 Hello and welcome everyone to Act In Flab. This is Act In Flab live stream number 20.2. This is our second group discussion on the paper, The Emperor's New Markov Blankets. And today's April 27th, 2021. Welcome to Act In Flab, everyone. We are a participatory online lab that is communicating, learning, and practicing applied active inference. You can find us at our links on this slide. This is recorded and an archived live stream. So please provide us with feedback so that we can improve on our work. All backgrounds and perspectives are welcome. And we'll be following video etiquette for group live streams like muting if we're not gonna be talking and raising our hands so that we can hear from everyone. We're down here on the bottom of the slide in the second discussion for The Emperor's New Markov Blankets and closing out April with this really great discussion. So check out this link, rb.gy slash kvnpyc if you wanna see what's upcoming for the different streams. The goal of today is really just to continue this awesome exploration and learning and discussion that we've been having and really appreciative that the authors are here to be talking about it with us. So we'll be able to ask some cool questions and anyone who's watching, feel free to ask a question in the live chat and we'll relay it to the group. Today we're just gonna be going wherever we need to go. We'll start with introductions and then if people wanna call out a specific figure or a specific question or especially reference a quotation in the text or just something that they've prepared that would be really great. So we'll just go around and introduce everyone and say hello, feel free to just introduce yourself however you'd like and give a short check in. And then after we have an introduction round we can return to these warm up questions and anyone can feel free to chime in on one of these questions such as what they're excited about. So I'm Daniel, I'm in California and I will pass it to Shannon. Hey everyone, I'm Shannon. I'm usually in California but I'm in South Dakota for the pandemic and I will pass it to Steven. Hello, I'm Steven, I'm in Toronto and I'm doing a practice-based PhD through the Solomon School of Applied Psychology and I will pass it to Scott. Hello, Scott David from Seattle, Washington and at the University of Washington Applied Physics Lab and we're working on synthetic intelligence and I will pass it to, wait, I lost my list. Who hasn't chatted yet? Dean, have you had a chance? Hi, I'm Dean, I'm from Calgary. I'm retired and I just like hanging out here because I learned a few things and I'll pass it to Chris. Hi, I'm Chris. I'm a postdoc at the Rural Universität in Bochum in Germany but I'm joining from Berlin today and I'm one of the authors on the paper and I'll pass it to Joe. Thanks, Chris. I'm Joe Duhurst. I'm a postdoc at the Munich Center for Mathematical Philosophy, also one of the authors on the paper. My background's in philosophy of neuroscience and computer science primarily, I'd say. Thanks. I'll pass to Yeller. Hi, everyone, happy to be here again. I'm Yeller, also one of the authors of the paper. I am currently based in Amsterdam although I'm supposed to be in Sydney, Australia. Yeah, happy to be here today. Today's a birthday of the Dutch King, so it's a special day. Happy birthday. Okay, I think that's everyone with the introduction so anyone, please feel free to raise your hand and start off. Some of the warm up questions are what is something that you're excited about today? Could be related to the paper or tangentially. What's something that you liked or remembered about the paper or from our discussion last week? And what is something that you're wondering about or you'd like to have resolved by the end of today's discussion? Maybe another question would be what kind of momentum or what kind of next steps would make sense? Emerging from this, whatever we end up resolving or not in the next conversation, where will that take us in our research or in our practice journeys, I guess? So, Joe, and then anyone who'd like to speak. Ah, thanks. Do I need to lower my hand again or is that done? It's just good. Good, great. Cool, so I missed the discussion last week so some of this may have already been covered but one thing which relates a little bit to something we were discussing just before we went live here and I think it's kind of a broadening of the points we wanted to make in the paper which I wonder about is what the, if active inference can be kind of arbitrarily applied to any given case, given the right kind of set up and parameters and so on, what does this mean for its future application? Is this, so far as it's too general or is this a strength and so far as it can be used anywhere really? And this, I think, is kind of lurking underneath some of the points we make in the paper. That's a more general point. Cool, so can active inference apply to anything and everything? Would that be a strength or would that be a weakness? Awesome question. Any other thoughts? Anyone wants to raise their hand? Yeah, Stephen. Yeah, I'd be interested to learn more about the way that the pearl blankets and the frist and blankets interpretations might be useful in applications and how that might inform the way that modeling's done. So that's kind of exciting for me. So that's one thing that I'm looking forward to. Great question as well. The distinguishing different kinds of blankets and even pulling back to like, what kind of metaphors will we use and how will we relate technical terms to different settings, different practices? Those are some big questions we wanna address, Scott. And then anyone else? Yeah, I'm interested in both of those questions or all that aspects. And then also the frizzle of the paradox of those two questions. So it's a chicken and egg question of frist and not frist and of active inference in a sense, right? We're observing something, we're projecting it out as a teleology onto the world and then asking if it's a good teleology, right? So in a way, if you go for a David Bohm idea that all reality is first a thought, that includes the realities that we don't create entirely but that we project. So one of the questions I have, so I was a practicing lawyer for 30 years or 25 years-ish. And when I see active inference, I think contracts. And so I think of it as biomimicry of what we can do socially and project further with real intention to control systems. So what I'm wondering is, and I think both are valid, but so ultimately is it a notion, I'll use the word biomimicry but obviously I don't mean in the physical sense but from us to learn from deep time, what we can do as biological organisms to better organize ourselves despite our social elements. So to get back to our biology in a way and maybe organize ourselves a bit better and more resiliently with reference to that. Anyway, so that's the frizzle I'm interested in is the both the cause being upward and downward with regard to active inference. Thank you. Cool. Connecting the micro and the macro, it's sort of sometimes it's laid out abstractly like multiple spatial scales that are nested within each other or time scales. But then when we talk about action and we have the example of sort of the person in a maze and you take a step back so that you can go around that's the kind of settings where we might want to apply active inference where the initial action might be non-linear or away from the goal in a sense but it's part of a policy trajectory that actually gets you where you want to be. So And the other part of it is with the Markov blankets being a Bayesian notion because of the Bayesian then you say, oh, when does Bayes effective? Well, that becomes a pretty broad application because I always say to people to get them in this space I always say, if you're in a dark forest with no flashlight, you don't run full speed ahead. You stick out your hand in front of your face and that's basically active inference. As I say, you want to know what's in front of you before you run into a tree. So anyway, good stuff. Thank you. Any other? Yep, Jelle and then anyone else? Yeah, so I think last week we ended the discussion on Shannon's point on what it means to have a folk folk psychological form of active inference. And I think it related to the tension between needing to have a very deep mathematical understanding of an increasingly kind of developing field of maths and applying these kind of more broad set of ideas to a particular kind of case. So I thought this was a point to explore further and think about how this would look like this more folk psychological way of understanding active inference using the concepts maybe without the nitty gritty details of mathematics. Shannon? Yeah, I was hoping that we could bring that up today and maybe ask Joe's take. And also sort of if we're talking about something like a first in blanket and that's supposed to be a real demarcation of a system or a pearl blanket, which is supposed to just be how we describe the system, then maybe in this, you know, folk psychological notion of active inference or talking about how a phenomenon appears to this person in a crowd versus another person in a crowd where both phenomenons are real, but depending on which one you decide to model, you'll get a different picture of how the crowds dynamics are changing. And we could model that mathematically, but just as easily someone in the crowd could have an informed notion of what it is that they're receiving and how it seems that the Markov blankets, however they are, are emerging around the crowds in certain ways. And you could start having a discussion that maybe you can translate from the scientists doing the fancy math and the person in the crowd who's just existing. Thanks, very interesting. Steven, then anyone else? Yeah, just building on that, I think I'm also curious because there seems to be, there's two sides. There seems to be this primordial soup which is referred to in the paper as well and the variational free energy and how that can self-organize. And then, and that happens at smaller distances and then these longer distances where you've got expected free energy because the organism is seen into the environment with sensory information and often it's using these partially observable Markov decision processes for modeling that. And I'll be just curious if those two domains of modeling which seem to come up, and I know there's more but those two seem to come up a lot in the literature how that pans out maybe with some of these questions of, is there an intuitive feeling of how the morphology is moving towards a target which could be that deeper sort of knowing of the body? And then there's this kind of partially observable decision process which is kind of cognitive sort of looking in the world and which ones are more or less like Markovian in some ways and how does that affect how we try and play with them which is or model them or play with them which I think Shannon was saying in a way is in a good way is exploring these ideas phenomenologically. Thank you, Scott, and then anyone else. So Steven's notion got me thinking about the notion of glass. In glass, they frequently talk about first order, second order, third order order. And so you have that amorphous material that's not semi amorphous solid, right? So it's not really crystalline, not really liquid. So it has these orders that are discernible through different framings of maths. And it made me think about Steven when you were saying that cognitive glass in a sense that we have this orders of order that we project onto disorder. And so those teleologies ultimately roll up into things like sovereignty, which is seen as perfect order in the abstract, right? The sovereign does not need to ask for permission or forgiveness because it is the quintessence, the fifth element, right? It is not earthly. So we have this metaphysics that happens at the bottom or top of the turtle sac depending on which direction you're going. And so I think that's kind of nice here. Hey, Blu, I think that's kind of nice here when we look at order and abstraction. And the last point about it is that, Kandinsky, the artist said that violent societies yield abstract art. And so one of the things I always wondered is, is abstraction the reverse of violence? So if we look at sovereignty as being a total abstraction, as I just described, maybe sovereignty is also total violence. And Mills said that the sovereign has the monopoly of legitimated violence. So anyway, those are a few ideas about, maybe we're looking at cognitive glass here with different orders of order that we can project onto the system. And maybe that helps us when we look at order as levels of abstraction, as you get further from the first and second order into the third and fourth order, maybe it starts to be something that starts to yield a more projective or teleological element. Thank you. Thank you, Scott, Dean, and then anyone else? Yeah, so I wasn't part of the tile formation last week. I was kind of in the background and then lost it, but I did go back and watch the whole thing. And I'm hoping that Chris has a few more things to say today about construal because that to me is kind of the inflection point. And he was starting to talk about some of those things and the separation between what did he have, model and target. I think at the end of the day, if we were to give a layperson or some, I think the term focus been used, I think if we give them a Markov blanket and a Pearl blanket, a lot of people will go, how does that change my basic question, which is how do I get the world? I think that's what all of us here are trying to do. And I think for a lot of people who aren't even aware of what a Markov blanket is or free energy principle is, they still want that basic question answered. So I'm really interested to find out today how the paper moves us to a place where that becomes easier, that the obstacles to getting the answer to, how do I get the world suddenly flatten and become, I don't know, more familiar? Thanks, Dean. And a thought on the cognitive glass and what you just said there. So framing it in big terms about sense-making. Sense-making about the world as well as about active inference is really helpful. And Scott, when you mentioned the cognitive glass, I thought about oblique. It might be called different things, different places, but it's cornstarch and water. And so when you slap it really hard, it's like a resilient boundary and your hand comes away dry and it's like you're hitting a piece of concrete. But then if you go very slowly, it kind of is like a liquid. So it's an amorphous solid like the glass. Now, the boundary between the air and the oblique is something that is real about the world. And so there's a distinction there, but it's almost like depending on how the boundary is crossed, not even like two different types of things. They're both touch, but it's actually the dynamics of the stimuli that dictate whether the response is going to be very stubborn and more like a solid or whether it will be very giving. And again, it's not like it's light versus sound to different sensory inputs. This is different quantitative differences within one interface modality. And it just shows how even when there's a clear boundary in the world, we think like between two phases, still the behavior of the interface can be very dependent on the type of interaction. So Scott, and then anyone else? Made me think, is it Scinarian all the way down? Or do you have myriad Scinarian vectors intersecting and we see that as complex? Steven, then anyone else? I think you could also see this question around glass because glass at a certain way of freezing or becoming solid, it doesn't crystallize, it doesn't break up, you can still see through it. So you've got this question of it being part of the medium through which information can pass because you can see through it. But there's also that question around that goldilocks zone when if you want something to rearrange and be the material like happens in the primordial soup, Kristen talks about there's a, too low a temperature, it sort of just becomes a solid, too high a temperature, it's going to dissipate and start to, so you've got a bit of the overlap between internal, external, solid, liquid, map, territory. It's kind of interesting. Yes, different phases of the mind for sense making, too rigid or too fluid. So Joe, then Scott, and then anyone else who'd like to speak. Thanks, yeah, so this just made me think that quite a few of these points for me at least come back to a kind of central thing. So Shannon's question about folk psychology which I'm very interested in, or kind of folk inference questions about teleology and questions about whether it's just something simple all the way down, which makes it look like it's complex. To me, I think there's an important question here about the role of perspective and kind of viewing of these systems going on. And the way I see it, a lot of the more teleological things that we project onto these systems, even like a boundary, perhaps, isn't really something we should take to be in the system that is rather just something we're kind of modeling the system as having. So it comes back to our target model question as well. So just a central question here of what we take to be really in the world and what we take to just be part of our understanding of the system of the world. Cool, thank you, Scott, and then anyone else? I love the conversation because everyone is moving so nicely and easily with such a facility back and forth between the metaphysics and the physics and appropriately. And I don't mean this is not a diminishment, right? Because we're demarking those edges. One of the things here that occurred to me in the last couple of notes is the notion of seed crystals. Because we are talking about phase changes in system behavior, I think. There's very often a threshold kind of trigger type of thing that happens. And so that notion that Stephen was bringing up about phase seems interesting. And that was implicit in what you were saying about cornstarch as well. And so one of the things I think is interesting is what do those triggers look like? And how are those encoded? And you're talking about popular encoding versus some biological encoding. So let's say I'm triggered into anger, but then that anger takes on the symbol of a certain person in a political party or whatever. There's different stages of phase change are happening there of solidity that are being offered to that glass-like structure of order that are being set up by those externalities. And it seems like one of the things, we just did a paper on misinformation with a group. And one of the things that's so interesting to me is it's not a single event misinformation. It's not a binary where, oh, I got misinformation. How could I have prevented it? Like, you know, getting hit on the head with a stick. There's so many levels at which you can address misinformation and intercede and intervene. And this feels like one of those exercises that perhaps we can start to identify interventions for information integrity by looking at, where might we put in seed crystals to put in phase changes at stages of that process to redirect it in ways that are more socially or productive or whatever the goal is of the group. Thank you. Thank you, Scott, Steven, and then anyone else? Yeah, I think this discussion also, it points to an important consideration bringing in the territory and the environment because a lot of these phase changes in the world out there. But also maybe unlike, say, the enamel of our teeth, which is maybe like glass in a sort of what you would call an equilibrium state. Okay, it's at a low energy state. It's just there, right? I don't expect it to keep moving that much. But we're talking about these non-equilibrium, steady states. So you've got this kind of micro states which things can revisit from which some are more probable from which there's attractor states which becomes these macro states which then the organism at different scales is visiting. So this is interesting because there's both going on. I think sometimes everything's non-equilibrium steady state but I'm told in the computer here, it's pretty thermodynamic equilibrium object. I'm in that world, even if my body has a lot of non-equilibrium dynamics going on. So that maybe plays out in some of these questions. Thanks, Steven. Yep, a lot of parallels, analogies, cross mapping between information and thermodynamics. And it's unsurprising I guess because there's a lot of groundwork there but it really is a fruitful intersection. Maybe we can just remind ourselves of some of the questions that we had from last week and then we can just continue walking along. So anyone just raise a hand. So these were like loosely structured by topic but we had big questions. How much mathematics is required to understand or work with free energy principle and active inference and just how shall we speak and work clearly on this issue? Then we had a few questions on a philosophical side. We recalled the papers of Mel Andrews and Ines Hippolito and Thomas Van S which were a footnote in this paper as sort of relevant contemporary work. So definitely on the philosophical side, I'm sure we could hear any thoughts about what next steps are on that line of research and this question of how to think about blankets for groups. You know, if it's groups all the way down then what's wrong with throwing blankets over groups but what kind of blankets and what kind of groups? Also blue if you want to introduce and give any thoughts that's cool too. And of course questions about what's real or not. We had a few themes arising about again about math and about how there is, there's multi-directional dialogue with active inference and other fields. So what happens when we cross over between different areas? Scott and then anyone else? So I want to talk about Markov flags for a second. And what I mean is the flag of Amsterdam with the three Xs. My understanding is from a tour guide that the Xs stand for fire, plague and flood. And so if that, let's assume that for a minute. And the idea was if you see that flag you better follow the club rules on preventing fire, plague and flood because that's what we do here in this area in Amsterdam. So to me, that is a Markov blanket because what you're doing is you're saying, hey, we are projecting this out. Come join the club. We are doing Bayesian analysis on how to prevent cholera and floods and fire in the city limits. And that's a, the signal is there on top of every building that has an owner who conforms to the rules. And so I wondered if we can do a Markov blanket analysis on the Amsterdam flag as a Markov flag expressing and saying all ye who come here, right? It's like Diogenes looking for an honest man. I'm looking for plague preventing and flood preventing people here. So to me that's a projection out in any other kind of active inference. We're projecting out through a flag instead of sticking our hand out to search for a tree. So that's active. And then I'm doing regulatory enforcement, which seems to me to be a body that is responding to non-conformity among its members. So could you comment on whether that's a possibility in terms of projection out of social elements of Markov blanket possible? Any Amsterdam residents is welcome to respond. Shannon, and then anyone else who raised their hand. A lot of Amsterdam residents, but yeah. So now we're talking about something like cumulative culture and our evolution of signs and symbols and how we co-construct meaning in space. So maybe the flag, if you don't know what it means, it's just the flag. You might not even know it's a flag. If you see it painted somewhere, it's just something that looks cool and has some shapes. But if you have been raised or taken a tour, I guess, in the culture surrounded by the symbols, surrounded by the meaning, all of those small instances of negotiating meaning and like we talked about active semantic inference. I can't exactly remember what that conversation was last week, but this seems like it could be similar sort of adding meaning. And if the flag doesn't mean to the person you're interacting with, what you thought it would mean to them, you take action in your world to make it mean the thing to them that it means to you. It's an interesting thread. I like that idea. Also, one thought on the flag before we go to an Amsterdam resident is like a blanket you think about under the blanket or inside the blanket, like there's a person or a blanket goes on a bed. But a flag is also fabric and it's flying in the wind and the borders it delineates are not on one side of the air versus the other side of the air, but as the flag is animated by the niche through winds, otherwise it's just on the flagpole, it delineates a social barrier or a social demarcation which is an encultured and an intersubjectively negotiated experience just like you brought up Shannon. So Yela and then Stephen. All right, thanks. It's fascinating. I'm not sure, I know the story yet from the three crosses. I think I'm different origin story, but maybe this is in the tour guide, so I'm not quite sure. I think this raises also points to one of the questions Joe was raising in the beginning, right? Which is in turn like, suppose we can give this analysis of the flag of Amsterdam as a Markov blanket and we can give an analysis of the origin of life in terms of Markov blankets and lots of things in between, right? So what does this multi applicability of Markov blanket say? It could, one is good point to a very deep self similarity that runs through nature at all possible scales or it could mean that the toolbox is in a sense to course grained to so that it applies to everything but you can't really make the relevant differences, right? I mean, in a sense you would like to think through what are the differences indeed through a kind of signs of group memberships in social situations versus more biological signs versus kind of self organization in physics. I think that's kind of the tension between the kind of applicability to everything and how particularly you would like your theory to be. Thanks, so Steven, then Scott, then anyone else? I mean, this is also brings that question, I think like we just mentioned there around where we put things in a Markov blanket and where things are part of this external states, hidden states, internal states because in some ways from the Markov blanket perspective the sensory states of the epiphyllia that the kind of sensory organs is what's been stimulated and it's inferring an external state in the world, i.e. the flag and that's having an additional inference that the flag has got this meaning at another level. So I'm not even sure what the whole thing's called in some ways because the blanket is dependent on connecting with external and internal but it's normally just that part of the whole active inference dynamic. So I think that becomes interesting is where what type of sensory input information is within the blanket biologically and maybe that isn't quite the same as what is thought of when we start talking about it and that may be good or maybe a problem. Thanks, so Scott, then Chris, then anyone else? This is fantastic, this discussion. So I realized all of a sudden, I thought, oh, wait a second. I just realized that some of these folks may actually be empiricists and may believe there is a testable reality out there. I didn't, because I'm not that person. So I made a living rhetoric and a persuasive speech as a lawyer. So for me, it's all projection. So I realized I was too diminishing of the question of whether there exists something out there to test. And I was looking at, and this whole thing, to me, this has been a candy store since a few months ago where I learned about active inference because to me, it's all about avoiding avoidable harms among humans. And to the extent that active inference is universally assertible, then it has a rhetorical power that's huge, right? So if we can get people thinking this surrounded by active inference blankets, and then I say, you know, you don't have an active inference blanket yet in the way you're doing business with other people in the supply chain or you don't have active inference blanket yet in the way that you're interacting among nation states, or you don't have active inference yet in your food insecurity supply chain or your cybersecurity, all these other things that I'm working on right now, contact tracing, blah, blah, blah, blah. It's information systems where because they're not within the same model, then they have conflict. And most of those are avoidable harms because they're just humans working out stuff that we were separated for 100,000 years when we migrated out of Africa. And now we're getting back together and trying to compare notes on what we came up with. So the question or this assertion is that it's interesting, to me, it's not that troubling a question of whether these exist or not. And I've read lots of philosophy and I'm down the road there. But ultimately, I say yes and on that question of whether it exists or not, to me, it's a provisional truth that's sufficiently functional. So I'm not no longer curious about whether it has an existence separate from the human mind. So take my comments with a salt lick, not a grain of salt, but a giant block of salt because I've realized I'm an old person and I just don't care if it actually exists separately. And I'm not saying that, to say this conversation doesn't matter, it's awesome. But the dynamics of rhetoric are such that we can employ the thing to do good whether or not it's real. So Scott has litigated realism versus instrumentalism. He's serving as the judge, jury, executioner, lawyer. Chris and then Joe, then Blue. Yeah, thanks a lot. So I think there's a useful way to think of the flag metaphor, so to speak, and to bring it a little bit closer, maybe to also what Scott was saying now and the paper itself. So the question that the paper is trying to answer, right, is kind of to what extent is the Friston blanket construct or the Markov blanket construct as used by Friston a conventional thing? So, you know, there are different kinds of, like you can have different views, of course, about the metaphysics of mathematics and so on and so forth, but generally there's some agreed rules about how maths works or certain branches of mathematics work. And that's where the Markov blanket construct comes from, right, it comes from statistics. And there's a clearly defined way of using it. So the convention for using the Markov blanket is very clear. It was introduced very clearly by Judea Pearl in his article and later his book at the break of 80s and 90s. And there's no kind of mystery or, you know, there's no question about how it should be applied productively. So whether or not you think that science is after capturing the reality as it is, or whether you think science only gives us approximate truths or some kind of, you know, approximate models of what's out there, that doesn't matter. What's really important for us, I think, and maybe the guys can also speak to that, is that for Markov blanket, there's a clear convention for how to deploy them. And it's very clear what they are doing and what they can't give you. Whereas for the way Friston applies the notion, first of all, he kind of softens the convention. The convention is less codified, we can say. It's less clear to what extent they apply where. And similarly, it's less clear what you can get out of it, right? So this is kind of the problem. So Stephen, for example, mentioned something about, you know, boundaries and the fact that whenever there's an interaction, these Markov, sorry, these blankets, let's leave the Markov part of the name out of it, these blankets appear and or they can be found there. And it's not entirely clear whether, according to Friston, there will be found at every, in every place where we have an interaction between two systems, for example, or something like that, right? Because the rules in which they are applied are not as clear as in the case of the original Markov blanket construct. And I just wanted to also to follow up on this very quickly, in the version of the paper that's under review now, we actually realized that perhaps in the draft you've read, we were a little bit too harsh on Friston because he does impose certain rules about, for example, how to individuate different kinds of nodes, but there's still a question of kind of what it means to individuate these nodes in a certain way. Thank you. Joe, if you wanted, then Blue and Scott. I'm fine, actually, thanks. Thank you, Blue and then Scott. So without specifying Markov blanket or Pearl blanket or Friston blanket, I want to touch back on what Yellow was saying about group memberships and also about what Shannon was saying about cumulative culture. So I think this is like a really interesting notion of like a Markov blanket or a blanket overlying a group and the affordances and like what it means to be part of that group, like age specific affordances or socioeconomic specific affordances under a given blanket and how and when this like cultural or group membership prior gets like instilled and also like do blankets then overlap? Like I'm of this race and of this age so these overlapping blankets, I just think it's a very interesting idea about group membership and cumulative culture and we might know that we're part of this group but we don't know all of the specific underlying back story about why we're treated this way or we're afforded these certain things. I just think that's a very interesting way to think about it. Thank you, Blue, Scott and then anyone else and also anyone who's watching lives. Welcome to ask us a question. So it's so interesting, a couple of things, Blue, on what you just said. So I do a lot of sewing and with the notion of quilting is really interesting in what you just said because they're actually a bunch of quilting museums that do ethnographies of quilts. And so it really comes back around, you had this quilt which reflects a person's history like when they did the HIV quilt each square representative person, right? So you actually have a physical membership instantiated in this blanket and then the blanket being the symbol of the community being knit together. I mean, all these knits probably is not the thing you wanna do in a quilt but anyway. Okay, the thing I was gonna mention in response to Joe, I think it was Joe's comment before about the, oh no, sorry, it was Christos about the application and the alteration from the paper that notion that you have this something that's running as a scientific and statistical concept that is brought out in the world. One of the things who, as I mentioned before, I work in an applied physics lab. Now applied physics is really funny because you're applying physics. So that seems like maybe it has something to say about what's going on here. And I realized, well, what you do in applied physics and what we do in our program, specifically in our program, we take technically feasible systems. So let's say that we could describe something that is technically feasible, the Friston notion is technically feasible and you can apply it in scientific context. But what I do in my program and what the lab does generally is then we take the stuff and make it ready for adoption in the field. So specifically in my program, I take technically feasible systems and I test them for whether they're bolts reasonable and bolts is business, operating, legal, technical and social. So if something is technically feasible like the Orion project where they were gonna use nuclear bombs to loft weighty objects into space, that's totally technically feasible, but it would be unreasonable socially because you'd pollute the entire atmosphere with radioactive isotopes, right? So it works in the lab, it's theoretically good, terrific. So what my notion is maybe what's going on here is the applied physics, the applied statistics. And when you get it out there in the field, it's can be seen, we can call them compromises or we can call them realities or we can call them affordances or we can call them Bayesian adjustments of scientific empirical views, right? Which when it gets out there in the reality, maybe the folk understanding is the reality of physics in the mind of humans. Maybe if the mind exists in society and the brain's just an antenna tuned into it. Anyway, just a couple of thoughts to play with. Thanks, Scott. So Stephen and then anyone else who has a question? Yeah, I'd be interested, Christoph, just further in what you said about what does it mean to individuate a node in a particular way in this idea of nodes? And I suppose at this larger scale, I'm seeing with these partially observable Markov decision processes being used in some of the modeling. They don't necessarily from what I understand talk about nodes in that way, but at the primordial soup kind of cellular level, they seem to often talk about nodes in that, maybe in the similar way or not, but I'd just be curious what your thoughts are about what it means to have a node, what kind of node that might be and what's similar or different between different ways that this is implemented. Yep, go for it, Chris. And I have figure two up or let me know if you want to go to a different figure. Sure, thanks a lot. I mean, those two, the last two comments and the questions even really matched nicely together, I think, because so, you know, I mean, we can talk about variables, right? I mean, in a graphical model, variables are represented as nodes basically. So, but what I of course meant is the differentiation of the differentiation between active states and passive or perceptual states, inner states. Yeah, exactly. Well, you can also use the loop, but yeah, internal states, external states, these things, right? So it seems to me that, and I think, I mean, not only to me, it seems to all of us that there's a certain degree of arbitrariness with regards to what is what. I mean, at least in the draft you've read, it seemed like that. To us, we then learned that after really reading a lot of Friston's papers very carefully, we actually kind of extracted the implicit rules for how to label these variables as active or passive in a graphical model. But this is still a very arbitrary distinction, one which is like the rules by which you individuate. So I used individuation here. Maybe I shouldn't have used the word individuate. The way you identify the role played by a given variable, that's what I should have said previously, is the rules that give you this identification are still very arbitrary and they are not actually anchored in anything very substantial. And this goes back to what Scott said about feasibility and applicability, because of course you can take this kind of way of partitioning a system. You can build a robot with some kind of AI system which will be built on the principles of the free energy principle and Friston's way of thinking about graphical models. And it will be perfectly good at doing certain things that you build it for. The problem is that when you're building a system for a particular purpose as an engineer, you kind of hold all the cards. You make all the decisions. You decide like how to, what kind of active states will the robot need or something like that. And that's a very different perspective, so to speak, from this perspective of, let's call it, what then it calls the evolutionary R&D department, right? So evolution doesn't really plan things. Evolution just throws things at the wall and whatever stick kind of stays and goes on to reproduce and all the other things kind of fall back. And in this sense, it becomes very problematic when we think of like Friston's grand project of explaining life as we know it, for example, as in the paper from which we took this diagram that's on the screen now, because it slowly becomes a little bit suspicious whether we can actually, whether the way he labels these nodes and the kind of the rules for labeling the nodes will be the same on, for example, different levels of organization, stuff like that, right? So whether this will be applicable to many different biological systems in the way that he hopes it will be or not. Thanks. And just to clarify, you all wrote that a spectral graph theory is used to identify the eight most densely coupled nodes which are defined as the internal states. So for those who think that given this soup of interacting particles, which are all the same kind, they're all following the same simple rules, yeah, eight densely coupled ones. Okay, those are internal. And then there's some downstream consequences with which nodes influence which nodes, but it's hard to argue that choosing the eight densely coupled nodes is simply, you know, true or internal when you can look at what it is and ask whether there's a different way to partition it. So Joe, and then Scott, and then anyone else. Yeah, thanks, Daniel. That's a very clear way of putting that point. So just to pick up again on something Chris mentioned briefly, there is a further individuation problem, I think, which is about how to identify which nodes you have in the first place or which variables you have. And this is a kind of grain question. So you can always, you know, look inside a node and split up into some other nodes or some other subparts. And this might not be a problem just that active inference suffers from. It's going to be any kind of modeling system is going to have this kind of issue, but it is going to make a difference to where the boundaries are drawn and what kinds of things you get. So that's just to say that there's already a kind of human involvement in the model and just choosing what scale you pitch your nodes at. Thanks, Joe, Scott, and then Dean, Gela, and Steven. So in that notion, love it. This is great. And that clarification about the nodes being assigned, in a sense, contextually given system state, it seems like if we start to, that may be a pathway to revealing the self other, or starting to unpack some self other challenges. So maybe a node that is once characterizable as internal, then characterizable as external is revealing something about the paradox of membership in groups. That the, in fact, it's not the node that changes, but the sensation of whether it's self or other. So you start to take that always paradoxal, is that Hofstetter stuff and everyone else who talks about when does a group become a, when does a group form, when does it have an internal voice, when does it speak an extra voice, all that stuff. And so maybe that analysis and that reassignment possibility starts to get into that. And maybe that also informs Blue's notion before, but the group, again, I'm calling it now the Markov quilt as of that last conversation, but that group identity blanket, where it's patched together, maybe those things are like Irving Goffman and things like that or mirror neurons, where you want something to be recruited by others as an internal state, you project it to them as an external state intrinsically. And so the communication, you're inviting others to encode their Markov blankets similar to yours, in a sense, or something like that. So that is exciting to me, that slipperiness of assignment, there feels like something that offers a pathway to analyzing those dynamics of group and individual identity. Thank you. Thank you, Scott. So Dean Yella and Steven and Blue. Yeah. So back to Construal and the question of how do I get the world? I think a lot of conversation today has centered on maybe a more clunkier expression, which is how do us get the world? Either us as flag conformists or us as red dots versus cyan dots. So even Pristin, Steven shared a video where he was talking to the Senate of the Institute. Even Pristin said that things get a lot more interesting when we get to the combinatorial level and they're not the same as when we're talking about how do I get the world? That there is a difference when complexity arises. So can we hold up Pearl and Markov blankets and get past the ore of this or that because we recognize that it's a little bit different when how do us get the world? That's a Construal question. And so that's what I'm kind of curious about. Thanks, Dean. Really powerful and a big difference between how do us get the world? How do we get the world? How do I get the world? How do they get the world? These are like really markedly different questions. So Yellow, Steven, Blue. So I was thinking of getting back to Blue's question and Shannon's questions on group membership, but maybe if somebody else wants to... Well, I just get talking on group membership. So, I mean, to what extent one, the kind of Markov blanket formalism applies to groups. I think it's an interesting one. And I think that's what... One thing that really matters here is really how you conceive of a group, right? I mean, we are a group of people that is spread out all over the world and you could think like a group in the more narrow sense, like a flock of birds or a school of fish. You say school of fish, I think so. And these are entities in the same kind of spatial temporal location that coordinate their behavior with one another and you can think as a predator arriving or so as a perturbation for that system as a whole. So in a way that entity is coordinating its dynamics with the rest of the environment, you might actually think, you know, this is a group that actually too weak, the active inference principles do apply. And you might in that sense might think as well that a group as a whole somehow has a Markov blanket. I think that analysis is more difficult to give if you think about group memberships. Also kind of do the graded group memberships, right? I mean, I think Blue was talking about we belong to multiple groups at the same time and not actively acting on being part of that group but just as a kind of label that I would endorse when you ask me. And I think there the application of Markov blankets is a lot more difficult and most more metaphorical. So I just want to say, so exactly how to understand groups there is important. If Shannon wants to understand kind of crowd dynamics or so that might be more applicable and if you want to really understand kind of social membership and graded membership and membership of different groups at the same time, then, yeah, I find it more difficult to understand how kind of the active inference framework can be applied to those kind of groups. Thanks. So Steven, Blue, Scott. Yeah, I think sort of sticking with that group idea that there could be the idea of morphology and the amount of constraints and affordances in the whole environment body organism dynamic that's constraining and where maybe there's not that constraint. So I'm thinking I know someone that she's trying to do some Markov blanket meditation processes and kind of thinking someone's meditating and starting to just sense their states in a maybe a way that slowly is like a cloud that self-organizes in a way. I could see a more plausible way where sort of something like not obviously anything the same as this kind of Markov blanket in a soup could be emergent. However, when you're in a more real cognitive environment where you've got these priors and the only way we're able to access it anyway is through self-reported data which a much higher level of dimensionality. It maybe is only in some of the temporal processes that we can get some analogies like but I think that it may be very different at different scales. So the way that constraints play out I'm wondering how we think about constraints and not having constraints. Nice question, Steven. Are the constraints hooked into this phrasing or where are the constraints and how does that shape how we model things across different scales. So blue and then Scott and then anyone else. So hopefully I won't go on too long but I want to touch back on what Steven just said and also what Yela said. When I think about how group membership goes into active inference like how can we relate these two things? I think about like specifically what is the likelihood that I'm going to take a ride on a yacht today? So my socioeconomic status doesn't really afford me the opportunity to ride around in yachts very frequently. So I'd be very surprised and I also live in the desert. I'd be very surprised if I was riding around on a yacht at three o'clock this afternoon like that just doesn't go into my model. So perhaps our group membership goes into construction of the prior and our model in ways that are maybe layered, right? Like an onion maybe. And so then I want to touch back on what Scott said because this is what really made me raise my hand about group alignment. And we were starting to talk with Kasper Hesp when he was on the stream about inferring the states of others. So I think this group alignment, I wonder about this a lot and specifically with respect to communication which we've talked about in relation to active inference and cooperation and building a collective intelligence. So when we think about whether the earth is flat or not, right? Like this at one time was like a hotly debated topic and maybe still is, but this is like, there's been a convergence, right? So we all now align or like most of the people in the world agree that the earth is round. So I wonder about this time to convergence and this has been used to measure information transmission and can we similarly use it to estimate collective intelligence? So like when we there's a time to convergence for a swarm or this might also represent like a phase transition, right? So there's a time to convergence and if it happens rapidly, like the idea is readily adopted that now the earth is round. So I wonder if these divergent views that we have as people in groups and divergent groups in a society, do they hinder our collective intelligence or do they like actually assist it? Because without having divergence, we could never actually come to the one true answer. So is it resistance that hinders the evolution of intelligence or is it something to think about? Thanks, Blue. Scott, if you're... Go ahead, Shannon. I just repeat the question you asked right there at the end. Yeah, I repeat it. Yeah, so I just wonder about our... the divergent views that we hold in society about whether the earth is flat or round. Does this divergence help or hinder our collective intelligence and the evolution of intelligence overall or is it resisting adoption of a new view? So perhaps there's resistance to change. Even when we see something is correct, like climate change, we can clearly see the effects of climate change but there's a huge resistance to changing. So is it this resistance that hinders our collective intelligence? Where is the evolution of our collective intelligence helped or hindered by divergence and views or resistance to adopting new views? Sorry, it was long. Yeah, Shannon. You want to give a thought on that at the end, Scott? Yeah. So if we were going to... This would be like whether a system is resisting its non-equilibrium steady state and like moving to a different steady state. And I guess that's a question I don't... Like I've been seeing pop up in a lot of the papers recently is how can you define not every system is going towards like a minimized free energy gradient or not every system is trying to maintain a stable state like systems dissipate and re-emerge and dissipate and re-emerge. So I don't know, is that just a way to re-ask that question but talk about something a bit more physical than collective intelligence. And I guess collective intelligence maybe is like collective agreement on a certain set of ideas or norms or something rather than like intelligence like smarter than another person. That's what we're maybe talking about here. So how does that system arrive on this steady state of what we think that the world is? Thanks, Shannon. And there's the convergence on shared generative model on shared mindset, but then it made me think of ant colony relocation and over evolutionary time colonies are shaped to appropriately relocate but not instantly and not just choose the first, you know, first day at coming back with exciting nest location. You can't go for the first one. So how to actually bear with the process of distributed computation or distributed intelligence knowing that it looks like disagreements at multiple levels, especially when it's not just as simple as ants relocating but cultural values and beliefs and everything like that. Can I jump in again about ants? Yeah, yeah. Oh, yes. There was a new paper and I don't know who was by it. I saw it on Twitter, but it was about ants. They travel across a scaffold and it's either it's like flat and so they just walk where they tilt it. And so some ants start to fall down and start to build the scaffold. And then when they build the scaffold, ants can all walk over it. So they're supporting each other. And I'm thinking I might be diverging from our mark of like it topic here, but I'm thinking that collective intelligence is like as long as the world looks flat, as long as the world is easy for us, you know, to navigate, then we can all go about our day. We don't really need to try to converge with anybody, but as soon as the world starts to be disrupted slightly, so we turn that scaffold, folks start to fall off. Like then we need to build something, build a shared collective notion of reality. And then the convergence matters. And so that could be something like as soon as there's a perturbance to the system, then it's useful to try to recognize boundaries or try to have some sort of intervention. I think Scott was mentioned to have a boundary that's suitable and helpful and productive. And that could be where it's fine to diverge all you want as long as everything's simple. And then as soon as it gets complicated or we start to like have harm for people, then we need to have a system that can result in some sort of convergence. Happy for someone to bring that back on topic. It's a great point. And I almost interpreted as how do we help those at the fringe who might be being harmed? How do we make that the stimuli and for realignment and for rescaffolding rather than keep on cutting off more and more of the failing parts of the bridge? They got to go. They got to go. Well, now we have no one on the bridge instead of seeing that those on the fringe and can be part of the recovery. So Scott, Christof and Steven. So Shannon, that was the right in the middle of the discussion. You don't have to bring back anything. That was perfect. So Admiral Rogers, the formal head of the NSA, once said don't let any emergency or disaster go unexploited. And the idea is that the disaster is that situation. People act in a disaster, you know, a hurricane, Sandy, whatever people, you know, Starbucks, they were putting out power strips for people to share, whatever, right? So it's exactly that. And the rhetoric, the projection rhetoric can cause a state change, right? The anticipation of a disaster when people are actually experiencing, they go, oh my God, this might happen to me or neighborhood watch, right? People kind of, you know, they go in and they want to protect property and things. So I love that point. And bringing it with the other points before, I think it was Christof that was asking maybe somebody else about, you know, that active inference in a group and does it dissipate in the activity? If you're just a member of a club and it's kind of lying dormant, is that really describable with active inference? And it occurs to me, you know, if you look at group dynamics, it may still be usefully described that way. And here's how I would think of it. If I, this is Irving Gothman stuff that I project myself out as I want to be perceived by others. So if I join a club, I may join a club because I want to be like, I used to be part of an entomology listserv. I've never been a, this is not about entomones cakes, but insects, you know. But I was part of a listserv back in the old listserv days. And I just was fascinating. I sort of wanted to be involved in it, right? So I was projecting myself out and I started making these connections with folks. So in a way, you can de-risk and leverage by being a member of a group even passively, right? I can, I'm extending out my feelers out there into the world by having that group membership. I'm de-risking by having that group available to me as me being a member, if it's a group that, you know. So, so I wonder if the dissipation of activity itself is an appropriate gauge of whether you're a member of a group or not. Maybe in certain leverages and certain de-riskings, there's a functional engagement that's maybe more passive by the individual, but still may stand for something in the overall goals of that entity. That's one point. And one other point is the leveraging de-risking groups. I keep coming back to eukaryotes. And eukaryotes, you know, it's a better deal. You've got multi-cell, you've got a lot more options. We've got mitochondria, we've got chloroplasts, we've got all the good apparatus, instead of having just these compartments where you kind of do stuff in the single-cell organism, right? So it's a better deal. And so maybe part of this thing with the groups, this goes back to Blue's notion is, is it, are we resistant? I think it was a blue, but we resistant to, is this resisting or facilitating the group membership? What's going on here? And, you know, the divergences and the differences. And I started thinking about thermodynamics first law. If you don't have, if you have isothermia, you can't perform, cannot perform work. If everyone has equal information, markets don't have trading. So there may be a benefit to not having, to having divergence, right? And maybe, in fact, the, that reality is, there isn't a facilitation or resistance, it just is. And that paradox is what we're, is the reality we're seeing is that it doesn't facilitate anything. There's no goal orientation. Individuals may have goals, but that's a more bounded goal than the super system of biology. Anyway, just a couple of thoughts that are gleaned from that last bit. And also just to remark once again, what a fantastic conversation. Thanks Scott. So Kristoff, then Steven, and then anyone else? Yeah, so I just wanted to, well, what did I want to say? I guess two things really. So one thing goes back to Blue's comment about like the fact that the fact, that if you have some kind of a complex network, there will be multiple mark of blankets in there. So it's completely feasible to understand, you know, membership to certain groups or something like that, or some kind of relationship structures to be overlapping in different ways. So I think it's conducive to this. But I think that we should not like, we're very focused now on this idea of, you know, blankets as a way of thinking about some kind of boundaries again, between groups perhaps, or between different kind of units of organization. And we're not, like it's always useful, I think, to kind of go back, you know, zoom out a little and not think about the blankets, because you have the blankets whenever you have some kind of a network of, you know, variables interconnected in some way. But what's really interesting is that in many cases, the behavior you will get from these systems, if you project them onto the social domain, will be very surprising. So one really surprising finding is something called the Tullman effect. It's with a Z and double L. And it's basically a finding that shows that if you have Bayesian agents interacting in a network, if the network is fully connected, they won't actually come to a consensus about some true belief. So imagine you have a network of, you fully connected network of Bayesian scientists and they're testing two hypotheses, right? And they're testing these hypotheses, they're increasing their credences and they're communicating their credences with each test to their neighbors. And what's really surprising is that the Tullman effect shows us that these scientists will not actually converge to a singular credence in the true outcome. Well, what is also surprising is that a less connected network will actually eventually converge to all the members having a very high credence in the most likely outcome. So this kind of, I just wanted to kind of throw it out there because I think that it's, you know, we're thinking now intuitively about these blankets as membership to groups and stuff like that. But the kinds of interactions we might get from using this kind of formalism and projecting it to a social domain might be quite surprising, right? I'm not saying it will be useless because perhaps the Tullman effect to go also back to what has been said about the flat earth situation and, you know, the belief in the roundness or flatness of the earth, the Tullman effect actually tells us that in a hyper connected society like the one we are living now, it's even rational agents might fail to converge to the most likely hypothesis, right? And that actually is something interesting that it's in a way the Tullman effect predicts the appearance of certain conspiracy theories in a hyper connected society because given the demographics, for example, of flat earth belief, it has been declining for decades until the advent of internet. So it only came to a sharp rise with social media when people started to communicate in a much more interconnected way. So I'm not saying that, you know, thinking about these systems in a social domain is bad, no? I just wanted to bring this, let's say anecdotal effect to say that the kinds of behaviors we will see these systems converging towards might be very surprising, right? Or they might not work exactly the way we think they do on lower scales of organization. Thanks for those points. So Stephen, and then anyone else? Yeah, that's some good points there, actually. And I think this is where active inference and the Friston Blanket has some advantages over the Pearl Blanket if it's about what action, there's a risk you see that information, you start to have this thing that it's all information driven rather than action policy driven. And by keeping it with the action policy, for instance, one of the things about having us on a planet in a solar system, which we used to have, well, now we know in a galaxy, in a Milky Way, blah, blah, blah, is that people have this sense of, we're almost meant to look at ourselves from space now, which is really only a proviso of modern technological perspective. People might want to be able to stay grounded in some way. There may be actually a psychological challenge with having to take this out-of-space perspective of us on a little ball going around the sun. So there could also be an action policy perspective here around flat earth gives you a way to re-ground yourself because it's like, well, I'm on a flat earth, so I don't have to take that, I don't have to go and imagine myself in space looking at myself. So there's a lot of interesting aspects to the action policy side once you start to have this scaling, and that could be interesting to keep that in mind. Thank you, Stephen. Joe, and anyone else? Yeah, thanks. Just something you said best, Stephen, when you say there's an advantage to the thriston blankets over the pearl blankets, do you mean an advantage for real systems making use of these things or an advantage for us as modelers? This is something that I really worry about sometimes that we slip into thinking about the practical benefits for a system of these constructs, whereas what we're really concerned with is the benefits for accurate, and in that context, both might be good tools, but when you use one tool or the other, it will depend on what you're trying to do with them. I would say more for how we look at it as a modeler, although the actual thing gets called a system and then has this kind of perspectival process. I think it's good to keep it in mind so that action doesn't get lost and it doesn't purely start to become information. Yeah, great. That makes sense, and I wasn't necessarily accusing you of this error, but sometimes I do worry that people get mixed up over who's getting the advantage, I guess. Who's benefiting here? But if you mean just that it's important to keep action as part of the picture of a model, so that's a fair point. And one contribution of this paper, and we look forward to the revised and published version as well, it's almost like if all we know is Markov Blanket, it's like, hey, hand me the screwdriver. Well, there's different shapes, there's different sizes. If you have the hex, it's not going to work for the Philips and different countries or different applications need different screwdrivers. So now we can, with these adjectives, ahead of blanket, we can start to differentiate. Hand me the Philips screwdriver. Okay, you're talking about the Friston Blanket. You're talking about the Pearl Blanket. And so we can be specific now because it's almost like enough air was pumped into the Markov Blanket metaphor. It was becoming too big of a tent, too big of a blanket. And so now we can start to be really specific and address those kinds of questions. When are we talking about utility for the system? When are we talking about utility for the Modular, which of course maps very closely to the distinguishing that you raised with inference with and inference within a model. So Scott, if Joe, if you want to give a thought on that. Yeah. Can I just... So we're already pleased with the reception this paper's had particularly among the Active Inference community for the reason you said there, but I think one thing we were worried about but people have seen is that there is meant to be this more positive message that these aren't necessarily bad tools. They're just different tools for different kinds of cases. So I'm really glad that people have seen that and kind of taken that as a helpful message. It's the yes and approach to the literature. It's not blood sports. It's not a knock down drag out with citations. We can just build upon each other and come to more useful good ways of communicating. So Scott and anyone else? Yeah, just two points. One, along those lines in law, one of the most effective strategies for disambiguating something is saying what it isn't. Right. And so even if you don't know what it is, you can at least start to say what it isn't. And so that's a handy notion along the lines here. And as the last couple of comments were observing, we had this thing out there that was very exciting. And we kind of, everyone is playing with it and said, oh yeah. And it's the same thing that's happened in privacy. The word privacy is just carrying water for all sorts of anxieties right now. The word trust has all sorts of... So the problem with language is we got a lot of subtlety and a lot of interactions. Interactions are increasing exponentially. True exponential increase in interactions globally is an assertion that I make at least. And if you think about that, we still have a vocabulary that's the same as it was way down the exponential curve of interaction increase. So how do we capture our subtleties in our interactions? Right. How do we expand our vocabulary? Well, it's exactly what we're talking about here is we take words. We use them by analogy in other areas. Right. We've already talked about blankets. Well, these aren't really blankets at all. And in fact, we get hijacked by the notion of blanket. We already seen that. Is it starting to use those... Is it a blanket or is it a membrane? Well, is it semi-permeable? Well, a blanket's not really semi-permeable. You know, that kind of stuff, right? So it's fascinating that we have... This is the appropriate and very exciting challenge to have to the taking it and running with it idea early on and disambiguation. And it means that there's a lot of energy, which means there's a lot of need for whatever it is that's being served by the concept. And so what you're... As modelers, this gets kind of interesting, kind of gets into Manhattan Project, kind of ethics of science, right? Because as modelers, you do have responsibility and you're discharging here by saying, hey, what does this mean for society? You know, should we develop nuclear energy unfettered with regard to whether it can be used as a weapon? Right? Because this is powerful stuff here. And so that raising of the notion and the questioning of whether it's appropriate to apply these things socially is going to come up very quickly as people do things like the rhetoric that I'm talking about and start to apply it politically and start to recruit biases using these concepts, that will be seen as a nefarious application of a scientific and mathematical statistical concept. And so your paper and the questions you're raising now are starting to lay that foundation of pause and saying to people in the future, hey, don't be so fast to run with this and just take a minute to think about the implications. What isn't this? Right? And so starting to ask that question can be as important and sometimes more important than what it is. Thanks. Thank you. There's a question from the chat. Sean Kelly wrote, are there special cases of Markov blankets that shield internal states a Markov shield? So would any of the authors like to just give a response to that? Yeah, Jelle. And then go ahead, Steve. I think I can give it a try. Of course, this really depends on what you mean by shielding. Right? And I think in the paper and also in the paper, we use a statistical notion of shielding that is pretty much equivalent to conditional independence. Right? So if you have three nodes and there's one intermediate node such that knowing the intermediate node will mean that knowing the third node does not mean anything, doesn't give you anything for knowing your first node, then that's a notion of shielding in the sense that the second node kind of epistemically shields off the activity at the third node from whatever happens at the first node. And this is an epistemic notion of epistemic and statistical notion of shielding. And also that one, right? I mean, shielding has a lot of kind of more everyday kind of connotations that are not necessarily congruent with the statistical notion. But if you mean the statistical notion, then yes, the very definition of a Markov blanket is that the things inside of the blanket are given the Markov blanket shielded off from the rest of the world in that statistical sense. Thank you. That was what I was thinking. An insulation, like a heat insulator is shielding it. And so that is kind of what these nodes are doing statistically. And then we're nuancing what kinds of categories these nodes might fit into. And that's a key difference between pearls phrasing and Friston's phrasing is the extent to which nodes are classified into different kinds. So Stephen, and then anyone else? Yeah, I think that question about how things when we scale up is really important. And in predictive processing and the modeling, you know, you have this hierarchical model of things going up to higher levels. And yeah, there could be this implicit suggestion that each of those steps or somewhere up there, it might go into other size blankets. But it may not be that blankets are needed as much to go another layer. It might be that there is some sort of process structure that is doing the job. At lower levels, it's a more active influence process. And if I could give one metaphor, there was a metaphor I'd put on about if there's a flag stuck in some mud and it looks quite stable, but it's blowing in the wind, okay. And then I have another flag, which is stuck in some mud. It's actually the top of an ant hill. And underneath the ants are all swarming to hold it still. So, and it's actually a little bit more still, but actually the holes more open. So in the first case, it's energetically, thermodynamically more an equilibrium is being held in a one place. The second one, you think it's in more of a thermodynamic equilibrium because it seems more firm, but actually there's a whole swarm of ants underneath trying to hold it in place. But none of them have got that intention. But looking from the outside, the system is holding the flag. But from the point of view from the ants, they're trying to enact some sort of action policy. And I suppose I wonder if there's something similar between the system as seen from the outside as a state and the kind of swarming that's going on underneath and that the two can get mixed up sometimes. Thanks, Steven. Always a great challenge and eternal question. What are the ants trying to do? But what you just pointed out there is really, it's related to this distinction they draw with these four different cases, the difference between a generative model, which is like the modelers perspective on something versus the generative process. And so those are two different generative processes that are giving rise to the movement of these two different flags. Let's say there's two different processes underlying it. Those are realist claims about what is giving rise. And then there's all different kinds of perspectives that can be taken. But if we mix up two people's perspectives or if we mix up a generative model and generative process, it's a category error. And if we're coincidentally correct, it would be miraculous because it'd be the wrong kind of thing that's being considered. So it is extremely important to prevent that kind of misinterpretation by really asking, are we talking about the process, the generative process or the generative model? And sometimes even just that distinguishing as well as this fourfold classification raised here, these go a long way towards helping model. So Scott, and then anyone else? Since we have Europeans on the line, I'd like to talk about Kant for a second. And I want to display my ignorance of Kant and the categorical imperative. So I complain about the categorical imperative because if can one person know enough, have enough situational awareness so that their good act can translate into goodness in the world. And again, maybe that's total misinterpretation of Kant or maybe just got it all wrong. But in the same way, I wonder about Markov blankets and this notion of knowledge, of an awareness. So that, because that one of the problems we see in the world is even a good actor, if they don't have good information, can do bad things like in the sustainable development goals of the United Nations, they put in a Gujarat power plant, was funded by a UN based financing agency and it cost them a billion dollars or whatever. And they were crowing about the success of sustainable development goals of energy and development were being pursued and furthered. But what they didn't talk about and they then got sued in court over is the fact that the local fishermen and local population was pollution and so it poisoned the local waters and the local population with the pollution. So they didn't do a sufficient analysis even in this situation where they're trying to improve the world and listing out all these goals of the world. That's their declaration, that's their Amsterdam flag of what they're trying to do with the UN, right? So, but nonetheless, they pursued one goal at the expense of other goals. And one of the things I'm interested in is your thoughts. You can go into kind of one and also correct me on my bad, non-interpretation there. But can you comment a little bit on awareness and knowledge? We talked about group and individual Markov blankets but can we talk a little bit about that notion about sharing of knowledge and situational awareness, just as a survival strategy. The ants were surviving in the Shannon's tilting bridge example before which was a survival thing. In Steven's flag projection, that was a little bit more the answer, a little bit smarter about their externalities in that one. Anyway, can you comment a little bit about how do we get, how might these analyses and these models and their implementations in reality and in the lab help us to understand the nature of group information flow? Thanks. Thanks for the question. So, if anyone has a thought, yep, Gela, go for it, and then anyone else. Great, I'll give it a try. I'm not sure if I can tie it all the way back to the categorical imperatives but that's, let's give it a go. So, I think the first thing that, the way I understand it, the kind of active inference framework is really a theory of learning for particular kind of environments. So in a sense, organisms start to learn to adapt to the specific niche that they are engaged in and then somehow slowly build up a particular model of that particular kind of niche that you can, of course, and that model you can use also for hypothetical kind of situations or counterfactuals or new kind of situations but there's no guarantee that this kind of previous history, I mean, if you enter a new niche or if the world changes, it's not guaranteed that your previous knowledge will be specific to the new situation or will be working in a new particular kind of situation. So I guess that indeed the kind of active inference framework gives you a perspective on situated situation-specific knowledge and not on the kind of categorical domain general knowledge and so this doesn't guarantee, indeed, that if you use your model to make inferences about new situations that you will end up on the right result and so that gap between what will be your best action according to whatever kind of imperative you set for yourself in reality and how it works out, that's always a gap in the knowledge that you have which indeed is related to the difference between the kind of the causal dynamic to the world, the degenerative process and the generative model of the world that you have. And so I guess that's indeed why, perhaps, I mean, speculating that different kind of ethical frameworks might fit better in such a kind of approach rather than the ontological one. That was very, oh yeah, go ahead, Chris. Yeah, I just wanted to follow up on this a little bit because we did the notion, you know, of the distinction between the generative model and the generative process was brought up before and now Jela alluded to it. And I think it's really important to keep in mind that the free energy principle, so the generative, maybe let's start differently. The generative process is supposed to be the process or the causal process that produces certain sensations in an organism and then the generative model is supposed to be on this framework producing predictions, right, which kind of explain away in the Bayesian sense the sensations that the model is, that the organism is experiencing. And the idea that's implicit in a lot of the literature is that, you know, you have the generative process on this side, you have the generative model that's trying to mirror the generative process. There's the sensory sheet in the middle and that these two things will somehow, you know, hold onto each other and they will be just perfect. There will be some kind of a perfect resemblance or mapping between them. And that's not entirely, it's not entirely clear that this is the most successful, evolutionary, successful strategy. So the generative model only needs to recapitulate the generative process to the extent that it's successful for in a given ecological niche or even social niche to ensure the survival and reproduction of the organism, right? And so that's one thing that I think important to remember about this, that, you know, the free energy principle from the get-go kind of does not necessarily produce agents that are perfectly, that will be always optimally informed about what's going on in their surroundings. So really, they learn just enough to stay alive. And of course, the inclusion of active inference gives them an advantage over traditional Bayesian agents because they can also go and try things out and learn in this way, but it still doesn't take us all the way, so to speak, to kind of transcend our model and get to know the process that's out there in the world. And I think this is also true in the social domain in many ways. So in the social domain, I think it's very plausible that the same thing happens. And we could expect that the same thing will happen for, I don't know, groups, for example, where groups will form certain models about how society operates or how the world is, which will be optimal to their given epistemic situation and their given social situation, but will not be objective models of how society or the world operates. Thanks. And what you just said there at the end about the cultural ways of viewing the world, it's known that many cultures, their own name for their own country or homeland might be like, this is the middle spot or this is like the main place. It makes sense. And it's unsurprising that it's that way. But then just like Scott said, we're all comparing notes. That's the phase of our test that we're in. And two people are mapping to a different territory with the same relative egocentric map. And so this question about how different egocentric perspectives and at which level those egocentric perspectives exist at, how those come into harmony is really an interesting question. So Stephen and then anyone else who'd like to speak. And yeah, this is awesome times. It's been a great discussion and we have 20 minutes or however much more people would like. Yeah, I just liked, I really thought there was a really interesting point there about when other approaches are maybe better than taking an ontological approach. And I suppose it, I think, and you're talking about the epistemological approach, I suppose that you get into this question of whether it's a more fuzzy, of interest in your thoughts on like fuzzy logic and a more fuzzy based epistemology rather than maybe the type of logical categorization of knowledge which ends up becoming ontological at some point. And that there is this idea of ontological pluralism that there's sort of their way around this, you know, or third wave, the third wave inactive consciousness has that. This idea that the Markov blanket, because it breathes with different regimes of attention, you sort of trying to get around that problem. But there is, it doesn't necessarily say how they do that. So I'll just be curious, yeah, a bit more. And the other thing if it's tying in with other areas is like, for instance, they're talking about affect now a lot and affect is this kind of integration, which then ties in qualitative research where you have axiology, which is kind of the normally this thing, which is sort of down there in arts based methods, but you know, the one where it's about how people feel about things. Well, now that starts to become part of a lot of what we're understanding about the world. So it could actually have quite an implication for participatory interactive kind of research methods and making them much more central in a way. So I wonder any of your thoughts on that about that, from the authors. Thanks, Stephen. If anyone has any thoughts on that, or it can raise their hand and yell it, go for it. I mean, so I guess what we do in our paper is a little bit before that step, right? So on the question of ontology versus epistemology, should you go for kind of ontological pluralism or not? I think what we try to say is that, look, here's a class of models that are typically used in literature as abstractions from a target phenomenon, right? And so if you want to understand those in kind of ontological terms, that's fine, but you need to give a story about how that abstraction relates to the target phenomenon that is there. So either if you go into metaphysics, you go for kind of more Platonase kind of routine. You say, oh, actually, this mathematical model is not an abstraction from the target system, but it actually, nature at its joints is somehow mathematically structured. Or you say, look, no, these are indeed abstractions produced by us and they're useful in predicting and interpreting the world, but they are not themselves the world. So in a sense, we kind of offer the kind of dilemma to anyone who wants to make a commitment that we're choice there between epistemology or ontology or ontological pluralism, saying like, look, you can go the ontological route and, of course, this is the way more challenging, kind of exciting route, but then you either need to buy these bullets that these models are not really abstractions, but somehow on the lie reality or need to give some other sort of story about how that works. And I think personally I'm curious to see what the response from the community will be, which of those options people will take. But I mean, I don't think either of us is necessarily committed to one of those which end up with just saying, okay, look, if you want to do it, then these are these are the rules, kind of, so to say. Maybe disagree with the rules, that's fine as well, but then give an argument for that. Thanks. Very interesting stuff. Anyone else have any thoughts? Yeah, Stephen, and then anyone else go for it? Yeah, I agree a lot with what you're saying there. I think there are some gaps in, and rather than saying there's a gap, sometimes it's just filled with the Markov blanket, if that makes sense. If in doubt, fill it with the Markov blanket. And I've been really curious, I'm actually looking a lot in terms of how organisms structure like the idea of peripersonal space and spatial inactive ecological approaches and the way that maybe there's a way that that space is actually, and there's a field called mental space psychology which has been a very fringe field. Actually, a lot of it based in Holland and that is, it seems more plausible now with this. So it was a much more of a practice-based approach, but I thought I'd share that anyway because I'm presenting something this Friday. But I do think there's something missing in terms of mechanisms that extend out beyond the body into the space and it may not have to be at all levels, Kristen's blankets in the traditional sense. Shannon, and then anyone else? I'm going to talk about brains. So I know we've been extending into the world and into the body and everything. And we've been talking a lot about group dynamics and maybe how groups behave or how we feel like part of a different group. And earlier we super briefly touched on mirror neurons. So these are neurons that fire when I observe in action, the same as when I do an action. And there's been a lot of work lately, not necessarily on mirror neurons, but on interbrain synergies, or you might call them interbrain coherence or cross brain coherence, where lots of dynamics in my brain are firing very similar to lots of dynamics in your brain. And this could be because, like in work from Ari Hussain on neurocinematics, this could be because we're watching a film and it's evoking similar responses because we're all attending to the same stimulus. But the film that we're watching, it could have a structure where the filmmaker sort of designs what you pay attention to. And so most people's gazes will fall in that direction. You'll have even more similar sort of interbrain coherence or unstructured where you're just sort of filming a scene and there's maybe some people over here, some people over there and your gazes could differ a little bit more. These structures could differ a little more. And this is all to say how, like the film is the environment in this case. And if we're talking about the rest of the world being the environment or like us in this room being the environment, our gaze is structured in a certain way by there's a blue box on my jitsie screen. So it's sort of directing everyone's attention at my face. But there's also some lack of structure. I don't know that we're seeing the same tiles in every space. And there's different ways that if we interacted, like if I interacted with one of you a lot in my daily life, like maybe Steven and I hang out and get coffee every day. And we talk about something every single day. So we have a propensity to pay attention to similar things in similar ways. And we've been specifically talking about Yella's idea and how much we agree or disagree with it or something. So we are continually like drawing our gaze to Yella to see if he understands or agrees with what we're saying. Like there might be this similar interbrain coherence between Steven and I that's different between me and blue because blue and I never talk. We're not even friends. These are all hypothetical, by the way. And these, so a question, it's not, you know, in your paper necessarily, but like if we're looking for Markov blankets in the world, one thing that we could look at is something like a interbrain coherence. And that might be something real in the world or it might be a pearl blanket, something that we use to describe how people interpret something the same or how their brain activities are synergistic in a certain way. And if you can be empirical, I know we were being rhetorical a lot of times, but if you can be empirical and look at these signatures of either, you know, conversation dynamics or brain dynamics and can relate them to how the world is structuring them, then that's a way that you can look for, you can still keep looking for Markov blankets in the world or find it useful. It's definitely, I think what we keep coming back to when we're talking about these blankets as a metaphor or these blankets as something real is that we can totally have both as long as we're specific about which one we're referring to. Thank you, Shannon, Jelle, and then anyone else. Thanks, Shannon, that's really interesting. And I think that there's a couple of things there, right? I mean, in either one sense, if two people have the same background, same experience or so, right, they might parse a particular kind of movie or scene in a similar kind of way, that if you put them in rooms separate from each other, you will indeed see a similar kind of neurodynamic going on in both cases. Maybe tied indeed to attention, where you pay attention, you gaze, et cetera. The other one is indeed kind of on-site, kind of interpersonal synchronization that might be happening, right? And I know there's this kind of second-person neuroscience kind of approaches that try to capture those kind of things in an experimental setting and see how, in what particular kind of settings, brain waves synchronize, et cetera. And that kind of second perspective reminds me a lot of the kind of discussion that there are an active inference on generalized synchronization, right? So these are the kind of examples that you get when there's two coupled clocks standing on the table and they tend to synchronize after a while. So this is a very simple causal mechanism, causal structure where synchronization just occurs over time. And when I first started thinking about Markov blankets, I used that kind of example as a kind of reduction at absurdum, saying like, look, if these things even have a Markov blanket, right, then it's an empty notion. Now, having worked with these philosophical signs here, I think need to be a bit more specific and indeed say, look, whether that thing has or has not a Markov blanket, that's a different question. But there is a point that there is a very simple structure of kind of causal dependence and causal independence such that if you would fix the table perfectly, the things would not synchronize and if you would let it slide a little bit and wobble a little bit, then they would, they do synchronize. So there is just a kind of thing like this intermediate factor that mediates the causal relationship between those two things. Now, the question is, is that do you need to have a Markov blanket for that? Does Markov blankets add anything there or is a kind of story about kind of causal dependence and causal independence enough? I think that, well, causal dependence there and causal independence such that if I would do this, then there would be no synchronization. I think that's enough. And then what do you call that? A kind of Markov blanket is a second kind of question. You need to translate it back to the kind of world of causal models and see how they relate to that. But I think there's a clear case to be made that this is parallel between the kind of cases you're describing and the kind of synchronization settings that are there in these kind of couple clock examples. Thanks. And one interesting aspect to kind of bring it back to this discussion about the internet, what kind of beliefs propagate culturally on what kind of connectivity. Just like Shannon was saying with the video chat and with the live stream, what somebody sees, it's not just like they're sitting in a different chair at the theater and they're watching the same play or they're in separate movies, rooms watching the same movie. And so cultural cues might lead somebody to look into a correlated position as another person. But when there's the customization, it's like each person is in their own theater, which takes it to another level, because it's actually different inputs and the inputs can be targeted in interesting or new ways. So it's kind of stretching the limits of our social cognition because it's not just a bunch of people asking what they just saw with a shared event. It's like trying to come to a shared understanding when people are seeing different events. So, yeah, there's some new fun things we're learning on the internet, I guess. Well, in the last couple of minutes, if anyone has some thoughts, I'd be curious to the authors or non-authors, where do we go from 20.2? Where's our momentum or direction or curiosity? One thing I'll bring up while someone's raising their hand is without buzzing in the middle of a conversation, how can we be really clear when we're intentionally or unintentionally slipping the difference between a Markov, Pearl, Princeton Blanket and generative model, generative process with a model, within a model. We don't want to police subtleties of language, but at the same time, somebody can really build a bridge that leads someone to nowhere if they're unintentionally not sure the difference between these concepts. So I'm just curious how do we promote rigorous and accessible norms around talking about these ideas, recognizing that we're all like totally learning this and formalizing the project around us. Jelle, and then anyone else? Just a small note. I don't remember exactly who mentioned this, but at some point we talked about two phases of research, and one is kind of taking an idea and pushing it along as far as you can, and then there's a new kind of phase where, okay, now we ended up in a completely new place. How are we going to kind of substantiate this? How are we going to formalize this a little bit? And I think this is the kind of, I guess over the last few years, active inference and also the philosophy of active inference has kind of been just another, I mean, also in my own earlier papers, kind of been running as far as possible, see how far we could get this and see what would happen as a kind of creative, kind of startup stage of research. And I think both in this paper and in a number of other papers that are currently kind of coming out, I think this kind of more kind of consolidatory phase is happening where we're like, okay, this is where we ended up. Does it really make sense? Do we need to make some more fine-grained demarcations? Do we need to tweak the conceptual apparatus to have it all make sense? And it's a different kind of work, but I think that that's really important for the next kind of phase for active inference research. So I hope these kind of discussions can contribute to doing that kind of work. We're consolidating Bitcoin to the key $20,000 level. You know, let it rest before the next... Really, yeah, yeah. Whatever metaphor works for people, but any other thoughts, because that's very nice yellow. Thanks everyone for sharing. This was an awesome sequence and it really helped take our lab and our discussion to another level. We would always welcome any of you to come back to continue the conversation, come for a guest stream, come as a visitor to be on any conversation you'd like. So thanks again to everyone and for listening. Peace.