 Hello, and welcome, everyone, to ACTIVINFLAB, live stream number 31.1. Yes, we are live. So, today it's October 19, 2021, and we're going to be having a discussion on the paper, Non-Equilibrium Thermodynamics and the Free Energy Principle in Biology. So, welcome to the ACTIVINFLAB, everyone. We are a participatory online lab that is communicating, learning, and practicing applied active inference. You can find us at the links here on this page. This is a recorded and an archived live stream, so please provide us with feedback so that we can improve our work. All backgrounds and perspectives are welcome here, and we'll be following good video etiquette for live streams. At this short link, you can see the past and upcoming live streams. Here we are at the end of October 2021 in paper number 31, and today we're going to be having an awesome group discussion. And then next week, we hope to have one or both of the authors join us. So, we'll be taking a lot of notes today and bringing up good things to talk with them on. And all of these are participatory live streams. So, if you want to join or if you want to co-organize with us, just let us know. Today in 31.1, we're just going to be learning and discussing this paper of Palosios and Colombo. And we're just putting out some perspective, giving some affordances for others to join in, and hopefully learning along the way. So, we have a few questions that Blue and others have prepared. Maxwell also prepared some notes to talk about. And we're just pretty much going to look through the questions that we have already written up. We'll add more. Anyone in the YouTube live chat who wants to ask a question, they're more than welcome to do so. And we'll just hop around wherever we need to go. We'll start off with the introductions. I'll just say hi and maybe something that we liked about the paper or something that stuck with us or something that brought us to this conversation. So, I'm Daniel. I'm a researcher in California. And I think something I'm excited about today maybe would be to look at this intersection of biology and thermodynamics and talk a little bit about the different ways we can approach that intersection. This paper has some formalisms, but not all the formalisms. There weren't figures, but there's so many graphical intuitions related to free energy landscapes and some other pieces. So, it just kind of cool to think about what are the different ways that we come to learn about the intersections of these cool ideas. And I'll pass to Stephen. Hello, yeah, Stephen. I'm in Toronto in Canada. Pleased to be here again. And I'm interested in this paper, the general topic of it, ergodicity comes up. And I'm interested to see, I often see these papers which have sort of free energy principle focus rather than say active inference focus. So, just to see how when that when that tree gets shaken, how that helps or hinders the sort of more process theory work, which is probably closer to where I'm interested in, but this is all very interesting. Thanks. I'll pass it over to Maxwell. Hey, I'm Maxwell Ramstead. I'm the CEO of Nested Minds Network, an active inference AI startup, and I am a fellow at the Welcome Center for Human Neuroimaging UCL in London. I'm looking forward to discussing this paper. I think there's a lot to disagree with. And I'll try to be collegial in my criticism. Beautiful. What more could we ask for? Let's start with the big question. And of course, other view, if you have a different take or there's other big questions, but this big question, which is in this paper, as well as the broader literature. What fundamental ideas from physics are utilized in free energy principle and active inference and how? And just what are the implications or tradeoffs associated with the use of this idea? Does it lend credibility to a theory to be grounded in physics? What does it mean to a theory for it to be grounded in a certain physical idea? Does it constrain? Does it prevent us from making the kinds of conclusions that we might want to draw about biological systems? So that's kind of the big question. Do either of you have any other sort of big questions that this raises or what is the alignment of like the vector of this paper with bigger ongoing threads in our research community? Well, I mean, I guess one outstanding question for me is what is the horizon of relevant literature in terms of time course that one needs to consider before, you know, trying to engage with and criticize this material. It's something that comes up, I think, and it might be useful to discuss. Maybe before we go into the specifics there, just when we're engaging in interdisciplinary or transdisciplinary scholarship or we're diving into fields where we might not have taken courses to the undergraduate or higher level. Just it's a really important question. How do we make appropriate claims? How do you take a little bit of a dip into the rabbit hole and come out with something that's useful rather than just being more confused than when you made the initial search? So these are big questions. Any thoughts or just how do we determine what literature are relevant? Like if we wanted to write a paper on something related to active inference or just in general, what do we do? Do we do keywords? Do we only look at the last five years? Are you asking the question rhetorically or do you want us to chime in? Maybe Steven first with a raise hand, but then just either. Unmute Steven and then yes. Steven, unmute and then continue. I'm probably not the best person because I keep digging and digging and digging, but I think it's a good question. I'm from a chemistry background a long way back and I always find it interesting as well because a lot of stuff that's used or developed comes from chemistry actually and uses physics. Like you say, we're in this kind of need to approximate an approximation. There's so many things that come up that are like new ways of describing stuff which makes total sense. I've read through the Free Energy Principle and Kristen's talk about approximation science and that science is ultimately an approximation. And then it takes me a couple of days to then get my head around that and it's like, yeah, that totally makes sense. But the implications of all of that are huge. The implications of being in non-equilibrium states itself disrupts a lot of the accuracy of certain ways that you can roll out things from physics. And are you rolling it out into chemistry, which is normally equilibrium chemistry? Are we rolling it out straight into biology? Are we rolling it out into something slightly different with all this modeling? So I think it does cause some challenges in terms of discussions because it's easy for other people to start telling the Free Energy Principle and Active Influence Community what they are and then start telling them and framing what they're not. And by which point, it's kind of hard to have a discussion because it's like the straw man's been set up. And I think that can be a little bit challenging for some conversations to move forward because now you've got a complex idea trying to, first of all, counter the fact that other people's paradigms have been challenged. So I think that itself, I don't know if that's a literature thing or whether that's a cognitive thing. But I think that that is part of the challenge that seems to be faced. I agree, Stephen. Thanks for that. And it's maybe a good segue, Maxwell. We can just start to hear some of your thoughts and all be typing some notes and maybe asking some follow-ups and unpacking because I know you'll have a lot to say on this. So maybe just go forward and I'll just type along and we'll see where we go. I mean, I guess that probably the thing that I find most puzzling about the paper is that it claims to be the first in the literature, essentially, to be looking at the connection between the physics and the Free Energy Principles. And I find this puzzling because it seems to me that the connection between physics and the Free Energy Principle has been really maybe the focus of active research in our community for like four years or so. So I mean, you know, Carl published or at least pre-printed his particular physics monograph in January 2019. And, you know, prior to that, we had all been working on, I mean, effectively the connections to physics. So I find this problematic because the work that we're going to discuss today focuses on, you know, older formulations of the Free Energy Principle where the connections to physics were maybe less obvious. I mean, as most of us in the community know now, the formal foundations of the Free Energy Principle have been subject to extensive debate over the last, you know, three years or so. Kind of Martin Beal and colleagues kind of started us off in this direction. And, you know, since Martin and a few other groups have stimulated interesting discussions through their very interesting and technically rigorous critiques of the Free Energy Principle. But really, it seems to me that this has been like a focus of our discussions for the last several years, the connections to physics and so on. You know, and where this has gotten us is the Bayesian Mechanics for Stationary Processes paper that was recently released, which accounts for the critical remarks that have been raised. And, you know, so really where we're at now is that we have formulations that start from non-equilibrium statistical mechanics and information geometry and end up, you know, saying, well, all systems that have stationary density dynamics and a Markov blanket will look as if their internal states are, you know, modeling in a loose way. In a more technical sense or in a more technical sense, encoding the parameters of probability distributions or Bayesian beliefs about external states. That's sort of where we're left with. And you get to this from the physics. So I just found that the remark was a bit puzzling. And the choice of literature was also, I think, a little puzzling for the same reasons. I mean, I have more specific thoughts, but I mean, you know, this is sort of for your first bullet point in response to the last one that we added. What's the relevant time horizon? I feel like a rapidly developing theory like the active inference framework and the free energy principle. I mean, if you're going to be producing tech, technical critical material, you should probably be working from the most recent. Versions. I mean, I find it puzzling, for instance, that the authors spend so much time talking about ergodicity when the word ergodic doesn't appear or just just about never appears in the particular physics monograph. So anyway, I just find that there's a kind of disconnect between what's presented here and the actual work on the physics of the free energy principle. That we have seen and, you know, I think that leads to several of the arguments being proposed, being a bit missing their mark to put it diplomatically. Thank you for the concise summary, Maxwell. So Steven with a raised hand and then we'll continue. Can I ask a little question as well relates to this is in the very first part of the abstract, it talks about the inevitable and emergent property of any ergodic random dynamical system at non equilibrium steady state possesses a Markov blanket. So that's a quote from the paper from 2013. The thing is, it says before that according to the free energy principle life is an in and then it quotes inevitable and emergent property, which is the quote, but I'm not sure that life is ever talked of as an inevitable and emergent property of any ergodic random dynamical system in terms of how it's referenced. And I'm wondering if there's a bit of a confusion in the paper there where it talks about this that quote that makes it as if that's what life that life is an emergent and an emergent property. And I think it's I think what we're talking about is adaptive reduction of surprise or Bayesian inference properties is an inevitable and emergent property, but I don't know if that would be life. But I don't know what your thoughts are on that or whether I'm misreading that. That reminds me of one of my favorite papers, what is life answering Schrodinger's question by Maxwell et al. in 2018. Because maybe you could even say something to this, what is the relationship between reducing surprise and life and do these recent technical developments in FEP help bridge the gap between inanimate and living cognitive or computational processes and where does physics play into that? That's very interesting. I mean, I think the FEP itself is not a theory of life just in and of itself, right? What it is, it's a theory to understand systems that are able to maintain themselves in non-equilibrium regime effectively. So, or I mean, more precisely, you know, stationary systems with a Markov blanket, which is to say the same according to the theory. So I've gone on the record as saying that the free energy principle can allow us to construct a theory of life. But the claim isn't just that the free energy principle in and of itself allows you to do that. I think, you know, the FEP needs to appeal to auxiliary hypotheses from, you know, biological and other sciences to really do that. I mean, it's a theory of thickness in physical systems that interest us and a subclass of those systems are biological systems. So, I mean, it is an aid in the formulation and probably the foundation of a formulation of a living system. So, so long as we agree that living systems are a subclass of persisting systems. And I mean, there's some talk in this paper that maybe this persisters, living systems as persisters might not be the right way to think about life. But yeah, I think it acts as a foundation. I mean, but then you need to add a lot more to turn that into a theory of life. I mean, if you read my Schrodinger paper carefully, for example, what we propose is a variational approach to neuro-ethology. We're not saying that the free energy principle in and of itself can just offer this. Like the free energy principle has to be used as a foundation from which you can build these variational approaches to neuro-ethology, to niche construction, to ecology, or what have you. So, I mean, that might just be nitpicking. So, I don't want to double down on that too much. It's good points and it's almost related to the definition or the distinction between like a discipline and a transdiscipline. If FEP is a discipline versus if it's connecting other disciplines, variational neuro-ethology, it's right there in the title. We're connecting across different areas and the framework of thickness is one of those enabling features that helps it be an integrative theory. But potentially it doesn't fulfill the disciplinary uses of theory. So, that's one interesting idea. I guess there's many other ways to go or things to ask. I guess I'm just curious. Again, like Maxwell, what do you think in the last few years? How have we clarified or learned more about where active inference and FEP are related to each other? So, how was that thought of in five years ago? Maybe when you were in grad school and then what do you think is sort of contemporary greetings, Blue, in terms of the distinction between active inference and FEP? And then welcome, Blue, and you can add anything you'd like. And I know some people want to mark the difference between the two. I think, personally, I think it's more of a sociological thing, because just about no one is contesting the validity of active inference models of cognitive systems, but there's a lot of pressure from different circles on the free energy formulation. So, I mean, it might be a sociological thing where it's just easier to say, oh, I don't work on this more metaphysicsy sounding thing. I just work on computational neuroscience, and that's straightforward. I think there's no warrant for that from a mathematical point of view. You should just, for example, look at the way that things are derived in the Bayesian mechanics paper. Yeah, I feel like active inference just falls out of the core Bayesian mechanics. I mean, I'd recommend everyone read through that Bayesian mechanics paper. It's really fantastic. It's very clear. One of the big differences between it and its previous free energy papers is that it's authored by, or at least there are two authors out of the four on the paper, who are proper mathematicians working on the mathematics of non-equilibrium statistical mechanics, effectively. So, you know, that paper really, I think, meets the standards of rigor in mathematics. You know, you have an existence proof for the synchronization manifold that's sort of at the core of the whole formal construct, the thing that links internal and external states, and then you have a construction of it. I mean, it's very straightforward, and having established, you know, the kind of core formal apparatus, it's unfolded, and by the end, you know, what you do is ask kind of, well, all right, let's split the blanket states up into sensory and active states, and active inference just falls out of, you know, making the active states subject to the dynamics that are unfolded throughout the paper. So, in my view, the connection is immediate. Yeah. I will grant that, you know, some of the computational models that we're using are a bit more, they look a lot more like just vanilla, you know, computational neuroscience models. You know, they're POMDPs and stuff like that, but what the math that underwrites it all is the math of the free energy principle, so I don't really, I don't really think the two can be separated. Thanks, Maxwell. And the Bayesian Mechanics paper we discussed in live stream number 26. So, it's, it is a lot to keep up with sometimes, but that was a very fascinating discussion. So, blue, and then anyone else? I was wondering if that was Bayesian Mechanics for stationary processes, the one we did, yeah. That's precisely it, yeah. So, I, sorry, I'm late on the scene, but I am curious about whether is the FEP a theory of life, or does it allow us to construct one? And does it only apply to life? Or can we take it in this kind of panpsychic direction, where even molecules like want to minimize their free energy and self assemble to become structures? That's a very good question. I would say that the free energy principle is not a metaphysics. So, it's a physics. So it's a formal framework that may or may not give us grip on the systems that we're interested in. I mean, you know, I, and I know there is some philosophical literature that presents it as more like, you know, a metaphysical statement about what life is. But I think it's more useful to think about the Bayesian Mechanics as just an extension of mechanics, effectively. The same way that, you know, statistical mechanics and quantum mechanics are flavors of mechanics, the Bayesian Mechanics is also, you know, and I realize that it's, you know, it's a little deflationary. It's a more instrumentalist reading of the free energy principle, but it's not completely deflationary in the sense that it's the model says that the best model of living and other existing systems is as a model of their environment. So, you know, there's a kind of double instrumentalism at play, I think, in the free energy stuff. To answer your question a bit more directly, I think it is, it can be, the FEP can be used to construct the theory of life. In and of itself, it's a theory of thinness. Yeah, it's a great set of questions to discuss. And I mean, this is only my view, like I don't mean to settle this once and for all, but there are people who want to defend a more metaphysical view on this. I guess the question is, how far do you want to take the commitment of the free energy principle? I mean, you know, that there are kind of two irreducible assumptions that go into the math. There's stationarity, which by the way is not the same as ergodicity. So, you know, I think a whole section of this argument just doesn't apply at all anymore. The one presented by Fombo and Palacio and Markov Blanket. So I mean, you know, if you think that, you know, systems that exist at some scale have to be stationary, and you think that, you know, systems that exist really do have a Markov Blanket, then you can be realist about the assumptions. But I kind of feel that it works whether or not you take the additional philosophical step of committing hardcore to the realist assumptions. You know, it works just as an, you know, an as-if theory, if you will. I mean, the whole thing tells you like living, sorry, systems that exist at non-equilibrium studies and have a Markov Blanket will look as if they're inferring external states. And I mean, yeah, I think this looks as if stuff is really important. And I realize that, you know, for some of us it may be a little bit deflationary, but I don't know if more is warranted from the math. I think that, you know, there's a, yeah, we just need to be kind of careful. Thanks for identifying those two key pieces, and maybe we'll have more to think about with the stationarity and ergodicity. So blue with the raised hand and then Stephen. So just a couple things, like when I say panpsychism, I wasn't trying to go into like the metaphysical realm. But because it's physics, because it's not metaphysics, can it then apply to life and non-life? Like can we just go all the way down? And so in that kind of way, I wasn't really implying as a metaphysical construct, but more in like the direct, does it broadly apply to non-life as well as life? And then also just a couple more things. So systems have a Markov Blanket, like Maxwell, that's you being entirely realist. Like I don't have a Markov Blanket, but can I be modeled with a Markov Blanket? Probably. So I don't know, like, you know, I don't curl up in my Markov Blanket at night when I go to bed. It's comfy. And then just lastly, like going back to the paper, this like, so I don't know, that was one of the main questions that I have for the authors is this concept of they say like systems are in dynamic equilibrium. I don't think that the non-equilibrium study state density is the same as dynamic equilibrium at all. And that was one of the main questions that I raised. And like dynamic equilibrium really is not, it's a chemical process, like of a reversible chemical reaction. So I don't, I really kind of missed that. That was one of the key things in the paper that I just didn't understand at all. Yeah, I mean, I very much agree with you on the points that you're saying, the points that you're raising here. So I mean, regarding the panpsychism thing, I mean, yeah, you can from physics now kind of say that any system that exists kind of looks as if it has a mind. You know, then you might not want to go full on the metaphysics of panpsychism, but it certainly warrants that kind of claim. I mean, in my view, this paper is making a few critical mistakes. And this is echoed a lot in, you know, a lot of the literature coming from more like poor people writing about this. I'm just going to cite, you know, two sentences from a paper preprinted this year. I won't, I won't name names, but you'll see the style of what I mean. To be as fair as possible will do this as in criticize the FEP by attending to the more recent published versions and variants of these claims. And then the next section starts literally the next sentence. The work presented in Friston 2013 is central to practically all the proposals comparing the FEP and inaction, and it just goes on and stuff. And I sort of feel like, yeah, the physics foundation has moved forward. Like, you know, Blue, you're pointing it out. It's all about non-equilibrium steady states and non-equilibrium steady state dynamics and, you know, stationarity. That's importantly different from ergodicity, you know, the nice things about the Bayesian mechanics paper is that it abstracts away from those assumptions. I mean, to be clear, the free energy principle at its core never really required the systems that you're studying, the ergodic. We assume an ergodic density in some of the earlier publications because it allows us to nest one step in the derivation, substituting the time average for the ensemble average in one of the really core derivations. But, you know, as Carl himself had been pointing out for a long time, you could finesse the math into, you know, something else. Like, there's no reason why you have to stick with the ergodic assumption. It was used to nest one step of the derivation and, you know, what lands the cost of and Greg Pavliotis and a company have done in this Bayesian mechanics paper with Carl is precisely to do the laborious work of actually generalizing the math to non-ergodic systems. So I mean, I think that's really important. The Bayesian mechanics paper, you know, we're in discussions with the people, active discussions with the people who produce the critical material. And I mean, at this point, I think everyone essentially agrees that Lance's new derivations respond to, you know, the critiques by Aglierra and colleagues that came out earlier this year and also the Beale at All critique. Remember that the core of the Beale at All critique is that for OU processes in general of the sort that Carl was using, you can't assume that the synchronization manifold will exist between internal and external states. In fact, what Martin and company have showed is that for a generic OU process, you can't, it does not exist. But, you know, that doesn't mean that you can't find very broad conditions under which the synchronization manifold exists. And that's what Lance and company have done in this new paper. Aglierra's critique, Aglierra at All, this is the Sussex Group also did a very interesting analysis of the math as it's figured in the particular physics paper showing that actually the precise assumptions that Carl uses in the paper would restrict the scope of the FEP to a rather narrow class of systems. But again, Lance just rederived all of the equations. So the formalism is now effectively different. And what you can really see from the math is that it applies to nonlinear systems and to non-aerogotic systems. And this leads me to another point. I think the assumption of non-equilibrium steady state is often not understood in the way that it should. So non-equilibrium steady state is a characterization of the density dynamic. It is not a characterization of the system itself. If I say that a system is at non-equilibrium steady state, I'm not saying that it is steadily in some state. What I'm saying is that the probability landscape that underwrites its dynamics is stationary or is stopped evolving. Those are two very, very, very different claims. So it doesn't follow from the fact that the system has a nest density that it is history-less. It doesn't follow from the existence of a nest density that there should only be one attractor. Like you can have metastability in an FEP system. You can have nonlinearity in an FEP system. You can have non-aerogaticity in an FEP system. That means that one of the three sections, the critical sections in this paper, are just completely off the mark. There's this new paper. That's some of the tenors of the inactive approach on precisely this kind of stuff as well. Unfortunately, the same mistakes are made. Aerogaticity in the context of the FEP really just means that the quantities that you're observing are measurable. It means that the average of your measures converges on average to a measure of the average, or of the mean that you're actually trying to calculate in a heuristic sense. It never meant that systems don't have a history. It never meant that systems necessarily converge on one attracting set. I really feel that that takes the teeth out of most of the formal criticism in this paper. The other critical parts of this paper talk about the challenges of modeling biological systems using state spaces. Those sections to me don't amount to much more than just saying it would be very difficult to do this. Therefore, it's not doable or something like that. I have a lot of friends who are computational biologists and I speak to a lot of people doing this. It's sort of like, well, some work would have to be done to write a paper to respond to this. What's the expression one person's modus tallens is another person's modus ponens? I sort of feel like, all right, well, you say this is impossible. I know that Formawee, Alec Chen is a really cool guy and he's brilliant. He and the Sussex team have recently, I mean this is two years old, so it's not that recently. I feel like this should be accounted for in this paper, but have recently automated state space construction in the free energy framework. We have active inference models now that can automatically assemble and can handle up to, it's either 10,000 or a million states. That's a lot of modeling power. I find the example that they discussed in a footnote there. However, this is on page 11. However, the size of the genotype space for a protein consisting of N residues chosen from an alphabet of amino acids of size M is N to the N. They say it's like 22B100 or 10 to the 130 possible states if you wanted to model amino acids. What they're not considering is the nested formulation. I kind of feel like nested nest densities should have been at the core of this argument and figure nowhere. Of course, if you think that you can just plot all of the states of the system across all the scales at which it exists in one homogenous state space, you're going to run into a problem. But you're not if you carefully construct these state spaces in a kind of nested fashion. A friend of mine, Van the Tower, works on, I pulled this paper up. He has a really cool preprint, a deep learning approach to capture the essence of Candida albicans morphologies. Basically there, what they do is start with a population of Candida, which is a fungus for those of you who don't know. And then you randomly generate a crisper and you introduce it into the population and you see how the population develops. And you plot each populational trajectory in a 30,000 hyperparameter space effectively. And you see the way that the Candida self-organizes into pockets depending on the kind of crispers that you put into them. This is not like fringe work. This is just routine computational biology where we're already plotting phenotypic expression in 30,000 dimensional spaces. This isn't like forthcoming work to be done. It's really just standard work. It's interesting how I think they modified the paper. We had discussed this in a live stream before and they nicely added some responses to some of the criticisms that I raised in discussion. I'd have to pull it out, but I said we're using this stuff in studies of cancer dynamics. I wrote in a paper, we can think of a biological system's extended phenotype as the set of attracting states of a couple dynamical system. And this is supported by studies of cancer genesis and propagation. I can link the papers and they say, well, these studies provide evidence that cancer can be fruitfully modeled and understood as an attractor in a dynamical system. But they don't provide us with evidence that all biological systems occupy an attractor. And I'm sort of thinking like, well, no, but this is a counter example to your saying that it's not doable in a real biological system. So if I'm being snarky, I feel like this is armchair biology of the worst time. You're telling me that what I'm doing is impossible while I'm doing it. And you're doing it by appealing to an old conception of the physics by talking about assumptions that aren't even part of the formalism anymore. I find that there's a lot to object to in this paper, both in terms of the substance and in terms of the general kind of strategy of kind of gotcha-ing. Now, does the free energy principle need to be discussed critically? Absolutely. And I should have caveat all this by saying that some of the most interesting and stimulating intellectual exchanges that I've had in my life have been with the people creating critical material. But, you know, the people that are generating this interesting critical material are actually engaging with the theory at its formal core. And, you know, I sort of feel like some of these things would have just been cleared up with, you know, a few conversations and looking at the more recent material. But now I feel like I'm just coveching at this point. Thank you, Maxwell. That's what these discussions are for and hopefully bringing into the open. It's just a multi-perspective scenario. To give you like a specific idea why this is problematic. Like, if you focus on ergodicity and not on the nested nest densities, then you're going to say things like, well, first you have to show that there's an attractor and then you have to show that the relaxation time is biologically plausible. And they claim that neither of these things are done in the free energy literature when literally, you know, the core 40-50 pages of Carl's physics monograph deal with. And, you know, Dan, I know we've read it all together during the last year, like, I don't know how they could have missed this. I mean, it's the physics monograph. This is supposed to be their topic. The core of the physics monograph is all about, you know, information length, effectively, right? And relaxation times back to non-equilibrium steady state and how all these things kind of work in a nested way to get you biological dynamics. So it seems to me like they are just completely missing their target by both by, you know, working out on formulations that are out of date and that introduce assumptions that are no longer at play. And also by just really ignoring, you know, the physics stuff, the free energy physics stuff that they claim to be criticizing, you know. We need to be just like if we were going to catch a baseball, we need to be performing anticipatory knowledge work, not merely reactive, especially to previous formats. Or if it's reactive, I mean, you know, react to the recent stuff. I mean, you know, there's been a concerted effort to ground this in the physics of non-equilibrium systems, like non-equilibrium statistical mechanics, information geometry. I mean, if you're going to be writing a criticism of that stuff, then you should probably be writing about that stuff is sort of my point. And, you know, and by the way, this often goes really badly when I point out that, hey, you know, you're working from a, you know, an outdated thing. They say, oh, well, you know, free energy version 25 from this morning and blah, blah, but I kind of feel like, you know, my friend, if you want to write a criticism of something, like you should be as charitable as possible with the target of your criticism on the one hand, and you should definitely not be working on versions of this that no one in the core community is still working on. And there is a dimension of, like, if you can't keep up, then I don't know, you know, I don't know who benefits from this. It's sort of like my questioning is like... Okay, so Steven and then Blue. Yeah, I think thanks for giving us a good insight there. I think the sociological psychological piece is present in all of this, as you say. And I think I've mentioned this before to you, Maxwell, actually, is the danger of systems, talking about systems. I think even if you were talking to people in the International Society of Systems Sciences, they often talk about systems, system dynamics, and everything else, things, complex systems kind of gets lumped into complexity, right? So there's a danger, because I don't think Carl necessarily would say FEP system, right? And that term system, I think, gets you into a lot of challenges, because the folk psychology of what a system is, is what we see it as from the outside. And, you know, these non-equilibrium studies, I'm wondering whether we say, Mark, he says blank, it's all the way up and down. But I'm not sure if it's more like thingness all the way up and down is safer than systems all the way up and down, because things can have all sorts of, you know, and they can have states and they can be what they are. And the question really becomes then, within all that ergodicity, which unlike physics, which tends to be about getting things to a certain number of decimal points of accuracy, Jen, it's a messy thing. You know, I was talking about that thing about approximation science. I mean, again, that gets your head around what that that makes suddenly you've got, well, the main thing is you just need enough of whatever this stationary extractable data is to be able to do something with it. And I'm kind of thinking that maybe because in the same way that in chemistry, if you have a beaker and you stir it up and basically the thing will have its dynamics, it will move based on the energy between the different things. And basically entropy is just basically the the force that's pushing it in the way that you don't expect it to. You know, we know it does this from this equilibrium state. We know that and entropy is just the thing in the middle. Now, no one has a bloody clue what's why. That's just a black box, right? Now, but that's because they're going from now when you go to these non-equilibrium far from equilibrium chemistry states and they're held, which you don't do when you mix up a beaker and just go from A to B. Now, we don't know how many non-equilibrium states these reactions go through, to be honest, in those cases now. But you wouldn't call the beaker a system. You'd call it a solution. You'd call it, you know, you might call the vessel and the chemical plant in which it reacts a system. But you just sort of kind of a beaker of stuff or things that's doing stuff and things, right? And then you sort of see what it does. And then it's like, what can we do that's useful? And normally what's useful is what energy shifts. But in this case, you've also saying the entropy can be useful. Not in the way most people would, what we could say the folk systems thinkers would kind of tend to think about it. So I'm just wondering if there's a danger of using the word system. Yeah, I would agree. I mean, I think we use the term system just because, you know, all of this grows out of dynamic systems theory in some sense. You know, Carl talks of particles, right? Like a system with a Markov blanket and does the free energy stuff is a particle in his account. So yeah, there is a, so there's a linguistic danger here because I agree with you. This is not what people think of when they hear the word system. A lot of our changes, as you pointed out, Steven, I've turned around that specific point. I think the term is important because especially in the context of relating the FEP to like the 4E approaches, I think it's not appreciated sufficiently that the FEP math is what happens if you take dynamical systems theory seriously. You know, so when I was doing my undergrad, I was reading the 4E guys and the 4E people because it's not just guys. And the 4E people were telling me, all right, well, you know, computer theory, you only get so far in the study of cognition, you have to do dynamical systems. Learn DST math and learn the physics of complex systems. So I did. And several of us did. But what happens is then if you continue asking questions along that direction, things arise like, well, how do you measure distances between states? And if you start asking that kind of question, inevitably, information and entropy creep back into the picture. Because that's how you measure distances between states, effectively, is in NATS and other information theoretic units. So I mean, this is to me the great irony of this whole, you know, FEP versus inactivism confrontation, which is slowly turning into like one of the big discussions in contemporary philosophy is, I don't think that the 4E people fully appreciate that the FEP stuff is just a prolongation of their thinking style, that if you if you really want to take seriously the core tenets of embodied cognition, then inevitably you are led back to information. There's no escape. It's really if you want to take their own framework seriously, there's no other place to go. But here, I unfortunately have to leave now. Thank you, Maxwell. Thank you for listening to me rant about this. I'll try to be there next week. And yeah, I mean, keep kicking ass. I mean, Dan Blue, I just love Active Inference Lab and I'm so grateful for the awesome work that you do keeping this community alive. Steven, it's always a pleasure to talk with you and I'm glad to see that, you know, there's so much, so much. Yeah, you're just there all the time and I just love your, you know, your your your contributions to the conversation. It's always a pleasure to talk to you. So big love to everyone. And I will be in touch shortly. Apologies for the snarkiness. It's just, you know, this is the kind of paper that really grinds my gear. I thought it was it was good to just point out a few of what I think are the issue. Thank you, Maxwell. See you soon. Goodbye. This blue will go to you. But this is, you know, it's a key difference between being an organizer and a facilitator or a participant. We're here to help scaffold individuals to participate in their perspectives. And that's the second order cybernetics. So things will get more contentious than we can even imagine. Someone just has to hold down the broadcasting. Make sure that hands are being raised and that protocols being respected because there are serious first level questions to resolve. But we can have precision in our second level policy. And that's going to return to a question in the live chat but first blue with a raised hand. Thank you. So I just wanted to, you know, say that a lot of what both Maxwell was saying and what Steven said just really underscores the need for consistent definition. Like how do we define a system and also like a consistent synthesis of the FEP like historically, where did it start? Where has it been? Where is it going? Like where are the landmarks? Like for example, this Bayesian mechanics for stationary processes paper, you know, has revamped the math and physics. And when we're making a new paper or a new argument, are we working with the current version of the model? Absolutely. Nice foreshadowing of some work that we're involved in. But if anyone is interested in sort of scholarship and a history from an empirical perspective on FEP and active, they should get in touch with blue or I because this is a project that we're super keen on. But that's a very important point. And that's why in our dot zero, we had like this slide with the active inference lab, the working ontology. So we're going in a participatory way, and we're working through these terms and we're finding where it's quoted in the literature, how it's used in the literature. And then we're adding notes and relevant resources. And we'll be developing curriculum that helps make simple declarative uses of these terms so that we don't get lost and we'll have the terms and we'll have the specifics of literature. And I'll be an awesome environment for learning, applying and communicating active inference. So before Stephen, you want to go with the question and then we'll go to the live chat. Well, I just add to that point you just made is I think it is really going to be important because we're seeing the confusions that can happen. Also, it's it's likely that there's going to be a kind of a multi ontology. There might be something in this whole field that speaks to how this changes just in the way that Carl Friston talks about when something is regimes through an infomax focus versus a Hamilton lease action focus versus, you know, a reward. So there may well be this commonality and there may also be a way to process when things systems, particles become, you know, derivations. And, you know, that's going to be interesting because there may be some nexus, whether it's called metaphysics or whether it's something more, you know, less less out there, maybe it can be grounded. And I think that that's starting to come up as we we hit, you know, that's transdisciplinarity for you right it's going to you're going to want to know what's in the discipline piece. So maybe there will be stuff which is kind of grounded in the biology world and grounded in the physics world and he's going to talk as best it can their language, right. And then there's going to be stuff which is in the applied world which ultimately starts to go transdisciplinary at some point as soon as you get out and get to any sort of scale. So I think that's going to be very, very interesting. Awesome. And both Steven and blue and many others join us weekly in dot edu where we talk about this. So let me read a few of the awesome comments from the live chat and ask a question from the live chat and anyone else can continue to ask a question. So there's a lot of discussion on the systems question. And if on said Maxwell, subscribe to our systems thinking course, we will discuss what the system is from the standpoint of systems engineering. So that's really important. Thanks, Yvonne, for sharing that. And indeed, two of the fundamental pillars or ontologies or transdisciplines of active inference lab are active inference and systems engineering. So those are the perspectives that we draw on and seek synthesis and application of. And we're all learning a lot and it will be updating our generative model in a way that changes our policy selection to be able to use words like system without going down the philosophy rabbit hole every time. And just having a way that we can talk about systems of interest like in comms our system of interest is the live stream. So how do we take a role based approach that embodies the best of our values and uses the term system. And we also get all these other pieces too. So good. Thanks for sharing that. Yvonne, Dave shared a lot of interesting information about life and about some of the history of a few different ideas related to conceptualizations of life. But the question that I wanted to get to was from Michael NYC who wrote, How would you model this conversation using this model? So that's a question we can think about. And it speaks to almost like applying because online teams are certainly an application that's important for different kinds of modeling. And I'll give a first thought on this. So one thing is when you said, How do we model it? The question is beautifully phrased because it doesn't say, Well, what does the theory say that this conversation is? That's setting up for an absolute hegemonic realist answer. How would you model it is going to be inherently different perspectives because different individuals and different years are going to have access to different tools and ideas and people. So they're going to model this conversation differently. So that's the first piece and that's moving from that sort of absolutist, modest realism into more of a multi-perspectival system. Now, what are some ways that one could model this? There could be a cyber physical depiction where one is very interested as we were in our 2020 Vatican at all paper in the information flows from a computational perspective. Like sense is the incoming states and actions are the upcoming states. So in the context of a video conversation, there's an audio visual feed in and there's an audio visual going out. So we could use blankets and interfaces and sense and action states to talk about just the cyber physical dimensions of the conversation. Another approach would be the hermeneutic or computational conversationalist approach, which Dave has connected back to a long lineage of work related to 1970s conversation theory. And he wrote, it provides a micro theory for nested mutual models focusing on the evolution and incorporation or rejection of predicates and their syntactic occurrence. Active inference work provides a corresponding needed macro theory focusing on secular evolution of the interchanges. So one could have a communication oriented model that focuses on the semantics of the conversation. One could make also as discussed that kind of a cyber physical very computational model. And there's other types of models that could be made. And then there's a million different implementations. So that's just the first pass, but it's a really important question. And it helps us connect the abstractions and the framework to the minute particulars, which are the system specific realizations that ground us in transdisciplinary and application. So Steven, and then any other questions from the live chat or blue. Yeah, I think that they're quite useful questions, though, is because that that question about how can I, I mean, ties into a couple of things. My first interest would be so OK, why would I so there's some but the other thing is how can I I how can I implicitly is also linked to. Is available to me like what information is possible to interact with, you know, that idea of this. So we can bring in we this modeling process sort of ties us in this sort of mid ground between how we actually want to talk about all of this, which I think is where the systems engineering is really useful because ultimately you kind of have to bring it into a system of externalize knowledge, which can be then communicated and moved around as in the papers and other things that can be done. But it's also how can you start to understand say the processes going on in the moment so it could be measuring what we're doing in this real time process and someone tracking that it could be a case of video in it and watching back the video with an interpretation thereof. It could also be, you know, if we were wearing sensors, you know, there could be just trying to model the dynamics of us as things, you know, there's a just trying to say how could I know what I can even try and to feed into a system. So I mean that one of the classic cases that's recently there with Ryan Smith, their work on interception they had to devise a method for recording. First they did it with the heart and they changed the way the heart was measured but they also did some recordings of the stomach and the stomach region. They had to find a way to get data to find out what could then be modeled. And was that a system they were modeling or was it almost the dynamics they were modeling and of the things, you know, and I think that's, I think, and just seeing how the dynamics change over time. And I think this is this is where maybe there's a lot of challenges that's going to be faced because we've had this conversation in this modeling world is, you know, jumping things into a system engineering world, which isn't only a non equilibrium philosophical world, it's quite useful for applied work because it does that ability to make something a system in a way that's intentionally that and it has system of interest just like Daniel was saying. And then you've got, but within that our system of interest is about going beyond the systems of obvious presentation, you know, there's stuff going on under the hood, which is not necessarily intuitively available at all. So I don't know what your thoughts on that Daniel, but I think it's interesting. Yeah, the only short thing I'd say is like, even if the system of interest, for example, is the mind of someone being educated. For example, in the curriculum of active inference lab or the audience who we're speaking to is our system of interest, or even live stream. Having your system of interest be the learner's mind or brain doesn't mean you're looking for the neuron level description. Having the system of interest be the live stream or the audience doesn't mean that we need to know every single person or how video encoders work. So it's an application oriented approach, perhaps even an action oriented approach that cuts through some of these questions or sidesteps or saves them for savoring. You know, after the project has been completed, after we've made the changes on the real world, it savers some of these truly excellent philosophical and important questions while also allowing us to act reasonably in the moment. And so it's always that North Star of a unified model of perception, cognition and action that's so motivating. So even as some of the formalisms do evolve, we hope that we can respect the spirit and the inertia of the research that we've carried out and also know when we can modulally swap in new elements of the framework. So blue, and then of course any questions in the live chat. So as a modeler, how would I model this conversation really like let's talk about how to really model it. And not from a philosophical standpoint like if this is my model of the FPP, I have a model of the FPP that exists in my brain. As does Daniel, Steven, Maxwell, we all have a model of what the FPP looks like. At the same time, we read a paper, have a conversation, take input from our, it's the action perception loop, right, my model exists. It's in my internal states. Through my senses, I perceive read here something that either confirms my model provides evidence for my existing model reduces my uncertainty about my existing model. I have some evidence input through my senses, then I update my model or if it doesn't jive or if it makes me realize something I mean always we're gaining information. So there's this epistemic value add through doing these conversations these live streams reading papers having these discussions, etc. So my my model is updated through these conversations and then holistically we kind of come to a place where our models either converge through this communicative process on something that resembles each other's models more similarly or not, right. We might have this divergence and then we go away with like this uncertainty like is my model wrong or or what was that argument about. I mean, everyone's been in conversations or discussions or heated scientific debates where you don't really come to a consensus, right. There is no general agreement or understanding. And so then you walk away with maybe some some maybe my model is invalid and then do some more deep thought about it. I mean that that's how I would that's how I would model it. That's just me in the Viacan at all 2020 paper with active inference and systems engineering systems thinking we use this font framework of formal documents ontology narratives and tools. And so one kind of fun parallel or a way that conversation and the real dynamics that arise can be modeled in active inference is like agree to disagree. Two individuals can have high precision and convergence on a narrative inference were agreeing to disagree. They can both be precise and agreed on that as then at the first level they can literally disagree. That's the whole point. So that is modeled in active inference as nested models just like Maxwell brought up. And so we have nested models and the second order matrix can be the same between two agents yet their first level matrix can be very different and none of those matrices are the system map territory. Mel Andrews active 14. So we've come back to these ideas and we'll continue to dance around them many many times because they're so fundamental that sometimes they're hard to see play out so Stephen and then anyone else. Yeah, I think we have to be careful to note that it appears as if. Information of what active inference tells it appears if knowledge is coming in and that it's being we see it in our heads so to speak and. And yet the this conversation is is we're we're attuning through the action of our info niche our semantic niche of talking and we're aligning what our actions. Give us in our externality be it through the diagrams be it through this document and the slides or whatever the boundary objects are that we can align to. Given us partly due to our own folks psychology over the last 150 years the sense and this is things really hard to counter this because this is kind of the way we tend to think you know that. We have it in our minds I so to speak so it's true there's like this generative model yet at the same time the generative model in. Is also part of what we're doing with generatively modeling each other's conversation and get in a sense of how I. Should say my actions of words none of those sentences were pre made in my head you know they it's an action policy that's close enough to my sense making so it feels like I'm on the right track right and until someone looks at me really strange. This is a carry on I've been in these conversations a few times a bit aware how to. Keep going right but the thing that's something that's quite useful to when we talk about this modeling is all we're modeling the action policies that are externalized and our generative model. Is as good as it is at being able to enact. Through participatory sense making what should happen next more than it is about having something that we've received. In our generative model. Thanks Steven that speaks to the interactionist versus instructionalism like is the FEP model that blue is describing this generative model of the FEP that's. In or embodied by all of us is it like a skyscraper made of glass and piping and there's 90 degree angles or is it. Some other type of metaphor that or even transcending metaphor and only realized through certain sequences of interaction so that's. Really interesting and cool yeah there's great comments Dave's thrown some biblical references in the chat and someone wrote I'm here first time in the live stream it's so good to listen all the cutting edge discussion about FEP keep up the good work so. Nice work Steven and blue but thanks also for joining you have tuned your regime of attention hopefully you're having some questions and thoughts come to mind. Just as they are for us were just on the other side of the video chat window and the participatory sense making side and how to mobilize that and enact it. We can have participatory sense making with people on the video stream those who make asynchronous edits and improvements to these slides like we still have to go through blues prior. Thoughts that were added to this slide deck and then from a history of science perspective and sociology like it's happening in a new way. One thing that was impressed on us I guess months to years ago forever ago in crypto slash act imp was like the time scale of these discussions going from peer review and annual conferences in person annual or multi annual cycles to like. The preprint and then we send the Twitter message to the author of the preprint could you give a guest stream and then the authors can respond like looking at what Maxwell wrote which was very important or said like Bayesian mechanics that was live stream number 26 so we did a dot zero contextualizing video. Respecting that people have many different backgrounds approaching the work from and then two of the commentaries bill and also Aguilaria Aguilaria were in guest stream number five and seven. So it's like we've heard from many of the important perspectives. And that's this participatory scientific sense making. We're all part of a bigger positive project that's transdisciplinary. It's not even just happening in the scientific literature, but it has to a large extent been because these are research research advances these are conceptual theoretical advances. And now we're all learning by doing and increasingly doing and applying. And so there's so much to learn and do so blue and then Stephen and then there's more great questions in the live chat so go ahead. So you're like giving shout out to us but I just want to like thank Daniel for all the scaffolding that he's provided in the active lab like to give us the kind of the baseline for these discussions to take place. And also, just to kind of go back to what Stephen was saying about, you know, about modeling and enacting really like, I don't see knowledge as coming in. In the modeling sense, I don't know that what you have is knowledge or what Daniel has is knowledge or what the paper contains is knowledge. All of that in my world is information. And I think that there's a difference between knowledge and information. And so like when I'm interpreting comments by Maxwell or Stephen or Daniel what when I'm interpreting that I take that information and transform it through my model right like so. So some of like take it or leave it or take what you want from this discussion and it's always like that I mean we always do that, whether people have to remind us, you know, to do it or not. Like you take away the parts that validate your model or that improve upon your model or that you believe improve upon your model in some way. So then when I have it and I have transformed it then maybe I can turn it into knowledge but I don't see coming in anything coming in as knowledge. That's just me. And with video chat it becomes super real like when there's lag all of a sudden it's like you're at the semantic level. And then it's like wait just the syntax I'm not hearing the words. So it's like we dip into this lower level and then imagine if the audio is working fine. You can totally hear non equilibrium steady state but then conceptually what does that mean. So these nested and overlapping layers of meaning so that's just very cool. So Steven and then there's a question in the chat. Yeah, this is a good point blue and I suppose it's there's also this sense that there's always got to be some belief in what you think the other person's saying. Like for instance, I believe the words mean XYZ because I know the English language, right? If we're all speaking in the language didn't know so I believe this word means this. I hear this word. These words together mean this. So there's always and I'm hearing it seeing it. I'm feeling it. The fact that my chair is comfortable changes things, right? So there is this sense that you know they talk about this in free energy principle is like it can also appear like, you know, so things as well as there being Bayesian inference that can appear like there's a Bayesian inference type process, which is what's happening. My models can be updated and I think what I'm really interested in is for some of these things to then to be capitulated as knowledge. I think there has to be and this is where the 4e work comes in is we bring this into some subconscious embodied method. Practices which we can read tap into by sort of eliciting what they're like maybe getting a metaphor it feels like the knowledge is coming from around me and I'm grabbing it some other people it feels like it's coming in as a sequence of very things. So my way of assembling the connections of my belief structure to the sensory information coming in will capitulate my modeling. And as best I can I'm going to be then trying to attune to the way everyone else is and we're all I'm sure it'd be different if we're all standing up during all of these talks. We might actually interpret and process a bit different if we're all in a pool having a cocktail, you know, I don't know. I think blues on a stationary bike about half the time. I have my stationary bike hidden out of view. But let's ask a question. Yeah, exactly. We could do a ride stream. We could do a pool stream, but you got to come to my house. Cool. I remember that. So let's ask a question from the live chat. And then we'll return to some of blues previously phrased questions about the paper. And that's what's kind of fun like in the dot zero, we get a lot of the details and the bulk of the paper and all the key ideas and figures if there are them. And then in the dot one, there's so many directions to go. And I remember early, it's like, wait, two times two hours group discussion on a paper. I'm used to academia papers being discussed for one hour, if ever, and people having very heterogeneous understandings or they did or didn't read the paper. So yet we find that there's always enough questions when we open the doors of the conversation to just like super fascinating direct directions for the discussion. So again, we'll ask a great question in the chat and then we'll kind of turn to some of these dot one questions because next week, hopefully we'll have one or both of the authors to be talking with. So Joseph wrote, how has active inference changed the way you view the world in daily life? That's a big question. Do either of you want to go first or I'm happy to go. I can chip in if you're one of the ways I've actually found it really helpful. And I sort of is changing the way I think about how I see the world. And also a bit change that I think about playing soccer. So the first one is there was a there was this sense that, you know, I'm seeing the world right. And it's all a bit separate from me with active inference. It can be quite nice to go out and think about actually perturbing the world and feeling the world as if you're feeling the world with your eyes club touched touching the world with your hands seeing the world and touching the world with your eyes. And I think that actually can help bring me out of my head, so to speak, in a metaphorical sense. So I found that useful actually just thinking about that the world has been something that's out there and it's very much something I'm sensing. And so that takes a little bit of the modernistic kind of concreteness. And then with soccer as well. It made me realize I used to always think as you see those things on telly like, you know, like someone's kicking a ball, you know, like the equations bit like terminates, you know. But now it's like, well, actually, it's, it's more like, I'm just, I'm just getting whatever my feeling is, right. And that whole other stuff, it doesn't matter. So then you realize that all these really great sports players, the really ones actually, they're actually they've, they've, they've climatized a feeling for what it means to like messy must have this feeling for what it means to kick. He doesn't need to have the trajectory of the ball in his head. Anyway, that helps us go through my crazy head. Lou, you want to go or I've. Okay, so there's a few angles and this could definitely be a worthy topic for another discussion but just to give a few angles. So first thinking how we view the world that was a visual illusion in the original question. And so it reminds me of the development of thought from seeing the world. That's the signal processing paradigm. Photons are pouring in and we're seeing categorizing classifying the world. Okay, then we have predictive processing. And that's like, well, you know, our visual field for most individuals, they see color in their entire visual field. And there's equal clarity in their visual field, even though there's an extremely heterogeneous distribution of resolution and color elements in the retina. And we have a big old blind spot. So color, clarity of vision and blind spot suggests that we're experiencing a generative model, a predictive model of vision. That's predictive processing and it applies also to other sensory modalities. And then that third step was where Stephen took it with the active vision, the epistemic foraging. And we know that even though we don't directly experience these either, there's eye saccades which move to the most informative part of the visual scene. So to me, it's about moving from signal processing 1.0 to predictive processing to active sensing and exploration as informed by epistemic and pragmatic. Value together. You know, eyes on the ball, but you got to also see more than that, generate more than that. So that's one really important part. The second one also parallel exactly what Stephen said, which is like, it made me think, even though I don't know about soccer, I don't know about the tele wrong continent. But it's kind of like the top down the video game model where there's like, I see it more and more in sports replays where it's like, it will pause and it will look like a video game. But that's not the action model of the player on the field. They're using much more of a personal flow and a feel and interactions and cues. And so they don't need to be having this 30,000 foot view. They need the peri personal perspective. So that is one piece and then just want a third one. But I think another there's many more we could list would be like the reward oriented world view. So reward maximization is the paradigm reinforcement learning is how neurons associate utility maximization is what investing is about in contrast with an uncertainty reduction paradigm, a precision optimization paradigm, where qualified levels of precision of nested models allow a lot of flexibility and a lot more realistic phenomenology for different models, because precision pursuit on an optimistic world model gets you utility. So I just think we get utility from a precision first world view. But from a utility oriented world view, we don't always get precision. And that I personally connect to how if we have like a tetrahedral based geometry of nature, we get cubes. But if we start with cubic geometry, we never get tetrahedra. So that's sort of where I connected utility is like bigger and a more general framework because utility falls out of precision on an optimistic generative model, but merely having utility. It's always going to be ad hoc or secondary where does curiosity and epistemic gain and all these other features of real decision making. So those are a few of the key pieces how it changes how I've seen the world blue. So besides from like having integrated fully into like my daily vocabulary, which is like, you know, we talked about things like the generative model like, you know, in completely unrelated offline conversations. So, so aside from like, you know, enhancing my vocabulary and building this new language of, you know, the FPP. That which is that's weird, right, like, okay. But aside from that, you know, one of the things one of the recurrent themes for me in active inference has been the idea like we touched on it in the mental action paper that we did, like Daniel you're going to have to pull the pull it up for me because I can ever remember you've got a better track. You've got your model is updated and configured with all of the appropriate live stream numbers, more so than mine. It's 28 it's 28 and it's only because of my extended niche, but continue. That's it. And so can you also query your extended niche because I can't ever remember the one where we talked about anxiety. Like was it if it was big five or sophisticated active inference or I always ask which one that one was but I can't ever that was a long time ago. But but going back, you know, touching on mental action and meditation and the idea that we touched on way in the long ago live stream of anxiety representing, you know, trying to project too many time steps ahead into the future, like, and having this relate to anxiety and and then like, depression is relating to, you know, fixating too many steps in the past behind you right. So this like model of like neuro typical things, it just always reminds me to come back to the present moment and I'm looking for an excuse to do that anyway just because of my inclination sort of Eastern religion and so forth. So, but but the reminder to stay in the present moment has now been like made through an appeal to my logical brain, like outside of a spiritual sense. And so that that kind of is is a cool, you know, effective of active inference, I think. Awesome. There's so many fun questions in in the chat. So I mean, I'll ask these, and then we can we can we'll still have time to return to the paper. So Mahal asks, okay, one, if you were to let your imagination roam freely, where do you wish or see the best possible real life application. So for that question, that's a question for us and the community and beyond. So we can maybe each just mention one thought that we're having. But, you know, if you see a possible future state, active inference of X, let's act infer serve, let's act first, update our models and serve and make the impact that we want to see. So this is actually an affordance of a question, because if two people are like, wow, we should do this, or could this be useful? If you can state it, you're like 90% of the way there. And we hope to develop the sort of transdisciplinary tools like active inference agent software and the expertise tools, people and ideas and the education. So that when somebody else who's not on this live stream, there's just three of us has that insight or sees it being applied in a new way. We want to help you scaffold that action trajectory, but blue and then Steve. So I guess I always come back to this to something that fascinates me fully is how information is integrated across scales. So how is the model of my, you know, DNA strands imported into my cells imported into my tissues imported into my organs imported into my sense of self. Right. Like so how does this and how do I import into the organisms that I'm part of like the active inference lab and so forth. So for me really like concretely realizing a mathematical way to find information passing up and down like is it salience or what else what other way this emergence. Like for me, that's like it. Like if active inference can nail that and this is like a fantasy daydream and I know people are working towards that but it's out there in the future. But if I could really wish one thing if I could daydream about my ultimate fantasy of active inference being fulfilled, it's completely that. Okay, awesome. Modeling nested biological systems something that we've seen qualitatively discussed in the literature like Ramstead et al. 2018 answering Schrodinger's question, variational ecology, variational neuro ethology, a lot of contributors to these works, but that kind of a model would be epic. Steven. Yeah, I'm really interested. Active inference. It doesn't have to do everything to be really a contribution it can it can it can fill certain gaps as well or give certain mechanisms that can give an insight either to attractively give a way to model or a way to just think about things. So what I'm curious about and I'm working on something called conceptual action sociometry which is an action method. I'm trying to, I'm really interested in how to work back from a high complexity context which would be too many dimensions to model from first principles. Rather than modeling the generative model per se, I'm looking at how to create an instrument to track the info niche being scribed by the agent. So it's almost like a trace. So, I'm really interested in how we can go both ways, as well as the sort of general extractable sort of principled work that can be show how generative models are happening and being applied across things like context of autism and other areas where there's different types of classifications. I'm really interested in kind of contextual work which will be outside of the ability to do full research but might be able to do an inquiry process and how to sort of just get a grip on that inquiry. So, be it people looking at their lived experiences and things like that. We're actually talking about that in the tools section. People are sort of giving me some advice or open to that once we get to the next phase. Awesome. Thanks for also sharing that and then I think the system just the one that I'll mention would be online teams or research knowledge inquiry and doing scholarship. And just at this point in time just facilitating the reduction of our uncertainty on the history of active inference and free energy principle literature and a few other areas because tools for research and sense making would be really awesome for a lot of different reasons. So cool. And then the second question from the hill and that sequence was what if we model our group conversation where we as humans disappear and where we are not exploring the universe but the universe is exploring itself utilizing us a beautiful thought question. And there's a lot to think about that so blue. Yeah, I mean I think that this is possible. Right. And, you know, we've talked about it in the computational boundary of the self the Mike Levin paper, like, ultimately, it's like dissolution of the cognitive boundary. And is there some like meta organism this goes to like the Gaia hypothesis, maybe not the universe but the earth, or maybe the earth is part of the universe. Maybe there's some ultimate like super organism that that would cognitive boundary contains all of us. And so just in the same way that like ourselves are working towards our optimal function given the circumstances that they're presented with right like given that assumption. Maybe we're also working toward the optimal state of the universe like every interaction that occurs between all of us is working toward the optimal state of the active inference lab or the optimal state of the United States of you know scientific research or who knows right. So any kind of, you know, umbrella category that we might fall into maybe we're serving the optimal state of that and really like I think it's quite possible and a really cool way to model active inference for sure. Another take on that from Bucky Fuller is this function first mentality led him to ask well what is the function of human and are we local sense maker what is our role. So those are big important questions. Let's return to 31.1 and kind of prime ourself for our dot two. Because in the dot zero, that's where like whoever's really excited about the paper, maybe whoever contacted the authors, or maybe one of the authors themselves or just a few people who want to learn by doing the dot zero, which was like blue and I for this one 31.1 the previous video, we go through the paper, the dot one, we unpack that sort of the bathtub phase. Everything is getting spread out a lot of sociological criticism criticisms or insights, few random questions from the get from the guests of our stream. A lot of different areas that a lot of threads get pulled out and then in the dot two, we try to have some synthesis because it's not the end of the discussion with the dot two but we have to move on from a live stream perspective so the dot two, especially when the authors can be there, we can get some more insights, we can resolve some questions that arose during the dot one and dot zero and the intervening times, so questions that didn't arise before we enacted this process. And then we get to ask those and then kind of use that as our jumping off point as we move on to our next phase. So here in the dot one, at the end, it's kind of like a fractal dot two. We're kind of trying to tie a few things together. We're turning to the text as a strange attractor so that we'll have a whole week to read, think, sleep, stationary bicycle, mind and body, relax, maybe do our day job. If we can find any time and be here, you know, next week or whenever works to continue the discussion. So blue, in this questions for 31.1, what were some quotes or some questions from the paper that you wanted to like learn more about or ask the authors about? Well, just as I mentioned, you know, even earlier in this live stream today, like the dynamic equilibrium of living systems like I'm not understanding that at all. Like maybe that's homeostasis, but homeostasis is not in my mind dynamic equilibrium. And I think that this, you know, Bayesian mechanics for stationary processes and the idea of a non-equilibrium study state density kind of supersedes this concept of dynamic equilibrium because homeostasis is not equilibrium. And so that's kind of a fundamental like it like bends in my brain a little bit when I when I've been reading over this paper. And also, you know, it says here we should pay closer attention to fundamental differences between physics and biology, and how these differences interact with importing concepts from physics into the modeling of biological targets. So I mean, is it physics all the way down because I always thought it was physics all the way down, right? Like, so I don't know that goes to, you know, like when I was in school, they always said, like, you know, biology is not as precise as chemistry and chemistry is not as precise as physics and physics is not as precise as math, right? Like, so there's this like tiered hierarchy of how, you know, fundamental your study can be, right? And so it, I think biology is based in the concept of physics, but but also chemistry, I feel like was really omitted and left out. And like in my heart, I'm a chemist, Stephen, I'm sure can relate to that. But really, like, so that's kind of one other thing I wanted to explore. And then the other part of the questions that I have for the authors are, you know, in active inference are the active states internal. Because they said in internal states include active states, but I've always thought of active states and sensory states as part of the interface between internal and external states as part of the Markov blanket. And so I just wanted to clear that up, maybe if that's a misunderstanding on my part, or, I mean, I guess action can happen internally, like as we've seen in the mental action. So that's definitely possible. But but not all active states are part of internal states, I thought. So I don't know, maybe that's some something that we can tease apart with the authors. And I'm looking forward to having them here and asking them these three questions, which are the ones that kind of for me, you know, don't align with my current model of the FEP. They were salient. They were different from your generative model. So it's a nice way to put it. And then I'll just show this one comic and then Stephen. So this is a classical XKCD comic. The fields arranged by purity. Okay, so it's a comic, you know, retweets are not endorsements, but you have this is a classic disciplinary hierarchy slash Ponzi scheme, pyramid scheme, whatever, you know, you have the physicist, let's just start in the middle. It's nice to be on top. That's the common pinnacle of the pyramid in certain paradigms. And then like, the physicists can make claims about lower levels. But then if the psychologist makes a claim about physics, hey, we need a sociology of molecules to be like, that's not science. But then the physicist talking about socio physics is like, Wow, that's so modern and exciting. So there's a sort of linearity, as well as potentially some sociological scaffolding that facilitates the diffusion of information and expertise and notoriety, like one way, but it's not always a two way street. And then the joke I guess in the XKCD is like the mathematician saying, Hey, I didn't see you all from over here. So again, just a comic now, playing with the media. What does it look like to I don't know, close the loop, or have these learn about each other's perspective, or there's a million ways that we could go counterfactual on this XKCD. What if I mean, that's the name of the XKCD's authors book, which is kind of fun. But how does active inference challenge this, either specifically or just how are broader ongoing sequences of actions like decentralized science, decentralized science, DCI, open science, open access, open source, global collaboration, transdisciplinary application, complexity science. I mean, there's a million ways in which this is kind of funny, because it speaks to an academic hierarchy. And maybe we've even met some characters who really embody these archetypes. But what are the archetypes going to be going forward? What are more accessible archetypes, and ones that are more true to form, and they can still be funny, I'm sure we'll have active memes just enough to warm the earth. But this is just one view that kind of speaks that disciplinary hierarchy. So yes, even then blue. Yeah, I think this, this is where this foundational shift is this this desire for purity starts has been really challenged. And of course, we're now we're in this post normal science era, where, you know, the reality is that we're going to have to make decisions and not just wait for the purity of the, the random control trial or whatever it is to make choices. So how can we be more systematic and rigorous and so then complexity comes in. And the challenges that we are facing is that a lot of psychologists and psychiatrists have become have also gone down the purity route. In a way, they're trying to like, the idea is they're all trying to be more pure. And the mathematician can be more pure than the others. But that question about being more pure. I think this is what blues points make a really good, you know, they come into question, like physics comes into question chemistry comes into question. When you talk about far from equilibrium states, even the idea of it in chemistry, they call it far from equilibrium chemistry. In physics, they call it non equilibrium states right because they can actually math and the mathematicians would call it Lorenz attractor states right. And, but in either case, you know, you only have to look at Lorenz attractor, the whole point is you don't know where you are in it, right. And, and not in the same way as quantum. Well, that's a question of whether you when you get into the fine details, we don't know yet, but it's still different to quantum, right. And I think that's where that quantum contextuality comes in as well. So, anyway, what's your thoughts on that? The thing that just came to mind and then blue is like, this is a linear, right, it's open ended. But sociologists and philosophers make important insights about the philosophy and history of science. And so the mathematician can say, yeah, we're the most pure or we're the most correct or whatever it happens to be. But they're nested in a sociological context in a historical and a philosophical context. So it is closed from a causal closure perspective, because it's not like physics have just broken out of the frameworks of thought. So how to move this line into a conversation that's including all of these disciplines in a trans discipline and including people who aren't in academia or research in that as well. So this is an interesting comic. And I usually like don't even talk about psychologists or sociologists because that's not science. Right. I'm just kidding. I really actually have a lot of respect for psychology and sociology. But it's just not something that's ever been like lumped into my pyramid before. So thank you for sharing that and seeing it in this way like maybe they should arrange by like the fundamentality of it like sociology without like social interaction, we would not develop things like mathematics right like there would be no need. So in terms of what's more fundamental like maybe sociology is fundamental and psychology like we're not even thinking about what makes us work. If we're not psychologically. Okay, right like if we're in a bad psychological state nobody's thinking about interest cells and DNA like that's there there's just no you have to have like the fundamental needs have to be met before you start to ponder things like what am I made of and how do my parts work right and and really like so so maybe in terms of mentality and then we are biology and and then you know so maybe in terms of the fundamental things sociology is the pinnacle of that. And again there's probably many ways in which different perspectives or active inference can challenge this or at least kind of continue this joke a little bit, but it's like pure active inference. It's really about sociological context. We really have a mathematical formalism. It's really grounded in the ways that people have been studying physics. That's one of the questions of 31 the paper is to what extent is it has it could it been grounded in physics, but we really have physics, and we really are talking about biology, and for comprehensiveness and and getting to the fundamentals, it is really fundamentally relational. So we're not taking this reductionist purity spiral. We're taking like a transdisciplinarity purity. And we're just saying this is our keyword. And to really get up what we're talking about we're really going to respect all of these perspectives. So that's kind of a funny challenge that you wouldn't necessarily hear arising from a disciplinary conversation. So Stephen go for it. Yeah, I think she's also one thing, you know, we say, though, that we see the importance of social interaction social inference and the challenges that I think actually this might be quite apt for where we had at the beginning with Maxwell's conversation is the perspective of the sociologists becomes what is ascribed to social interactions. You know, the psychologist is what's ascribed in the mental models that we have in our head, the biologists are ascribed in the categories, the chemist is ascribed in the steady state, you know, the. So when you get into this agent based modeling, which again I think is is more familiar to the newer work in biology, but it's something that I had to start to learn about, you know, and looking at these dynamics and modeling actually. It starts to go under the ground of all of these. It doesn't make them invalid because they are useful when you can take a perspective on the systemic work. But we're not taking perspectives in this and the whole ground starts to become shifting and that's really interesting. And here's another kind of a subtlety actually. So sociologists. Okay, just starting there. That's our prior, you know, D coming in from the left. Sociology is applied psychology. Okay, so we're going all the way, you know, chemistry is applied physics. It's just like it's interesting. The conversation is going that way. I'm just seeing another comic where it could have been like, physics is applied math. Chemistry is applied physics like who is saying which field is an application of which other field. So there's a lot of ways that conversation could play out. And by imagining the counterfactuals, the semantic, the rhetorical counterfactuals, and then asking, well, what is concordant or disconcordant with our generative model and why. That's a very rich starting point for transdisciplinary discussions because the whole inter meta transdisciplinary discussion. It's been decades, centuries, truly with different words. So how are we going to look at this with a new light and maybe even integrate some of the truly recent developments in tools like performance computing and interactive notebooks and ontologies. These didn't exist one or 200 years ago in the same way they do today. So how are we going to respect qualitative and quantitative and different backgrounds recognizing that we're never going to get the end of the story. Even if you got 100 PhDs in any one of these fields or lived 100 years in a field outside of academia, you still wouldn't be at the end. So how are we going to take a radically different approach to learning and applying maybe active inference could have something to do with it. So in the last few minutes, anyone else in the live chat could just ask a final question. Otherwise, let's just sort of recap the big aims and claims of this paper were about the foundations in statistical physics and dynamical systems theory for the free energy principle. And what's on the line? What's on the table for whether the theoretical underpinnings are adequate or not for FEP and ACTIMF? Well, it's the practical utility as well as the theoretical adequacy. So that's what matters. They claim that their work is starting to fill in this gap. And whether we think that there's any any other pebbles in that gap or the gap is, you know, a meter across or a kilometer across, there is a differential and we can all positively contribute to that through discourse and through contributions to the literature. And then the main argument of the paper and one of the big ideas that will be really cool to talk about is that the foundations and physics allow the FEP models to be maximally general. To describe across systems. But in the other words, these assumptions also decrease the ability of FEP models to include enough factors to provide biologically plausible representations for living systems and the way that they are in the world. So, you know, big idea, physical and physics grounding of FEP and what that says about the theoretical adequacy as well as the practical utility. And then this idea of generalization. And if we over generalize, do we sort of leave behind some of the features that we wanted to explain in the first place. So that's a little bit of the recap. And I'll just pull to our kind of closing slide and ask these questions and then Stephen and blue, you can give a last thought. What would a good understanding enable? What are the unique predictions and the implications of this research or of what you thought of after learning about this research? What are the next steps for FEP and active inference? What are the goals of this or your research? And then what are you still curious about? What is your intrinsic motivation that's going to keep you reading these papers and contributing to the discussion? So, Stephen or blue, any last thoughts? Stephen first and then blue. I think in terms of this paper anyway, I think that this really does highlight this need to understand the difference between particles, voxels, things, systems. I think this, because these types of critiques keep coming up right at times, I know it has a risk of kind of tying the free energy principle community in some knots. And so I think that it's a challenge. And I think that that that may be part of where this goes. And at the same time, I think it's a good provocation. And there is some questions here about how and why those confusions arise. And so I think that might be, you know, I'm curious about learning more about that because I think that's going to help in other ways in doing this work. Thank you, Stephen blue. So it's awesome to me, I think that I'm talking about the general general generalizability sorry that didn't come out well, the generalizability and applicability of like the free energy principle. I really like that that's also like a trade off between like epistemic and pragmatic value. Like it reminds me of that balance that we see in the FEP and so I think like the ultimate goal is to resolve that. Like how do we strike the balance between epistemic and pragmatic value? How do we strike that balance between generalization and application? Great point. We want to describe many systems and describe them well and walk in with a scaffold, but minimal constraining priors that prevent us from seeing some fundamental or even essential attributes of systems. And then that general and specific connecting that to learning and doing to epistemic and pragmatic values or gains, which hint hint we have included in the act in formalism as part of our policy selection approach. So Stephen and blue, plus everybody who asked truly excellent questions in the live stream will see you in an organizational unit meeting or the discord or some other way just this is your affordance to participate and contribute. So thanks again to everybody and we'll see you next time. Bye.