation does not use statistical distributions. It uses the symbolic and the logical inference and now we're gonna move back into the distributional space and it'll be awesome to see similarities and differences. So thank you, Connor, to you for the presentation. And thank you Daniel for the introduction and for inviting me as well as big thank you to the other organizers of the third applied Active Inference Symposium. Yeah, I'm really happy to be here to present. So my name is Connor Hines. I'm a PhD student at the Max Planck Institute of Animal Behavior and I'm also a researcher at the Versus AI Research Lab, R&D Lab. So I'm gonna do something a little bit, I guess, unconventional for people in my position. Like I'm a junior researcher coming to the end of my PhD. So usually when I give a talk I would present on my own research, like what I've been up to for the last 10 years working on. But instead of that, I'm actually gonna talk about, given the motivation of the symposium, I'm gonna talk about something that's more of an overview or perspective on the current state of the field in multi-scale Active Inference or multi-agent Active Inference and what in my opinion, we need to do to move forward as a field. I think that's very resonant with the kind of motivations and the title indeed of this symposium. So I'm gonna give a general analysis of what multi-scale Active Inference is, why it's important. I'm gonna provide a brief analysis of its formal basis as it currently stands. And then what we need to develop in this kind of sub-discipline of Active Inference, multi-scale Active Inference, to really make it rigorous and really to actually reap the benefits of what it promises. So generally Active Inference has been used a lot to design agents that can solve problems, plan, and just generally emulate behavior that we deem intelligent, which includes things like risk-sensitive decision-making, intrinsic motivations to resolve uncertainty. And finally, from a more scientific standpoint, the ability to firm a process theory about how biological brains actually might work. But in a lot of the theoretical work on Active Inference from the last 10 years, or 10 plus years really, there's also alongside all the kind of practical building adaptive agents, there's a claim that Active Inference is inherently or intrinsically multi-scale from the very get-go. It is a multi-scale framework. It's not just about building single agents. So it's really, whenever we write down a single Active Inference agent, what we're implicitly implying is also a nested hierarchy of Active Inference agents. Both below and above. So colloquially, you'll often see this in papers as the idea that there's Markov Blankets all the way down. Markov Blanket is a statistical structure that's very kind of intrinsic to the definition of agents as they are defined under Active Inference. So I'm not going to get into defining that, kind of assuming that there's a more disciplinary audience there, but I'm sure other talks, for instance, imposing can provide a better clarity. So yeah, Markov Blankets active inference all the way down or all the way up. And at any given scale crucially, the free energy minimizing dynamics or the Active Inference dynamics are kind of claimed to be aligned with or parallel to the free energy minimizing gradients at the level below and above. So what the claim is that as agents are doing their thing and doing Active Inference at one level, it both entails and is constrained by Active Inference processes of the macro agent that they're participating in. So I'm a cell that's part of a tissue, as well as the micro agents that comprise them. I am a free energy minimizing cellular agent comprised of organelles that are also minimizing free energy. So this kind of constrained, neat nested gradient descent on free energy is part of the story of multi-scale Active Inference. And it also crucially assumes that these dynamics are aligned, correlated, co-operative across these different scales. So I should mention that there is a formal argument made more recently, I would say the last five years about how this is possible. And it relies on an apparatus from statistical physics called the renormalization group. This basically allows you to analytically identify shared symmetries, energy and conservation laws at different scales in a given system that's comprised of subsystems and subsystems, so on an infimium. So there's a formal argument specific we made in the a free energy principle for a particular physics monograph in 2019 that applies the renormalization of apparatus to multivariate stochastic differential equations that are kind of the equivalent of agents. So you can apply that framework to certain sorts of coupled stochastic differential equations that exhibit Markov-Blanket's squash-coupler instructor, and you can kind of prove analytically that there are going to be nested systems of Markov-Blankets and that they're all, in some sense, minimizing free energies at their own scales. So I'll get more to that argument later, but I just wanna mention that as I define multi-scale Active Inference, that there is a formal argument that's related to it. So this slide I just put together to demonstrate the idea of nested free energy minimizing processes visually. So at a given scale, we can think of an agent as occupying some point in its free energy landscape indicated by this red orb, which represents, say, its configuration, its beliefs and its actions, and it performs Active Inference and in doing so, minimizes its free energy. So it changes the position of that ball on that landscape and that is all we need when we say Active Inference. That corresponds to the agent doing inference and doing action and kind of getting to the fixed point of its local free energy landscape. The multi-agent case is simply when we add more of these processes. So there's other agents, usually assumed to be similar agents, and the word similarity, let's put an asterisk on that, and they're all sitting at different ports in their own free energy landscapes. They're all, the position of their local red ball is maybe in a different place. So the claim of multi-scale Active Inference is that as we link these multiple Active Inference agents together, so they can actually exchange information like observations and actions with each other, what we automatically get is some kind of super agent that is also minimizing variation of free energy and is in some sense an emergent or supervenient Active Inference agent. And I say the word we automatically get a super agent with an asterisk because there may be some conditions on that mapping from local to global that have to be elucidated. So we'll come back to that in a bit. But in short, I think the definition of multi-scale Active Inference is very eloquently put in this paper by Rafa Kauffman, Panof Gupta and Jacob Taylor. I think Rafa Kauffman is actually gonna be on the panel later. And this line from their paper is a really nice, I think just summary of what it is. So I'll just read you that lad. The upshot is that in theory, any Active Inference agent at one spatiotemporal scale could be simultaneously composed of nested Active Inference agents at the scale below and constituent of a larger Active Inference agent at the scale above it. In effect, Active Inference allows you to pick a composite system or agent A that you want to understand and it will be generally true that both that agent A is an approximate global minimizer of free energy at the scale at which that agent reliably persists and that agent A is composed of subsystems, a sub-I that are approximate local minimizers of free energy. So that is the claim as I'm going to continue evaluating in this talk. And I think it's just a great reference point to make. So okay, that's what multi-scale Active Inference is. Why is it important? Why do we actually care about that? Sounds feels awfully nice and beautiful visually, but why is that important? So there's a ton of actually really important implications of this, both for the engineering and the natural sciences. First of all, the namesake of this symposium, I assume inspired by this recent paper by Carl Friston and all about enacting ecosystems of parenthetically shared intelligence. So this is the third applied Active Inference symposium. So to really make this resonate with the applied aspect, let's make this very concrete. If we can figure out this multi-scale endeavor then we can actually engineer distributed systems of multi-agent intelligence, where local agents and doing their own little local Active Inference processes are also cooperatively instantiating a global agent that's also performing Active Inference. This has huge computational potential, of course, compared to kind of the state of the art predominant methods for artificial intelligence, which are deep learning, which really is about propagating global information through an entire computation graph. So although you could argue back propagation is local in some sense, it's really not local in the way that multi-scale Active Inference promises to be local. So if we can figure out how to actually engineer multi-scale Active Inference, it will have really tremendous implications for the study of artificial intelligence, just from that pure engineering standpoint. It'll be cheaper in one word. It'll be computationally energetically memory wise, cheaper, a lot cheaper. Secondly, from a kind of more natural sciences motivation, which is kind of where I'm coming from, I'm doing a PhD in biology. So I'm interested in questions about actual real systems in nature. Just the idea of being able to get super specific and rigorous about phrases like emergent intelligence, emergent computation and collective intelligence, super organism that's often thrown around when talking about social insects, right? These are terms that you hear thrown around in many different scientific disciplines that deal with multi-agent systems, network systems. But none of these terms, to my knowledge, have really rigorous or precise conditions. Multi-scale Active Inference is a kind of framework that's in the position to provide those rigorous definitions and conditions. So from a scientific standpoint, it could really lend a lot of potential and usefulness for other scientific disciplines. And finally, another pragmatic motivation, there's loads of fields that are obsessed with designing and engineering systems where local selfish individual behavior can lead when networked appropriately to some desired collective outcome. And these disciplines you really wanna figure out how to engineer that properly. So this goes from the design of financial markets and trading systems all the way down to how do you design a multiplayer video game. So that's kind of just motivating why is multi-scale Active Inference even interesting? So then the question of course becomes, is the multi-scale Active Inference claim actually true? Are all multi-agent Active Inference systems comprised of and themselves comprise nested hierarchy of free energy and minimized agents? A glance at a smattering of other scientific disciplines that specifically deal with multiple agents, multiple interests, collective phenomena like coordination, group behavior, collective intelligence? A glance at all those disciplines would naively suggest that the answer is no. So there's things like frustration in thermodynamic systems, game theory, the very existence of zero sum games and Nash Equilibria, bandwagon effects when we're talking about social networks and opinion dynamics, sacrifices for the common good, which we see in different contexts in biology like in the context of kin selection, but also in the context of arguably cell death in the tissue. These are all basically plenty of systems where local constraints and global constraints or desires or free energy gradients, whatever you wanna call them, coming to direct conflict. So the obvious example that I enlisted at the top of these bullets is the idea of geometric frustration that we see in Ising systems with very low temperature. So these Ising models basically describe lattices of ferromagnets that are happy when they're pointing in the same direction as their neighboring ferromagnets and they can be in an up or a down state. So basically the magnet can be pointing up or pointing down and these global systems are defined by a global energy function and the whole system is in some sense trying to minimize that global energy function. But sometimes you'll find cases in these collective systems where this little spin in the middle cannot be happy because they're getting conflicting information from two neighbors. I wanna be pointing up in blue like the agent on the left, but I also wanna be pointing down like the agent on the right. So this is a system that's collectively finding some fixed point of its global free energy, but it's actually leading to a local conflict where this agent is not at a point where it can do anything to make itself happy or minimize its free energy further. So just from even the basic glance at ferromagnetic lattices, we already see instances where local and global gradients or local and global optima are not aligned in the right way. So given all this, the burden of proof for multi-scale active inferences is still on us. So we need to show that collective active inferences generically do align again across scales. And maybe if we can put an X across the word generically and it's not some automatic condition, if they don't, then at least we have to establish exact data conditions in which they do. So anecdotally, we do actually have some conditions. There seem to be some kind of basic ingredients to get collective active inference to work. So one is that we basically need agents to exchange actions and sensations across some kind of Markov blanket. This is not really a condition. This is almost more part of what it means to be an agent. So having Markov blanket separation between agents is just another way of saying we have multiple agents in our system rather than a single agent. If you're violating a Markov blanket property, so internal states of one agent are not allowed to see the internal states or external states of another agent, then you're kind of cheating because you're kind of saying it's actually one agent and what you're doing is information sharing within the brain of a single agent. The second condition, which is something that's often elucidated more anecdotally and not really formally, is this idea that agents need to have some kind of shared narrative or shared hidden states or sensor space in their generative model. So I've worked a lot on collective active inference systems just simulating agents and trying to get them to do interesting things together. And my intuitions and experience do agree with this basic fact. If the agents don't have any similarity in what they're representing or trying to achieve, then it's kind of like trying to fit a square peg into a circular ball. So this is really nicely elucidated in the one of the earliest cases in this paper, a duet for one by first and then Fred from 2015, where they show that for two agents to really align, they kind of have to have a shared generative model and then you can get kind of this nice synchronized behavior. Again though, these things, like what does similarity mean? What is a shared narrative actually made formally mathematically? Those things have not been elucidated yet. So right now, a lot of the building of these collective systems is based on our intuitions and kind of engineering things in using some vague guidelines like, oh, they should have a shared sensor space, but there's no mathematical conditions or guarantees about what degree of similarity is needed between two agents models to get the intended dynamics. And finally, we have to have at least some agreement between the generative model of each agent and the generative process, which is really the behavior of the other agents generating their data. So this is kind of related to the previous point about having shared generative models, but just to be very specific, the physics of the space that transfers your actions to my observations, that physics can't be dramatically, crazily different than how our generative models represent those physics. So if we took two fish with the same generative model of each other and they normally would school together in a fish tank, but we throw them in a volcano or shoot them out into outer space, they won't school together because then the generative process is so dramatically deviating from the way they're representing that physics, the way their generative model is constructed. So these are, again, just ingredients, kind of guidelines or anecdotes, but there's nothing really rigorous behind these conditions. They're more like a list of best practices. So now let's get onto actual rigorous stuff. So the first real rigorous attempt to show that multi-scale active inference generally works is in one section of this free energy principle for a particular physics monograph from 2019. So let's use this apparatus I mentioned earlier, the renormalization group operator to basically show how one can successfully course-crain multivariate stochastic differential equations that admit sparse coupling between their state variables. So the main result in my mind that connects these renormalization results to multi-scale active inference is the fact that the Lagrangian of the system at one scale can be expressed as a function of the Lagrangian at other scales, and that applies in a scale invariant fashion. That is the main output or the main benefit of using a renormalization group apparatus. So you can kind of think of Lagrangian like the generative model. It's a physics term, but it's related to the generative model of the agents that comprise the system and therefore also their free energy. So in terms of active inference, it means that this reasoning of the renormalization group can be used to smoothly move between the models of individual agents at one scale and the model of a collective or larger agent at a different scale or a smaller scale for that matter. And the nice thing about it is general for all kinds of dynamics and thus generative models. It doesn't depend heavily on the form of the stochastic differential equations that form your system. The issues with it is that there's not a global link to basic mechanics and active inference. It's still all done in the traditional physics formalism. So we don't actually have an explicit link to local inference and global inference. Although if you know the connection between the Lagrangian and the generative model and the free energy, then you can make that connection, but it's not actually made explicitly for us in this part of the monograph. It also requires the assumption that the generative model and the general process are identical at the local level. That's related to how the Lagrangian is defined. That's also a restrictive assumption that's probably not realistic in my opinion. And then finally, there's something about kind of spatial temporal segregation of scales. So we need to make assumptions about how fast random fluctuations are at one scale relative to another scale in order to justify kind of course framing or forgetting about certain states as you move between scales. And that's also something that you could argue current researching to collective dynamics challenges, that assumption about how fast noise is at one scale relative to the next. So now I'm going to discuss quickly another small contingent of active inference research that is attempted to address this mapping between local and global inference processes. So what I want to kind of just generally say with this presentation and to our community is that the kind of approach taken in these two papers, first of all, active inference model of collective intelligence and spatial systems, just collective active inference, this is one of the types of research I think we really need to move active inference, multi-scale active inference forward. So I'm not trying to be too biased because I am the first author on the second of those papers, but I'll also be the first one to point out the limitation. But benefits-wise, I think these approaches are really important because they formally relate a local generative model at one scale to a global generative model at a different scale. So really tie how did the parameters of one model relate to a course grade model? And these are really good steps in the direction of a formal theory of collective intelligence that goes from local intelligences to global intelligence. However, there's still issues with these. One of them is that they only deal with inference at the global level, not active inference. So both these papers concern with a bunch of local active inference agents that cooperate to form a global inference agent, like a passive Bayesian agent, rather than that active inference agent. And it's also unclear whether the systems studying these papers are actually very generic, like the results are generic to study collective intelligence in general, or they're nice formal arguments, but they're only applicable to these specific systems. So we still don't have something that's even more kind of zoomed out and abstract than these, which tend to be a little bit case specific. And finally, the actual scale transcendence that we're doing in these papers is still relegated to really one step. We're not doing the full multi-scale, infinite-scale regression that's something like renalization of promises. So that's kind of the current overview of what I think are the most promising directions in multi-scale active inference. And I'm aware on time, sorry, five minutes. So I'm gonna quickly try to go through what I think are really promising directions to push in terms of multi-scale active inference and kind of intuition partners that I think will help us study these systems in a way that's different and also actually better informed by other research disciplines. So the general idea that I'd like to put forward is that misaligned gradients can actually be a good thing. So it's actually sometimes good when local free energy gradients are misaligned with global gradients. So sometimes the global system will actually do better if the local systems are performing worse. So this is something you could call multi-scale conflict where the free energy minimizing processes at one scale are actually doing bad quote unquote, but it's because they're being driven by some higher scale process that is doing well. So rather than trying to always avoid constructing processes like this, I think this kind of frustration to use the analogy from statistical physics can actually be an inspiration for what we should investigate further because it actually might be key to facilitating optimality at different scales. The reason I put this forward is because there's loads of research, just recent research in the last several years that are suggesting that actually making local agents more frustrated or more unhappy might coincidentally or not coincidentally lead to better collective or global outcomes. So this is expressing various forms in various bodies of work. One of the biggest patterns I've noticed is the study of collective behavioral systems over the last several decades. The idea that local noise and local dysregulation can often facilitate global coherence or global coordination. And where multi-scale active inference has something to say in my opinion is in framing this benefit of local frustration in terms of a misalignment of free energy gradients. So it may be that actually temporary misalignment of local free energy gradients from global ones may facilitate the descent to fixed points in the global free energy landscape that satisfy everyone at all scales. So I'm basically expressing an idea that's been known in various communities like stochastic optimization and stochastic resonance theory for decades. But I think we as active inference practitioners have a new and potentially useful perspective to shed on that, using the language of active inference and free energy minimization and Bayesian inference in general. So instead of thinking of accelerating optimization by just adding noise to the system, we can think of exactly how to design local generative models such that there's an optimal misalignment of local and global gradients or locally global generative models in a way that facilitates everyone in the long run actually facilitating or minimizing their free energy. So yeah, that's just kind of like something I'm putting out there. I'm investing in it now in my own work but I have no real results on that but I just want to put that out there in this venue because I think it maybe will inspire other people to think in a similar way. So just to conclude now, multi-scale active inference, I would say is still largely based in theoretical or philosophical descriptions and illustrative simulations but we're still lacking a formal theory. There's some theory in terms of the renormalization of arguments of the monograph but they're still in my opinion, a bit underdeveloped, a little under demonstrated and will align some restrictive assumptions like the fast and slow fluctuations, the identity between generative model and process. There's a few more recent papers, those two by Kaufman et al and then by myself on spin glass systems that have attempted particular proofs of multi-scale Bayesian inference systems in particular situations but their generality is still not known and not proven. So what I'm kind of trying to conclude with is by saying we need to incorporate findings from other disciplines relate to the role of noise, conflict and frustration in facilitating not subverting collective intelligence or collective coordination. And I think we can really benefit by looking at these other disciplines to help us build a really powerful formal theory of multi-scale active inference. And finally, I think we need to set the goalpost for what counts as a formal proof of multi-scale active inference. And once we get there, once we're saying, okay, this counts as proof, this is satisfying how can we use that to actually do the hardest part in my opinion, which is engineering actual multi-scale active inference systems that are intelligent and minimizing free energy at multiple scales. Yeah, so with that, I'm going to conclude. Looks like I'm just on time. So yeah, thank you again for the invitation to present. And I'd like to thank a bunch of people who are listed here and beyond who have influenced my thinking and kind of my opinions. Yeah, if there's time I'm happy to take any questions and close. Awesome. Great. Talk. I'll just give a few seconds if anybody wants to type in a question. Also, it's really cool like Aswin in the previous session was highlighting PMDP and just the way in which we enact the collective intelligence, different people seeing a paper where an analytical formalization is introduced and then there's still so much work to get it to the package and then so much more work to take it to the last mile. And I think your presentation really checked a lot of those boxes. I'll just read a question and then that will just be an appetizer for our continued discussion. So Marco Lin asked, do you expect the inferentially connected dynamics to exhibit behavior akin to theories of multi-body systems and to what extent can we transfer insights from that multi-body of work? And then second question just for our thinking and learning from Marco. Have you explored integrating work with integrating work on self-organized criticality with multi-scale active inference or other frameworks who can provide more flexible frameworks or assumptions for a generic notion of multi-scale dynamics? Great questions. I hope that we can continue. Have you back any time or just continue to collaborate in the ecosystem? So thank you for the epic talk Connor and good luck finishing your PhD. Thanks a lot, Daniel. Yeah, talk to you soon. Talk to you soon. So thank you for that. Greetings. All right, well, our next session. Hey, Burt, greetings. Great. How are you doing? Good, good. Very well. Our next session is with Burt DeVries, Dmitry Bagheev and Bart Van Erp. It's gonna be called Towards User-Friendly Design of Synthetic Active Inference Agents. And I know a lot of people are super excited to see this really practical and cutting-edge work. So to you, Burt, and just let us know how we can support. Okay, great. Well, is my audio good? Yep, sounds good. Okay. Then I'm gonna share my screen. I hope I picked the right one. I don't work. It's Zoom quite often. Looks good. Looks good. All right, super. Well, thanks a lot, Daniel, for hosting this symposium. I've been watching some talks. It's really amazing. And we really feel privileged to get a chance to present ourselves. So we are also, just like a few others before us, interested in developing a toolbox for Active Inference. And so this picture, or it kind of shows what we're about or what we're interested in. So here's a lady on the left-hand side. And I'm gonna try to get a laser pointer. And she has this idea about rewarding behavior for a vacuum cleaning robot, right? So she's writing down, she has a textual expression. Move around the apartment, apply suction until the floor is clean. Do not touch objects and when don't return to the dog. So that's not so hard. I'm gonna rate that with one star out of three stars in terms of difficulty level to specify that. But that's not enough to program this robot, right? Because what she really needs to do now is to specify a generative model. And that, you know, there's effectors and actuators, right? The robot has to move around, apply suction until the floor is clean. So there's sensors, probably a camera. Do not touch objects. So maybe there has to be object recognition. This is a really difficult task to come up with this generative model here. And on top of that, she has to specify this kind of rewarding behavior in terms now of probability distributions of this generative model. So very difficult, I'm gonna rate that with two stars because the next thing she has to do for this model is to specify the inference procedure to do actually active inference and free energy minimization in real time for this complex model. And really that's almost impossible, right? Only a few specialists can really write a procedure for a variational free energy minimization in some very difficult model. So what we are about, what we've been working on is to try to automate the inference task. So get rid of the three stars. And yes, she will still have to specify a model but in the long term, we try to get away from that. So in the long term, we hope we will get a toolbox and now we're talking five, 10 years, right? Where a textual description would be enough to specify some initial model with an initial prior and everything else is just automated inference. Learning of states, parameters, structural adaptation of the model even maybe based on her feedback updating the prior. So that's long-term. For now, we would be very happy if we could just automate the inference task. So why is it so difficult to specify inference for an active inference agents? Well, we have so many competing KPIs, right? We want to do this for large model scopes, not just for ABCD models, but maybe there's also continuous variables and hierarchical models, right? Must be very user-friendly. We really don't want her to worry about robustness of her code. We don't want her to worry about whether two variables have conjugate relationships, right? Adaptivity, we want to update states, parameters, maybe even the model, the model structure has to be low-power because these agents often run on edge devices, right? So they run on their battery-powered. That's to be in real-time because you can't learn how to ride a bike if there's no real-time reasoning. And on top of that, you actually want to minimize variational fruity, right? You want to do it at least as good or at least in the neighborhood of if you would do a manual derivation. And some of these deciderata bite each other, right? If you want to minimize variational fruity, but you have to do it in real-time and on low-power, yeah, that kind of bites each other, right? So these are difficult KPIs that we want to, yeah, they're all important. You can't just take one out because then the whole system wouldn't work. So when you read papers on active inference, you often also read, and now we implement variational fruity minimization and that can be done by message-passing on a graph. And I want to clarify first why it has to be done by message-passing on a graph. I'll do that by giving a very short answer and then do an example. The short answer is that Bayesian inference involves computing very large sum of products like what you see here on the left-hand side. Here's a product, AC, AD, BC, and then we sum them in, AC plus AD and so forth. This is a sum of products. Now, we know by the distributive law that this here on the left-hand side can also be computed as on the right-hand side. If I multiply this out, I get A times C plus A times D and so forth. This is a product of sums and they are exactly the same thing. The only difference is that to compute the left-hand side takes four additions, sorry, four multiplications and three additions. To compute the right-hand side takes two additions and only one multiplication. So on the right-hand side is much cheaper to compute than the left-hand side. Normally, when we write down marginalization and Bayesian inference, we write things down as in the left-hand side. What message passing does on the graph, it will automatically convert that into much cheaper to evaluate product of sums. And I'll give an example of that. So here is an example model, f of seven variables, x1, x2 through x7. And this model happens to be factorized, fA of x1, fB x2, and so forth. Now we can draw this factorization as a graph. And what we do, and this is called a Forney-style factor graph, what we do is for each factor, fA, we allocate a node. So fB gets a node and fC gets a node. And we associate the variables in our system with the edge. And an edge is connected to a node if that variable is an argument of that function. So fC is a function of x1, x2, x3. And that means that fC connects to the edges, x1, x2, x3. And fD is only function of x4. So fD only connects to the edge, x4. So what you can see in this graph is nothing but a visualization of the factorization assumptions that we have for this model. Now, if I'm interested in a big marginalization task, I integrate out overall variables, but x3, so x1, x2, x4, and so forth through x7, I'm interested in this. Then taking advantage of this factorization, I can rewrite this sum, this basically this sum of product into a product of sums as below here. What you will see here below, this computes exactly the same thing, but I've made use of this distributive law. For instance, fC contains no x4, no x5. So I moved it over the summation sign to the left. And fB also doesn't contain x4, x5, x6, x7. So I moved it all the way to the left. And when you do that, you are left here with an expression where I only sum over two variables. And here I have to sum over six variables. And here over two and here over two. So you can imagine if each variable, let's say x1, x2, if each variable has 10 interesting values that you need to sum over, then I have here the original marginalization problem. I have 10 to the power of six, so a million terms. And here in red, I have 100 terms. And here I have 100 and here I have 100. So here I have 300 terms and here I have 1 million terms. So it's an enormous reduction in computational complexity when we make use of this distributive law. Now it turns out that if you write this out, you can associate these intermediate factors with messages on the graph. It's just an interpretation, a visual interpretation. It's as if FC receives a message from FA and FB, receive or FC receives a message from FB and computes an outgoing message, mu x3. And the same thing for FE. So FE receives a message from neighboring factors, FD and FF and computes an outgoing message, X3. So what you see here is that the entire marginalization process can be represented as basically computing a few messages on a graph and multiplying some of these messages with each other. And that's how you can do Bayesian inference and also how you can do variational free energy minimization. So this works in factorized models, but I would say even stronger, if your model is not factorized and you have a lot of variables, there is just no way you can do proper inference. So any serious model is factorized. Like the brain is almost sparse, is almost empty. We have about 10 billion neurons and each neuron connects to a few thousand other neurons. So if I would draw the graph, that graph is almost empty. It is hugely sparse. And so there is no other way to do inference in the brain than by message passing. So that's why message passing just because it's more effective than anything else. Now, then the issue is, which message do you compute? How do you compute messages? Because there are different ways of doing it, right? And we also read in active inference papers, you can do this by variational message passing or expectation maximization or belief propagation and variational Laplace and all these terms. It turns out that there is an umbrella framework for all these message passing frameworks. And that umbrella framework is called constraint better free energy minimization. And I will try to illustrate it by, well, by this slide. So here I have this graph. This is just an example graph where my generative model is basically factorizing F A, F B, F C, F D and F E. And I've also written that here. So this now is the variation of free energy. Now, I haven't made any assumption on Q of X. So Q of X is still Q of X1, X2, X3. It's just a joint overall variables and it doesn't have any factorization assumption. It makes sense to also assume that the posterior kind of follows the factorization assumption of the prior, namely of the generative model. If we make that assumption and that means we're gonna make the assumption that Q X is also now a product of Q A's of X of A where Q A's of X of A stands for beliefs over nodes. What I mean by that is that Q of B is a posterior belief over this node, meaning it's a posterior belief over the edges that connect to this node. Just like F B is a function of X1, X2, X4. That's, if you will, the prior or the generative model. Then Q of B, the variational posterior for this node will also depend on X1, X2, X4, and on no other factors. If you just do that, then you will count some of the variables double because X1 is part of the belief over F A, but also part of the belief over F B. So we just have to discount that by dividing by beliefs over edges. That means that I make now an assumption that my posterior beliefs is divided into local beliefs over nodes and local beliefs over edges over variables. This will make things a lot simpler. In fact, if my graph is a tree and it is a tree here, and I would do message passing on that tree and I could, suppose I could do that perfectly, everything is linear Gaussian. Then I get perfect Bayesian inference. There is no approximation. So this is a good assumption. Sometimes it's still very hard to compute a message because even the single messages that come out of these nodes, there's still integrals or summations. And in particular, the integrals may be a problem. We may not have an analytical answer. So what we sometimes do is add additional assumptions. We'll say, well, the posterior belief over F D, I can't compute it in general, but I'm gonna just assume now that it's a Gaussian. That makes it easier. Or we can make an extra factorization assumption and say the posterior belief over F B, which is really a belief over the joint X1, X2, X4, is gonna be broken into independent belief over X1 and belief over X2 and X4. These additional assumptions, if I impose them as well, this is what I recall now. If I all substituted here in Q of X, I get what's called a constrained Bethlehemm energy. This is the same Bethlehem as in the Oppenheimer movie. This is Hans Bethlehem, where it's named after. And so we have a graph now that is highly factorized and we have local beliefs over nodes and over edges and they're indicated with red. And we have additional constraints in green. They could be Gaussians or mean field constraints or other constraints. And now we will assume constraints that make it possible to compute all the messages. And now we can just automate this. By making different assumptions, we can turn this into expectation, maximization or belief propagation or hybrid forms thereof. We can turn it into any relevant message passing algorithm that you've heard of. So this is a very nice umbrella framework that basically encompasses everything. And there is, we've written a pretty large paper on this in an entropy journal where you can read all the math on how this works. So we've talked about why message passing, namely because it's the most effective way of doing inference. And we've talked about which messages to compute, namely we turn our variational free energy into something called a constraint better free energy and then we can compute messages. The only thing that's left is, well, when do we pass these messages? What is the sequence of messages? Which one comes first? And this is where we see a lot of papers, right? You have to write control flow. What's called control flow? You have to say, okay, here's my algorithm for active inference. First, I specify a model. Then let's do inference for every time step, collect a new observation, update the state, update the desired future and so forth, compute expected free energy, select the policy, et cetera. This kind of program, the problem with active inferences is that there is nested for loops in here. Here's a for loop and here's another for loop. And for each of these policies, I'm gonna have to go into the future. So I'm gonna have another time loop. So it is for loops in for loops in for loops. This will completely explode in terms of computational complexity. So as a result, some very clever people have written very clever algorithms of doing this much faster. Sophisticated inference, branching time, active inference, dynamic programming, EVE are recent proposals for doing this very clever. In the end, all of these proposals come down to a particular message passing schedule. Once we commit to message passing on a graph as our inference procedure, it's the only thing that's going on. And all of this sophisticated inference and branching time, active inference, all it does is it schedules the messages. It says, first this message, then this message, then this message. I don't mean that as a slight to these algorithms. They're very clever. And as we've seen in the presentation by Asvin Paul, you get huge improvements if you go from regular inference to sophisticated inference. But it's good to realize that these algorithms just specify in the graph which message comes after which message. So here's an example. Here's an example of a graph and a message sequence. Here's message one, then message two, and message three goes up. And then we go from FC to F, F and here's message five, and then we go to FE. And this could correspond this sequence to dynamic time programming EVE or sophisticated inference. There are a couple of problems with this approach basically with having the user to specify a clever algorithm. First of all, you have to be a specialist to do it, right? Only these are very clever people. So that means that if we let it, we leave it to say to an engineer in a company, it's well, it's the high probability he's not gonna get it right. That's very unfortunate. But there is another issue and that is that in a sense it's a global variable. In the message passing schedule, all nodes are visited because if a node would not be visited, then we shouldn't have it in the graph. And that means if one node crashes, basically the message passing schedule is invalid. I have to reset my system. And if you fly a drone, if it's deployed and it's out in the field, and a node crashes, a transistor burns out, and I have to totally reset now my system. I have to compute a new message passing schedule. Then you're not doing inference and you're drone flies into the wall. So this is not robust. And it also for a very, the same reason we may actually want to take out a node. We may want to prune a node. We want to do structural adaptation. And we can't do structural adaptation because we have to reset the system, recompute a message passing schedule. So this procedural style where an engineer, our priority specifies which message comes after this message has some disadvantages. It's not very robust. And if you want to do it very clever, you have to be really a specialist. So a better system is what we call reactive message passing. And it's very related to what was in the first session called the actor model. Keith Duggar had a nice presentation on the actor model. So what we will do is we will say we will not have a global message passing schedule. The engineer will not specify anything anymore. The inference code that an engineer will have to write is just say, react to any free energy minimization opportunity. In other words, there is no inference code. It's completely automated. And we will replace this global message passing schedule by local triggering inside the node. So each node is now just an autonomous system that's interested in minimizing its free energy. It can do so by sending out messages. And when will it do so? Well, it receives messages. I mean, it feels when it looks at these messages and it feels like, oh, there was an opportunity for me to minimize free energy by or expect the free energy by sending out a message, then it will send out a message. And each node will do so by itself asynchronously. So you get parallel distributed processing or concurrent processing as Keith called it. In principle, you could play this game on many computers at the same time. And so you get tremendous advantages. First of all, you don't have to write difficult code. Second of all, you can do multi-threading or you can run it on multiple computers at the same time. And so there's also robustness advantages because even node crashes, then there's nothing that stops the system from just finding another path, right? If this node crashes, this path from here's message three, this path now doesn't work. So I cannot send anything to FE anymore from X. Well, then I just sent a new message here. Why not? It's like when water falls down a mountain and it zig-zags its way down into the value and you halfway put up an obstruction, it just finds another path, not the preferred path. This has to find, well, the second best path because the first path has been obstructed, right? And that's what's gonna happen in this system as well, right? That's just how nature works. It tries to find the best path, the easiest path and if that's not available, then we do the second best path. And that's also what you can do with reactive message passing. So you can prove nodes, you can do structural adaptation and it's far more robust. And you can also get, let's do chance encounters with other drones, right? Drones that get close can start communicating with each other and when they're far away, they stop communicating with each other. And this is no problem because you can basically change nodes can change on the fly, who they communicate to and who they want to listen to. And so that's the way nature works and also how it works when we do reactive programming and reactive message passing. So in summary, we're interested in automating inference in active inference agents, right? Because it's an operation that's basically only for experts and this active inference technology is not going to be successful unless we get more people if we democratize it and we get competent engineers being able to develop good agents, right? You shouldn't have to be a top specialist in the world to develop an active inference agent. Now, in order to automate inference, you must do message passing and I've talked about that for efficiency. I've also talked about which messages to pass. Not necessarily do you have to follow this framework but constrained, better free energy framework is very convenient is an umbrella framework that basically goes over all the interest in other message passing computations. When this is passing, reactive message passing, it's fully automated so you don't have to write any code anymore. In principle, you can do parallel distributed processing. It's robust to structural changes. You can learn new inference pathways. So lots of advantages here. Now, how do we do it? I like to introduce a toolbox that we've been working on called RxInfoR and we do that with my lab here at the university. I'm here in Eindhoven in the south of the Netherlands and we have a lab. The lab is called Bizelab. Here are postdocs and assistant professors and PhD students. And we've been working on this for many years and some of these like Albert and Ismail and Thais have written dissertations and our best work, we have consolidated that in a toolbox. And the toolbox is called RxInfoR and you can, if you wanna have a look, you can go to the website, RxInfoR.ml. And RxInfoR works in the way that I've just discussed. It does message passing. It tries to minimize constrained, better free energy. That means it can come up with all kinds of message passing algorithms. It will do it in a reactive way and it will try to do it in real time and low power and all the KPIs that we're talking about. Now, it's of course not done, but it's functional. And I'd like to show some demos and I will leave it to Dimitri and Bart who are two advanced PhD students in my lab to show the demos. So I'm gonna stop sharing. Awesome. Thank you, Bert. Great talk. Sure. Can you hear me? Yeah, yeah, yeah. Okay. I will try to share my screen. Okay. So you should see it now. Looks good. Okay. So yeah, hello to everyone. I'm Dimitri Bagayev. So I'm a PhD student in Bioslab and I'm from the University of Technology. And yes, I have a small presentation about actual software developments in Bioslab. And yeah, so over the past few years we have significantly improved our tools. And basically my entire PhD was dedicated to implement this idea which Bert was talking about. Like implementing the variational reactive message passing. And this presentation I just wanted to show you what you can actually do using this theory under the hood. Okay. So basically in order to automate active inference, we need to automate Bayesian inference. And we have already a lot of solutions for that such as STEM, Pyro, NumPyro, which is funded by Google, Infer.net is funded by Microsoft. Turing is in July by IMC and many, many, many. So and basically these solutions are really, really good. So and they're really good at prototyping as well. But our goal is eventually to be able to deploy this kind of systems, not just prototype. And we are really focusing on this particular properties for this automated Bayesian inference. So it must be low power, adaptive, real time, scalable. It also must be user friendly at the end and it must support a large scope of models if you want it to be useful. So and in Bayes Lab, we want to build such a software with such nice properties. And it's always about trade-offs, right? So we do something better in one particular domain and maybe other software libraries, they might be better in a different domain. So, but we are really focusing on this particular properties. And so yes, I will reiterate a little bit there's presentation. So how do we achieve this? So we have, imagine we have an environment and we have an agent and the agent a lot takes some actions and the agent, basically what he needs is to come up with some sort of good enough probabilistic model of its environment in order to do Bayesian inference. And in our framework, we encode the model as a factor graph, which not only models the observations but also actions and desired future. And this approach allows us to decompose the complex relationships within variables and hidden states into some kind of structure and local blocks and it's not a black box anymore. So, and the model itself may have some sort of background motivation, interpretation and may encode your prior knowledge about some particular physical system. And the locality of these blocks basically allows you to scale to millions of variables and hidden states. It allows you to pre-optimize it maybe or maybe use like some sort of different approximation strategies in different places. So it allows a lot of very nice properties as well. And we use reactive message passing to run actual variational Bayesian inference. It uses reactive programming under the hood to minimize the approximation to the variational free energy. And yes, as Baird also mentioned, it's very much related to actor model. And basically in our X and Fur, you can think of different nodes as actors themselves. So, and they have basically one single purpose is to send a variational message that minimizes free energy. This is a very short and very high level description but it is essentially what is happening under the hood. So we are not treating different agents which interact with each other as actors but we also treat the actual components of the underlying model as actors themselves. It's like a very hierarchical structure. So this is the main central idea of this inference. So here, for example, first example we can do an inference in a dynamical system. This example, which is quite old already, I think it's like two years ago. So we track a position of the object given some noisy measurements which are indicated by green dots. The actual real signal, we cannot observe it but we just can plot it, is shown as blue and the inferred signal is shown as red. And the data set is infinite. The inference end just reacts on it and does not assume on any particular data size, simply reacts in the observations most possible. Yeah, I'm actually not sure how smoothly Zoom shares my screen. Maybe you can see it's a bit lagging in the animations. I'm not sure because maybe Zoom does not share it on a full frame rate. And also on the right-hand side, you can see how we define models in our framework. We use Jula as a programming language. And so basically this is everything that you need to define this particular model and run inference on a data set. And actually I like literally spend more days to plot it instead of inference. So inference was the easiest part for me. Plotting was way much harder to relate to user friendliness. So, and we actually have plans to improve our model specification language, make it even easier. So for now for technical reasons we have some auxiliary statements in the model specification language but we are working to improve that and make it even easier. This is another example, which is similar to the previous one that uses much more complex and linear dynamical system of the double pendulum. And the system is chaotic. And we can observe on this small part of it a lot of noise also indicated as a green dots. And nevertheless, given good enough model, you can infer the other hidden states with pretty much high precision and the code needed for that is also relatively short. We also have examples with active inference agents that interact with their environment. So the left up shows a mountain car problem, very famous problem. The left bottom side shows an active inference agent which tries to control the inverted pendulum from falling in windy conditions. It reacts in the wind. We also have a demo of an agent that controls a pendulum in an ever-changing environment. So on the right side you see a pendulum with an engine and engine has limited power and the agent itself needs to reach the goal and the goal is indicated as a red circle. So, and basically in this demo we can change the environment in real time and see how the agent reacts. So you can change the mass of the pendulum on its length or the amount of noise in the measurements or we can change the goal, we can change maximum engine power, et cetera. So the agent will still try to infer the best possible course of actions in order to reach its goal and it just never stops reacting. It's also actually possible to restrict engine power such that it will no longer possible to reach the goal. But the agent will still try. We have other cool demos with smart navigation and collision avoidance which are still under active research and the code for them is not available publicly. It will be soon available. But for example, in this example, we can define a set of agents with their boundaries and a set of their destinations. And we can see how they try to resolve their roots all together and we can have some static obstacles in the map. We can see how agents can find their most optimal path in order to reach their goals and avoid any possible collision. And it's also not necessary to have static obstacles. The obstacles themselves may move. So in this demo, we have hundreds of agents that navigate through a map of obstacles that move from bottom to top where the circles are obstacles and agents are depicted as small dots and they need to go from left to right, basically avoid any sort of collision. And as I also mentioned, we want to perform efficient and real-time inference but we also do it with low power, low performance devices such as Raspberry Pi or CoolPi as an example. And we have some results of successfully running the Bayesian audio source separation, for example, on CoolPi. So it is actually possible. We also run active, we try to run active inference agents also on CoolPi. So as the aforementioned inverted pendulum. And as I mentioned, we also need to have a large model scope. And basically RX infer has not been designed to solve any of the aforementioned problems specifically. We have a large set of different examples in our repository, different models, different data, different inference constraints. We have examples for linear regression, to the Markov model, auto regress model, hierarchy models, mission models, Gaussian process and so on. This approach is very versatile. And for example, if you compare it with sort of a conventional software libraries, where you, let's say, have a library that solves the Kalman filter, might be a very great library, maybe super fast, have top performance, works great and very reliable. Super good. But then you are constrained by this particular model, Kalman filtering, right? And you can't really change it much. In RX infer, we are free to define our own models, which we can pretty much easily define a model that essentially would act equivalently to Kalman filtering equation. And so basically in the demo that I showed before about object tracking, it was essentially a Kalman filter that was written in a probabilistic model. So yeah, that was my small addition to this presentation. So our software is free, some MIT license and it's open source available on GitHub. Yeah, and we would be happy, thanks to be able to present where we would be happy to answer any questions. Thanks. Awesome, awesome. All right, I'll just ask a quick question in the chat. Marco asks, sorry if I missed it, are the collision avoidance demos real-time adapting to other agents' behavior or is it a collectively pre-computed path? So basically they are not super real-time, they're kind of fast to compute this path like maybe five seconds or so, but we're basically working to make it real-time. So we know what is the problem, we know where to improve and we will make it feel that, yes. Almost, almost like that. Next question, do you have some comparative data with other methods and just more generally, what kinds of benchmarks or when you're talking with industry in different settings, what are people like looking for that killer app of active inference or what are they looking for their key measures? So I personally have a big paper about comparison with sampling-based methods like HMC and also in my PhD thesis there will be a comparison with nuts, also other sampling-based methods. So long story short, sampling-based methods cannot really run this kind of sophisticated inference in real-time. They're very time-consuming, they do not really scale well to large problems which is really needed for active inference agents because if you have like a large environment, very complicated, you will have a lot of unknown variables in your model. So, yeah. So there is a paper that compares and basically we show that, yeah, our approach scales much, much better. So I personally run on just a regular MacBook laptop. I run the model with two million unknown variables. And it was like quite, quite possible. And if sampling-based methods, you may find yourself in a model with like 100 variables and then you wait like two hours or something. And then it turns out that your chain did not converge. Or something like that. Oh, yeah. It's people commenting in the chat like how far message passing and factor graphs have come and so to bias lab and to BERT at all. We definitely appreciate this exciting line of research. I mean, there's so much to learn there and sometimes looking at the equations, it can seem like it's like written in stone and just sort of the beginning and the end is variational free energy. But then in your presentations, you're really showing like, no, we are hands-on. That's where we get the interpretability, the modularity. That's where it really is implemented. And it's like an information logistics challenge. It's not like an esoteric philosophy question at that point. No, no, indeed. I mean, I should say it's taken us, I mean, we are no geniuses, right? So our lab exists more than eight years and you see all the people in the lab. It's taken us many, many, many years with lots of wrong directions to get this to work to where it's now. So it's a very long path. But at this moment, I'm pretty confident that at some point in the future, and I don't want to say in three months or one year, but we will be able to write a toolbox that will allow people to design a generative model and just press a button and forget about the inference. You don't have to worry about the inference anymore. It will be fast and automated and that will happen and it will happen within a few years and maybe somebody else will write even better toolbox, but I'm pretty confident that even our toolbox will be able to do that. So, and I think that, yeah, you know, people talk about, so why don't we have the success of deep learning and generative AI, right? Well, they have to success because of big data, availability of big data, big computers and toolboxes, TensorFlow and all the successes. We don't need big data because agents collect their own data in the field. We don't need big computers. Active inference agents, you know, they manage their power resources, but we need a really good toolbox because programming an active inference agent, programming the inference by hand is just not doable. So we need a really good toolbox that really automates this. We hope ArcSympher will be one of the first toolboxes to do that. I am sure that other people will also be working on it and better toolboxes will come about. But I think the optimistic message is that it will happen, right? And once we have a toolbox like that, then we can actually, a large community can start building agents and we can actually show deployable agents in the field, right? That they work and they work better than the reinforcement floating agents or whatever is out there, right? So that's, I think that's a very positive and hopeful message. It's what we expect, it's what we prefer. Yeah, yeah, yeah. Any last comments from either of you? Comments from us? No, no, no, we just, I'm just very happy to get the opportunity and yeah, I want to, so, yeah, everybody can download this toolbox. I think at this moment, you still should be a programmer to work with the toolbox. And I hope you're friendly because, you know, it's not totally polished in the way that we want, but it's coming, right? It's coming in the next years. There will be a good toolbox for almost everybody to use. But people that are interested, even people that are interested to work here at Bias Lab, we have an open position for PhD students. So we're happy to receive emails from people that are interested to have to work with us. Thank you. Demetri, anything in closing? No, just thank you again for our possibility to present. Super nice to be here. Cool, yeah, well, later in the year, we will be discussing your two-part recent work. And so we're gonna be getting a lot into the details and I hope that people in the Institute and the ecosystem will be as excited as we all are. So thank you. Thank you, bye-bye. For the next interval, we are in the section hosted by the Active Inference Institute on our recent paper, Active Inference Institute and the Active Inference Ecosystem. So two short messages. First, if you're in the live chat, I'll post the link right now to the paper and this can be more of like a conversational interval. So if you have any questions about the ecosystem or the Institute or any of the stuff that we're talking about, let's make it a conversation. And then second message, if there are any co-authors on this paper who are listening to this stream, then message me on Discord or post in the YouTube live chat or just check your calendar and I expect some others to join in as we go. Welcome Pablo. We'll be hearing in a little bit from you as a presenter. Is there anything that you wanna just bring up or reflect on before I share the screen with the paper and we just walk through it, explore it and hear questions? I'll bring in the paper and again, anyone watching live, please just feel free to write any questions and I'm just going to bring it so that Pablo, you're not gonna see the paper but it's on the live stream. Well, here we go. This is a paper that had an awesome co-authorship and even broader circles beyond the listed co-authors of collaboration and contribution. So just like everything that has happened and will happen in our work, it was massively participatory and a real learning by doing and a real enactment moment. The first version was released three days ago on Saturday, August 19th and I'll briefly just list all the authors. First, we had the institute as the first author so that the citation can always remain like institute at all, institute comma active inference institute. And we also had Andrew Ghire, John Boyk, Libra Burrian, Matthew Brown, Archie Corday, Scott David, David S. Douglas, Pablo Fernandez Macchiere, Daniel Freedman, Holly Grimm, Avel Gwinnon Kalu, Maria Luisa Aineko, Virginia Blue Knight, Alexandra McKaylova, Ali Ramjou, Adil Razy, Jakobs McAul, Renan Tamari, Dean Tickles and Alex Vyakken. So this includes everyone from founders and officers and board of directors and formal participants on different projects and different roles as volunteers or facilitators or interns and also people who are continuing to join the fun and show up and makes the institute in our field what it is and what it can be, what we want it to be. There's a table of contents for those who like to click through and just broadly there's gonna be a preamble on active inference but outside of the preamble this paper is not conceptual, it is not technical, it is not philosophical, although perhaps it can be read in all of those ways but it's quite a logistical, historical and vision-oriented position document. After getting some of the active inferences out of the way in the preamble, we're gonna turn to the institute itself in terms of our vision, values, principles and how we see moving through the challenges that we're facing right now onto the specific history of the institute, our current organization model in terms of our governance and leadership structure as well as the institute units, the organizational units that support all the work we do. We then turn out to the community and ecosystem growth and development and go through some details of different roles that are stakeholders and participants in our space and some of the tools that we use today for managing information and message passing. We talk a little bit about our communications plan internal and external, both sides of the blanket and point towards some of the types of support, infrastructure and administration that we do and that we will offer for the active inference ecosystem. We frame that in terms of continual deployment, continual development and quality performance and growth evaluations across multiple scales, close with some discussion and then provide an appendix that overviews a variety of the open source services that the institute has provided over the past. So let's just jump in. I'm actually gonna jump into the open source services first and overview them before we jump back up to the top and look at the big picture because this is really to appreciate the work of so many contributors who stepped up in the uncanny valley, in the gray zone, in the ecosystem as it's forming, as it's always forming and really made it happen. And so many people we know are out there scribbling or taking notes or printing out papers by Friston at all or struggling and asking questions. Are there applications of this? How would this matter? Why would this matter? Is there a code version of this? How do you connect the dots between these equations? Those are what we ask every day and so when people can combine their regimes of attention into a synthetic intelligence outcome, then we go so far. So the first category of works that the institute has developed are educational. We've done video live streams since the end of 2020 till now, more than 370 live streams. All live streams are curated in terms of their transcript and we increasingly are developing automated and semi-automated methods for transcription, translation, sub-captioning, language modeling and so on. We've had textbook groups actually now as of now we have three completed cohorts of the active inference textbook group, the 2022 textbook. We have two courses that are being offered that are produced by external content developers and both of these have only started in the last three, four months. First is Chris Fields' Physics as Information Processing and second is the works spearheaded by Avel and co-organized with Kairos Research, active inference and the social sciences. Both of these are like awesome and great learning experiences. As we heard from earlier in the symposium we're working with Sanjeev Namjoshi on his incredible textbook although almost to call it a textbook is to either think more broadly about textbooks than one has or constrain what it really is that he's doing because it's such a unique real research development full spectrum effort. So I'm really looking forward to seeing how this plays out in 2024. Until then for those who want to get involved with the testing that we do and we've developed the active inference ontology which is a critical resource for the ecosystem, it enables us to center accessibility, rigor and applicability across languages and settings by knowing what it is that we're talking about when we talk about active inference. On the research side, we've also developed the ontology in a research direction in terms of research on and about and with active inference itself. As far as software goes, our primary development first off has been the documentation and connecting the dots around some of these other incredible packages that others are developing. For example, just in this symposium, we've heard from some of the core developers and open source contributors to PyMDP, the Python package for generative modeling as well as just a couple of minutes ago from Bert de Vries and Dmitry Begayev coming from the Julia language, Rx and Fur side. And both of those are some of the leading packages for applying active inference. So we do a lot of work to connect those dots and make it more accessible. Not everybody is going to jump in with all six feet to a programming language. And what we've done with active blockference with the work of Jakobs-Mechall and others is to wrap generative models from active inference inside of the complex systems modeling framework Cat-Cat. We've developed notation interoperability systems that actually let us pull back one layer from the implementational details themselves. For example, of Rx and Fur or of PyMDP so that we can bring it halfway to where the ontology and the visualization and the natural language exists and then use language models and other techniques to deploy it in an implementation fashion. We've developed a host of research publications. We support people in all of their different stages of their career, different seasons of life and some really creative work has been directly produced through an institute project or somebody who knows may have gotten good ideas or good feedback just by hanging out, being in the game. And we've performed literature meta-analyses to help grapple with the scope and the quantity and the quality of what is unfolding before our eyes or before our antenna. Again, even as we only so partially sample this ocean of information. We're active in standard setting and continue to develop our qualification standards for active inference that could be the basis of education, research and professional development. We provide a variety of mentorship and facilitation service type activities including volunteer program, internship program, distributed facilitation and applied active inference symposia. These activities just as all others that we're describing above are open source and they're also open to your participation. So if you can follow some of those suggestions by JF time boxing and prioritize and pay attention and be active, all those fun things that he said and you're looking to do it in a semi-structured way with a volunteer program or in a more structured way with the internship then I hardly encourage people to get involved back now. And in terms of governance activity we have a scientific advisory board as well as a board of directors and together with the officers those groups constitute the formal governance structure of the Institute but there's also so much more to it and we look forward to understanding what ecosystem governance can look like in the coming years. So welcome Dean. I just summarized the work itself. That was the pre-amble and now if you'd like to say hi you are Pablo before we jump into more sections of paper or look at more questions that people have asked how goes it or what was your experience of working with the paper or what brings you on to this strange stream at this opportune time? Basically why I'm here is because I for the first time even as an acting board member with the Institute I was able to get a better sense of the kind of depth that we aspire to and I'm not sure whether or not we are going to be I really don't know what directions we're going to be going in the next coming months, weeks, years, et cetera but I have a better sense now as a result of participating in this development exercise of what the potential is that I probably had at any time and I've been with the Active Inference Institute for a couple of years. So that's saying something I think we had to get to this point to be able to write this document and now it's acting as a bit of a springboard. Now, which way we go and how we land I'm not still certain about what that looks like but I think that's part of what the really interesting thing is. We seem to have moved away from the doc and now the real journey begins. So that's my take on it so far. Moving away from the doc, DOCK, pushing off, using it as a springboard, continuing the adventure, leaving the pre-print behind continuing to version it as a living document and also wayfinding outside of the doc. Pablo? Oh wait, I don't hear you yet Pablo. Sorry, oh, I was mute. Hello, it was a great experience to go through the document and review it and read it and thinking about it. Actually, it's a living document which is super exciting and I keep taking notes and hopefully act on it on the future and it's been a super experience to find so many people with the same interest some values that I really think are gonna be the basis of organizations on the future. So I think we are doing something very cool here. Cool, all right, well let's just walk through some of the sections in the 25 minutes that we have. Again, the paper is being shown on the stream but Dean and Pablo, you don't see it. So we kicked off with a preamble. We wanted to introduce active inference as briefly as we could and point towards the fundamental first principles and physics based nature of active inference and initially point towards its integrative capacity in principle and then in the second paragraph there's about 104 citations in this paragraph and that's where it's really only a sampling of the domains, 21 domains here. So it's a place where we can point to where is active inference being applied or how is active inference being applied or how is it different or better than another approach in a given setting which isn't to say that it does everything in all settings for all people at all times and all of that but this is to say that here's where one can look to find out more and even just building this and bringing it all together we see so many opportunities for epistemic resources in the ecosystem and having real time literature and analyses and all these kinds of services that we wanna provide. Starting on page five, we turn to the Institute itself and we write as of August 2023 the active inference Institute is a registered nonprofit organization tasked with identifying, establishing and managing the sustainable implementation of administrative and governance functions to give components of the active inference ecosystem coherent forms and reliable channels of communication. Two, publishing and licensing protocols that establish open and fair use and effective dissemination of community products within and beyond the ecosystem. Three, services at the scale of individual humans and the community at large so that stability is protected while risk and uncertainty are minimized within the ecosystem and four, organization and operation of cyber and cognitive security systems that ensure productivity, inclusivity, accessibility and safety and discourse in collaboration. Figure one gives a graphical representation just one of how we are now and for those who wanna dig more check out the 2022 active inference textbook where you'll also find a golden ring and we kinda took that graphical motif and ran with it, had a lot of fun and a lot of interpretations there. So just here looking at figure one Dean or Pablo just wanna add anything. Yeah, Dean. So one thing about the building of this document with a whole bunch of authors it's often not made explicit but there are two aspects to learning about yourself. One is to actually throw some stuff down on paper because you believe certain things need to be included and then the second part is the consolidation process and I think one of the most interesting consolidation processes was this graphic because I think what it does is it tries to condense as much as possible and yet still remain coherent around what the relational aspects that we are trying to make available are as opposed to this is what we've captured now look what our graphic designer came up with. So if you've had a chance to kinda look at this and really analyze it and think about the layers and think about okay, so what's the background here? What's the foreground? Does this representation allow for a swapping? Can we have things that appear to be background or fundamental basis stuff that could actually be things that we are addressing in the moment versus those things which we look at as particulars that in fact are generalities. So again, it's one of those things, I don't know, Daniel probably knows whether it's a where's wall those thing or some sort of artists that came up with some strange twist on things but the bottom line is spend a few minutes, have a look because you'll wanna go through your own building of what these relationships are and you'll also wanna be able to consolidate. I wasn't there when this representation was built but I do know that I wanna be able to both look at it do a comparative analysis to what I think the institute should be and then be able to consolidate. Awesome, yeah, just a few notes on this and to give a visual overview since it may be clearer to see this then go through the pros that essentially describes it as much later on. The golden ring represents the active inference ecosystem and I think we could talk about golden rings for a long time to come but coming in from the ecosystem on the bottom, bottom for the reader, not from the other side of the table are the projects and all the projects are participatory endeavors. They're happening behind on and in front of the golden ring depending on how you look at it and that is the appendix and all of those different projects that we talked about and all of the iterations and new threads that I know that we're gonna see in an hour and I know that we're gonna see in a month and in a year and in a century and those projects are being supported through two units kind of like a department but an organizational unit one centered on education and one centered on research. The education unit is called EDUACTIV and the research unit is called REINFERENCE. So right there you kind of have active inference case that wasn't clear. Coming from the institute scale supporting the unit scale are a variety of functions such as funding and partnerships, communications and hierarchical processes such as the construction of enabling systems and at the institute scale other than administrative functions we have the officers, the day to day operators of the work, the board of directors which is the formal governing body and the scientific advisory board where we really wanna call out broadly and have people who may have or may have not thought that they would find themselves in a scientific advisory board for example in 2024 this is your invitation and we really want to have a vibrant scientific advisory board who serves liaison type roles with different sectors and just gives periodic super valuable high leverage insight and then I guess depending on how you think about it as all things we have just a few terms they don't just represent bricks in the wall but it's more of like the fibers we weave and some of these are already things that we do every day courses, educational groups, events, live streams ontology, research, others of these are things that we know will be coming into place in the coming years and we know that the tapestry is just gonna continue to expand so it's really exciting. On page six we have another figure and this is just another way to represent some of the functions that we serve in the ecosystem this isn't the only final or a fixed representation of the functions that we have or the only way to slice and dice or any of that but just to connect some of the dots between the functions that we do and want to serve for the ecosystem. So starting from the broadest let's just say where the listener is where you are right now is in the space of awareness and that happens through communication. Some people who become aware through a search or a language model or a video suggestion or a friend or happenstance or synchronicity or however some fraction of the aware will pursue education and about education we have and will continue to produce the materials, documentation, ontology courses live streams and so on. Education is produced and rests upon a comments it's an epistemic comments it's a verified information environment and that is really the space the wildlife preserve as I know Dean likes to refer to it not jokingly at all. And underneath the commons is the support that does sit between a participatory commons and a formal governing kernel that support in the coming years may go many directions but we know it's gonna involve all of our favorite things like hardware, software, information, behavior, narrative and more. And then we have the regular institute scale governance partnership, sponsor and donor relationships such that we can first find a type of stability or sustainability and then more actively play a regenerative role in our ecosystem we wanna dispense micro grants to people and we want to explore paid internships and be able to meet people where they are in their active inference journey and things have been happening to that end and through many means since 2021. On page six, Pablo or Dean wanna add any thoughts? I'm super excited about the governance, the education on the game and tools. In an hour, we're gonna see the first active inference game. This is very fun. On seven, we describe our vision, our values and our principles and just to read the top level on the values and principles they are active inference and exploration, integrity and inclusivity, dynamic internal modeling, anticipatory behavior and continuous development. On page eight, we talk about some priorities and challenge areas. These are challenges across multiple scales. They're challenges that individual people face in their journey to learn and apply in so many different backgrounds. These are challenges of national interest. These are challenges of global emergency as well. These challenges are educational in nature, for example, related to scientific literacy and workforce development and just general professionalization and changing times. On the research side, there are some priorities and challenge areas related to the fundamentals of cognitive science and grounding the cognitive science in physics, getting that integrative first principles framework that we expect and prefer. About information science and diverse intelligences, things are faster and different than they have been and active inference is perfectly poised for sense making and decision making in that space. User experience, accessibility and socio-technical design. This is a major open question across different settings. So it's not uniquely held to the Institute. However, we really do believe and really will commit to exploring new kinds of ergonomics and onboardings and experience so that we can come through and be how we want to be from the first time somebody enters the game where we can highlight accessibility, meeting people in the language that they prefer to listen to and speak at the language, in computer language or at the math level that reflects somebody's preference in that moment as well as being able to use augmented and synthetic intelligence systems to ramp up and even push the frontier of what augmented people and other kinds of collective intelligence systems can do. Cyber and cognitive security. These also are broader topics. We've worked on these topics from an active inference perspective many have and we think that this is gonna become increasingly relevant as things do become more and more challenging in certain ways out there. Scaling the active inference ecosystem. We've heard about this from the narrow and technical side. We had estimates of the computational complexity of sophisticated inference about how many computer operations have to happen for a given calculation to play out and that's sometimes called scaling active inference. Like, okay, it took one minute on one gigabyte of data. So how many minutes is it gonna take on this computer with this much data? Scaling the active inference ecosystem though is a broader question and that's where we come to more of the social fabric and the real sense of belonging that we all want to have and co-create and provide for others in the ecosystem. And so this kind of scaling is really transdisciplinary. It's really skin in the game. It's as real as it gets. And applying active inference, some applications we've heard from today, some applications are cited in the early pages of this paper and more and more come through every day and we support at the Institute various applications of active inference. Dean, what can you add in there? Okay, my mute is off, good. So here's something that never came up in the writing of the document but I think it's kind of interesting and I'm curious what you guys have to think about this. 175 years ago, an institution came informed together is called Smithsonian Institute. Now, I don't know if 175 years from this conversation today, there'll be the equivalent of whatever the active inference institute is, 175 years forward. But I think it's okay to start thinking in those terms. It doesn't mean that we follow in their footsteps necessarily, but I think it means that we have a long-term view and that the things that we write about ourselves now have some shelf life, have some real sense of what we can't, there are things that we can and we cannot change. But if you look at the front page, the landing page of the Smithsonian Institute, the first thing that they say is welcome, which is what we say about the active inference institute. And then they say the Smithsonian Institution is the world's largest museum, education and research complex. We are a community of learning and an opener of doors. Join us on a voyage of discovery. So in three sentences, they kind of sum themselves up. What's the formula for not just existing for 175 years, but being a force? And I think, again, Daniel, whatever you decide to highlight in the document today, what I think we're really trying to do is say, do the words that we put out now, will they continue to resonate with the people for decades to come? And again, I don't wanna sound all hand wavy, but I wanna hold up an example of a place that really did set itself up to be something that people could return to again and again and again. And I don't think there's anything wrong with aspiring to having that kind of place, that kind of respect. And I think that the things that you're pointing to right now, hopefully set the, we're preparing properly. So that's what I'm taking away at this point. We know about that preparation measurement cycle from Chris Fields, don't we? We go on, talk a little. Oh yeah, please Pablo, go ahead. No, no, definitely. I completely agree. I've been thinking about that for a long time. You want to build and act for the best of the future. And I think to reflect that on the paper, it's super important. So yeah, definitely up to do that. Yeah. Way ahead, we describe some paths that we want to take and the history of the Institute. We describe our origins, but I actually wanted to come to figure three and four on page 11 to speak to a building. One of the most amazing and encouraging things has been the all ages, all time zone, active build, contribute mindset. An earlier version of myself, I might have thought like, hmm, people who do variational Bayesian methods and machine learning, they're going to have no issue on boarding to active inference. It's just going to be like as simple as saying, yeah, it's a variational Bayesian model with action. So it's just signal processing on the inbound that's just control theory on the outbound. But if you're familiar with variational Bayes and variational free energy, this isn't changing it that much. Oh yeah, we use a particular partition and it's related to the free energy principle, but don't worry too much, you'll get there. I thought that that would be the path for so, so many and future analyses will reveal what they do or don't reveal, but what we found in our enacted work as it really played out that there was inclusion and unique contributions to conversations ranging from philosophy of science to fundamentals of statistics and meta science and the role of quantitative studies and social systems, all these amazing topics and people who just scampered so rapidly through the prerequisite, what could be seen as prerequisite from a disciplinary perspective was facilitated by the recognition that one was already in the space and in the way that they needed to be to be an active inference. And that was not written out before and even now the traces are barely recorded, but it will become to be understood more fully. And to your point, Daniel, I think especially what the pie chart shows is the actual spectrum aspect of this. So I know that right now that maybe the neuroscience and the psychology piece take up a bigger portion, but I think one of the things that we're trying to promote is the idea that you can come from just about any background as opposed to, oh, all paths lead to, like as you said, an earlier version of you said, oh, okay, well, I can see where everything is gonna, where the confluence is going to. No, I think what this demonstrates is that it's actually becoming more distributed the longer we continue to exist. Yeah, this figure, of course, I'm super happy with everything that we were able to do, but this figure is like almost a little bit like a a paper target that we're just going to blow out later because it was a fraction of only people who had formal affiliations listed, which means only people for whom we had done full literature-based polling of their information. So it doesn't even represent the actual participation profile. This is more like the invitation list for live streams. So we are gonna see another long, a long tail that covers the part that's not the long tail, I believe. The rest of the paper continues along, possibilities, challenges, next steps, more about the institute organization, various other pieces of information that potentially eager stakeholders or participants or donors or liaisons or other people would like to learn more about. It's there to engage with. We want people to hop into our Discord and continue the conversation or email us or otherwise get involved. It's why we called this symposium in acting ecosystems of shared intelligence at the, of course, relevant suggestion of Carl Friston because we really do wanna enact it. So if you feel like you're on the sideline, waiting to jump in or you're on the beach and people are out there surfing the 100-foot wave, then maybe it's time to get more involved. So I will leave the closing word first to Dean, then Pablo, and then we'll go to Raph's presentation. Well, just real quick. Again, if you want a consolidated document that speaks to the now, but it also points to a direction that we, none of us really know is certain, but it's also really exciting. Have a look at the doc. Pablo? Yeah, I'm here to make sure that I act as one of those people of all backgrounds and perspectives that has taken advantage through one year of learning and acting on this institute. And hopefully you see you in one hour. So thank you and have fun. Thanks, awesome. Thank you, thank you, Dean. Thank you, Pablo. Welcome, Raph. Also, come on through on the paper. So, how are you doing? I'm good. What about you? Pretty good. Other than this brief interval of white light, I'm just hanging out. Fantastic. So I'm trying to figure out here how I can, I'm using a different slide software so I need to make sure that I can pull this up in the right way. So give me just a quick second. Yeah. Great. Cool. Just to read off what's coming while you figure that out. So right now we have Raph Kaufman and this presentation is gonna be GAIA, an active powered network for planetary scale sense making. Following this 30 minute presentation we'll have Evelle Guinean-Carlou on embedded normativity. That looks perfect. That's perfect. Awesome. All right. Thanks Raph to you. Thank you. All right. I was gonna try to see if I could blow this up to full screen, but. I got it. I got it on the stream. Oh, looks good on the stream. Go for it. Cool. So I'm just pulling up my notes here. All right. So let me just test that this moves. Did I just move the slide? Great. Cool. All right. So as they said, hi everyone. It's a huge pleasure to be here and thank you so much to Daniel and everybody else at the Institute for organizing this. I'm going to talk about what I think is the single most important application of active entrance, not as a detriment to the others, but it's sort of upstream of everything else. And that is how do the 8 billion of us humans make sense of what's going on in our planet and all scales? And how do we use that understanding to survive as a species? So this presentation here is a bit of a technical presentation and it has a lot of content. It's going to move pretty fast and assume that you're familiar with some concepts, but I'm going to try to leave time for Q and A at the end. So first of all, I'm going to talk about the motivation behind what we're doing in particular the concepts of the medic crisis and the third attractor. Then I'll give an overview of our network architecture and the application so far and cover the many remaining challenges. And last but not least, I plan to convince you that this is the single most important thing you could ever be working on and I'm doing this because we need your help. So that's a taster of what you have. And so as I said, let's jump right into it with the motivation and what greater motivation could there be? If you haven't heard of the medic crisis, it's a term that it's originally from Dr. Who, but it was borrowed by a thinker called Daniel Schmackenberger to talk about the total risk for humanity and the biosphere that is posed by the combination of three factors. The first being the increasing interconnectedness of everything from climate to food security to national security to biodiversity and everything in between. The second being the risks associated with self-evolving technology including runaway artificial intelligence. And the third being the increasing omnipresence of coordination failures personified as the God of coordination failures, MOLOC. And so the combination of these three factors as hypothesized by Schmackenberger will lead to either one of two attractors, one of them being chaotic breakdown of all structure and the second being a reversal to oppressive authoritarian control. And obviously both of those qualify as bad in my book, I assume in yours as well. So the question of course is, can we actually find a third attractor that is positive sum as opposed to negative sum? And rather the question is, can we design it? So what I claim and what we as the God of consortium claim is that yes, we can and we can design this attractor and it's the design goal is basically to build resilience and the stabilization into our bio-social economic system. And to do this subject to multi-level system constraints like the planetary boundaries, the desirability, meaning it has to be something that people actually want so they will help it happen. So it needs to preserve standards of living and so on and it needs to be feasible, it needs to be achievable from our current initial conditions. And of course there's a lot of design complicators in the mix, some of which partial observation, there's lots of information asymmetry going on. There's a lot of uncertainty about how the pieces of the world system work from a scientific perspective. There's obvious computability constraints associated to this. And last but not least, you can't really control what happens in most cases, we can only create incentives, recommendations and nudges. So easy, right? Of course it's not easy, but we think we can do it if we apply the sum of the principles from cybernetics that we already know and apply them in a decentralized positive sum composable way. So what we're talking about here is a decentralized network a decentralized hybrid AI human network for planetary scale decision support and automation. It features, it recognizes that we will have human and artificial agents interacting by the billions in an open network. Its goal is to facilitate both confidence or model alignment and coherence or goal alignment at multiple level of scales up to the global scale, meaning the survival of the system as a whole. It will need to feature built in incentives and governance and it will need to be privacy preserving intrinsically. So the below listed participants are participants, founders of the consortium I'm gonna talk about it later. So this is the framing, this is the motivation for what we're doing and now I'm gonna give a whirlwind tour of the architecture so far and I'm gonna cover mostly what we've built the version of this architecture that we've built so far thinking about this explicitly and then I'm gonna talk about how this connects to other architectures that already exist in the world just very briefly. So we've been framing this as an architecture for building decentralized digital twins which are local models of real world systems that run in the network. And the role that such digital twins play is that they help us understand the costs and benefits of strategies and projects that happen in the real world. And we, in our organization, Digital Gaia we started out caring primarily about agro-ecology. So how do you actually make better decisions meaning better recommendations, negotiations, valuations, investments into a farm, into a forest, into a bio landscape to make it better, to make it more likely to survive and thrive. So as a practical implementation it needs to tackle several real world challenges. First of all, it needs to handle distributed non-IID data, data is sparse, it's heterogeneous, it's private, it's potentially unreliable. It needs to enable localized, collaborative forecasting and planning, personalized planning and privacy preserving. It needs to accept the fact that it won't be able to directly control but only recommend it as we discussed. And the outline of what we came up with is this concept of the Gaia Network which is a mesh of AI agents called natural entities or ENDS for TORC that act as proxies for real world systems. Each ENDS runs on an engine called Penguin which it uses to inquire, learn, plan and allocate resources. And ENDS communicate in a language called the Gaia Protocol. They use this language to independently interact with their environments and each other. And if we do this right, I'll show a bit more on this later but if we do this right, the Gaia Network's behavior approximates a single composite agent that handles, that interacts with and couples with its global environment. Cool, so just to give you an overview of how this works behind under the hood, the Penguin engine is really the heart of our architecture. And the core loop that is mentioned here is a typical active inference loop but to make this work in very heterogeneous and distributed context, there's lots of work involved. So just some highlights that I think are gonna be particularly interesting for this crowd. First of all, so as I said, the same core engine needs to be able to work for very heterogeneous context. So each agent has access to a library of generative models. And these models are specified using a Python based probabilistic programming DSL and they declare what context they are appropriate for using a shared ontology. Using that ontology, Fangorn can then select the most appropriate model to use online given the configurations like project configurations and the kinds of observational data available. So using the model that's appropriate for a form or for a forest, what kind of form or kind of forest, what kind of ecosystem, what kind of, you know, their context energy project. I'm not gonna get into the examples right now because there's more on the applications later on. So we end up with a network of heterogeneous agents, but that's okay. They actually form a single model in a formal sense because they're all connected hierarchically by a system of hyper priors. And this means that Fangorn needs to support inference of generic hierarchical generative models. And we actually do this automatically using approximate posteriors with short throw constraints. And so we have this global model that connects all the nodes and describes general facts that apply over multiple contexts. And this means the whole network needs to jointly perform inference over this global model using a peer-to-peer or federated algorithm. So initially, this way this works initially, the network sets the global posterior to be the same as the global prior. Each node then independently performs local optimization of the variation of free energy and then passes the parameters of the new global solution to the peers in the network where they use it as a hyper prior. And this works over a broad range of topologies and other pretty general constraints. And by the way, I'd like to mention that there's a lot of overlap and interaction between the stuff that I'm talking about here and some of the other presentations that you're seeing today. And I wasn't able to reference them explicitly in this talk, but I'm sure you guys are gonna see the connections. Cool, so what it turns out is that the shared parameters that represent our global model are effectively stored by the network, stored in a federated parameter store, meaning that each node stores their own local parameters, their private stuff about their own local context and multiple nodes stored copies of the global parameters. And this architecture, because it supports heterogeneous local models, it actually allows us to introduce another kind of nodes which we call antmoot nodes. These are nodes that aggregate scientific and empirical information using a meta-analysis model. And we actually are able to use that meta-analysis model to constrain our project-level models, or ants, using quote-unquote imported knowledge that wasn't directly observed by them, which is left to right in this picture. But this actually also goes two ways. So empirical findings from the projects, what actually works on the ground for this particular project also informs the antmoots posterior, which is right to left. And so we actually have a complete cycle of science and empirics all implemented using hierarchical active inference. So finally, just among the global parameters that network learns by federated inference are the precision associated with data coming from different data providers. And this gives each node a way to consume data without trusting ground truth and basically to learn together which sources they can rely on for what. This is just some of the more core features we're working on integrating many other capabilities into this architecture. Not gonna go through all of this, but basically what we're finding is that we have all the building blocks to turn this into a full-featured decision network complete with its own internal knowledge economy, meaning that contributors' contributions get attributed and rewarded in proportion to the free energy reduction that they afford. Cool, so jumping into applications, as I mentioned, our main focus as an organization has been on modeling agroecological projects like regenerative agriculture and agroforestry. And this exercise is all the constraints that we've talked about so far. So again, I'm not gonna go into full detail, but just to give an overview of what is the anatomy of generic model for a farm or forest. So we started at the bottom right with observables that are things that we can measure about that ecosystem, about that local context, like the plant count and size, vegetation indices from satellite data, and so on. These are linked to latent states which form a nonlinear dynamical system that's parameterized by the policy, the things that modulate the latent states, like agricultural interventions and practices. You can think of the timing of planting and harvesting, what kinds of crops or trees get planted, what products like pesticides get applied or not anymore. As well as covariates, so external events like weather, physical risks, wildfire, drought, and so on. And so there's some scientific and measurement parameters that go into this as well, things like the effect sizes of the various actions and covariates. And these local parameters are conditionally dependent on a set of global parameters, which as we discussed, they link information between different ends and different sources of knowledge. And so these same global parameters are used in the meta-analysis model and they actually establish the link between what is found to happen on the ground and what is reported in the literature. And so I'm not gonna go into this because I think I'm gonna run a bit long, but basically the concept is that we have these project nodes that correspond to many farms and they're each on independent nodes in the network. They're also connected to this common meta-analysis model and together all of these nodes, they incorporate academic studies, expert knowledge, and locally relevant context. They do so privately, so only processing the local data locally and through this parameter sharing, they actually jointly and iteratively estimate the global posterior in a cycle forever for as long as there is new data about the project, as long as there's new studies and new expert knowledge. So all of this is happening basically in an online way. And again, the topology can vary but the principles remain the same. So just to show a little bit about how this works, this is just two epochs of the global posteriors. And this was just a location and scale for two parameters, the slope and the intercept of the relationship between vegetation index and the tree biomass in one given hectare of forest. And as you can see this converges pretty quickly. It's comparable to local only methods. And just to highlight how this can be used to, also the limits of how this can be used to estimate source reliability. Here's a really successful example at the left. This is pretty straightforward to use this to estimate reliability of satellite data. And for other cases like very local, very context specific data, you actually need more data or more data from more independent sources that we had in this demo here. Another very useful use case of this is as we discussed, we don't have direct control over what happens at a farm. We can only recommend and nudge. Often we also like we get, there's a node gets stood up, a net gets stood up for an existing project that has already happened. And we wanna find out what actions we're taking in the past. And the same generative model can be used to estimate what happened in the past and the exact example it infers when planting started and when harvest started for different fields. So bringing it all together, one of the applications that we've been using, we've been working on, it actually uses all this machinery together to validate claims about projects performance and to estimate future performance. We've deployed this in real world projects. This one, the examples here in the scene this photo are by relevant to agroforest reforms in Colombia. And we found that this actually helps the project developers and their funders to come to a shared understanding of how successful the project is, how much it's worth in terms of its impact and its fundability, whether it should be scaled up or needs to change course, what is the quality of the strategy versus other potential strategies for natural regeneration and so on. So it's actually useful on the ground for decision making in this context. Cool. So finally, I wanna talk about the end state and going back to our original motivation of solving everything, I forgive you for thinking this is nice, but hey, we've been kind of in this downward spiral towards the meta crisis for a long time and what you're telling me doesn't even come close to solving everything. And you're right, there's a lot of moving pieces that need to be built and connected, but here's the good news, we can do this. There's no mystery here and we don't need perfection. I love this quote from Edward Folbrock that says that to define a policy, we don't need exact empirical measures or optimality. If one jumps from an airplane, it may be nice to have an altimeter, but what we really need is a parachute. So there is only so much that we can get in terms of dimensioning returns, what we should be refocusing on is creating and refining solutions everywhere and connecting them everywhere. And you might be thinking about how does this actually scale to global goals and how does this actually get used for scalably building into the fabric of decisions in our society. So about scaling to global goals, I again won't be able to get into the details, but maybe Connor mentioned this in this presentation earlier. We have actually some proof of concepts that we can actually build active agents that are made of active agents. So we've shown in this paper here that a collective of interacting active agents is able to perform approximate inference at the ensemble level. And that this further that this happens through specific mechanisms that are mediated by cognitive capabilities like theory of mind and goal alignment. They work together to improve collective performance by eliminating ambiguity and through actively exploiting diversity. So that's one. And the other question is how do we actually build this kind of stuff into not just these ad hoc decisions, but into really the fabric of decision-making and that is our global economic system. So talking about here work that I did with Casper Hesk, this is a game that highlights the ability of, and it's just like the ones that we just described to infer through the system state and use it to credibly incentivize long-term positive outcomes in an externality-happy scenario. This scheme, we were able to show that it works even with adversarial bot strategies and even with high uncertainty and the intentional misinformation. So you can actually get these ends to work with budgets and incentives to drive decision-making that internalizes externalities and compensates for occlusion and generates long-term positive outcomes. That's not all. There are other solutions that have been around for longer than hours that apply similar principles to global decision-making problems in the real world at scale. This is one case study from our partners at Cognizance that they developed for the COVID pandemic. And this is not using active and formalism. This is a combination of traditional recurrent neural networks for prediction and evolutionary optimization for prescription. And it's actually able to achieve some pretty amazing stuff. It discovers a Pareto front between prevention of cases and economic costs. It's also a showcase for hybrid human AI intelligence. They did an initial version of this model. Then they launched an X-Prize. They got a bunch of contributions for other models. And then they did a meta model that discovered a better Pareto front. And even better, this tool has actually been used to advise policy in the real world. So it's actually very exciting that we're working with them and connecting the dots because we need to connect the dots. We need to bring it all together. This illustration here is just, I think a subset of the effects of the climate change crisis, which is just one aspect of the meta crisis on humans, on our society and the economy. And so we actually need to be able to deploy solutions that work across many different domains and to connect these dots. This means two things. The first is interoperability, aligning decisions across context so that we don't get unintended consequences. And second is reusability. So being able to transport and transfer learning and structure across context. So for instance, using that evolutionary approach from the COVID example in our agricultural context or in many other possible contexts. And this takes a form of libraries of components, APIs and so forth. So that's what we're working on. So we're launching this consortium. We just put up the website a few minutes ago and we have the goal of having a minimal but functionally complete implementation by the end of next year. And this is an open project. We are working on this at our organization called Digital Gaia, but we're also launching this open consortium that welcomes contributors from anywhere, both active entrance experts and non-experts. And we need a lot of contributors. We need all of you because this is the most important thing that we could ever be working on together. And we need your help to tackle the remaining challenges. So if you wanna help us build the planetary brain, please go to gaiaconsortion.org and or reach out to me, you know where to find me. And together we can design, build and learn what it takes to achieve planetary regeneration. Thank you. Well, Raph, thanks for the great presentation. I'll post the link to Gaia Consortium into the YouTube live chat, but let me get a few quick questions in. So here we go. First question. This is from Marco Lin. I agree with the critical importance of such a project. How will you deal with the challenges of scaling given the dependence on real-world entities like humans? For dependents on real-world entities like humans for major metacrisis domains such as geopolitics, sociocultural fragmentation and other particularly hard sense-making domains. First of all, hi, Marco. Yeah, thanks for the question. And I 100% agree that I'm not claiming that this is sufficient. I'm just claiming that it's necessary. We do have a lot of energy that, a lot of ideas and a lot of energy that we wanna put into exactly what you described, how to weave what we're doing into the existing incentive landscapes, whether they're market incentives or policy incentives, but ultimately this needs to be an enabler, this needs to meet people where they are and that we need to be cognizant that change is not gonna happen overnight, at least not right now, right? Typically these things take a while to mature and then they snap into place. And we're already finding some of this happening in some of the domains that we're working on where nothing happened for a very long time and then bam, hockey stick. And we believe that the same can happen in pretty much anywhere. So we're very optimistic. Great. What are currently the greatest challenges for Gaia as an organization? I love that kind of ambiguous question, like is that the Gaia consortium as an organization or is it Gaia, but let's hear both. What are the greatest challenges or frictions for the consortium and for the bigger picture too? Also, I'm not gonna talk about the technical stuff for a change, I'm gonna talk about the more meta stuff. It's actually very related to what your previous presentation, Daniel, about the paper about the institute talked about. And it's really related to building out this common, what we were calling the common cognitive kernel for sense-making and decision-making. And relatedly, breaking through this culture of sort of suspicion and third force and getting people trying to stake their claim to something. So we are and always have been extremely open, but it's kind of hard anyway to build trust. So we are aware that it's going to take a while to build this trust with everybody. And I think the only answer to that is we need to be as participatory and transparent as possible. Also being humble and knowing where we haven't, we don't have the answer and asking people to help us. Awesome. Any closing words or thoughts? Well, if there are no further questions, again, I thank you all for paying attention and I'm going to be reaching out a lot more. You're going to hear a lot more about this consortium. As I said, it was just form. Did you just put up a website in time for this presentation? And by the way, thanks so much to Mahi and of course, Steve and everybody else who's been part of forming and getting this ball to start rolling. We are very early stage, very formative and still very client. So we definitely are, I just gave you my perspective on a bunch of these things, the shape of it that has come from all work on this so far and even not all of it, just a piece that we could cover in 20 minutes or so. But we are very plural, we want to engage with you on different paths forward, different opportunities to build things, to experiments, to connect the dots, to learn together. And ultimately, we really just want to provide a platform a convening space, a protocol in the classic sense of the word of a shared language, a shared means of communication, a shared means of understanding to find together what we need to do to actually build this third attractor. Awesome. Call Gaia Trimtab. Thank you. Exactly. Thanks so much, Daniel and see you guys soon. Bye. Bye. All right, what a cool presentation. Next up, welcome Avail. We have Avail, Gwenen, Kalu with embedded normativity. So thanks, Avail, to you. Hello, can you hear me? Yep. Good. And can you see the slides? I only see the top left of the slide. You only see the top left? Yeah, just, yeah, try again, thank you. Okay, what now? Yep, looks great, thank you. Okay, good. So I am Avail, Gwenen, Kalu. I'm interested in a multi-scale collective organization. This is why I came to Active Inference to begin with. And the core question I've been asking is, basically when we look at human organization, there is creative evolution, there is a challenge through time of the specific patterns that are enacted. And this is very hard to account for. And I have been trying to pin down more and more specifically what exactly is Gwenen and how to model it. So this specific work is a work I wrote in the context of what my PhD, of course, which is funded by the XCAPE project. A project in cognitive archeology that's interesting, interested, sorry, in how the archeological landscape, shape, cognition, and vice versa. The title of it is Anomated Innovativity, an Active Inference of Agency, an Active Inference Account of Agency BM Flesh, because you will see very quickly why. So first, we will go through the existing work on encultured cognition in the context of Active Inference. Then we'll try to expose the concept of embedding activity, which we introduced here, and what it means. And then we'll look at the more semiotic question around the externalization and utilization of normativity in cultural landscapes. So first, encultured cognition in Active Inference. So we know Active Inference, I will go quickly over it. I still need to do it because there are onlookers that are not familiar. So Active Inference say that agents continuously predict the specific cues, sensory cues they will see, and they predict the action or policy they will not take. And they basically enact an ongoing stream of expectation that they reconstruct given a calling signal. So this is best I could inform in this schema which shows the agent as the medical system which engage through environments, through action and cessations. And all of these specific domains, they minimize specific function that is called free energy and whose minimization basically means there is a proactive process going on. And a core property of Active Inference is that it dissolves agency as we typically understand it. We don't have agents that have like internal monologue that drives them to take decision. We have a conscious anticipation of what's going on and whatever an agent anticipate itself to do, it will do. And so this goes for computationally satisfying but very deflationary notion of what is going on with agency. So the question then comes, what is the relation between an agent and the cognitive landscape in which it evolves? I will present very rapidly again the framework of skill and agility pioneered by Brunberg, Kiewoszstein and Riedwilts and which says that skill agents which we understand as agents that understand what is going on around them, they will basically experience the world as the landscape of afferances which is opportunities for actions. And they will, their cognition pretty much reduces to the experience of this landscape and adaptive engagement with it which basically maximize the grip of the agent as on the system. And essentially for reasons we will not go over because it's math, the very existence, experience of this landscape, it is equivalent to the existence of constraints of our attention, the attention that the agent gives to thing which translates the embodied expectation that are two time embodied, but the money expectation that the agent enacts. And so this becomes very interesting when we cross it with the existence of enculturation in humans because the basic experience we have of afferences of what we can do, it is shapes, culturalities shaped by interaction wherein we learn with other humans about what we do, what we can do and what we can do. And this means that the basic human experience, it's thoroughly encultured, it's embedded in what one should call the cultural landscape, landscape of cultural affondances and it is embedded in cultural and material regimes of attention. So why is to pay attention to things? And this is a figure that comes, I believe from this year on the thinking throughout the minds where it shows basically all the action and station cycle when we have several agents, they build a common environments that has epistemic cues on how the world is like and when we are encultured to pay attention to things in a certain manner that makes it so we perceive and enact this landscape. And so the encultured constraint of all the flow of attention because of what we saw just before the deflation of our account of agency, it means that the very expanse of the landscape, cultural landscape that is collectively enacted, it effectively integrates intentionality and it effectively integrates normativity in a way that clashes strongly with classical account of both. So given this state of the art, let us look at what I mean specifically by embedded normativity. So embedded normativity is a kind of normativity, I think you have got it. So normativity, it means basically the criterions by which we judge things to be good and bad. In the context of biological system, they will look for specific, I don't know, chemical concentration and they will avoid other. This is the phenomenon of chemotaxis. This is pretty basic. And this is a form of normativity. But for humans, we will typically think of laws, cultural norms, culturally shared way to judge behavior. And again, at least in the context of flow and choice theory, we typically think there is an agent, he has access to information and we'll think, what should I do? And he will judge rationally about what is good and what is bad based on a collection of values which are internal to the agent. This is not how it works. What I call embodied normativity is by contrast, regimes of normativity, where it is the norms and value that guide the agent's behavior. It is embedded in the material and cultural environment. So basically the agent projects a specific set of norms and value in the very experience of this environment. And it will perceive the locus of normativity as the environment and not as a form of internal judgment. Let us look at what can exactly how a normativity work. It will be more concrete. So first we can materially constrain people to do things. So for example, there are roads, you can go off roads, but it's harder and you can't really do that if you're a chariot. So when you have a specific pattern of roads, which is constructed by cities, you are pretty much constrained to follow the roads materially. And for this we will talk of structural embedding. And there is another regimes of normativity where we build on environments that will basically, to you suggest you into adopting specific norms and specific values. For example, escape thinks that vertical patterns in archeological culture, they are somehow embedded within what we'd call vertical norms. So hierarchy and certification in social relationship. And for this specific form of embedding, we can talk of semantic embedding. And what is must be very clear is that we don't have structural embedding on one side and semantic embedding, there are basically two mechanism, complementary mechanism in which normative, sorry, embedding normativity occurs. If we look at the city, so this is impression of a Sumerian culture city-states, we have a matter organization that prompts specific kind of behavior. So for example, we have a bridge and this bridge allow you constrain you to go through the river in a specific way if you are on foot. But this material constraints, it is perceivable. So by seeing the bridge, you will also see that you can navigate the world in this specific way. So you don't have a separation between semantic and structural embedding, we just have two ways in which constraints of our behavior that are embedded in the environment work. And something that is a core to the account of embedding normativity is again, that it is not internal, but we need to go into more detail about what we mean by it not being internal because of course, we're talking about agency, agents constrain, what agents do, it's not outside. What it is is embedding normativity is a regime of normativity where the norms and value, they stem from the experience of fordance. For example, if I see a fire, I will not think about, hey, could I put my hand in it and get my hand burned? I think it's really painful, I won't do that. If I even think about putting my hand in the fire, I will immediately feel bad. The normativity it's embedded in my experience of the world. I feel like the fire or more specifically, the possibility of putting my hand in the fire is intrinsically bad. I do not feel there is a possibility that is neutral and then I judge it should be bad. But of course, this judgment does not occur outside even if I feel like it's about my fordances. It emerges from the engagement I have with the material and cultural niche in which I find myself. And if we want to untangle what this means precisely, we will have to go into more detail into the what I call the externalization and internalization of many normativity. So let us look at something very simple, how normativity work for embodied agents. So embodied agents, they do things with hopefully in a niche that is material. And this niche will be marked by the activity of my agents and this marks they can then act as constraints either through structural cementing and bedding in the agents' environment because they basically signify that in the past I wanted to do that, maybe doing that is good. For example, we have centuries with, I don't know if century is the English word, but we have path. If animals that have legs, this is actually important, work on a specific path, they will compact the earth and it will become proper to grow anything, even grass. And you will have a path that will emerge that are pretty visible. And I'll tell you, hey, not only this earth is compact and allow you to walk on it like without a fort, but also this has been widely judged to be a good outlier for navigation in this specific environment. So the very experience of a path basically invites you, it affords, you perceive the affordance of walking through the path. And this is something that again can result from your own activity that can be returned into the environment by your own activity. And so this begs the question of how, okay, there exists embedded normativity, but how do we embed normativity? What is the activity of embedding normativity like? And the word I will use for that is externalization. We perceive the world because we externalize specific causes of sensation into an imagined objective external world. So when we see the path, we see a path. We do not see compact earth that we then compute to mean that other things have walked on it. We just see, hey, there is a path. And this is because we have externalized a specific cognitive kind, which is the path. And this is an argument that several people made. Here, Lena Steiner, which she found out well, worked on the production of devices that help blind people see things through, I believe, an array of repository device on their shoulder. So the core argument they make is that if you could give people that, they will start seeing things. They will start perceiving things as being before them rather than having vibration on the shoulder and then thinking about, oh, what did it mean? And so the basic argument is that perception is just picture of what is the structure of an environment, given the existence of a specific coupling device that relates action to sensation and vice versa. And this critically relies on the internalization of specific, we could say a lot of things that I will say priors, specific priors in perception. So missing the path as a path relies on me having the prior knowledge that there is such a thing as path and I should be expecting to see them around. And so you have duality of externalization, in the production of embedded normativity. I can only see norms out there if I've internalized, integrated a specific apparatus of perception that entails that there are indeed those specific norms that are embedded in those specific afferences. And so this produce a more general question of semi-optics, basically of information theory. How do reading and writing and learning activity works? So when I externalize, I treat the world as having specific natural categories that I pretty much imagine. And I can use them because I mentioned them as an excellent memory that allow me to read and writes normative cues. This can be, you know, letters, pure signs, but as far as I'm concerned here, all signs are at least implicitly normative there about how you should interpret the sign. And this means that we can experience specific landscape of afferences and we can write things out there. We can write normative cues by a behavior that will then constrain further behavior, but nothing tells us that we can communicate through this. The communication, it works because of the internalization of prayers. If a given community of agents have internalized a specific set of prayers, that is a set of prayers that enable externalization of normative cues, then only it can use the environment as a way to write and read normative cues as an embedded normativity thing. And this communication, it can be read in terms of information theory. Of course, it can be read in terms of semantics, but what is critical to analyze that the process of internalization, it is there is no trust system, there is no private key that allow you to be sure that the way you internalize norms is very specifically the way in which the MLM activity was encouraged to begin with. There is a loss or there is ambiguity or there is, I'd rather say it like this, an active demarch of reconstruction of given prayers by the agents. And so this gap between the intent of other communicants and my intent by understanding of what I meant by their active engagement with the world, it builds a tiny gap in information that is somehow enough to drive the opening evolution of cultural landscape. And this is something that I do not think is embedded in active inference. And I think is of major interest to research, to the research, at least in the evolution of, open-ended evolution of cultural landscape. So to sum up things, I just presented the paper that is called Anomalydormativity, and that speaks unsurprisingly of embedded normativity. First major point is that if active inference is correct, then our notion of agency is dissolved essentially in the notion that we can perceive and enact cognitive landscape, umvalide if you want. And given that this is the case, we will experience the norms and value that guide our behavior as a property of the cognitive landscape we experience, the affordance that we can perceive. And this specific regime of normativity, I call embedded normativity. And the third thing is that you have a critical thing going on, you have a semi-autically constraining thing going on in the externalization and internalization of regimes of embedded normativity. Because those two things, they respectively enable nomad, embedded normativity to exist, to have a causal power. And they enable also the open-ended evolution of the norms that are embedded in a given landscape, in a given material landscape. And therefore, the open-ended evolution of the associated cultural or cognitive landscape. So thank you. I will enroll my references slides in case someone has the, really needs to know what the reference is and is ready to move around the video for this. And do you have questions? Awesome, thank you, Avail. While people are typing any questions, could you maybe just give a little introduction to the Active Inference for the Social Sciences course? Like, what brought you to want to facilitate this effort and where are we at and where do you want it to go? What do you want it to be for people? Oh, unmute and then continue. Good point, thank you. The Active Inference Institute organizes and carries research, organize a course on the Active Inference in the social science that is based, essentially in explaining the basic demarch, the basic hypothesis and the consequences of the cultural landscape formalism that I hinted at slightly in this presentation. The goal of the course is to give a basic ontology that is an interface between complexity science, let us say physics written large, cognitive science and social science to afford interdisciplinary integration in the social science and in the study of the evolution of human societies. And I think it is cool, but I'm biased because I've organized it some, maybe you should ask Daniel, which is absolutely biased, whether the course is cool. I've been enjoying it a lot. We just finished last week with my section on collective behavior. First we had the introduction lecture from you, then Ben White on the basics of being an active agent and we have Lorena and Mao's sections upcoming, as well as your more dedicated session. Yeah, it's been a really fun experience and it's gonna provide a really useful missing piece in the ecosystem. I guess just one question I have. How do we take some of these topics you're bringing in about the embedded normativities and think about them digitally? Like are we talking about conceptual embeddedness in information environments and or are we talking about the ergonomics of sitting at our desk or how do we bring that kind of pleasure path, stigmergy on the ground into the digital? Are we talking about something that's embodied and embedded in the information space or are we still talking about bodies? So if we talk about humans, we talk about bodies with the brains that do things in a material environment. But your question is, I'm not sure I understood this. How can we bring this specific concept in the study of digital phenomena like Twitter, internet? Is that it? Yeah. So you have basically the whole study on the affordances you have. So the recipient of design has incorporated the WILO goal, the national influence in its structure. And most of the study in the digital design center on what platforms afford or do not afford. And you have a richer stance when you take into account not what you can do with a specific interface, but what the specific interface invites you to do given a set of cultural return large. It can be self-learning. But given a set of power that you share with the platform in a way that is formal but indirect. And so yes, I would look at it in terms of user interface design. I think this concept specifically is useful for our interface design. Then you have implicitly the question of the flow that platforms such as, I don't know, Twitter, Facebook. How specific structural environments bring specific collective dynamics that might be a bit richer, not richer specifically, but that might go in a direction that the user interface question do not go into. And that can still be modeled in terms of constraints and therefore in terms of embedded normativity. Is that on point? We'll go to a question. Yeah, that's interesting about what the interface does and then what it invites. It might invite things in order for that invitation to be fulfilled that has to actually provide that possibility. But it might invite things that aren't relationally possible. Like there might be a wall which invites climbing to a given observer yet they cannot climb it. And then on the other side is a capacity that isn't invited and that is the adjacent possible of tool use where something that was being used for this purpose, it's like all of a sudden it just fell out of someone's hand. They grasped it a different way. It invited a different engagement and then that can become re-entrenched and that points to that sequential making and breaking of new spaces that your work has featured over the last few years. Yeah, so exactly this work has been developed with while thinking about tool use, while thinking about the cognitive ecology, while thinking about the experience of cities, the way urbanism from specific behavior. But if I can get back on the internet thing, I'd like to make it more specific. So the interface design as far as I know is focused on interface. That is not surprising. But in doing so it is focused on specific affordance what you can do. You have another layer in ecological psychology which is the layer of solicitation. So for example, Twitter does something very specific which is when you log in, it shows you the number of notification you have. Then it removes it and you have to wait to know which were those notification and you have to wait now, one second, not one minute, but a random amount of time. So this is what they call dark patterns. They are patterns in design that are explicitly designed to create addiction. So this is a layer you can't afford in affordance because from an affordance perspective, there is just moment you can access the information and then the moment you can't and then the moment you can again. There is nothing going on. If you don't look at the motivation and the way a specific structure environment is perceived, the semantics maybe, the meaning that given an escape as for an agent. And even if you look at this, you are still looking at the platform in a way that is embedded from the specific patterns, that specific patterns of things that people do on the platform. If you have Twitter without the aggressive curatees and the way to aggressively signal that you are in whatever it is is going on, you don't have Twitter much, you have Mastodon maybe. That has very similar affordances and very, very, very dissimilar solicitation. So this is a way to take a multi-scale account of things and when you take a multi-scale account of things, you have to look at the semantics basically behavior, how specific coordination occur at specific scales around specific communication, specific message passing. And once you look at how specific message become meaningful to specific agent, you have a problem basically because this is subjective, this is constructed through the activity of the agent and this is whatever is the range of time you are looking at, you will have the construction of new possibilities of adjacent possibles as you put it. And yeah, this is the big question that I think we are missing at the moment in active inference. Awesome. Well, thank you for this presentation. Thanks for the social sciences work and we're all looking forward to seeing how the projects play out. Thank you, goodbye. Farewell, Abo. All right. The next presentation is with Pablo Fernandez Maquietto. So welcome back, Pablo. Hello. I'm here. Okay. I did. Take us on an adventure of curiosity. Yes. I'm gonna start sharing my screen. Let me know if everything is all right. Should be perfect. Yep. Cool. So this is one of my favorite memes and probably most of you know it. This reference is the cave of Platon and to me it was kind of a fine-tutive truth that now with active inference, seems like we are closer to demonstrate that that's how it is. So the first time that I start act and serve at the Institute, I just... This was kind of my feeling or my face. It was looking into something that it was complex, scary and look interesting. So I went and why I decided to go on a adventure and go through it in a world where people is driving by strong incentives that motivate them to reduce uncertainty in the systems they are part of. My approach is through games, which are gonna go later, but I really on this talk want to encourage all kinds of persons or people that are curious and I want them to act and get involved in the community because you may find that for now most of the actors are scientists and very well unrepentant people and you might think that you are not a part of it, but very far away from that, the Institute and corrects the opposite. So I'm super glad to have been through this journey and I wanted to share a little bit of my background and then we go to the game. Here's the second and last meme. The world you were raised to survive is not longer exist. We are on a digital revolution and with this digital revolution, it comes a bunch of different things that are evolving and hopefully bring us to a better world. Okay, who I am, I'm Pablo, my name is Pablo, but more define what best define me is my action and my experiences. I think role game plays and sports and movies are super important and as a kid I was dyslexic and the system didn't feel my necessities. So I really hated a school, but I love learning and it's something that I've been doing all my life and it's curious how a system that is designed to give you knowledge and to help you with the learning. In my case, it didn't work at all. And just that's a little starting point. I believe in the copy, transformation and combine all things and on my adulthood everything shifted and I became successful professional and I love IT teams. What I do for a living is a product manager and what I do on my pet projects, site projects, call it as you want is related to gaming and gamifying organizations and experiences. Nowadays, I trust more my intuition which is something that it really was broken during my childhood and you will see a little bit more about that. Here are some things of my interest, licenses, like open science, which is the case of the Institute, philosophy, that's the first meme, education, governance, family, art, love, communication, technology and nature and a bunch of other things that I will be happy to share any other time. On one of the questions that Daniel asked me for this talk, it was what was exciting for me about all this, what, right? And for me, I always, when I give talks, I like to serve a view of the last 100 years and the industrial revolution and how the ones that were on this world before us did something right. I'm sure that we all know big mistakes are big things that have failed or that could be better through iterations, but on data, it's very promising and very good. I think it's inspiring to have the basic education, how it grows, the vaccination, what it did for us, democracy and how poverty dropped drastically, how literacy it grew and because all these things, child mortality just went very low. So basically, the future of education is what makes me be excited and how I ended on this ecosystem. It's super random and I'm gonna explain it in hopefully two to five minutes. I, as a product manager and very interested in technology, I heard about blockchains back in 2015 and I didn't understood anything at all. And then back in 2017, I heard about the STOs, I did not disturb them, I did not play that game, but it was raising a fun thing to consider which is digital assets and ownership of those digital assets. For now, we didn't had a technology that could scale and that was good enough to ensure that the digital assets that you do own were yours and you had the freedom of transaction on them. I did understand that very fast and because I'm a game player and as a game player, you always have your different assets and do things with them and as well, I have been collecting music and all kind of digital assets. I bought this very famous collection of punks sold them very fast before I could be a millionaire but during 2021, I decided that I wanted to build something but then I went through this beautiful project that was the first art show on Ethereum which one of the artists, it was Daniel Friedman. So I as a curious person and a person that has been working with artists, I always like to know what is the driver of those artists, what they do and try to get in their walls to understand better and to get as much information as I can before I decide to encourage or collect or just always deeper on music or art for me is very nice thing to do. So I discover these cards and education was one of them and I went through Flickr which is where it was the first at least that I found a place where it was and the license it was a creative commons license of not commercial use, but it was an open license which I've been a part of creative commons community for a very, very long time since I was working with musicians and I think it's one of the best things that happened to the internet. These are some of the pragmatic things that the audience cares about. Hopefully we will have some of these on our games AI for sure is gonna be there. Free energy principle is gonna be one of the things that hopefully we help to minimize and to make it funnier games or happier games or games that help organizations to be better. The human factor, it's super important for me and implementation and action I learned by doing and it's one of the values that the Institute encouraged more. So I just jump on things and try to understand them through doing and sharing as I'm doing now. Here are the epistemics and collective is one of them that are as well super interest. Agents, of course, practical and curiosity which is the venture that I have the feeling that I'm being involved all my life and that I encourage always my two girls to be curious and to open things and look behind the doors and just be an active. Here are some values that I'd like to point with our exploration, the culture of curiosity that I did mention a lot. Continuous learning because that experience on the school I think I just keep going and keep going and it's something that sometimes bad experiences could turn out in good habits. Learning by doing as I already told integrity and inclusivity, honesty, accountability, a big one, trying to be professional on everything that you do, for me, I started as a human being curious. People of all backgrounds and perspectives are welcome to the institute. I could be very happy to be the link and to speak with someone that has doubts or if I could help, I think I could help on the very early stage and if you're not a very technical person I think I could help there and yeah. And as we are learning by doing, I would love to everyone that's on the stream, I'm gonna share this link, which is a new game that we are testing and I'm excited to share it with you. It's the adventure of curiosity and here, Daniel, I'm gonna copy and paste the link. It would be awesome. Okay, there's a tiny typo on your slides. It just ends with a curiosity. Thank you. But I'm gonna paste it into the YouTube live chat. Okay, thank you. Yeah, I told you I'm dyslexic. I have all kind of typos. My life. Sorry for that. I always try to improve it. Let's, okay, so you sir. Yeah, I put it in the YouTube chat so people can head over there. All right, yeah, could you tell us about this platform and this is the first active game? Yes. We started thinking on how to do a game a while ago and we decided to go with this picture that explains how it works. Actually, it's simple and you have four pieces, the internal states, external states. And in the middle, you have the sensory state and the active state. So, yeah. Here is the mission. You have some instructions that you can read and it would be fun if you go through them and you can go either way. And here we are gonna go to the sensory state where you're gonna listen to a music where some notes you can censor that are not in the famous song of Forelis. So, you have to take notes and you can go and you can see that we are on the mark of Langkid, which is so everything very Dalinian, the Dalibut painted. And here is your active part of the game where you have to do something, which is add the different notes that your music professor is telling you don't miss because you're probably on the external states are trying to play the music. So, you have to infer that you are missing some notes and go to the active state and start walking through the right notes. If I miss the note, I just follow into the nothing and come back into my internal state. So, probably would have to go back and censor again and once I take the right actions and the right notes, I can go and solve the puzzle. And the idea here is to have a tool, a gaming tool to explain or so you can experience how this active inference and free energy works. And once you go to the end of that adventure, which I'm not gonna finish because it's something that players should do, you're gonna have a door where you have to put the sequence in low keys of what were the missing notes and it's gonna allow you to come inside the treasure room where you will have feedback that you can send us and we will be very happy to have it. And then if you have a wallet, you will be able to have a digital asset. If not, we will add information to help people to go from web two to web three and to get those digital assets. And here to be developed would be future adventures that I would be very happy to do. So this was pretty much the presentation and hopefully we have new players in the coming days. Hope you have fun and I really, really encourage anyone to discover and play with all these ideas and thoughts and experiments and things that I have the gut feeling that are gonna revolutionize how we understand the world and how we can improve in it by reducing uncertainty. So thank you. And I don't know, Daniel, if you have any comments or something in this last 10 minutes. Oh yeah, we can definitely talk for a few minutes. I'm seeing many fun comments and questions in the chat so people can post it. Yeah, the gameplay is smoother on even a normal computer. It just like, it may look choppy on the live stream but it's quite a nice playing experience. Like maybe just a little context like how did you get into building these worlds or what kind of tools are these? And we obviously don't see them in today's mainstream educational and research offerings but like what could that even be to bring together this kind of immersion and role playing with education research? Yes. I've been playing role games and role games. You have very diverse kind of players. So I was playing with guys like me like have had a struggle through the school and guys that end up being researchers like yourself because they were very good in those environments. And for me it was like we are sharing a lot of cognitive values and information and they are really having fun with me. So what the academia is telling me is not what I'm feeling in outside academia. So that through the years it became more and more clear that was one of the drivers to have a very diverse kind of friends. And that one of the most important things for me to learn it was having fun and to be a good professional I have to be enjoying what I'm doing and to enjoy what I do it has to have that gamified feeling of it. And so in 2021 I was playing with a friend I just went to try to explore that road because blockchains what can do is have a persistent sea of objects and attributions to those objects and to those wallets that you can add and build on top of them. It's like having a huge database for a game purpose and then you figure out that again is kind of for our organization. So you need rules, you need to serve values, you need to have like what we are building now which is the CC on CEDO which you agree with the players what's gonna be about and how it's gonna work and we are building this fantastic storytelling of how a kind of world where it was in the early 20s compared to what's happening now in this imaginary world that's complementary to what's happening in real life. And we are just exploring those areas for three years and once I got into the institute we started sharing these ideas you send us to do system thinking course that it was super cool which we explore and learn and this document that I can say it was very important for us to understand and we find out the necessity to be able to computize what we are doing. So we start focusing more on ontology, on all the breaks that you need to understand to build the skeleton of something that could last for decades which is the idea and we are trying to build it so everyone can copy paste fork it and do whatever they want with the game adapt it to their own organizations and use it on their favor. And we are just exploring and having fun. That's awesome and there's so many pieces there it's true about the fun and the curiosity those intrinsic motivations like being six hours deep into counting ants in the desert or 5,000 pipette pushes deep or whatever it is in the trenches and in the last mile and it can be easy to lose sight of but when we lose sight and when we put out the fire and the spark it's a cold dark night and no career or status or extrinsic valuation is gonna provide like simply the joy and the deeper meaning of like two people going on a little adventure running around this space being inside active inference rather than active inference being like this high speed train that's leaving anyone behind or you're on the outside and it's taking these Swiss interns and you're getting thrown off. It's like it is cool and air conditioned in this space and people play and watch a lot of games where they wields violent arms and not even like saying whether that's good or bad it just it is a genre to wanna have a camera either in a character's head, first person perspective or a camera in a second or a third person perspective which you can switch between here and like those kinds of topics about like egocentric versus allocentric navigation and our ability to have a perspective swap on ourself and be able to take that kind of like out of body situational awareness in real spatial settings and just to be able to cruise through this space and it's like, yeah, there's no right way to boomerang from extra, you can go from external to sensory you can go here or there. It's an appetizer and a fertilization that like we've really just never seen. Okay. Yeah. Thank you very much, Daniel for this. Any last thoughts or comments? No, just enjoy while we are here, which is I think our mission. All right, thank you Pablo. I'm sure everyone's gonna really enjoy this and let's also of course continue to develop these games and continue to offer these on ramps and have times where we're meeting up in these spaces and playing around. Thank you. All right, very well. All right. All right. Welcome, Mal. This will be the final of the presentation sessions in the symposium. This is Mal Albarasen with the presentation shared pretensions in active inference. So thank you, Mal, to you. And you're the first to join with an AI steward slash guardian. Awesome. Shall I share my screen? Yes. Let's do, let's do, let's do. And it's just a second. I'm looking for a window. I think this will do. Okay, so can you see the screen? Yep, looks good. Thank you. Awesome. So unfortunately, I imagine that you are seeing the entire screen and not just the presentation. I apologize about that, but I got it too. It's cropped. I got it. Okay, wonderful. Thank you. All right. So welcome to the presentation on shared pretension under active inference. We're going to cover a lot of topics. Shall I introduce myself a little bit? Or have you already done that? Awesome. So my name is Mal Albarasen. I'm a PhD student in cognitive computing. I work in active inference in the interaction between social sciences, philosophy, consciousness science and artificial intelligence. We're going to start a presentation on pretension, an active inference. So I'm going to define a lot of terms. It's going to deal into philosophy terms that are very deep. And so let's get started. If you will just give me one small second. So shared pretension is a term derived from phenomenology and it refers to the shared anticipation of future events among a group of individuals. This philosophical study of experience and consciousness becomes particularly interesting when we consider it in the context of active inference. When we combine these two concepts, we get a powerful tool for understanding how groups of individuals can coordinate their actions and work towards a common goal. This is the main focus of the discussion today. So imagine a group of people planning a trip. They all share a common goal to have a successful and enjoyable trip. The shared goal forms the basis of their shared pretension. They all anticipate future events related to the trip, such as booking tickets, packing their bags and visiting various attractions. As they start to take actions towards their goal, they're also engaging in active inference. They're making predictions about what will happen next and then adjusting these predictions based on what actually happens. So for example, they might predict that a certain flight will be the cheapest based on their past experiences, but when they actually check the prices, they find that another flight is cheaper. So they adjust their prediction and book the other flight. This constant cycle of prediction and adjustment guided by their shared pretension helps the group coordinate their actions and make their trip a success. So we're going to delve deeper into the concepts and explore the role of past experiences known as retentions in shaping our current knowledge and future predictions. Let's explain first the concept of retentions. Retention is a key component of time consciousness that retains the past trajectory of a temporal object. It's the preservation of the content of a now-past, hyaletic datum in our present consciousness. A hyaletic datum refers to the raw, uninterpreted sensations that arise in our consciousness. When we retain these data, we are essentially keeping a record of past sensations and experiences that can inform our present and our future actions. In active inference, we can formalize retention by thinking of moments as binning of states which are bound to a hierarchical state. This way, retention is seen as the binning which becomes the current state in forming the next expected state. To further connect retention with active inference, we can consider the way in which retention plays a role in updating an agent's belief and generative model. The past experiences influence the current expectations and actions, thus shaping the agent's understanding of the world and the future. So let's consider a practical example to illustrate this concept. If you had a good experience at a particular restaurant, you might choose to go there again in the future. Your past experience influences your future decision. In hustling temporal phenomenology, we would think about much shorter timescale, but if we extend this concept, we can see a much larger structure wearing the retention as simply the past state. It's like having a mental note that says, this place is good and this is how an example of retention works in your everyday life. They're not merely representation or memories, but presentations of a temporally extended present, which distinguishes them from recollections. This involves the presentation of some already constituted content in conscious experience. Retentions also play a crucial role in mapping past knowledge onto current beliefs contributing to the development of shared narratives or goals that we will delve into in the rest of the presentation. They can be former now points and the current awareness of the past and far sedimented retentions can be engraved in scripts or deep temporal priors or more fluid deontic cues. This phenomenon corresponds to a specific application of message passing between agents in the present moment and in the sensorium. Retentions and protentions may also emerge jointly within a group forming a collective understanding that transcends individual perspectives. Protentions, as we saw, refer to the anticipation of future events based on past experiences and current knowledge. It's thus forward-looking. It's the forward-looking counterpart to retention which is backward-looking. Protention is the anticipation of future trajectories of a temporal object. It's the process of looking ahead, predicting what's to come based on what has been and what currently is. So in active inference, it could be formalized by considering the future states that are expected based on the current state and past states. Protention is thus the anticipation guiding the next action. We can consider the way in which protention plays a role in updating an agent's beliefs and generative models, influencing the actions and thus shaping the agent's understanding of the world and the future. So for instance, if you see dark clouds in the sky, you might anticipate that it's going to rain and decide to carry an umbrella. This anticipation of future events and its pretension, it's also going to determine whether or not you're wet when you go outside and therefore, whether or not you expect to be wet. Your brain is constantly making predictions about what's going to happen next. So let's put this in the context of temporal flow, which ties together those concepts of primal impressions and we will see retention and pretension. Temporal flow is the structure that represents the continuous progression of time from the past through the present and into the future. Agents are constantly revising their generative models and they do this by integrating the observations, which we could refer to here as primal impressions. This updating process is a balancing app between maintaining the predictive density of the agent's current beliefs and adapting to new information. Pretensions and pretension structure time for a given agent creating a framework that shapes their perception and anticipation. Different agents or systems will exhibit varying qualities and quantities of retention and pretension, which resembles something like Mike Levin's concept of cognitive light cones. So given the presence of pretension, we can anticipate a potential horizon, not merely as an arbitrary expectation. Pretensions are thus the point at which agents' beliefs about the world are updated to match their current observations, which leads to a more accurate understanding of the environment and allows for more effective, goal-directed behavior. Primal impressions are the fulfillment of a pretension, which is the time-bound nexus of perception and existence. It is resolved, sorry, it is the resolved probability that redefines a local gradient. It's in primal impression that reality collapses. The primal impression is thus modeled as the essential Bayesian cognitive moment. It is the juxtaposition of prior beliefs or retentions with ongoing sensory information in an unfolding future-oriented present or pretension. So now we have a description of temporal thickness from the phenomenology, which is compatible with generative modeling in active inference agents. For Husserl, this flow of conscious experience is primarily made up of sensations that well up in a raw, uninterpreted form that we called earlier, hyaletic data. So this data is the now phase of a temporal object. It is formatted according to the idetic or cognitive laws of consciousness. This is specifically relative to inner time. We can understand how temporal flow structures, how agents are able to make choices, act to self-evidence, connect the passive structure with the active component of decision-making or goal pursuit. Individuals have drives and teleology insofar as they have preferences. These preferences are protented goals reached towards by what we would call potentially the fiat, where the fiat constitute the will to act towards a desire. So the fulfillment of a pretension by fiat drives the agent towards a teleology across a trail set. In a retensional extension, more information is amassed. So the longer your retention, more information can be caught in a moment. As individuals amass enough information, they can then find a structure to the information which becomes a map rather than a territory. It is computationally more efficient than the sum of all the information amassed. This new state of information may lead individuals closer to a shared goal. The information amass turns into an emergent goal of the retention, which grabs focus at the present moment because it seems to bring signals to the model that the pretensional goal is closer and is computationally simpler to retain. So sharing of retentions and pretensions among group members in a social context can facilitate coordination and the achievement of a shared goal. By aligning their beliefs about the past or retentions and expectations about the future or pretensions, group members can better predict each other's actions and intentions. This allows them to work together more effectively. In some cases, group members may have retentions and pretensions that align naturally due to common experiences or understanding, even without active communication. However, group members often actively communicate their retentions and pretensions, aligning their beliefs and expectations through dialogue and interaction, thereby enhancing their ability to predict each other's actions and intentions. Retentions and pretensions may also emerge jointly within the group, not confined to any individual agent's own vault, forming a collective understanding that transcends individual perspectives. This complex interplay of individual and collective dynamics can be explored further in hierarchical generative models, encompassing resonant cognitive models, communication, and even joint emergence. The structure of time consciousness encapsulated in retention and pretension thus plays a significant role in the emergence of shared pretensions. As agents retain and process similar elements from their environment, they can co-compute shared pretensions by mapping each other's cognitive pathways and building upon one another's trials. This process is formalized by Bayesian updating. The relationship between the topology of pretensions and the expectation gradient, as well as the role of retention in mapping past knowledge on the current beliefs, plays a role in development of shared narratives and goals. We can use this to get a deeper insight into the processes that drive the emergence of collective behavior and the development of shared understanding among agents. There is no better way to do it than using active inference by extending generative models of individual agents to account for the beliefs and goals of other agents. This is the Kintu theory of mind. It enables them to anticipate and adapt to the actions of their peers by offloading to the group. So we discussed a little bit earlier shared goals. Pretentional goals are future-oriented objectives that individuals or groups strive to achieve. They're not just mere desires, but they're actively pursued through the act of will that we called earlier fiat. The fulfillment of a pretension drives the agent towards a teleology, that trail set we discussed. The focus of a prediction entails less surprise and therefore the agent self-evidence is by seeking to accurately predict their surroundings. Agents select the series of actions or policies which generate transitions between states that produce expected observations. So through the choice of policy, an individual can reach their pretended goal by minimizing the expected free energy that a given policy affords them. This seems pretty central and simple in terms of active inference. So farther pretensions are represented by the policies of longer time depth. In a hierarchical model, this corresponds to a policy tied to a higher level action in a slower time scale. The choice of policies driven by the expected free energy of that higher order policy or the entire multi-scale generative model is based on the agent's preferences, allowing an agent to conduct golden-acted behaviors across temporal scales. So this is where we can cash out Jeff Yoshimi's trail sets where agents assimilate actions into the anticipated continuity of objects in their surrounding world. This dissimulation ensures a smoother understanding of the environment and helps agents adapt effectively. Near pretensions act as the near present contextual proximity and the far pretensions rely on passive synthesis of deontic cues which are sedimented retentions in the environment. The preferences of the agents aligned by policies correspond to attracting sets which in the socio-cognitive setting can be instantiated as narratives. Having shared narrative is a necessary aspect of any exchange between artifacts, people, institutions that have some kind of attracting set, generalize synchrony which we will discuss a little bit later as well. So a team working on a project has a potential goal to complete the project successfully. The shared narrative might be the project plan that guides their actions. It's like everyone in the team has a shared story about what they're trying to achieve and how they're going to do it. The relationship between current states and future observations is captured within the B matrix transitions. This matrix represents the probability of transitioning to a new state based on the current state. So the B matrix transition and the A matrix is utilized to determine which new observations is most likely to occur given the current state and potential transitions. In the Husserlian framework, the expectation gradient structure plays a crucial role in defining the depth and fulfillment of pretensions. The structure represents the varying degrees of certainty and confidence agents have in their predictions about future states. The topology thus is inter, and the arrangements is interrelation, sorry, there's an interrelation between the pretensions or the arrangements of future and expectations which is proportional to the expectation gradient. The higher the gradient, the more detailed and accurate the pretensions are likely to be. The potential horizon represents thus the static states of the expected futures of the temporal object while the continuation horizon is the amalgamated trail of the temporal object. Potential and continuation horizons are carried in protensions while the imminent horizon consists of the primary impressions. So agents can coordinate through co-construction where agents or individuals interact and collaborate to build this shared understanding and shared goals. This process is deeply interconnected within a complex web where agents typically coexist with other agents or subjects. Forming the collective often refer to as multi-agent groups, swarm ensembles or sandwich. By aligning their beliefs to the signal found within their niche, they can better adapt to the environment and develop a shared understanding of their surroundings. Sharing information with other agents is beneficial as it allows them to trust one another and pool their resources. This collaborative approach is preferable to a conflict-driven scenario in which every other agent is perceived as a potential source of risk and surprise. By sharing information, agents could collectively construct a more accurate representation of their environment leading to more effective decision-making and action. Through these interactions, shared narratives and goals can emerge from a co-construction of the world and each agent's generative model. So these narratives and goals are the result of dynamic process that in fact can be mathematically modeled through the equations that we introduce in the paper that this presentation is based on but that I will present a little bit later. They capture the process of mutual learning where agents adapt their generative model based on the information gathered from their interactions with others. Agents may start off with slightly or greatly divergent generative model. So in order to minimize surprise, there will be a natural alignment within a group as agents adapt to the evolving dynamics of their environment and social contexts. For example, again, consider this team working on the project. Even if you have your own understanding of the model or of the project, you'll interact and share information and learn from each other and then adapt the model. Over time, you'll develop a shared understanding of the project and a shared goal for its completion. This is an instance of co-construction. It's like building a house together where everyone contributes to the final structure. The emergence of shared narratives and the learning and recognition of them may be made more efficient through certain mechanisms such as language. In an experiment illustrated in a paper that Carl was on, two agents share the same sensorium, ask each other questions. I think it's a paper on 20 Questions in Emma Holmes with a shared narrative entailing a common language. This illustrates the market efficiency and the minimization of free energy simply by having a shared language, a common ground. So shared potentials are common, future anticipations or goals that emerge among groups of individuals through co-construction. It allows them to better anticipate each other's actions and intentions and work towards a common goal. It's strongly tied to the conscious efforts of individuals to work towards that goal and to effectively create an organizational structures where we're engaged in a conversation in an exchange, assuming that this exchange is meant to align our generative models. So think of, for instance, the football team. Again, they have a shared potential to win the match. They anticipate each other's moves and adapt their strategies based on the ongoing game. Everyone in the team has a shared vision of victory and they're constantly adjusting their actions to make it happen. So individuals infer each other's mental states by observing cues, which they can attribute to and causes. They use their own model or models that they have learned to recognize about others. This process is known as generalized synchrony entailing mutual predictability and can be interpreted as a collective minimization of free energy. In a group setting, individuals create cues in the world that direct other agents' attentions towards the same signals. They're models of the world aligned by sharing similar goals and similar environments and by reinforcing patterns of sampling reality. So we could try to represent this through polynomials and morphisms. Each agent's interaction with the environment corresponds to a morphism and the collective interaction of all agents can be represented as a polynomial. So a morphism, so if a polynomial p is defined, a morphism between two polynomials p and q consists of a pair of functions f and g such that f maps p of one to q of one and g maps p of i to q of the quantity one. This morphism must satisfy the commutative requirements meaning that the following diagram commutes p of one, p of i maps to q of the quantity i and p of one maps to q of one. So we could see how we could create a framework to further explore the cooperative and competitive dynamics among agents. For multiple agents indexed by i with boundaries p and share environment boundary q, their interaction can be represented by a morphism of polynomials, which allows us to understand the shared potential in multi-agent systems and provides a foundation for further exploration of cooperative dynamics. We allow each agent to predict normally its own behavior but also the behavior of its companions and the environment's response to their actions. So the dynamics of shared productions are really useful here and can be properly cashed out mathematically. So recursive cognition and prediction on the shared environment of the agent's actions and response to each other's actions. We can use Sheep Theoretic and Tobos Theoretic approaches to understand multi-agent systems. A multi-agent system is a system where multiple individuals or entities interact. So for example, a traffic system can be considered a multi-agent system where different vehicles are interact. But Tobos is a category that has both spatial and logical structure, allowing for the expression of logical propositions and deductions within it. So we can use these approaches to construct a coherent understanding of the world from the perspective of multiple agents. The internal universe of agents, which can be seen as mathematical objects, containing beliefs, perceptions and predictions of individual agents are represented using Topoi. These internal universes provide a foundation for understanding and representing the complex interactions and shared understanding among multiple agents in this dynamic environment. We try to construct a consensus, Tobos, by patching together the internal universes or Toboi of multiple agents. And the idea is to use the Sheaf Theoretic and Tobos Theoretic structures to build a shared understanding. So we can thus derive the synchronization of goal-oriented actions among agents over time and leverage mathematical insights for category theory and generative models. This connection is the mechanism through which the group synchronizes over time and ultimately achieves their objective. Agents within the system might possess varying levels of certainty in achieving the objective. Some may choose to expand less energy temporarily, decelerating their progress towards a desired outcome. This variation in certainty and approach leads to the emergence of different roles within the group, with some agents focusing on computational tasks, while other monitor outcomes are engaged in different actions themselves. The nested and distributed structure of generative models in active inference allows for the integration of both short-term and long-term goals. Agents with similar or compatible preferences can coordinate their actions by aligning their policies and working towards common narratives. So this alignment reduces the overall expected free energy of their joint actions, increasing the likelihood of achieving shared goals and facilitating the sharing of goals and expectation among group members. They can learn, which allows them to learn a central role in shaping agents' preferences and strategies over time. Agents can gather new experiences and update their generative model, they refine their preferences and methods for achieving goals. They can influence the formation and modification of shared narratives by having different perspective over different parts of a goal. Here, the concept of general life synchrony is manifested when all agents have learned compatible or harmonized aspects of the overall generative model. The synchronization is rooted in the retentional and potential dynamics of agents and can be more accurately described using mathematical models. So in conclusion, we tried to shed light on the concept of shared potential goals and multi-agent systems. We've emphasized the importance of understanding this concept as it plays a crucial role in the coordination and synchronization of agents' actions and intentions. In the future, continuously highlight the role of cues in the environment and how they can be simple like marks on the road or more complex like symbols and language. The cues become carriers of passive synthesis, far pretensions and far retention and become a way to minimize free energy as they minimize the amount of any individual has to compute. It also allows us to delve into the concept of other evidencing in which the environment is in effect the self and others must learn to navigate in terms of others and self. We've provided mathematical models to an equation to analyze the dynamics of shared potentials more accurately in terms of category theory. We've explored the dynamics of shared potentials and potential in hierarchical generative models which reflect the complex interplay of individuals and collective dynamics. We've encapsulated the structure of time consciousness and how this allows agents to co-compute shared potentials by mapping each other's cognitive pathways. So our approach basically begins an examination of Husserl-Temperl phenomenology integrated with the hierarchical active inference model, polynomial morphisms and polynomials. This enables recursive cognition and prediction of not only an agent's actions but also the environment's response to these actions. So I think this comes to the end, Daniel. I think I'm literally at 30 minutes. Thank you. Impactable. Thank you, Mao. Yeah, great presentation, awesome topic. So thank you for joining and looking forward to seeing how the work continues to develop. Farewell, Mao. See you. Bye. See you. All right. Okay, all right. The next session is gonna be the final session of this symposium. It's been a really incredible journey and I think this final session is gonna be great as well. So we have a whole panel here and people can use their cameras even though I'm not if they want to. Welcome everyone. I'll just introduce the facilitator, Kurt Geimangel. So Kurt is a legend in the active ecosystem and has done incredible work with a variety of researchers and curious individuals in the space. So we could think of no one better to host this conversation and bring it all together. So Kurt, I'll pass it to you. You'll keep an eye on the live chat. Anyone can call it if they need any supports. Otherwise, thanks everybody for joining and we don't see you yet, Kurt, but. Yeah, it would be all right if I left and then came right back for whatever reason my camera's not working. Sounds good, sounds good. Okay, I'll be right back in about 25 seconds. Well, welcome everyone. We will get started in just a second. So odd. I apologize for that. Okay, well, you can hear my voice at least. Yep. Give me a moment, I apologize. Boy, okay. Well, we'll do what we can. Okay. So you all can't see me, but that's all right. So we have this all-star panel here today, which to me means my invitation must have been a mistake, but I'll take it. Thank you to the Active Inference Institute and I'll go around this virtual Zoom table and introduce everyone briefly. Carl Friston is the Welcome Principal Research Fellow and Scientific Director at the Wellcome Trust Center for Neuroimaging and Professor of Neurology at University College London. He's also the Chief Scientist at Versus. Anna Lemke is Chief of the Stanford Addiction Medical, sorry, Medicine Dual Diagnosis Clinic at Stanford University. Her popular books include Drug Dealer, MD, and Dopamine Nation. Rafael Kaufman is CTO of Digital Gaia and the on board, sorry, and is on the board of directors of the Active Inference Institute. Bert De Vries is a professor at Eindhoven University of Technology, where he directs the bias slab research team and also works with industry. Guillaume De Ma is an Associate Professor in Computational Psychiatry of the Faculty of Medicine at the University of Montreal and the Director of the Precision Psychiatry and Social Physiology Laboratory in the CHU Saint Justine Research Center. He's also affiliated with Mila or Mila, which is Quebec's Artificial Intelligence Institute and Other Art Science and Consciousness Service Initiatives. My name's Kurt Jemungal and I use them, I use my background in Mathematical Physics to analyze various theories of everything that are proposed. These include both the theoretical physics side of grand unification with gravity and dualities, other schemes, as well as attempting to understand what constitutes consciousness. You can find the podcast by typing in theories of everything onto YouTube or whatever podcast catcher you have. So my question to everyone, the initial question is, what have you been working on in the past few months and what excites you about it? We'll start with Bert, you seem to be smiling and your name sounds like mine. Okay, so, all right, so my name is Bert, and yeah, I lead a lab in Electrical Engineering Department and so our lab is called BISLAB, as you mentioned. So we are interested in Bayesian inference in general, but more specifically in doing it as fast as possible. That has lots of applications outside active inference, also in signal processing and other control systems, but definitely, of course, also for active inference. So half our work over the past few months, but even over the past few years has been on developing a toolbox to support or to get as far as we can go with trying to do real-time Bayesian inference or real-time energy minimization and trying it in applications. So that's the work that I try to lead a team of PhD students who do the real work, of course. I work around with a cup of coffee, but that's what they have been working on. Great, and Professor Lemke, please. Yeah, well, what I've been working on in the last year or so is spending a lot of time thinking about how it is that a faith framework in particular surrendered to a higher power improves people's lives. And I've been really struggling to come up with a way of talking about this that is inclusive, thinking about it, talking about it. It gets to a kind of core piece of it that I'm interested in, which is not so much whether or not God exists and what the proof is or not of whether God exists, but really what is it that changes in people's lives when they undergo a spiritual transformation, when they surrender to a higher power? Why do their lives get better when they do get better? And I'm really new to the active inference world. Daniel Friedman is the one who's introduced me to these ideas. I'm really excited about the ways in which the whole active inference model might help me at least think about what's going on there. And so that's what I'm working on. It's a little weird, but that's what I'm interested in. Great Professor Friston, you're muted. Yes, that's going to be the title of my next book, Can't You're on Mute? I think that's a retro from the good of the future. I generally work on what I'm told to work on. So I've just been trying to list what I've been working on in the past few months. Interestingly, it actually starts with Daniel Friedman and his fascination along with Axel Constant for evolutionary explanations, speak to many of the scale-free issues that we've been hearing about in the prior session. So that's what I was working on, really a scale-free approach to understanding temporary nested processes as free energy minimizing processes as a kind of active inference and how that would lead to a variational synthesis of evolution. And then I was told to work on belief sharing by Mao and colleagues versus and worked on that in the context of synthetic language and to try to establish a really simple proof of principle that communication was an emergent property of any free ensemble free energy minimizing system. Just to open brackets, just to make the observation, it struck me time and time again during the session today how important communication is. Whether we're talking about Burt's passing messages or responding to your, you've got a new posterior or we're talking about the kind of communication that Mao was talking about, one of everybody's been talking about. It does seem to be a really, and indeed the Gaia notion of overconnectivity and too much communication or the wrong kind of communication. It does strike me that's quite central to both the theme of the workshop and the specific presentations that we've heard. And then in the past few weeks working on stuff, I think that Burt would be more interested in, which is fast and frugal message passing schemes that enable the evaluation of expected free energy not to suffer from the limitations imposed by deep tree searches. So possibly Ashwin has spoken about this earlier on, but you're using ideas from dynamic programming backwards induction to try and take the pressure of the message passing to get fast and efficient evaluations of the expected free energy in sort of single unit, sort of single agent active inference. What have I been doing this week? Can't remember. That's what I've been doing. Great, Rafael Kaufman, please. Just a moment. And can everyone else hear? Okay. What about now? Yeah. Great, yeah, thanks. It's contagious, I think. So I've also been drinking a lot of coffee as I've been caring for my now nine month old daughter. So. Congratulations. Thank you. So aside from sleep deprivation, what I've been doing is I told you guys about this in the session earlier today, but in a nutshell, I've been working on what we call the Gaia Protocol or the Gaia Network, which is this open approach for building a common engine. You can think of it as a common runtime and a common language with the building blocks of shared decision-making in the context of the Metacrisis and getting to a set of building blocks for decision-making that don't kill us all. And that's kind of it. Rafael, how much coffee is a lot of coffee? Well, it's espresso, so I don't think it's quite comparable to American coffee. All right. And Guillaume Dumas, please. Yes, sure. Can you hear me? Yes. Okay. So in the context of actual inference, I have mostly focused on, well, the two big topic of my lab, so precision psychiatry and social physiology. So on the precision psychiatry side, Nielke Boyten presented earlier in this symposium our work on trying to model theory of mind, being able to evaluate through digital phenotyping the level of sophistication of interaction in patients with neurodevelopmental disorders. And that's connected with also our interest in the social mind and how we can connect our understanding of how we deal with other people but with also ourselves. And so that connect with other work at the Miele with deep learning architecture with higher order function in humans where we're trying to combine different theories of consciousness such as global workspace and attention schema theory. And in the context of attention schema theory, there is this kind of recycling of self and other mechanisms. So that's one big work. And on the other side, I'm very interested also in multi-agent systems. And so with the Nathalie Castle, we're gonna join us soon. We've been working on creativity and the emergence of cultural norms in multi-agent systems with the idea of applying that to climate action and also a new project that is starting here with the NFRF of a big Canadian project on indigenous knowledge and how we can think about new narrative around AI and typically a less solipsistic and individualistic way of dealing with AI. So that's the two main field of research. Can you explain what you mean by less solipsistic? Less solipsistic when it comes to the AI? The Californian view of AI because the narrative is mainly driven by US and California right now tends to circle around optimization, profits. I mean, even open AI states that AGI is about optimizing profitable human tasks. They put like profitable in their definitions. But in the case of indigenous systems of knowledge, we are more talking about community sustainability and the view already only in the application is less individualistic. But also in the cognitive science point of view, we have like the hardcore computationalism that I would say the brain is just like a computer in a very restrictive sense. And so AI, like typically those transformers that are massively used right now are kind of like brain in a vat in a very silicon sense while we are embodied systems that are constituted by our interaction with others. And so like, in a way, all we can think about artificial intelligence in this kind of more co-constitutive way and through a developmental and cultural lens. Okay, now speaking of the brain as a computer, we frequently hear in these circles the brain as a predictive machine. So where does this the quote unquote brain as a predictive machine have its limits? And also, is that to be interpreted as implying conscious experience is a predictive machine as well? Okay, and if not, why not? So I think Carl, you'd be a great person to start this off and then we'll go around the table again. Yeah, I was mindful of Mao's presentation the past half hour and a notion of temporal thickness. And I'm sure that's got a lot to do with the necessary conditions for the kind of consciousness and guessing you're referring to, that freedom from the moment. And if one reads prediction in its sort of psychological or pretentive or pretension nature in terms of being able to predict what will happen in the future, I think that that's gonna be a sort of a key bright line between things that do not possess a certain kind of sentience and things that do. And that bright line just rests basically upon, well, from the perspective of active inference having a generative model that includes the consequences of your own actions in the future. So just by having consequences, you're now talking about the future and the consequences in the future now become random variables, therefore you have to infer them, which leads you directly to the notion of planning as inference, which means that bright line is just a difference between things that plan and do not plan. And I would guess that that's where you probably want to start in terms of foregrounding the role of prediction as being an aspect of self-organization and it's sort of reading under active inference that characterizes conscious things from non-conscious things, namely the ability to plan. Does that make sense? I have follow up questions. We'll get to them at some point. But for now, Anna, do you have any comments on that and Bert as well afterward? Well, you know, I'm new to this field, so I'm just familiarizing myself even with the language, but I can respond to the question the brain is a predictive machine in terms of the work that I do clinically in the kind of psychopathology that I see when I'm working with patients, I work primarily with narrative, the stories that they tell about their lives. And one of the recurring themes I've seen through my work is that when patients tell stories in which they are perpetually the victims of other people's actions or the world, they tend to then, that tends to be a predictive model for them so that they will then go out into the world and unconsciously create scenarios which will perpetuate their victimhood. And that part of getting into wellness is to stop seeing themselves as entirely the victim of other people or circumstance instead, begin to appreciate what they contribute to their life problems. So I guess when I think of the limits of the brain as a predictive machine, I think in some ways, one of the big limits is that it's a very powerful predictive machine that actually allows us to subsequently shape what actually happens to us and or our perceptions of what happens to us, which then can perpetuate a false narrative. And I'll just give one small example from my own life. I've had my conflicts with my mother and one of my beefs about her is that she's a very poor communicator. And that whenever I email her, I either get a cryptic response or no response at all and it drives me crazy. And about five years ago, she sent an email, asked me some questions, I wrote her back and responded and it was clear to me that that required yet a response from her, which I never got. And that then perpetuated my narrative about her as a very poor communicator and many other negative things. And then about three months after I sent that email, I found the email in my draft box. So I had never actually responded to my mother's email. I hadn't actually sent the email. And that was for me just a personal, wonderful example of the ways in which we can, our actions can actually be manipulated to support our models and perpetuate falsehoods about the world that we live in. Right, and I have a quick question, Bert, just for Anna before we get to you. So use the word narrative there. How are you defining the word narrative? Is it the same as model? Is it a sequence of events? Like what is the specific definition of narrative? You know, I never really thought about it in those terms, but when I talk about narrative, I'm talking about the stories that people tell about their lives because that's sort of my data. And also about models because what I've discovered about self-narrative is it's not just a way to organize the past, it also becomes a roadmap for the future. That's the language that I use, but I see it maps very nicely onto your all's a language of modeling the world. I see, Professor DeVries, please. Yeah, it's clear that the ability to predict is the essence of intelligent decision-making. Yeah, I have a different background, right? I'm not a psychiatrist. So I think about these things in different ways. The thing that I think about when I think about prediction is, I would assume that it predicts far ahead that things get less accurate, right? If I predict that I want to go three quarters around the roundabout, I don't care about the centimeter where my car goes when I'm around there. I just want to get in the right lane. And I would assume that the brain doesn't take much, it takes less computations if you care about things less precisely, but that's very hard in a computer. We have, this is what kind of kills us in our, I think in our current way of implementing active inference. When we want to predict deep, we don't care about the accuracy, but we don't know how to do it much cheaper when things are less accurate. And so that's still a, that's a problem that we need to be working on. It might be a key to building or to scaling up active inference agents if we can actually compute messages that we want to know, that we don't care about if the LS precise, that we also spend much less computation on them. Thank you. Now, Professor Kaufman. I am not a professor, but I'll answer anyway. You just call me Rafa, Rafa. Yeah, so there's so many really interesting avenues here and it's not often that I get the pleasure of discussing this kind of stuff with this kind of diverse crowd. But I do, I mean, I've been interested in these questions for a pretty long time. And I think what comes to mind is like how active inference as a lens, for instance, it enables us to get another sense of what non-dualist views on consciousness are saying when they get us experientially to notice the difference between what we actually perceive and what is between the various different processes or things that are going on in our head and our tendency to lump them all together into the same, okay, this is my experience that I'm immersed in. So noticing, being able to notice the difference between the process of perceiving and acting and what it says it's automated and the narrative or the various narratives that are going on in parallel in my head, whether I'm aware of them or not and our ability to just look kind of seemingly arbitrary mental levels on top of it to try to make sense of it all and to force or static experience of having multiple frames going on at the same time makes sense of that experience in a way that aligns with our presuppositions about how things are. So I think that's super interesting. I also wanna say I like that, like this view of consciousness as being kind of defined by having a narrative or an internal model that's about self or at least is about self and that leads as Carl said to the planning as inference, that's actually like super deflationary in a way that people are not used to thinking about it and so it's an exciting, it's opening the door to all the sorts of exciting interdisciplinary conversations to be had on the basis of, at least I feel like better, less talking past each other than we've ever had. So I'm excited about that. Aaron, how is it that you're using the word deflationary there? Carl, I wanted to ask you about that when we had our podcast together, but Rafael has a similar question. So you said deflationary view when we have narratives of self, do you mean it leads us to something that consciousness is much more than? Sorry, much less. Yeah, I mean it's not necessarily much less than in the subjective sense, but it's much less than in the scientific sense. So one example of one not saying exactly that, but maybe something close to what Daniel Bennett calls hetero phenomenology, which is basically the statement that the sum of what one can say scientifically about consciousness is equal to what can be studied and modeled and theorized and communicated and learned and agreed upon on the basis of objective of what people, including you, but also other people say about their experience, which is, and of course, like what can be measured up neurocorrelates and whatever, which is not necessarily the same as our experience. So what we think is our direct experience of being conscious, right? So one way to look at it is that the science of consciousness doesn't necessarily have to put the primacy on people, the preferences, people saying they have quality that we thought in that sense, it's the inflationary, right? It's, you have to explain why people believe they have quality or not why people have quality because it's not necessarily a scientific truth that people have quality, right? Professor Dumas. Yeah, so well, on the limits of various parts of machine, well, we should be always careful to not move from one map territory fallacy to another one, for sure. So we need to be skeptics and avoid to refide the methods. And well, this world symposium show how this metaphor is super useful and fruitful, but I can see like two main limits or at least things where we should be careful. So following what Professor Lemke said, like in psychiatry, I think the looping effects and the way people picture themselves can be very unpredictable. And so the way we think about the brain as a predictive machine, think prediction for patients would have different signification of what we mean by predictive machines in the context of active inference. And in general, also like the, I can see that certain psychologists or anthropologists would have a big appeal all of a sudden for active inference and pretty different principles, but taking the words as directly what they think the word mean, where it needs to be, we need to be careful in the way to communicate it. And typically following also what Professor DeVries said about prediction as being super important for decision-making, I think like we should be careful about the weirdness of cognitive science and how it doesn't necessarily expand to a non-western educated industrialized country where maybe the cultural value is not about optimization of your decision-making or your profit. And so that's the first thing where I think we should be careful. It's more like a matter of how to communicate the theory in that case, not necessarily a limit of the theory itself. But the second one is about the equivalence or not with other frameworks. So like I can see how we had the Mao Al-Basin talking about category theory, and I'm very curious about right now among the different frameworks out there, how can we define what is equivalent or not with active inference to be able to see those limits. Maybe we are living what happened with quantum mechanics in the early 20th century with different interpretation, with different tools to model quantum mechanics and here with artificial intelligence and cognitive neuroscience, we have also all those frameworks. And I think the limitation would be then to not take this interpretation as the only one but try to have a cross talk and a differential diagnosis of which one is the best for explaining what. Okay, now my role is the moderator, but I would like you all to speak to one another. So I'm gonna ask a question that will serve as, well, you each will speak on it, but as the other person is speaking, as sorry, as the other people are speaking, just think, okay, is there a question that I have or is there a comment that I have? So the question is what are some of the recent advancements or breakthroughs in your respective fields that you find particularly promising? And well, recent, let's say 2020 till now. So we'll start off with Rafael, you're smiling and you look like a champion. Well, I was thinking what is my field, right? Cause we're very much in the last mile of applying the wonderful stuff that y'all come up with and making it useful on the ground. But so I'm extraordinarily excited about research like what Bert is doing. And I feel like as you all were saying, we're also starting to see some convergence on different ways to get to the same, at least to the same shape of answers. And in some cases, even to the same results. So one example of this is work that Chris, Chris Fields, Carl and others have been doing on the quantum, basically explaining quantum information theory and how it relates to the free energy principle, looking at the free energy principle as the classical limit of quantum information theory. And I don't pretend to understand all of it, but as somebody who's coming from a quantum physics background, just being alive to see this kind of convergence happening, which as I mean, as Guillaume said, we've been at this business of like both classical, what the hell's going on for over a century now? And it's nice to be at a time where again, we're starting to talk less about each other. And that's from a theoretical perspective, from a practical perspective, I think just finding that we have all the building blocks to create a fast, interpretable, reliable and aligned decision-making systems, which includes AI systems, autonomous systems that have all those characteristics. And that it turns out that all the things that the reinforcement learning and the neural networks people thought were very hard or impossible are sort of in the realm of what we can do. Looking at things from an active inference lens and vice versa, a lot of things that active inference models haven't incorporated so far. It turns out if you toss in a neural network approximator here and some boroughs and other massively parallel computation techniques, it also becomes feasible. So it's converging. Yeah, for me, what's remarkable is that there's so many domains of what we thought was exclusively human or would be exclusively human for decades that in just the past couple of years, robots or computers seem to be just as good if not exceeding us. And I don't know if that's promising or worrisome, but anyhow, Karl, please, if you don't mind answering the question and then we'll go Anna, then Bert and then Dumas. Actually just reflecting upon one of your observations, I think it's very difficult to identify one thing. In a sense, what is impressive is the diversity of advances and applications. I just say that because that's what I was thinking over the past six hours. It's listening to amazing presentation after presentation and just noting, found diverse. And yet there's this common thread, this common commitment, basically satiate our curiosity and using the tools that naturalize that kind of sense-making and curious behavior and communication that inherit from either maths or category theory or as rough notes now quantum information theory, but just to pick up on a couple of things which are relevant to this conversation. So the work with Chris Fields, it sounds lovely and exciting to bring quantum mechanics into active inference, but that's not the move that I think Chris is really wanting to champion. The move I think is something that we've all been addressing in one way or another, which is leveraging the scale-free aspects of this principled approach to self-organization and hopefully self-organization to some kind of generalized synchrony. So that's where the quantum information theory gets into the game. It is scale-free and indeed it will go further and say it's completely background free and everything is constructed. It's just, so I think that that's a lovely move because once you've gone scale-free, you then start to ask deep questions about how you couple one scale to another scale and in a sense, ecosystems is just that. How do the delusions of an ecosystem? How is it constituted? How is it co-constructed? What is the structure of it? These are all the questions about how one scale links to another scale. So I think there have been lots of advances in that direction in many, many different fronts and you can read that either in terms of coupling difference of spatial or yeah, spatial scales, but probably more importantly, sort of temporal scales and you see that wherever you look, you'll just come back to what do you mean by a narrative? It is exactly, I think, as I said, it's just a plan. It's just a story, but notice a story that has a temporal aspect to it. I have narratives about being a good person, being a good father, being a good scientist. I also have narratives about, I want my cup of coffee or I have to go look after the, I don't, but Raph has to go look after the child. So yeah, we've all got narratives at very, very different timescales and of course if you're, I just came back to Bert's example of, I'm an autonomous vehicle and I'm sentient and we're five years into the future and I have to drive around the roundabout. What temporal scale and what kind of temporal course-gaining to define the narratives, necessarily narratives would be appropriate for that kind of situation and the ensuing planning. So to my, so a short answer to your question, I think there have been many, many advances. They have, I think what they've had in common is basically transcending either different domains but in particular different scales of application. I also agree with the notion, well, another I think important pragmatic advance is something that Bert mentioned which is democratization of this technology. I think Raph also hinted at, this is the time you start using this for the common good. So I think, you know, things like RxInfer and PyMDP and I didn't know about the guy, a project but it's as though there's been great advances there as well. So this kind of democratization, I think it's really important, this sort of socialization where everybody can play and start to sort of not to talk past each other. I think that's a very important advance. And did you say the word scale invariance or scale independence? I said scale invariant, I actually said scale free and I shouldn't have done that. I meant, I said scale free. So the idea that you can apply exactly the same mechanics and literally, for example, say from Bert's perspective, the same kind of message passing at different spacetime scales or at different levels in a hierarchical model. So I've actually got a question for Bert in terms of reactive message passing because reactive means that you don't have to prescribe the scheduling, but in addressing the problems or the issues that entailed by having to specify the scheduling of talking or message passing, you're bound to deal with time and in a scale invariant context or nested with nested scales, for example, you have to deal with the separation of temporal scales. So I think there's gonna be a very important generic question which technically Bert will have been thinking about furiously for the past few years. I think implicitly we're all gonna have to be addressing soon, which is how do you put the timing of your messages when you make a move or when you listen to a patient or when you actually pass the message on a factor graph? How are we going to be able to put the separation of time scales into the architecture in a way that speaks to this scale invariance? The gear project, for example, how do you integrate live feed from traffic flow sensors with fluctuations in the climate? These kinds of this kind of data comes at very, very different temporal scales and yet has to be assimilated and modeled in a way that is also, I think, has to take due courtesy to that separation of temporal scales. Bert, could you please recapitulate the question for the audience and then begin to answer it? Okay, yeah, the issue is if you have, I mean, active instance agents are nested agents and the higher levels supposed to operate on a larger temporal scale, but they're also working at a lower resolution. If you look far ahead, you don't care as much about precision, not that the centimeter level, when I go around to roundabout, I don't care the centimeter where I land, but for the next few milliseconds, I do care about because it may mean the difference between getting in the ditch or not. So I don't want to send, so the higher level, I want to look very deep ahead, but I don't want to send every millisecond message to go to look, let's say, a minute ahead because so I need to space it out a lot, but then I may miss things. So preferably you would just send inaccurate messages, but that only works if you actually have a method to also, let's say, use less computational power to compute a less accurate message. And we're not good at that yet. What we are thinking about is there's a new field, or a new field, but there's a field called probabilistic numerics, where we are used in math to just compute everything very precisely or as precise as possible and do not care about how much computation we spend on it. So in probabilistic numerics, I hope we can leverage this for message computations. I would like to spend, let's say, proportionally less computational power on the accuracy of a message. One way possibly would be to consider a message, a latent variable that has an uncertainty by itself. But I don't have a totally clear answer for Carl because we haven't solved that either, but there is a problem in, let's say, what we do on our, I mean, our computers, weren't so completely different from computers that, let's say, from the brain, that, yeah, we are spending, we're spending just too many computations on messages that in the end are very, very inaccurate. And that's a problem in what we do on our computers currently. Yeah, so I don't know how we, I mean, what you want to do with this on the higher level, if you wanna look maybe 10 times farther ahead and spend about the same amount of computation on the lower level, that's sort of our goal. And, but we don't have an answer for that either at the moment. And Anna, so please feel free to comment on or ask a question to anyone. Yeah, well, I mean, I don't have anything to contribute, unfortunately, to how computers work, but I can tell you that this idea of temporal scales is something that we face often in our work with patients, for example, addicted patients are very focused on short-term rewards. And it's a fact, in fact, their ability to control how they feel in the moment that is partially what drives the addiction. So when I try to adopt Urall's language of minimizing surprise or minimizing free entropy, that's one of the things that people are trying to do on a short-term time scale when they become addicted. So I have a young woman who is addicted to nitrous oxide, which has a very fast onset of intoxication over the order of seconds and a very fast offset. And she says that that's exactly what she likes about it because she's controlling it second by second. So when we work with patients to get them out of that short temporal horizon, we actually rely more and more on action and having them change something in their lives, namely abstain from their drug of choice for long enough to kind of completely reset their brains and to allow them to see this longer temporal horizon. Because when they're chasing this short control, they're actually not able to see themselves in the longer narrative arc of their lives. So that's what comes to mind for me. I don't think it's gonna be helpful to people who are trying to build computers, but that's the kinds of interventions that I make with humans. I have a question about that. So you mentioned control in the short-term, and you've also mentioned that you study the positive effects of having a higher power in your life or surrender, okay, I kind of gave the punchline away by saying surrendering to the higher power. What I was going was, okay, what's the association between seeing yourself in the largest timeframe and a higher power and then also, is there something that is akin to giving up control when you look farther and farther into the future? That's such a great, yeah, that's really at the heart of what I'm very interested in because it's a real paradox, right? It's this kind of locus of control within ourselves that really in modern culture, we think is a great thing. But when that's taken to an extreme, and one example of that is addictive behaviors, it's very bad for people and for communities. And so what can pull people out of that is this kind of surrender to a higher power, giving up that locus of control, locating that locus of control outside themselves, not necessarily like in a theistic sense. One of the things that they talk about in an alcoholics anonymous, for example, is you don't have to believe in God, it just doesn't, it's not you, you're not driving it. And so I would be very curious from the perspective of you're all's understanding of active inference and the free energy principle and how the brain works. Why is it that sort of embracing our inability to control what happens in our lives can actually be the very source of healing, especially embedded in this really kind of over controlled, yeah, I would even go so far as to say endemically narcissistic culture. Like I'm really curious, I don't want to take the conversation in a direction. Please, please, that's a fantastic question. So if anyone has a comment on that, please. I have some thoughts. And I think this applies both at the personal level and at the global level. And I think it has to do with the, my quote from earlier brought by Edward Fullbrook that if you're falling from a plane, yes, maybe an altimeter and some instruments might be useful, but what you really need is a parachute, right? So we have this, we have this presupposition that or whatever like framing we have, we have operating in our day-to-day is gonna be, okay, this is the right framing and it's gonna, it tends to be this, this rational scientific framing of, oh, okay, very, very near consequence for most stuff. And it turns out that even like if you inspect our day-to-day behavior, bigger, more complicated models are not necessarily better, which is where we get the success of heuristics under bounded computation, bounded rationality. And if you scale it up to 8 billion humans interacted in resource-constrained planet, or the possibilities, but also of challenges, then you just, you have to, we have to like drastically lower our bar or yes, how much material we have, but also even like how much, where does information gathering reach, diminishing returns, where does modeling reach, diminishing returns? There's a whole literature on in the business world about the expected value of information, how much you should actually invest, also in science is also known as optimal experiment design, where basically, acknowledging that you have limited budget in terms of how much you can act and how much you can probe and how much you can, how much time you can spend thinking about stuff. And I think what we're doing when we feel like burned out or exhausted from overthinking, we're innately feeling that okay, we've gone too high and we need to give ourselves a vacation, give ourselves some free time here. I think that doing this principle in a principled way, we're just not just like taking the heuristics and the signals that we inherit from evolution, but actually being able to figure out collaboratively and with some rigor. Okay, this is, we don't need to know exactly how much, how many degrees the world is gonna warm by 2020-100 in order to know that maybe it's a good idea to start doing something about the amount of carbon, the atmosphere or whatever, right? And so I think this leads to this idea of a real knowledge economy and of things like abstraction as a service. How can you actually build in this kind of like sophisticated and sophisticated translation layers that take some of this burden from ourselves as individuals and even as organizations, right? And just put it out in the world as value-added services, which is what they are. And Guillaume, if you have any statements or questions or retorts, then please feel free. Yeah, thanks. Now I was still also thinking about the recent breakthrough post-2020 in active inference. To me, like there were like theoretical progress that were very interesting. We heard already about the formalism maturity at the mathematical level between multi-scale and skilled free aspect. I really liked also the development of a more multi-agent perspective of active inference. It's very interesting, like especially like for instance, the seeing multi-agent systems, we have just heard of like even society as well, like as many as one system or many as many subsystems and how we can use maybe those formalisms to also deal with policymaking. That's very interesting venue. And the emergence of norms, culture and ideas also because one thing that I'm still struggling with in the case of active inference models is like how to get outside the checkerboard. You can put a lot in the model a priori and how you make the model, creating new stuff that is not backed inside at the first. And on that like one very interesting breakthrough was the application of active inference to morphogenesis. I really liked the work that has been done on that. And on the technical aspects, I think the two main focus that I love are all the deep active inference and how to scale up active inference because it's a strong limitation for the adoption of the formalism. If it's not scaled enough compared to other framework like deep learning. And then the link with empirical data. I really liked, for instance, the work of Ryan Smith and how to connect with the extra physiology and clinical data. And I think it's also something very important to anchor the theory in the real world and have like falsifiability and empirical validation of those models. And what are some of the applications of active inference to morphogenesis? So, well, I'm not the expert here. Carl would be the best to answer that, but I was referring to the work with Michael Levin and how a model can help to create a sort of embryogenesis. So maybe Carl, you can explain better than me. Please. Yeah, I don't have a reputation for explaining things very clearly. But yeah, that's absolutely right. It was just at work with Michael Levin and colleagues showing that you can get quite expressive and bimimetic pattern formation and movement of different cells into an organization, often described in terms of morphogenesis, simply by communicating your beliefs. So I'm coming back to this sort of cross cutting theme of communication. So if you just broadcast your beliefs and you're a little fell and you have, there are a little ensemblance of cells and they all have the shared generative model that includes if I was in this position, I would sense that. Now, they all have exactly the same generative model, the same predictions, the same expectations. And they're all broadcasting their beliefs about where they are. And the free energy minimizing solution is just when they're all in a place that they, such that they receive signals that they would expect to receive when they're in this place. And of course, if that is the same for everybody, there's only one arrangement where each cell finds its place. So basically it's just knowing your place is an emergent property of making the world mutually predictable through communication. So it's a deflationary account of morphogenesis. But the game, I thought, you know, one of your units is precision psychiatry. So I thought you're going to talk about precision. So I'm going to do a game now. I'm going to, he talked about morphogenesis. I'm going to talk about precision now. But I think there's, that's a really nice way just to pick up on themes which everybody's just mentioned. And in particular, the pathology of precision. And by precision, you can reprecision in the sense that Bert was talking about in terms of do you use an unsigned integer or a double? How much can I coarse-grained my numeric representation? Or you can use it in terms of coarse-graining in the sense of the renormalization group. You know, just chunking things in hours or years as opposed to milliseconds and minutes. Or you can look at it in terms of a statistician describing the reliability or the inverse variance of a signal. And of course we have to estimate that when I say we, I mean your agents and statisticians and act accordingly. And certainly in the work of Ryan Smith on addiction, much of the sort of mathematical explanation for these addictive locked-in OCD-like phenomena rests upon a failure to get that coarse-graining, that precision estimation right. The reason I wonder whether it would be just worthwhile revisiting addiction and psychopathology or certainly pathology of behavior from the point of view of getting precision wrong and certainly assigning too much precision to low-level processing, which is what Bert wants to avoid doing. It just strikes me that that kind of story also has mileage in terms of why we are in a state of paralysis when it comes to climate change. Because I also noticed, Guillain, that you talked about climate action. I've never heard that before, but that seems to me to be the important thing. Why isn't there any climate action? And Anna, it'll be like you go into the clinic and find some of the Parkinson's disease. Why isn't this person moving? And interestingly, the computational explanation for Parkinson's disease is assigning too much precision to the evidence that you're not moving before you move. So if you can't ignore the fact that nothing is changing or your prior beliefs, your predictions that I'm going to stand up or going to initiate walking, don't get a look at it because they are immediately canceled because you've assigned too much precision to the lower level processing. And I'm just wondering whether that kind of pathology is exactly what Ralph is trying to reverse by having models at hand, recommendations at hand, provide a more coarse-grained view of things, a deeper view. Anna, would you like to comment on that? Yeah, I love this idea of the pathology of precision because I think it manifests in so many different ways, not just among my patient population, but I think it's almost like a cultural sickness in a way, the ways in which we seem obsessed with certain types of data and we're missing the big picture. So I'm gonna think more about that. I'm gonna read more about it. I really appreciate the discussions. Interesting for me. And Anna, when you're referring to the pathology of precision, do you mean so in a more conscious sense that we're over-evaluating something that we don't need to? Whereas Carl, you're referring to it in an unconscious sense, like the brain is putting too much precision on something because in Parkinson's, it's not like you consciously are putting precision in a place. Well, I think you could look at it sort of like in both cases, when I think about addiction, people aren't doing it consciously. It's that they really do see this as adaptive and healthy and also even on some level, they can't do otherwise and they're not able to see the true impact of their drug use on their lives. They're genuinely not able to see the negative consequences. I mean, that's what contributes to getting caught in that vortex. But I mean, you could also see it as part of what's happened culturally, like for example, like the whole wellness industrial complex, the way that we now count ourselves through all these different devices. And if we could just count our breathing and count our heart rate and take supplements, then we would somehow reach some levitating state of precise wellness. And I don't know. I mean, just kind of, this is all kind of new ideas for me. That's interesting. So you believe that we can be inundated with health data and that that's detrimental to us. Oh, absolutely. I see that all the time. Uh-huh. So for me, I used to have, oh, maybe I'll take this out of the, well, this is going out live. Okay, let me be, let me figure out how to say this diplomatically. I used to have a device that would measure my heart rate, let's say that, and my sleep. And instead of improving my sleep, it led to me becoming obsessed with it and then noticing, oh, I didn't sleep well. I must not be feeling good today as well, because apparently there's a high connection between how you sleep quality and how well you feel. That's exactly it. Plus added to that the other layer that I should be able to control it, right? So with all of this data and information, I should be able to, I don't know, reduce free entropy or whatever, to reduce surprise as you guys talk about. Mm-hmm. You know, and I think that's obviously, that only goes so far and then can actually contribute to our misery because why aren't we all levitating like the Buddha or whatever when we have all these tools and we can pay attention to all this data? So... I see this also with people who have productivity tools. So, and not only that, but mechanical keyboards, let's say. Some, the reason why I don't have a mechanical keyboard, even though I think I'd love it is because I know that, why the heck do I care about the clicking sound of a keyboard? But if I got one, then I'd be like, well, what's the difference between clicking sound A versus clicking sound B versus clicking sound C? And I'd become obsessed with the trappings of productivity that is sharpening, so-called sharpening the ax rather than cutting down the tree. There's this phrase that is apocryphal and it's said, it's attributed to, I think, Lincoln, which is that he'd spend 80% of his time for half the time sharpening the knife than cutting the tree. And when you just, if you just think, and then this is just echoed in productivity circles, but it just, it can't be the case. Why would you spend so much time sharpening your ax? Like, look at anyone who cuts down a tree with an ax. Most of the time, they're not doing that. Anyhow, so if anyone has any comments on what was just said, please. Can I just talk about what I find exciting in my field about axi... Sure, sure, yes. I read, it was a long time ago, but there was a paper and it's about that it says, well, active inference is not really a scientific loop because it's biased. And I read that and the sound of the paper was kind of, so it's not good, but I think active inference is maybe not a side loop. It's an engineering loop because there's bias and we need bias in engineering. We need to make stuff. We need to build stuff. We need to have a bias. So it's an engineering design cycle. I see that everywhere around me. And active inference could be a complete breakthrough in engineering, right? The fields around me are signal processing. I am myself in a signal processing department or group and everybody builds algorithms. I mean, in active inference agents, it's inference over states. Then the floor below me, they build control systems. Well, it's inference over actions. Then other people are working on machine learning. It's inference over parameters. Active inference could be, and you'll like this, Kurt. It's a very deflationary view on engineering because everything is just inference. And so rather than building algorithms, everywhere, if we become really good at implementing energy minimization, we will be able to build a great engineering design cycle and we'll be engineering better machines for medical procedures or for other things that are important. So it has a tremendous application potential in engineering. In engineering, I think in many fields, people have sort of drifted in different directions. Control theorists have, I mean, they do almost the same thing as signal processing people, but they speak a different language now. And, you know, so, and signal processing people, it's a completely different group for the machine learning people, but it's all information processing. And this brings it, it's filled, can bring it together. So I think it's, but the thing is that in order to, to make it successful in engineering, we need to build an application that impresses, right? It's not like a tic-tac-toe thing. That it really needs to impress people. It needs to be better than, you know, some other control systems. But once we do that, I think there's tremendous application potential because there haven't been enormous breakthroughs in signal processing and control. The last big breakthrough, I think it was Kalman filtering, and this was 1960s. And I mean, it's kind of funny that the essence of what we do in active inference is also Kalman filtering. That's, I think there's tremendous opportunities for what we do here for engineering. So that's why it's exciting to me. Anna, do you mind expanding on what you said about acting and therapists should be helping their patients with that? Oh, sure, just so, I mean, one of my critiques of mental health treatment today is that there's not enough, there's not enough encouragement of patients to actually go and act differently in the world as a way of gathering data. Instead, it often ends up being kind of this world building between therapist and patients, not necessarily ultimately adaptive in the world. So I was really just kind of responding to what Raph was saying, that we need to act in the world. I think that's more true now in modern rich nations than ever before because we are so incredibly sedentary and interacting. Of course, we're interacting with a virtual world and that's good and bad, but I mean, I think we need to be actually acting in the world. And so there's different forms of therapy as you know, there's talk therapy and then there's also cognitive behavioral therapy or psychotherapy instead of talk therapy. But cognitive behavioral therapy, as far as I understand, focuses on the actions. Is that incorrect or? Well, it also focuses on the cognitions, on the cognitions and I mean again, treating addiction like you're not gonna really get that far with cognitive behavioral therapy or anything that's focused on just emotions and cognitions. People have to go out and actually try, stop using their, stopping their substance or their addictive behaviors and gather data from that experience and then come back and process it. Mm-hmm. So Anna, in your field, and this question will go to everyone, but in your field and what you study, where is the largest gap that you would like to see closed? Well, I mean, we're facing a huge mental health crisis. Now, we have more and more young people coming in with depression, anxiety, suicidality, addictions of all sorts and these are not necessarily people who are struggling by virtue of trauma or socioeconomic disparity. These are people who have really privileged lives in many instances. So it's really a puzzle. What is going on for people? And I think a big part of it is the fact that people are not having embodied experiences. They're not having experiences in the world and also the experiences that they are having are these kinds of very quick fixes and fast pleasures. So I think the co-created sort of models through healthier communication that allow people to feel part of a community and also to have like truthful, co-created narratives, trying to use the language here. I think that's really important. So for example, one of the things, remember I mentioned what he's excited about, one of the things I'm excited about in the field of addiction medicine is mutual support and the proliferation of things like alcohol is anonymous but also other mutual help groups, a lot of them existing now online and the way that people are together creating healthier narratives and acted together to counteract a lot of the unhealthy narratives that I think are driving a lot of decision-making today. I wanted to know, is there a correlation between the rise in mental health or sorry, mental illness or mental health issues, whatever we want to call it and a certain trait of people, is it affecting the population the same? So the whole population has raised 20% in terms of how many mental health issues they have per year or is it affecting people who deal with abstractions more and more? So for instance, we're talking over Zoom and some people study abstractions just like us and then there's some people whose work it is to do something physical like running or swim. Like is it affecting everyone equally or are you noticing that there's some broad trend? Well, the broad trends that are out there are just correlational but the more time that people will spend in the virtual world, the more likely they are to suffer from depression, anxiety and other mental health problems. People haven't really been able to narrow that down to specific content, but they have been able to save just the sheer amount of time that you're spending online increases your risk for certain mental health, poor mental health outcomes. Okay. Now, Carl, if anyone has any comments or questions, please just raise your physical hand. I can see that. And okay, Raphael, sorry. No, I was just gonna say that I think another notable trend is that and I just saw somebody say this on YouTube just yesterday that young people are disproportionately affected by things like climate grief because they're the ones that are gonna be alive to deal with it. And I think that applies more generally that, I know Peter Sanghi already like 30 years or something ago wrote about the inescapable network of neutrality. Reality that what we do affects each other, right? And we built, we took advantage of this huge resource buffer that's called the biosphere and earth to pretend that it didn't for quite a long time. And, you know, got a lot of mileage out of it, but now we're at a point where there's a whole generation of people that are coming to grips with the fact that I'm gonna stop myself from saying a square word but oh my God, we need to, we actually need to change everything about everything about everything that we do and we need to do it fast. And by the way, it's not just what we do in, it's not just what we do out there, it's also what we do inside how we get ready for how to show up for life internally, right? So no wonder that it had that impact of myself, I'm built with anxiety and a lot of other things. We've had conversations about what are we doing bringing a daughter into this world and all these kinds of things. I think it's only natural that it's coming to a head in this way right now. Neal, you have your hand up and I can't see you. Yes, yeah, it's connected with what has just been said. So, and your initial question about the gaps that needs to be closed. I think like a scale free model of health and mental health, particularly would be great. Like we are in silos in biomedical research and the fact that someone is having depression can come from interacting genes as much as interacting people and also is related to climate change and so on. So how we can have a new health systems that doesn't deal with those silos and integrate those different scales to me. It's like really a big challenge, but a challenge that current models and work that we can see going that direction. And I'm very interested about that. Bert. Yeah, well, in my field in engineering, active inference is not understood because almost all papers are written by neuroscientists and they're hard to read. So I was really happy to hear today that I think it's Sanjeev Namjoshi who was writing an engineering book on active inference. So that will, I think it will really help that together with the availability of good toolboxes for implementing active inference should make a lot of engineers much more enthusiastic about active inference because it's not something that is not understood at the moment in engineering circles. So I hope that the book will be good because I'm enthusiastic about that. And Carl, where are some gaps in your research that you'd like to see addressed? Oh, there are more gaps as a whole empty space out there yet to be explored, but in terms of what seems to be emerging from the session and specifically the past few answers it does seem to be important to have this very generic just to take Bert's sort of line that this is just one deflationary simple and probably the right way to understand stuff and to make recommendations or to describe people's actions possibly to themselves in a therapeutic context. And as such, it should be push button technology and it should be democratized and socialized and I think that's the challenge practically and one may ask why would you want to do that? For me, there are two clear imperatives. One is very abstract and it's not really my within my comfort zone and the other one is in my comfort zone and the one that's outside my comfort zone is this notion of interactivity and hyperconnectivity and the metacrisis that we heard about. And again, also refer to this in terms of what he was trying he was trying to distinguish between a Californian notion of optimality and another kind, another way forward. And to me it's a stark contrast with growth is good versus sustainability and of course the maths of the free energy principle is just about sustainability. It's just the description of the physics of systems random dynamical systems that self-organized to some non-equilibrium steady state and that is what we are. So for me, there's something deeply if you like apt about the free energy principle and it's common research as active inference in application to ecosystems and lived ecosystems and realized ecosystems. So if those basic principles can be brought back into globalization, into the market, into fintech, into social media, into politics, into climate action, I think that would be a good thing. I'm just mindful of this struck me in a number of the presentations today. If you remember before Bert was saying if you look at the brain, which is a really lovely example of a self-organizing system to a non-equilibrium steady state then it's empty. And what did he mean by that? What did it mean you're empty headed? What he meant was it's incredibly sparsely connected. Now that tells you immediately that a pathology of connectivity is hyperconnectivity over connectivity, which immediately, well, it made me very alert to the presentation of the metacrisis that one of the first three things that was underwrote them, the metacrisis or the current crisis where we're contending with is a destruction of that sparse, delicate connectivity that defines thingness and defines ensembles of things technically in terms of Markov blankets. So if we want a world in which lots of different things can coexist in some kind of generalized synchrony in a sustainable way, unique sparse connectivity and the pathology of the thing that will destroy that is over connectivity. So it seems to be very important that we get that into play in terms of machine learning, artificial intelligence, politics, fintech, climate change. The only way it's gonna get there is epistemically by equipping people to actually build their own little models and ask their own questions. You can't tell people this, they're gonna learn and they're gonna learn it for themselves. Just very quickly, because I'm sure, yeah, we've only got a couple of minutes left. The other agenda which I'm more familiar with is exactly Anna's and Guillaume's agenda, which is making this work in the context of Neurology and Psychiatry. So if you can democratize and socialize this way of describing things so that people can now build models of that particular patient in the other use of precision psychiatry, I suspended Guillaume's unit was called after. So to personalize medicine that is really personalized in the sense that you actually have your digital twin of your behavior. And then you've got that, you optimize your digital twin to become a model of your patient. And then you can start to do experiments on that model, behavioral interventions or even share that model very much in the spirit of CBT with the patient and say, look, this is you. This is what would happen if you went out and did this and this is what would happen if you went out and did that. That to my mind and indeed that was the initial motivation for much of this work was actually to build observation models of psychiatric conditions to work out both the pharmacological and physiological basis and the disruption of the overly the pathology of precision and message passing on the factor graphs that are our brain even though they are very empty on the one hand but also get that behavior that that key thing that Anna was talking about that active engagement with the lived world into that model and hence active inference. And just to conclude that activity, that sort of physical engagement, that embodiment that sort of forese and everything else. I think it's really coming to a head now in terms of people's after the large language model after the chat GPT moment, that the bounce back has been what's missing. What's not there and of course what is not there is agency and embodied engagement with the world. And that's why I think there's still a lot of work to be done in bringing artificial intelligence red as active inference to in a way that matters to people who can make a difference, which is basically everybody, but specifically politicians and doctors and the engineers and the like. So you all now have 30 seconds to 60 seconds to speak directly to the audience. What is the message that you, what closing message do you have? You're speaking directly to someone who's listening. They're a curious person. They're interested in active inference. They also want to lead better lives, hopefully and do something propitious. So what message do you have for them? Anna, we'll start with you. Ash, I'm just gonna say what pops into my mind right now is that one of the things I have learned from my patients who are trying to get into recovery from severe addictions is something that they call the set aside prayer where they set aside all of the notions that they have about how the world works and try to be completely open and receptive to information coming into their minds. And I think that's just a wonderful, a wonderful frame or concept for all of us living in the world to periodically just take a moment and take these models and just say, everything I think I know about the world, I'm going to temporarily suspend it and I'm just gonna be open. And when that happens, we can be present and learn in a way that it's not possible when we're just trying to reinforce our models. Great, fantastic. And Bert, then we'll go Carl, and then Rafa, and then Guillaume. Well, I just enjoyed today very much. I thought there was a little bit clearly of both for people from let's say psychology neuroscience but also for engineers. So if you haven't watched some of the talks, go look through the schedule because some of the talks were really good, I think. So I really enjoyed that. And then, yeah, but should I tell people, go work out, do a lot of sports. It's good for you. Oh, yeah. Sorry, I've got to make a joke, but I can't because it wouldn't happen. So I'm gonna use my 30 second just to thank Daniel and his team for this, you know. So if you want something to do, you should go and watch the live streams and get involved with this ecosystem. I hadn't seen that paper being presented before but I was really impressed with the sort of the active influence institute and its openness and its welcoming attitude and vision like the Smithsonian. So if you want to pursue these ideas, get involved and if you haven't got time, just make sure you attend next year's active influence institute celebration. But thank you, Daniel. Bye for now. Yeah, so I'll second what Carl said and follow on with it's an invitation not just to participate in the active influence institute but also the invitations participate in building this Gaia attractor, this new way of doing things that acknowledges the value of growth and also the value of sustainability joins it all together in this thing called regeneration and it really is a collective effort, a collective learning effort. So, and this also by the way also applies to the panelists as well. I think obviously what Bert and Carl are doing, it has immediate things that have to do with what we're after but one of the main things that we discussed, that we keep discussing is also like this, the intersubjectivity and the importance of being able to operate well as humans together and that connects directly to cognitive science, psychiatry. And yeah, so everybody that wants to be engaged and be a part of building a better world should be thinking about what am I doing with my as an individual or as an employee of an organization or as a researcher or as a leader or as a family member, what am I doing, how does it contribute to this new non-equilibrium set of state? So yeah, that's kind of it. I probably blew through the 60 seconds, but here it is. And Guillaume. Well, I would thanks also all of you for the discussion and the organizer for what they are doing. Indeed, like the work of the Active Inference Institute is very laudable and interesting from an open science perspective. They are really embodying that. So big kudos to them. And well, Berthe Fritz was saying like to do sport is good for you. I'm not very good at sports, but some say that science is a team sport. So at least have a good team perspective when doing science and being kind to each other would be the best advice to everyone. Well, thank you all. I also would like to... You get yours too. Somebody else has to come in from outside the Markov blanket though. Well, I wanted to just thank you, Daniel. Thank you, Daniel and Raphael. And well, and also Carl and Anna and Bert and Guillaume. This was tremendous, not a fun. And well, I hope I get to speak to you all individually. And for, as usual, I have way more questions than we were able to get to. Thank you all. Thanks, Greg. Thank you. Thanks, everyone. Farewell to the panelists. You're all welcome back anytime. Thanks, everybody, for watching or rewatching. Right now, we're about to head over to the Discord and hang out and talk a little more if you want to and then stay involved, get involved. So, till next time. Bye. Bye-bye.