Friston provides a profound mathematical foundation for biological agency, framing life itself as a persistent effort to minimize informational surprise. It is an ambitious unifying theory that elegantly bridges the gap between fundamental physics and the intentionality of living systems.
Installieren Sie unsere Erweiterung an, um sofort in jedem Video zu suchen
Agential Topics: Karl Friston's Free Energy Principle, May 6, 2026Hinzugefügt:
Hi everyone, I'm Charlie Munford and this is Agential Topics. Agential Topics is sponsored by Agential Biology Institute, which is founded because we think there's a missing causal layer in biology that gives cells an intentional direction as they develop. And we're building the infrastructure to study that missing causal layer. So, aential topics is where we get together and try to surface in this conversation the best ideas that are out there about cognition and agency. And uh today we're welcoming um luminary Carl Fristen, author of the free energy principle. And uh he has a few words to say and then we'll have a discussion.
>> Thank you. Um, let me share my presentation and then explain why I'm doing a presentation. I think that mean the >> Okay, I think it should should have permissions. All set. Great.
>> Right. So, I had um well, first of all, let me say um it's a great honor to be able to present to this group. It's also a bit intimidating after witnessing the last the last discussion. So I just want to remind you I am a physicist and psychiatrist with no foundational training in cell biology, evolutionary theory or philosophy. So if you can be gentle and understanding uh with my responses to your questions, I I would I would appreciate that. Um I had anticipated on after discussion with Charlie not doing a presentation but um I think there was some agreement that perhaps it would be useful to scaffold the discussion with a brief 20inut presentation. Um so what I've done is prepared 20 slides um picked out from a number of um graphics that could guide the conversation but specifically to address several questions that Charlie sent me and an additional question that Stuart asked um um a couple of weeks ago. Um so I'm going to um go through this um presentation just covering the themes that um would um I be useful to address um in relation to the questions that were asked. So I'm going to go through it very very quickly without not in any depth but just to put out there the different applications of the free energy principle but that might be usefully referred to in terms of answering the specific questions. So I'm going to go through the statistics of life you know the role of Markov blankets um with a special focus on um scale invariant applications of the free energy principle uh under the reormalization group because I think that's going to be relevant for some discussions and then a couple of applications in relation to natural selection and action selection in cognitive science and ethology. um and then just returning at the end to uh the importance of the scale invariance and the relationship to um dissipative structures. So um when a little box appears in this uh beautiful pale green that's a question from Charlie um cut and paste from uh an email um that I hope to be able to speak to. So the free entry principle is effectively um a variational principle of least action that gives rise to a particular mechanics and um and that's going to be called a basian mechanics that has um as its commitments exactly the same commitments as quantum statistical and classical mechanics. Um the only distinguishing thing is that careful attention is paid to um the notion of self in the spirit of self-organization, self assembly, self-information, self-organized criticality and so on. Um and that's usually um or it starts with the notion of a markoff blanket that enables one to individuate uh the states of self from the states of everything else. Um and um this is borrowed from the work of Judea Pearl uh where that separation or individuation rests upon a conditional independence between some internal states and some external states. say the rest of the states of the universe when conditioned upon a markoff blanket and I'm going to further divide the markoff blanket or the the markoff blanket is further divided into effectively inputs and outputs in the sense of sensory states and active states that are operationally defined in terms of parents and children. So the markoff blanket is defined in terms of the parents, the children and the parents of the children and the sensory states influence the internal states but are not influenced by the internal states. And similarly the active states are influenced um influence the external states but are not influenced by the um external states. And with that partition we've got a definition of self uh in terms of the states that constitute a particle person or population. Um internal states uh if you like exchanging with external states birectionally but in a particular way through the mark of blanket states as sensory active states. Um my favorite example is the brain where the brain can have all sorts of internal states influences as actuators or autonomic reflexes and active states that change the external states that could be the body um that then reciprocate by changing the sensory states that in turn influence the internal states. So you got this circular causality in play which um usually is uh in um the treatments I uh present uh is read in terms of sort of action perception cycle of a particular kind. Um, and then I normally ask people to forget about the Markoff blanket and then we just rehearse basic physics. Um, and then put the Markoff blank blanket back into play to see what kind of mechanics emerge.
And the physics that is usually rehearsed is um the starting point of most of physics which is basically a random dynamical system or a launch line equation expressing the dynamics flow or rate of change of states as some lawful function of where you are in phase or state space plus some random fluctuations. And then from this we can interpret the trajectory or path through state space as a probability density over the states given the system in question say me conditioned upon me. uh and then the argument goes well we know a lot about the density dynamics in terms of the rate of change of this probability density in and of itself and I've expressed that here in terms of a plank equation could also call it a master equation or time independent scroian equation uh and because the systems of the kind that we're interested in um keep revisiting the same state that they were always in otherwise the the uh state would just diverge off into infinity. Um we know that the probability density in and of itself does not change with time which means that the foca plank equation the density dynamics has a solution and that solution is really interesting and allows one to express the dynamics of any system that possesses this pullback attractor or characteristic set of states in which you find it. um in terms of a flow on the gradients of the log probability given the system in question. Um, interestingly this flow um can be subject to a Helmholtz Hodgej decomposition um and separated into some basically orthogonal flows a gradient flow on the log probabilities as the system if if you like counters the dissipative effects of random fluctuations and an orthogonal solenoidal or curl flow which I think may be relevant for our arguments. Um, so first question, walk me through this.
This is from Charlie. Walk me through the strange loop. I am a prediction of what happens around me in part because of my predictions. Is that it? Yeah, absolutely. Um, but in an even simpler sense, we start off with strange loops.
Um, in the sense that when this solenoidal component becomes state dependent, this path or trajectory um becomes chaotic. And um it means that um although we keep revisiting neighborhoods of states we were once in because those are characteristic states of the thing that I am um we never actually um repeat or revisit the we never tread in the same river uh twice.
Furthermore it means that these kinds of systems will necessarily show oscillations by rhythms life cycle reproduction. It can be no other way. So at any scale um what we're talking about are systems that have this sort of recurrence uh um recurrent kind of aspect endowed by this Helmholtz uh decomposition.
It it also to a certain extent in a in a slightly naive way answers Stuart's question where does novelty come from?
Well, the very fact that we're dealing with these um strange attractors um means that uh in the future of any trajectory, everything is novel because we've never been there before. So, baked into this formalism is a really trivial kind of novelty, but it's a very constrained novelty because we are dealing with systems that have characteristic states or an attracting set technically a pullback attractor. Uh that's just a slide just to um lend intuition to this um decomposition into gradient flows down the plug hole here and the solenoidal or circular flows that have to accompany this in any um non-equilibrium um uh system. So um yep.
So that's another um one view of strangeness in that we are implicitly dealing with systems that have stochastic chaos where the chaotic aspect uh rests upon a strange attractor. Um what is what about the markoff blanket? Um so what we're doing now is putting the markoff blanket back into play under this very generic formulation of self-organization uh to these pullback attractors on n on equilibrium steady states. Um and observe that these this decomposition applies to um all partitions of that markoff blanket uh partition uh including the internal states and the active states. But by construction, by just individuating or differentiating self from non-self, by construction, these gradient flows do not depend upon or the gradient flows of the internal active states do not depend upon the external states. They only depend upon the blanket states and more specifically the sensory sector of the blanket states. So now we've got a um a mechanics that can be read as perception and action at some level whereby it will look as if the internal and active sometimes called autonomous states of anything say me will look as if it's trying to gradient ascent or climb or increase the log probability of um the sensory states conditioned upon me the system in question. And this I think is um or is it often sold as just a unifying formalism that many people would recognize from different perspectives and a few here. Um so what um what is the log probability of the sensory states conditioned upon me? Um, it's just those sensory exchanges with the world that have value for me in the sense that they characterize the kind of sensations I would um expect to have.
They are literally part of my attracting set. Uh, and therefore they are valuable. And from this you can interpret this as a statement um that underwrites things like reinforcement learning, optimal control theory or in economics expected utility theory. Um, the negative of this is known as surprisal. of information or more simply surprise um and via the Fman Jensen bound. It's a thing that is it is um upper bound by a variational free energy and this can be read in a number of different ways. The infax principle, minimum redundancy principles or maximum efficiency principles and of course the free energy principle. The average of that uh free energy or self-information um is uh entropy, the time average, which means it looks as if the internal and active states are look as if they're trying to minimize on average the entropy. And of course, that's the holy grail of self-organization as exemplified by synergetics. And if I was a physiologist, it's just a statement of homeostasis. It's just keeping um my essential variables within bounds that uh that then maintain my integrity in the sense that I'm limited to my characteristic states. And if I was a statistician, I would interpret this as um something called basian model evidence. Basically, the probability of these sensory data given me as a model of how those data were generated. Um and from that you can appeal to things like the brazen brain hypothesis, evidence accumulation and predictive coding and the like. So Charlotte's question, what is the opposite of the consequences of the FEP? What would a self be like if it maximized free energy? Um so the answer to that question is it wouldn't be a self. In fact, it wouldn't exist because the free energy principle is basically vituated from its environment and possesses a markoff blanket. It can always be read as minimizing free energy. So if it didn't minimize free energy, uh then it could not exist. So the free energy principle is a somewhat tortological description of self. It's just um a way of describing self-organization um by appealing to the existence of these conditional independencies and therefore the sort of teological reading that you uh that you put on top of it.
The FEP solves a grounding problem in a different way. Um significance is rooted in surprisal not in survival value.
Right? How does a an evolving self-replicating system come to care about surprisal rather than survival?
Are these two goals ever at cross purposes? No. By definition, stipulatively, the um surprisal is just the negative value by definition stipulatively. Um so they can never be at cross purposes. They are the same thing. They're just um inverted.
Um an example here just to give you um an illustration of self-organization uh where we just simulated lots of macroolelecules all coupled in ways where we have markoff blankets um um that determine the interaction or individuate each macroolelecule. But the objective here was to find another markoff blanket at a higher scale um in terms of um unpacking ensembles of these macroolelecules in terms of this particular partition to demonstrate the emergence uh spontaneous emergence of a a little sort of viral particulate thing here. little virus with a little psyllium. Um uh here uh internal states, blue active states interestingly underneath the sensory states that are pushed into the external uh states um just to demonstrate the emergence of this kind of um self-organization.
And I show that example just to illustrate um that there are markoff blankets at different scales and then ask the question how do you jump from one scale to the next? Um and this um is a I think a really interesting um question which actually speaks to a a deeper question which is um what is a state? Um and the the argument pursued in sort of um fuller treatments of the free energy principle sort of follows the following um construction. So what's a state? Well, now we've defined something in terms of markoff blanket.
We can say that the state is an iggon state of a particle's mark of blanket.
Why? Well, because that's the only thing that I can see from the outside. I can't see because of this conditional independence what's going on in the inside. I only need to see the outside the Markoff blanket and I can summarize the state of a particle by the principal states of of the uh of the Markoff blanket. But um in order to define the state, I've implicitly defined a particle. So what's a particle? Well, a particle is a set of blanket and internal states at the lower level. And then what is a state? A state is and so on all the way down. So there's a you if you like you elude the question what is a state or indeed what is a particle and what is a state of a particle um by um um recursively uh going to finer smaller and smaller scales or conversely going to higher and higher scales. Um and you can formalize this in terms of the renormalization group in terms of uh regrouping and um reduction operators.
Uh so I can start off with any particular scale.
I can partition this into um blanket states and internal states and then I can take um the um um move to a higher scale and then take the iggon states of that um and reduce the dimensionality and then just start again. sort of generate increasingly higher scales through using the apparatus of the normal group implying that there is uh there are markoff blankets and markoff blankets and markup blankets um all the way up and all the way down. And here's just another illustration of this. Um, one thing that emerges from this analysis which I think is potentially interesting as you do so you um through the reduction operator you um you effectively throw away fast fluctuations that are on the inside and the internal states when you sort of group things together at the scale above. So effectively the dynamics of systems as you increase their scale also slow down.
So by construction um the scales at a higher level more coarse grain scales have more deterministic and slower dynamics than fast um fluctuating smaller uh smaller scales and that I think is going to be um relevant uh relevant later. Um so one application of that which is um a more serious address addressing of Stu's question about uh novelty as it emerges say in an evolutionary context. One application of this um reormalization group formulation or application um of the free entry principle at uh different scales would be on some readings. Um um an account of um the link between an extended genotype and phenotype in in theoretical biology or evolutionary theoretical evolution.
um where you can think of slow dynamics at an evolutionary scale um as Markoff blankets of uh certain kinds that prescribe constraints on faster um phenotypic states um that both are conforming to exactly the same principles but at different uh at different levels. And it's interesting in this instance just to think about this in terms of um top down and bottom up causation that couple the scales together so that the um the genotype if you like at the higher level specifies say epigetically the dynamics and the um the uh characteristic sets and the implicit attractors for the evolution of a phenotype at a high scale. But the likelihood of or the negative free energy or the the marginal likelihood or the model evidence we talked about before um in um that scores the quality of this phenotype living in this particular context um now subtends the um um the bottomup causation in terms of the grouping and averaging um that you get um in terms of the population dynamics. Um so on this view if you come back to the fact that both levels you're um you're never quite revisiting exactly the same state there's a um in virtue of these um thisa stochcastic chaos when the stochastic chaos is read at the genetic level you will get novel phenotypes um um and that will be another another instance and I think a slightly more interesting answer to Stuart's question about where where how do you explain the emergence of novelty where it's just a free energy minimizing process or if I invert that um a basian model selection u process where the model selection rests upon the basin model evidence read as natural selection to produce uh different things. Kate neighbor has uh she has been a former guest and may attend the meeting and has criticized the FE for being able to apply to pendulums, crystals, etc. which he argues erodess explanatory power. Uh is this so? Um I don't think so. I think I think you have to um that worry um goes away when you realize there are different kinds of things um that are um specified by the statistic the uh influences among the different um kinds of states that constitute the markoff uh blanket partition. Um and one can imagine this um you know for example if the active states were an empty set you would have um a uh a good model of a crystal and possibly a pendulum. Um and um one can I'm not going to do this but one can certainly work through different permutations and arrangements of these dependencies that underwrite the notion of self or a thing um and generate a whole um ontology of different kinds of things. But one which is particularly interesting in relation to novelty of a different kind is if I were to take this sort of over this is the the minimally sparse dependency structure that would admit a markoff blanket but I can make it sparser um and if I focus on one particular um uh spification um it's the link between the active states and the internal states that would be appropriate if I was dealing with multisellular organization. uh not necessarily organ or uh um organisms or large things that have a sort of deep hierarchal structure. So they got markoff blankets in the inside which means that the internal states now are not directly influenced by the active states. So if I remove that arrow something very interesting happens.
If we can always read the internal states as effectively inferring the causes of the sensory states, if the internal states don't have direct access to the active states, then the active states become a various cause and that basically can be cartooned as follows.
So it now looks as if these sort of larger deeply structured um systems, things selves um look as if they're inferring their own action. And that allows you now to describe should you want to um the self-organization of big things with this deep kind of structure as um planning in terms of planning as inference as if they're inferring and planning um sharing a kind of intentional behavior simply because it looks as if they're thinking about the consequences trying to infer what they're actually doing. So there's now a divide between the actual real variables, the active states and the um random variables um that are being inferred under this u um what is known as a generative model. And that introduces an interesting um twist here because now we have to take the variational free energy um functional um and apply it now to active states and this induces something called an expected free energy or the free energy that I would expect if I committed to this course of action. So, I'm basically now trying to tell a story about action selection and a different kind, a different natural kind that would not be um could not be the that is sort of um more expressive and more like me relative to a crystal. Um so, um what is this expected free energy? Well, um, it's something that endows behavior with an intentional aspect, a motivational aspect. And many people argue that it it is necessary to be an agent in the sense that you are actually selecting among different paths into the future that renders you agential in an authentic sense. Um so here the question you argue that some organisms are more agential because their predicted time horizons are longer which means the loop includes counterfactuals.
Are there really any living organisms that have a time horizon of zero? I think that's an excellent question. It really begs so gets in the heart of what is an agent in this formalism. So an agent in this formalism would be something that has beliefs about its own action. But notice um that the in order to um select the right action to infer the right action it has to know the consequences of its action and those consequences live in the future. So the implicit generative model has to be future pointing. So it has to include these counterfactuals.
Um and you um Charlie is absolutely right that you know the temporal horizon the depth can vary but just having a future in mind or an implicit future in mind I think um is a necessary condition to be an an authentic agent. If you had a time horizon of zero you would be a thermostat or a virus um which would be one step up from being a crystal. Um but you wouldn't have any intention or motivated behavior because there would be nothing to select. Uh you're not selecting among um counterfactuals that underwrite your action selection or policy selection. Um and the the other way novelty gets into the game um here is that when you look at the functional form of these energy functionals that um underwrite or describe um the um action and perception um I I won't go through this but just just to highlight the fact that you know there are different interpretations a statician would like this kind of interpretation machine learning person would like this kind of interpretation but this Um the these different ways of expressing these uh the free energy functional is nice because you can now think about the corresponding teology or interpretation when you take the expected value of these under what you think you will um experience or see or sense in the future if you did action A. And interestingly the complexity and accuracy become risk and ambiguity and these two terms become intrinsic and exttrinsic value. I just wanted to focus on that because um if I ignore the extrinsic value, the value um that these sensory outcomes have for me, my prior preferences um then this intrinsic value um is known as basian surprise in the visual search literature or mutual information uh in expectation.
Um but it it it's more interesting because it uh basically scores the epistemic affordance in terms of the information gain. So it is the quantity that describes expiration. It is the if you like the mechan the part of the uh imperatives or motivated motivations or descriptions of apparently motivated behavior that um talks about information seeking behavior and novelty seeking behavior. So baked into this formulation for and only for certain kinds that have this large deep um hierarchical structure um you will have necessarily a a novelty seeking aspect uh aspect to it. Um just for interest the um if I remove different bits of uncertainty we come up with things like path interval control from engineering or um resensitive control from economics and then finally if I take all uncertainty off the table we end up with um expected utility theory and sort of pragmatic um accounts of goal directed behavior. Um so walk me through the solution of the dark room problem. Why doesn't it apply to agency by definition? Um by definition if agents are those things whose behavior can be described in terms of minimizing expected free energy they are necessarily maximizing the uh expected information gain. Why are they uh why does this preclude the dark room problem? Because the first thing I do in a dark room is to switch on the light because that has minimizes my uncertainty because it maximizes um my information seeking capacity. That's the action that I would do. So the epistemic affordances offered to me will immediately take things like me out of a dark out of out of a dark room.
Um finally um just come back to this scale invariant thing which which I thought I think is interesting. Um so this is more for sort of the sort of physics as opposed to the agential um um biology at hand here but I I think it's interesting in in it nicely relates to things like uh prior's treatment of dissipative structures. Um so I've just situated now um the basis mechanics I've just been talking about um in a series of uh different mechanics that you get as you go from the very small to the very big. And if you recall earlier on I was saying that as you go from uh lower scales to higher and greater or larger scales the random fluctuations tend to get averaged away and things slow down.
um and that from the point of view of this Helmh decomposition means that this gamma this gradient flow um is progressively suppressed as we go from the small to the big. So in the limit the quantum scale we have no solenoidal component because it's completely dominated by the random fluctuations. Um and in that context you can um treat the the plank equations as a time depend independent scr equation and then elaborate quantum mechanics from that.
Um if I have lots of things together then I've now got um statistical mechanics um that can be articulated in terms of fluctuation theorems and the like. But as I get bigger and bigger and bigger, the balance there's a Goldilocks regime in which the dissipative and non-disipative parts, the um the gamma and the Q or the solenoidal parts start to uh play together. Um and that's the regime in which the felandry principle is most interestingly applied to biotic kind of self-organization that has this sort of basing mechanics when I put the markoff blanket into play giving you the self-evidencing and basing mechanics.
But if I go too far I get too big the um random fluctuations go to zero and I'm just left with the selenoidal dynamics and that was just Newton's laws of motion classical lrangeian mechanics. So when I get to the scale I summarize dynamics in terms of the position of the moon and the sun and the earth or large massive bodies then there is no intentional or genetic behavior anymore because we've passed through the scale um at which there is the um u an ad mixture of dissipative and non-disipative uh dynamics um and that sort of I um nicely fits in with um sort of um you know Priagene's early observations only when a system behaves in a sufficiently random way may the difference between the past and the future and therefore irreversibility enter its description. Um yeah and then oh how sorry these are questions I haven't um found a slide to address. Has anyone successfully built in an active FE robot and what's its personality like? Um so the um the strict answer is well no. Um but people have built robots based on the FAP but they don't have much of a personality. They just own fridges and things. Um however um it is applied in the context of computational phenotyping and psychiatry to try and build a um apply the free energy principle active inference to a digital twin and then twiddle uh or um tweak the um parameters and the prior beliefs of the digital twin. So it reproduces the same behavior of uh in an experimental context as your patient and that basically means you can now phenotype or characterize your patient in terms of their prior beliefs. Um so in a sense it's sort of phenotyping their personality in terms of what a priority they believe in a very limited situation. that limited situation um is if you like prescribed by the experimental paradigm, the economic game or the choice behavior you're trying to elicit that you can actually build a generative model for. Um but it has been um there have been articles written about um how you could apply this to revolutionize robotics and there is interest in um sort of um interactions between different um agents um that uh do involve um sort of effective veilance emotional and possibly personality like uh personality-like traits. So, um, I've been talking incredibly quickly. Um, and the font hope that you're two minutes per slide. That's 40 minutes I've been talking. I'm sorry about that. Anyway, these are the people who ideas I've been talking about. And of course, thank you very much for your attention. I'll stop my sharing.
>> Wonderful.
Thank you so much, Carl. That was so incredibly efficient and um, illuminating.
My goodness. And thank you for so carefully going through my questions. If I had known you were going to take them so seriously, I might have I might have thought about them for a lot longer.
But anyway, that was that was really beautiful. Um, so we'll have a a question and answer now. Um, I want to let you guys know that Carl has to go at a quarter um to the hour. So, we only have a little bit over 60 minutes to discuss. So, um, and since he's already included the answers to Stuart's questions and my questions in his presentation, I'll sort of put myself and Stuart at the end of the line as far as questions. Um, so the way to do this guys is to click on the bottom of your screen under react and raise hand. And that'll just put you in the queue to ask a question. And, um, remember that you're typically muted when you're not talking. So, when you start your question, um, be sure and unmute yourself. And after you finish your question, I will um probably mute you again. So if you have follow-up questions, try to keep them to a minimum so more people can talk, but um you know, it's inevitable. So uh just try to, you know, be efficient and um thank you guys so much. Off we go. Um I have Henry Henry H.
Oh hi.
So uh I really enjoyed the talk. So you mentioned that the different scale you actually use the same formula in terms of mechanism, right? For example, you use all the same u mark of blankets as intermediate for different level of thing.
However, as we know in the uh biological system when the divided into different level actually is not a simple self organization principle anymore because maybe during evolution at the very beginning that play a very important role but as soon as this biological process is established so the constraint is kicked in the constraint is a biological process right so you do not lead every time to emerge and you just follow simple biological rule. So in that case so-called the u I like the you the con constraint lover actually constraint loyalty is not happen at the system level constantly only in the event of micro evolution occur and be selected upon. So the rest of them is no longer evolutionary problem. I I keep emphasize this point because they become biological problem. So I I would like to say you know we know biologically they have different law but maybe the physical law is the same. So that's my question >> right that's a challenging question. Um I think the simplest response is that I the fringi principle is trying to um equip the natural sciences with a sort of physics of biological systems that should also obviously entail evolution.
Um but you know just to pick up on a couple of important points. So you're absolutely right that the the the the um the formalism the functional form of the dynamics is exactly the same at each level. So that's a requirement of the reormalization group that you got the same functional form for the dynamics or the same lrangeian at each and every level. It's just applied at different scales. So that's you know that's absolutely crucial which means that you are compelled to describe the biological process uh processes that subtend evolutionary de development uh with um using exactly this uh this formalism.
Um, so I wouldn't draw a line between the um the variational principles um that underly self-organization and the biological mechanisms and principles that you um that you were describing. I would hope that the um the the the physical description in terms of these variational principles is apt to describe exactly the processes you describe. The sort of one I I I can't remember all the details of your question, but sort of one interesting um um aspect of that is that you know the way that the phenotype uh behaves in relation to its eco niche um has consequences for the level above. Um and that um that is read under this sort of um application of the renormalization group as meaning that the free energy the negative free energy um or the pathog of it can be equated with the model evidence or the marginal likelihood. So you can now read the model evidence which is a characteristic of this particular phenotype here as a likelihood that you will find it in this particular eco niche. And when you read it like that then the um the you can read it as adaptive fitness and that means that it will now contribute to the gene pool um in proportion to its marginal likelihood which is going to be in proportion to its model evidence which is going to be in proportion to the path integral of the uh free energy principle and then what that means from the point of view of evolution.
What evolution looks as if it's doing if I was a physicist or a statistician is effectively doing basian model selection read as natural selection. Um, and one way you could put a teology on that should you want to, which means that basically this evolutionary uh process is basically seeking evidence for um the kinds of phenotypes that exist in its environment. and it does so just by hypothesis testing and accumulating evidence in the phenotype um that this is the um the thing that is most likely to be found in this uh in this environment. So that um would be one example of a mathematical story that I am hoping would be um um sufficient or at least um um would accommodate a a more mechanistic or teological um take on um on the biological processes. Does that make sense? Is that what you were um >> Yeah. Yeah. Thanks. Yeah. you know I I will ask later on the question. Yeah, thanks.
>> Okay, >> great Mary.
>> Hi, thanks thanks so much for this really stimulating talk. Um my question has to do with uh the correlation of your work with um deacons absential causation. Um, in active inference, nested marov blankets are coupled statistically with higher levels inferring the hidden states of lower ones, but biological levels seem coupled to material leaks. Cell leaking CO2, biopotons, exosomes, and tissue is organized around what the cell cannot close locally. So, this looks to me like Deacon's absential causation. The higher level is shaped by what the lower lacks um and as opposed to the basian influence of her hidden states. And I'm wondering if the the marker of blanket formalism um claims these are equivalent or is it agnostic about the physical substrate of the coupling. Thank you.
>> Right. Yeah. I did say I haven't got any foundational training in philosophy. So you may have to unpack uh some of those uh some of those concepts. Um uh >> well basically sorry I I've um I've written the nine requirements of um a holon a biological holon and one of them is that they all have leaks and lacks.
They have the laxs that deacon describes but then they also leak and that coupling um is what uh uh allows each level of the holon to um to fill in what it's not able to create on its own. What it's not able to close on its own. I hope that makes it simpler >> right a little bit. Um well I'm clearly going to have to read a bit more to give you an informed answer to uh to that question. But certainly the the um the filling in I was muted there an echo. Um >> that was my fault Carl. I'm so sorry I messed up.
>> No problem.
Yeah. Um so the yeah the the filling in um does certainly remind me of um the role or one way of describing basian inference um in the sense that if you've got a um the internal states of anything uh it could be a sort of the internal states of a mitochondria or it could be the internal states of a cell or it could be the internal states of a brain at different scales. uh will certainly do a filling in um in terms of optimizing its uh implicit beliefs um um about the outside of uh or the other side of the Markoff blanket, the hidden states that are hidden behind the Markoff blanket through this process of inference to the best explanation. I mean I'm being very teological and anthrop anthropomorphic here. Um but you can certainly read self-organization at any scale as doing a kind of filling in that inherits from having um uh a compressed representation that is um has the greatest evidence in relation to to the data. Um and that and that um interpretation inherits from the fact that the the log of the evidence can always be expressed in terms of accuracy minus complexity. If you remember one of the slides which basically means that the the best explanation or the best phenotype or the best sort of um intracellular dynamics is that which minimize provides an accurate account of um the um the sensory states to which it is exposed but as simply as possible. Um and part of that simpl um minimization of complexity and part of that simplification means that you've now got a compressed representation that effectively can be used to fill in. But I got a horrible suspicion that's got nothing to do with the filling in that you're talking about. Could I So who is it? Mary. Yeah. Could Could I ask you just to really dumb down answer just ask your questions if you're talking to to um to your children?
>> I always have to rewrite my questions in what I call 7G 7th grade for myself. So here it is. My question is um uh you do a mathematical modeling of the relationship between the um holonic levels the nest on nested levels of the marov blanket. Um I see that um every level every holonic level in life um is lacking things and it's and it's leaking things. So um and again this relates to deacon. So for example um uh in a a tissue um like the lung tissue it's leaking um um carbon dioxide but it's also lacking certain nutrients that come from the blood and so I see that every holonic level has both leaks and lacks and it's only able to maintain um a a closure in terms of its agency if it has things coming into it and it's contributing um materially and other and also thermodynamically and it's also contributing to the other levels and that I see this is the coherence between the coh the the um holonic levels and I was just wondering how that relates to the mathematical modeling of the relationship >> between the levels that's it >> oh well that that I understood that that was brilliant thank you very much >> sorry about that I I get um I think I get a little over enthusiastic about the language when it could be simple Right.
It's always good to learn to learn the right the right words. Um no well in a sense that is the role of the Markoff blanket. Um so you know very often I think people conceive of the Markoff blanket as some sort of insulation that isolates you from the world. It's not.
It is the interface. It is exactly though that that sort of structure that allows for exactly the sort of the leak and the lack to um to be mediated by a birectional exchange across the mark of blanket. It is exactly that and you know if I was reading the dynamics from the looking it through the lens of um thermodynamics this would entail an exchange of energy and entropy across across >> Exactly. So you know the markoff blanket is is not something that carves things uh and separates things. It's something that glues something together with exactly. So in a sense the bipartition of the blanket into active in sensory states is exactly this um um this complimentary l lag and and uh yeah >> and I'm glad you brought up thermodynamic because it seems like it works materially energetically mathematically and in terms of prediction which brings in Rosen.
>> Good. Yes. I mean, you know, one way of um one way of articulating the um that Helmho decomposition um is in terms of basian filtering which is um the most popular in instance of that is actually predictive coding and then you get into the whole world of predictive processing. So there's a really intimate relationship. I mean I repeat these are sort of teenological interpretations you know where you are committing to a particular teenological interpretation of these dynamics and the dynamics are just what they are um >> I'm all for teiology I believe that one of the nine requirements is kannotus or desire so >> about it a good another two good words there I did recognize those and of course that speaks to the expected for the energy and the that's why we call them prior preferences you know we shouldn't really. But yeah, that's that's exactly our level seeking.
>> Nice. Awesome. Thank you, Mary. Richard, >> thank you. This is great.
>> Thank you, Carl. Um, you know how sometimes when you're watching something on YouTube and you think this is going a bit slowly, so you put it at one and a half times speed. When I listen to the recording of the presentation that you just did, I'm going to be looking for something that slows it down by a factor of 10 so that I can take that in. Um I uh was interested in um understanding a little bit better the potential symmetry between the inside and outside of a system. Um, in one of your answers to Charlie, you said that if a system wasn't trying to minimize free energy, or perhaps you didn't quite say it like this, but if it wasn't successful at minimizing free energy, uh, it would, it's as though it wouldn't be a self, which means that the identity of the blanket would not be stable, right? that the the the blanket wouldn't be what you thought it was. Which seems like it's another way of saying that um there's a dynamics here where systems find selfhood by allowing their blankets to move until they find a place where the blanket is stable. until they find some sort of symmetry between inside and outside where they're doing a good job. The inside is doing a good job of modeling the outside, which must mean somehow that the outside is also doing a good job of modeling the inside. That there's a symmetry there. And that's the position where the blanket is stable and and the self is a self and before that the self is in transformation, right?
that not just that the insides are changing but the identity of the self is changing and that symmetry um if I'm if I'm right it's a sort of a a different way of a different starting point a different uh turning the story inside out instead of saying imagine that you have a self what does it want to do it wants to model the world it wants to make things predictable another way of thinking about it is imagine that you have a system with parts that are in relationship where the identity of one part self and non-self is not yet stabilized and what happens in that system is that the relating finds a finds a an interface of symmetry is is that a an equivalent way to think about what you're saying am I have I got the right end of the stick >> yeah no absolutely in fact um you know if I was um trying to explain the pre- energy principle at that level. I I I'd certainly start with the second that that would be the preferred u the preferred one. So I mean the free principle um it you know it's almost taught logical um you know in what it is. It says if something has attained a certain kind of synchrony with with everything else um then it can be described as behaving like this. It can be described as seeking evidence. It can be described as having preferences. It can be described as predicting. It can be described as self-evidencing. Um it's not saying that to self-evidence is um the reason. It's not saying that to survive you have to self-evidence. It's just saying that things that survive are defined in terms of um self can be described as self-evidencing. So I prefer the sort of the more deflationary um aspect. Um well I say I prefer it is easier I think to um motivate that perspective and just to sort of um celebrate your your sort of notion of this um symmetry between the inside and the outside and there are multiple ways we could talk about that because you know if that's right as you are trying to predict and understand and model your environment the environment is trying to predict and understand and model you when you get into things like desire paths niche construction and then you sort of put the um put other things like me into uh the niche. So the niche now is actually all things like me and then you get sort of cultural niche construction and more importantly a move from a generalized synchrony or synchronization of chaos between of a of an asymmetric soul between me as an individual and an environment that doesn't look like me to um a um an identical synchronization um where you and you are my environment and I'm your environment. So very much at the moment you know we are in discourse um we um you're my world I'm your world um and we are sufficiently similar in our sort of generative models and phenotypic construction that and I have a very well- definfined Markoff blanket courtesy of zoo um that uh that we can now find our free energy minimizing solution. So if I was talking to somebody didn't know anything about self-organization or basic statistics, what they would read from that is that you are finding a um the um you're finding the dynamics that are collapsing onto the synchronization manifold. So that would be a dynamical systems perspective on the symmetry and the synchrony that you're talking about. So technically a sort of generalized synchrony or synchronization of chaos that lives on the synchronization manifold that is the free energy minimizing solution. So after a period of time then we will come to synchronize and we will estab blanket that you know the the you know the um this um um manifold that sort of um separates us but also joins us in some in some fundamental way. So from that perspective um you if you just integrate any set of random differential equations uh introducing um sufficient symmetries between the inside and the outside you will get exactly the the synchronization of generalized synchrony and indeed identical synchrony almost that we are currently engaging in. So you know we are in conversation um and you know um apart from the turn taking we're actually rehearsing exactly the same narrative as I try to understand you you try to understand me and then we communicate it so we are in synchrony in some sort of you know at many many many different levels the final point which I think your question speaks to is this notion of transiencece that we have to find it you know we're not just given the mark of blanket job done um we come into being and at some point we cease to exist. So the Markoff blanket is always in flux from the point of view of the level above. So from the point of view of the scale above uh markoff blankets come and go. There is no agotic steady there is no urbanicity and there is never uh a truly non-equilibrium steady state. It's never attained by the time you get towards it. You've disappeared.
You've literally dissipated. So the failure to um to look as if you are self-evidencing just means that the Markoff blanket has dissolved. That just means you are dissolving be absorbed gently back into the universe. You've ceased to exist as something that can be disambiguated from from the universe.
And that happens at every scale all the time to everything. Um, so you know, as I was talking to you, I'm sure I've lost about, you know, 230,000 postsaptic specializations in my frontal lobe. But that doesn't matter from the point of view of my brain or me. Um, but from the point of view of evolution, of course, I'm going to last for, you know, you know, a blink of an eye, you know, over the few decades that I'm allowed. So we we we all come and go. These markup blankets form and and and dissipate and form and dissipate. Um um so it's only while your markoff blanket is in play and you are identifiable as having characteristic states in virtue which you have to have in order to have the markoff blanket uh partition that these rules apply. um when they don't apply anymore you have to move to the next scale just to you know to talk about what is concerned but yeah just >> so if it's the case that at a given level of organization the relationship between agent and environment finds a a kind of generalized synchrony when you have a um nested system with many different levels of organization my intuition is that those nested levels in order to talk to each other in a meaningful way to not just be noise to one another that they would need to be harmonically related to the frequencies at at other levels. Is that does that match your intuition too?
>> Yes. Um and um but but but in a here now we will have an asymmetry as you move from one scale to the next scale.
>> Yeah. Yeah, the there certainly the the the the um and literally the imaginary part of the um those um values that I talk about when you do the iggon reduction for the for the reormalization group operator they are literally the frequencies they detain the frequencies of the harmonics but you only you only retain when you go to the higher level the slower ones. So right it's not just that the higher levels are slower frequency but that they're particular multiples of the wavelengths you had at the others at the other levels. Yeah, they are literally linear. Well, they're locally linear mixtures of of values that literally are. Yeah, absolutely. But the the the um the the when you apply the um reormalization group, you you truncate the harmonics if you like or the iggon uh the iggon functions so that you're just retaining the principal ones that have these slow frequencies and slow times that don't dissipate quickly. Plus the rest of them have very large negative real parts which means that they disappear almost immediately from the point of view of the scale above. So they become random fluctuations. So all that's left are these slow um slow harmonics that are the sort of the modes of the system at that level at that level of description.
Yeah, absolutely.
>> Excellent. Thank you.
>> That was beautiful.
Okay, Wayne.
Hey Carl, uh thanks for uh thanks for uh giving us this talk. And there's a million questions I get asked. I'm just going to go to one of the ones I uh put in the chat. Uh I'm wondering uh how you know are you familiar with uh Robert Brosen's work in anticipatory systems and life itself and uh perhaps some of ah Lou's uh derivatives of that where he he sort of uh clarified things and went a lot further along the same lines and uh do you think uh like how do you think uh FE and active inference might fit into that frame? framework.
Um, and do you think they're compatible?
>> Right. Um, again, you're going to have to forgive me. Um, one of my mentors, Jerry Adelman, once called me an intellectual thug because of my lack of scholarly background, so I I don't I know I know a lot of of things, but only about a limited uh limited set of domains. I'm not absolutely sure what you're referring to, but I did pick up on the words of anticipation and certainly um that would be um you know a central pillar um of the predictive processing reading of of active inference that can be attributed to these uh um more deeply structured kinds of um phenotypes or or multisellular organisms or particles particles in in general. Um and um you know so I if it's well perhaps you could do me a favor could could you could you just sort of summarize what the main insights were from this literatur so uh I mean Rosen made a bunch of very impressive arguments. Uh so in anticipatory systems uh he sort of derived what uh what properties a system would have to have to be able to anticipate uh and you know so he came up with something he would call rows and complexity uh where you know the the your system is going to be self-referential for instance so it's not it's not going to be amendable to computation for instance right secure touring computation would fail uh for an anticipatory system. U in life itself he makes some uh some deeper claims based on some category theory arguments and his MR systems but I don't think I want to really go into that because it's going to end up being perhaps too much for I now recognize Rosen. Yes. No, I have I have come across uh Rosen before and um um I don't know his work deeply, but I do remember thinking at the time this is exactly right and and internally consistent with the um >> um with with the free energy principle and I also remember thinking in fact I thought then um this is also very reminiscent of early cybernetic formulations um of things like the good regulator theorem and um the law of requisite variety that you that sort of echo this. Um, and I think the the interesting thing you just said there that this can't run on a Turing machine with the sort of um um you and the self-reerential aspect of course um you know the whole argument I was trying to make about what defines an agent or what is agency. Um it is um it is having a generative model of the consequences of my actions. So it's implicitly self-referential. Uh it's not your actions. it's just my private future.
Um, and I've got a model of my private future which is quintessentially anticipatory in nature. So that particular aspect of um self-evidencing in these deep structured things seems exactly consilient with um um with um with Rosen's um u um um notions. Um the um and why can't it work on a Turing machine? I think that's an interesting thing that um um >> well halting problem right that's I mean >> really you you get you get endless loops >> yeah um and there is um an orthogonal argument um which is much more sort of dull-minded but practically possibly an interesting one that um your sort of cheering uh cheering machines um um are exactly the kind of thing that can't exist under you know under a natural science formulation of the kind the free energy principle aspires to um in the sense you've got sort of you know infinite memory and infinite compute and the like. So it's you a beautiful construction. I'm thinking here of things like universal computation um and the like um and so on of induction. U I mean it's a beautiful mathematical thing but the whole point is you can never realize that in reality when you actually look at self-organization in real discip uh structures. So that's that's certainly one. Um >> yeah, when I was thinking the physicist who made this this kind of argument, he was actually uh he was actually uh critiquing Roger Penrose's book uh on his strong argument again his argument against strong AI and his argument was that you know he he actually does a lot of coding because he's a uh kind of he I think he's a electrod or QED physicist and he and he uh he said you could never actually construct one of these things.
He he did a whole lot of talk in this that you know it's great you know it's a great theoretical argument but you actually can never construct this can be done not for the reasons that that you outlined because you didn't have that to work with at the time 30 years ago maybe but >> it reminds me of um the the wonderful thing about sort of um my understanding of um I repeat universal computation or current excitement about uh about um taking this algorithmic complexity to its limit in terms of that it is uncomputable. So people talk about universal computation but by definition it can't be computed. Um so that always makes me smile. Um but there there's a there's another interesting take on that which um is due to Jeffrey Hintter which is a notion of mortal versus immortal computation. Um so his argument here is that the software of the kind you you could run on a one human architecture and you know potentially a Turing machine um is immortal in the sense that it can be run on every machine. But of course that's not an apt description of real life computation or real or or agents uh in in in generic uh in general in the sense that they are examples of mortal computation which he would read um as substrate dependent. So the computation is in the substance is in the substrate itself. So when you switch off the the mortal computer it dies. You can't switch it back on again tomorrow.
>> Yes. Like 11's poly computing. uh type of thing. Yeah. Yeah. To me the I would my view is that computation in the touring sense you know at least the finite touring sense is one of the things that living systems can do. What we've done is we've abstracted that and we're sort of reimagining that as as the only thing they do which is that was kind of what Penrose's argument was also is that what we do is more than touring computation.
uh what type of computation it is, you know, we can't say at least yet.
>> Nice. Thanks for >> Yeah, that's wonderful. I just want to make sure I don't miss anyone. Um Alicia, did you have your hand up?
>> Oh, go ahead. Have Stuart ask his question. Okay.
>> Yeah, Kaden. Kate and Stewart, you guys are up next. Thank you.
Hi. Um, I was the one with the handout because you put Stu to the back of the queue, but I'll share.
>> Um, Carl, I want to say it's such a delight to meet you. Um, I the more I understand your language and what your model is saying, the more I realize that you are a common kindred spirit and um, uh, one of my new heroes. However, when I first encountered your stuff, I was kind of horrified because um I I'm I've been studying the emotional system for 40ome years and uh one of my big beefs is emotion is pretty much missing from everything. Uh so um I always thought that any kind of machine metaphor, even one as good as a computational metaphor, is still a machine metaphor. But this one is really pushed right up against the exactly the kinds of differences that we experience as embodied agents uh fully embodied agents. So um I've backed off on my that is I still think that it's too neurosentric. I still think that it's ignoring all of the deep embodied introsceptive processes that come from the bottom up in terms of um stochcastic this and that, but it's much more orderly than that. Um what I want to say is I came at your work through Mike Leven's work. Um, and I'm wondering I'm wondering how um the concept of bioelectric fields that are kind of synonymous with your what um Markoff blanket um how you're rooting that maybe in quantum field theory in terms of the fact that the markoff blanket has internal information and external information. So uh that's one of my questions but the my most important question has already been raised by uh well several times and it's the question of identity. Um the the idea of self as an automatic you know self generating this or that is great but the the concept of identity I think is really important here because the marov blanket is essentially the partition the boundary between internal and external.
So, um I I think that there's extra work that can be added on uh based on what the emotional system is doing. And I'd love to talk to you about it someday, but for now, I want to um talk about resonance and brain waves because if we're going to look at the brain, let's let's do it. So, you talked about how the the slow wave ien states are kind of the reference point and in my work, I talk about that as the resonance at the edge of chaos. and how we're actually perceiving these things and participating directly through our emotional sentience, what I call the the self-regulatory sense. So um my question for you is about the d the various brain waves and how the theta wave which is the very slow actually when we're sleeping or when the mind is dipping into the unconscious if you will um that has been associated with emotion and it's also been associated when it's disconnected from a gamma frequency with psychiatric disorders. So that's the pointed question I'd like to ask you about how how this uh jibes with this the psychiatric dimensions of it. But I really would love to talk to you sometime about how emotion and veillance in particular can enhance this model and make it more lifelike.
>> Wrong question.
>> Yeah. Yes.
I I could spend an hour unpacking that question and so so many intriguing uh aspects to it. I I don't want to I won't jump to to the specific um the last question because I think it's worthwhile just re-emphasizing the importance of emotional inference and intraceptive inference and I think you know if you'd if you got something like a Neil Seth um you to one of these sessions um so Neil Seth would be like me but is much more sort of beast machine invested in the intraception and the embodiment um aspects of it um but you know at the end of the day he would still I think be appealing if he if you forced him to the sort of predictive processing gloss on this kind of self-organization self-evidencing so I think it's really important that you know this physics of sentience is not about um visual input in a sort of you know um um machine learning vision uh sort of machine vision context you know 99% of the useful information about me comes from my body so you the the intraceptive inference is absolutely crucial and of course the you know again if you if you allow me a sort of teenological um um description of what's going on you know if I am self-evidencing where I'm going to be looking um for evidence is my body it's literally my gut feelings um because one of the most important things I have to infer is my state of mind at the moment and that presupposes of course that I have the hypothesis or um the representation that I am me. But let's assume that I'm sufficiently deep and structured that I have at least a minimal selfhood and more importantly a selfhood that has a granularity that allows me to recognize that I am in this state of mind. I am embarrassed. I'm in love. I'm frightened. I want you well I'm thirsty. that's not strictly speaking in emotion, but it's pans pansy in uh um in its uh way that it speaks to um interceptive inference in emotions.
So, you know, that's probably one of the most important jobs of this self-evidencing is just to recognize what state of mind I am in and implicitly all my attentional intentional sets that are now going to color and shape my self-evidencing and my actions and behavior in the future.
The the sort of you know you also mentioned veilance and effective. So there is a little literature on effective inference under the free energy principle and also emotional inference and intraceptive and to me they have very distinct meanings you know um effective inference is just is it good or bad and that usually reduces to the level of uncertainty about what I I'm going to do next. So you know there there are many sort of probability distributions associated with um active inference readings of neuronal dynamics.
Um one of the most important I think is basically um the belief distribution um the belief structure over what I could do next from which I have to select and if I am uncertain about what to do that seems to be often associated with angst and anxiety and negative veilance and then when suddenly oh I know what's going to happen to me I'm going to um you know have a cup of tea uh get a reward get a fruit juice if I'm a key and so I know the the future and of course that future is conditioned upon what I am going to do even if I'm doing nothing that's still a very active thing that I select then that uncertainty resolves and I feel good and that has a positive veilance and there's an interesting literature coupling that to dopamine uh release you can sort of simulate these things so uh I'll stop talking about that because that wasn't your question but I just wanted to acknowledge that you know um intraceptive sources of evidence and the embodiment is absolutely crucial. You can also apply this uh this theorizing to the immune system. So it doesn't have to be you don't have to be neurosentric. Um you can be >> um you can also be um you can look at different parts of of the body and and you know tell a similar story. But just to come back to your specific question um the the um the the theta um and gamma and brain waves and frequencies um then um that's why I was sort quite careful to emphasize the role of that Q that selenoidal um aspect because what that means is you can't have this kind of biotic self-organization without frequencies at all frequencies. is everywhere all the time. This can't it can't be done otherwise you'd have a point attractor and you'd be dead. Um you you using them so you are frequencies important? I mean well yes absolutely because if if there weren't frequencies they wouldn't be you wouldn't be there. Um and also these frequencies um you know in that conversation with Richard uh there's a certain nesting uh which is important and you're picking out I think a really important sort of frequency band gamma and theta in the brain which is clearly um um an important consideration when it comes to things as you say like your research. I would argue that the theta rhythm is particularly important though um in a in a special way. Um, you know, we were talking before about sort of, you know, a reciprocal exchange with the environment. Uh, and in our conversation, my conversation with Richard, we talked about turn taking.
There's a similar kind of turn taking that occurs in active vision, active perception. Um, in the sense that, um, I will act upon the world. I will palpate the world in some way. I will query the world, ping or prompt the world in some way. and then I will listen to how the world responds um and then do my belief updating through this sort of free energy minimizing process that seems to have a saltatory aspect that occurs at theta frequencies. So for example um if I um look at the way that I actively sample the visual sensorium through psychic eye movements I do that at the theta frequency. If I'm a little mouse, I'm going to sense my world by moving my whiskers at a theta frequency. I am talking where the phone names are emitted at a fre at a theta frequency. I am listening at that frequency. So, it seems to be quite fundamental for animals uh um you know of of roughly our size. that this is quite a fundamental kind of frequency which sort of suggests that the gamma frequency is what's happening in between these um sense making episodes after you've moved or positioned your um effect your sensory organs in the right kind of way to gather the right the right evidence going to resolve the uncertainty or respond to those epistemic affordances at every scale. uh which suggests that you've also got this sort of oscilly dynamics in the belief updating uh process itself. Um and the interesting connection with schizophrenia uh in my world is that um it may score the fidelity or the precision of messages from lower levels of the hierarchy in the brain to higher levels. um and most formulations that I'm aware of in terms of predictive coding and the um and the basian brain of schizophrenia is that it's a disorder if you like a sort of functional disconnection that is due to um a failure to modulate the gain or the precision of message passing amongst different hierarchies. And the you know the link with gamma activity in this instance is um that it it looks as if um that gain depends very much upon the excitation inhib inhibition balance that in turn depends very sensitively on fast spiking inhibitory neurons that that that orchestrate and create this the these um fast synchronous uh exchanges usually expressed in in the gamma range wi within which the uh the gamma figures are nested within the theta as you do your belief updating as you sample the world at different levels. So that was a sort of very specific answer to your question which I hope spoke to your particular interest uh but it was also uh I did manage to slip in a bit of a shell for embodiment and emotional interceptive processing.
>> Thank you. I just one final quick comment. Um is there positive emotion in your model?
>> Um yeah well any any resolution of uncertainty would certainly be um have a positive veilance. Um in my model uh well in the model when people like Ryan Smith or or and others apply th this active inference to emotional inference.
Um emotions are then arise as a construct. Um they so there are some people who have you know I say a kind of alexinder a dyslexia of emotions and they just they're just good on bad and other people have a very fine grain nuanced um representation of self a self model that then um so the the structure of the emotions may be very um specific to each individual um and the emotion I repeat is a hypothesis about my current state of mind um which is something I have to recognize in a James Langian kind of uh kind of way.
>> So would it would it be correct to say that you're saying veilance comes from the top down and not the bottom up?
>> I don't think I'd say that. I think whenever asked these questions the inevitable answer I arrive at is it circular causality. um you know I have to recognize I'm in a particular emotional state which means it's a completely bottom up process but the whole purpose of being in this emotional state is that I can top down contextualize attend to the right kinds of things you know invoke the right kind of sensory attenuation um or indeed act in a particular way so it's both bottom up and and top down >> thanks I would love to chat sometime go ahead Can you hear me?
>> We can >> absolutely can.
>> Yeah. Carl, um it's a deep pleasure listening to you. I have begun to understand your work. You've thought so long and so hard and so carefully. It's you you present an integrated formulation um that is years of your thinking and it's going to take me some time to take it in. I want to raise two issues and I hope I could state it clearly for this is for all of us Carl not not just you and me. Uh one is the conditions for open-ended evolution. So in the biosphere, life started something like four billion years ago on this planet uh with with uh first the proariots for many many many years and then we got ukarots then multicell organisms and we get redwood trees and tigers and leopards and uh bacteria and fungi.
We have not achieved open-ended evolution anywhere in the field of artificial life that my old friend Chris Lankan started. It keeps getting stuck.
I think it's getting stuck because it's algorithmic. But nevertheless, we have not. So in the evolution of the biosphere in a profound sense, new possibilities arise.
And in the evolution of life, we have always had to find new opportunities or new affordances.
So we're non-equilibrium self-reroducing molecular systems. So we have to eat. So to eat, we have to find something in the environment that's useful. And the problem is what constitutes useful for me, namely new relevant variables. So the fundamental worry with that I have with with what you've done which I think is wonderful is you've pre-stated the variables. One begins with a pre-stated phase space which is exactly the move that Newton makes. One has position and momenta and one begins with the boundary conditions of the system that define all possible combinations of position and momenta that defines the phase space.
Then you have laws of motion among the variables, Newton's laws and then you have initial conditions. You integrate the laws to get a an entailed trajectory. So theme one is that evolution seems to create new relevant variables. Um cardiac ejection volume was not a relevant variable three billion years ago, but it's a relevant variable now. How is it the case that the evolution of the biosphere creates ever new relevant variables? Question one. Question two, can we deduce the new relevant variables from what the biosphere was two billion years ago? My claim is no. So Andrea Roelly and I wrote a paper called a third transition in science. It's journal of uh uh journal of the Royal Society interface uh April 14, 2023. And I'll just say it's there if you want to read it. The second issue I want to raise is a different notion of identity. I will not refer to quality or experiences at all but let me say it briefly. Um I think it's useful. So what is work? Well, it's force acting through a distance. Atkins says no work is a thing. It's the constrained release of energy into a few degrees of freedom. Okay, fine. Think of a cannon is a boundary condition constraint. The cannon ball inside and the powder between the cannon ball and the the base of the cannon. The powder explodes and you don't get a spherical explosion. The gas escapes down the bore of the cannon and it does work on the cannonball.
So, okay, no constraints on the release of energy. No work.
And then I not being a physicist but I trained in medicine. Uh you may have too. Um I asked where did the cannon come from? Well, it took work to make the cannon.
Well, the cannon's a boundary condition.
This is very strange. Newton's got boundary conditions. He doesn't tell us where it comes from.
So, no constraints, no work. And at least sometimes no work, no constraints.
Somebody made the cannon. I got a little further. work can construct a constraint.
In 2015, my Montilla Mateo Mosio proposed constraint closure. It's a set of constraints that depend upon one another. I almost misheard them. And so I'm going to be very careful. My notion of a constraint depending upon a constraint is that constraint one constrains the release of energy that constructs constraint two. It's not just depend upon that constraint one is the boundary condition that constrains the release of energy that constructs constraint two.
Gonoshanazi has a nine peptide autocalytic set. Each peptide liates to have fragments of the next peptide to make a second copy of that peptide around all nine. The system is collectively autocalytic. So it's a contean whole. The parts exist foreign by means of the whole and it achieves constraint closure at the same time.
Peptide one binds the two half fragments of peptide 2. Therefore lowers the activation barrier for the reaction. The reaction occurs, work is done, a peptide bond is formed. Gon set literally constructs itself. I want to point to a new idea, the union of catalytic and constraint closure. I think that's a definition of life. Um, and it's a new union. It's a union of catalytic con and constraint closure. The system literally physically constructs itself.
There's no notion of experience here, Carl and the rest of us. The Leicia URO I think may find this happy. Um it it literally constructs itself. So I think that's a notion of agency and identity without yet uh appealing to qualia or experience. And the this system could be a nonlinear dynamical system. It could be edge of chaos. Therefore, it can categorize its world and act in its world. Um, and I think this system is capable of open-ended evolution because in a contean whole, the function of a part is that subset of its causal properties that sustains the whole. The function of your hearts to pump blood, not make heart sounds. But a part that's a physical thing can be co-opted so that other causal consequences come to be of use in a new function. So your heart can evolve to be an earthquake detector. And that's not deducible.
The there's no deduction of alternative uses of the same thing. And I think that yields for us open-ended evolution at I think this is probably right curl on all of us. And the question is in what sense can we mathematize it and it's not so obvious. So it's a huge whole turf and I'm fascinated to think how it ties into the way that you have formulated everything you've formulated which seems just beautiful. So I I'm I'm kind of puzzled and thrilled to be listening to you. I think that's a comment but in any case it's an opening of some kind. Thank you guys.
Well, thank you again. It would take me another hour to to respond to all the invitations uh implicit in that question. I just want to say it's a real honor to speak to you. Um before I was saying that Jerry Adelman um was calling me an intellectual thug and to cure me of that he made me read all the good and greats including people like Ernst May.
Uh and that's when I came across your works. I've I've always had a bit of hero worship for you.
>> Oh gosh, this is great. Yeah, this is >> you take one another out to dinner.
>> Everybody else comes along our treat.
>> So that's your ideas have been under underneath a lot of this notion of self-organization for decades now in some form or another. um in relation to uh you I can't answer those questions but I can certainly touch upon the two big themes where can you predict where the relevant variables will emerge in the future you know in terms of open-ended evolution uh and I agree with you entirely no um can you um um and the sorry I'm trying to formulate or paraphrase the second question Um, how can one mathematicalize?
Um, I should just also slip in. Uh, at the same time that Jerry Adelman was accusing me of being an intellectual thug, he also accused me of having mathematicsis, which he used to abore, but you're you're inviting me to indulge in mathematicsis. Um, so the way that I would indulge uh in that in that instance would be as follows. Um, I think that sort of the catalytic aspect and the closure aspect. um is um if you were to write that down in terms of a random dynamical system or you know more simply just the sort of differential equations you're speaking to um um a fundamental nonlinearity um and you know any sort of um any influence of the influence of one thing or another of a catalytic sort basically is a second order interaction. So as soon as you do that, you're introducing high order terms into the dynamics.
>> And if you remember before I was um careful to say that as soon as that solenoidal part of the dynamics becomes um state dependent or second order in the states, you now get biotic kind of self-organization. So the way I'd respond to that is the mathematical space that you would demonstrate you have to have this catalytic kind of closure um in order to be self-organizing in you know some meaningful biotic sense uh would just would just be to show that with and without those high order dependencies in the in the serenoidal part that the non-discipative part um is um necessary to actually reproduce the kind of dynamics that that we've been talking um the the closure part and the roll of constraints I I thought was really interesting as well. Um but purely um from um the perspective of a mathematical constraint. So if you wanted to um again indulge some mathematosis in terms of constraints um what I would immediately go to is the um the fundamental ro role of constraints in observational physics uh due to ET Jane and he brought to the table this notion of constrained maximum entropy and ultimately maximum caliber principles where the caliber is just uh a measure of the likelihood of a path through some state space and the like.
Um but the operation the heavy lifting is done by the constraint. Um so it's the constraints on the maximum entropy that actually give things their functional form and their shape. Um and when you write down so the free energy principle is dual to the constrained maximum entropy principle. But what the free energy principle does is it sort of makes very very clear what the constraints are and the constraints just are those afforded by the generative model. And if the generative model implies the existence of a markoff blanket, then it would necessarily then um mean that you have um there has the constraints. The constraints um are almost stipulatively um required for specification of the generative model. or read another way, the genetic model supplies the constraints that define the kind of thing any self-organizing system is. So you you you can't have a self-organizing system unless you can write down its pullback attractor. Its pullback attractor creates now a probabilistic generative model. That generative model is the constraint that coupled with the maximum entry principle now describes the evolution and the dynamics of the system at least as it is observed. So I think you know there are easy moves u whether they're pleasing or satisfying or not is is is another question but there are easy moves to accommodate both the autoc catalytic and catalytic aspects closure aspects and constraints within the maths of self-organization of the kind that people like priagene or etan would would recognize um um and interestingly also speaking to the previous conversation about oscillations and dynamics it, you know, all rest upon this sort of high order solenoidal dynamics that characterizes which you don't get in 20th century physics. You know, they don't deal with that. They just deal with equilibrium with point attractors.
It's only when you get into non-equilibrium physics and self-organization that these things uh really hit you in uh hit you in the face. So, I hope there was something else I wanted to tell you, but um I've I've forgotten what it was now. All sorts of interesting things came to mind. Was that sufficient? Um the the first one >> yes no it's not sufficient no I just wanted to acknowledge the unknowability of the relevant variables that will emerge in the future um you know this this is a real thing in uh machine learning at the moment um and it's basic also I would imagine a real thing in economics in terms of radical uncertainty just not knowing the space of possibilities how can you have any uncertainty over a space where you don't even know it's support. Um, >> exactly. Exactly.
>> Yeah. And in in machine learning, this is a big thing in terms of causal discovery or structure learning. And it's a really difficult problem because you can't use standard gradient descent um to explore a space of possibilities that you've never ever witnessed before, hypotheses that you would never witnessed before. So the structure learning problem is a a really outstanding problem and it's nicely illustrated by the sort of benchmarks that the machine learning AI community bring to the table to to show what's the next direction of travel. So there's a something called the ARC 3 challenge which is all about basically trying to get machine learning people and large language models to do abstract reasoning which is exactly exploring things that have never se been seen before. the true novelty. Uh that's a you know it's a really a really interesting problem. Uh you know that I repeat you can't solve um you can't solve by sort of uh gradient descents.
You you have to move into the world of discrete hypothesis testing and structure learning you know you know in the way that you and I do it as scientists and and theoreticians. It's it's a a very deep problem. So I agree entirely. You you you can't predict the unknown. You just have to wait patiently and uh admire it when it happens.
>> Well, this is lovely. And uh uh I'm I'm wondering, you have a bunch of people waiting and I'm just wondering about Alicia URO somewhere back there if she could say something. This is her turf as well, but other people have their hands up and I'm not the I'm not in charge, Charlie. You are.
>> Five minutes.
>> Oh, it's okay. We But speaking of constraints and closure, um I promised Carl that we would end at a quarter to the hour because he has another meeting.
So I just wanted to um you know say if Carl, is that constraint still with us?
>> Oh, I I only need 10 minutes. So you got another six minutes. If if you can do two questions in six minutes, that'd be wonderful.
>> Oh gosh, that that may be a challenge.
Um but but let's give it a try. Uh, we've got Henry and Bill in the queue.
Um, and >> oh, and sorry, and Alicia as well, um, was in the queue and then was out and back. But Ste just queued you up, so maybe there's something you can say quickly.
>> I just have a quick question. Would you be particularly upset if I tried to detologicalize uh the con the markoff blanket by saying ontologically it's a sort of constraint regime or overarching constraint regime that integrates all the various dimensions of the of of an organism metabolic endocrinological cellular biological ical cultural and so on into an overarching constraint regime and what I find that I don't understand exactly how it works in your view is how does that constraint regime uh come to be and my argument would you be that you have enabling constraints that are typically uh the autoc catalytic recursive processes feedback operating against the background of context independent constraints like gravity or or uh you know cos the cosmic light cone or whatever. And then they they in they interact such as to create an interdependent constraint regime but always always local and temporary.
And Katherine, would you be very upset if I said at other levels of organization, maybe some of those constraint regimes might get tagged. And I'm thinking of emotions and feelings as the tagging.
That's a terrible. I feel like I'm coping out, but it's a form of tagging in terms of veilance, tagging in terms of feelings that certainly minimizes a surprisal because all you have to worry about is danger. Get out of it or pleasure, go towards it. Uh that sort of thing. And so it's an emergent property that tags constraint regimes at different levels. Anyway, that's my thought. Would you be terribly upset if I called the markup blanket a formalization of a constraint regime?
>> Um, well, that's an easy question to answer. Not at all.
>> Made my day. Thank >> you.
And I I was mindful um of of your presentation a couple of weeks ago um and how consistent that was with this sort of scale separation of scales with you the kind of constraints that don't change with time and sort of the laws that we operate under and the ones that are and your your mention of sort of the transients the temporary that exactly was what I had in mind when talking about sort of the separation time scales and the transients of these markoff blankets operating on these contextual constraints. Yeah. just to follow and the interface of course is the embodiment of that Marco Frank that negotiates what gets in what gets expelled I wouldn't say lacking and and and needing I would say um trying to achieve a metastability according because it's an open system with matter energy information coming in and being expelled as waste or waste or whatever thank you thank you very much Well, guys, we we may have to bring it to a close uh with Carl and but um with a great round of thank you and applause.
>> Thank you so much. That was incredible.
>> Thank you very much.
>> Yeah. And I just want to say, you know, Carl, this this uh meeting is part of basically one long conversation that's been going on for about two years. And so I' I'd like to invite you to, you know, be part of that conversation and to to to come back and and join us, you know, just as a participant. Um, and and that's how a lot of us have gotten here.
So So, uh, thank you so much.
>> Well, thank you. That was wonderful.
>> And I'll keep the meeting open. Um, so maybe Henry and and Bill, we can, you know, we can continue to discuss those of us who have a few more minutes. And, uh, thank you so much.
>> Thank you all. Bye. Bye.
Thank you.
>> So uh so yeah I actually just have one quick question because uh uh so according to the FE model so we should consider biological system actually is judge them we are judging the overall capability of use information for correct choice.
So, so if that is true then the have two issue we need separate one issue is the biological system the capability of process the information right because the environment information in you make the right choice but at another hand so the the circle information self contact also very important so that's is the idea of why the species will be replaced right So from the you know bacterial you know ucel and multisellular every time no matter how successful the species are they will be replaced so that actually kind of address the question Dr. Koffman is asked is what is the ultimate driving force for the evolution. So we thought maybe is increase information complexity that is the major goal of the evolution process. Every time they place different species because every time they placed it turns out the system is more complicated and therefore they actually have more powerful to predict to make the choice.
Right. So therefore we can say at the very beginning they have limited this but with the brain formation everything that become more maybe in the future the future selection we have much more involvement for this direction human may design we go to or we can we can design use totally different mechanism to support our nutrition system whatever for example. So therefore the if that's the case we really need the thinking about how the you know information contact increase the complexity whether or not you know I want to ask the speaker he's ga the email so so to how to you know to thinking about those idea because the simple question is is bacterial or human which one is more effective in terms according to FE model to dealing Right? Because this is relative speaking but also is historically totally different. So so when people talking about cognition everything we have to come back to the original historical ecosystem rather than what we know right now to imagine the how the system is working. Yeah. I will stop here. Yeah.
But anyone else you have answer we can discuss.
>> Well Henry I'll offer a partial answer.
Um I would wanted to mention this to Carl. Um we had a paper being refereed uh we just accepted by bios systems talking about the epistemic limits of cellular intelligence and uh one of the issues that came up is our application of marov blankets to help explain those epistemic limits. Um the it's it's very confusing to a lot of people that Markoff blankets uh are present in all physical entities and what distinguishes living forms as you're asking what where where's the distinction in the living process um Francia and I came up with an idea of a biological marov blanket and that links to uh discreetly links to the plasma membrane and Katherine from Kate's comment about biofields, our critical biofields, and that links further to our concept of the phenome is a sensory organism of the cell. And so we believe that a biological marov blanket, a a distinct type of subcategory of marov blanket distinguishes prebiotic from biotic states. And the key is that the plasma membrane is itself discriminating and acts as the the interface the beginning interface of cognitive selection within the cell and that's that is the foundation of or of theformational complexity that you're talking about Henry >> did I understand him correctly to be saying that if if he didn't say this explicitly but I think I thought I understood that what he was saying the boundary between life and non-life is just sort of a level of recursion Yeah, >> the degree the degree of recursion of of predictive recursion is what makes >> Yeah. Uh I'm I'm sure that he would have said that. Uh I he said something else that's really very important that kind of flew um could easily fly by. A Markoff blanket and boundary is a sensory boundary. He made he made that as a distinct statement.
And of course that that caught my eye because we're talking now about a new concept of biological markoff blanket which is ex explicitly linked to the cellular syn to its actual sensory apparatus its totality of its ability to sense external environmental cues and from that information transits the external membrane as the beginning of this process and it gets measured internally. from which info computation and veilance derives and so we really have a strong linkage of how living systems work if we if we if we make the assumption that FEP is a very useful theoretical construct and it's actualized in biological terms at the level of the plasma membrane as a first interface.
So, so I agree I mean after the first interface and then the gradually build up use different mechanism that's why so you know for the congregation study we really need you know from the beginning and gradually every time the complexity increase and insured by specific molecular or physiological or evolutionary mechanism as a constraint every step to fix them up. So we can simply say because the coding become more and more more and more not complexity to deliver such a you know the accumulated interface as as such. If you >> I was a little I was still a little bit um puzzled. He had a he had a pretty direct answer for the dark room um problem and he was saying well the predictive system flips on the lights but um I I'm still I know it's super basic but I I'm still puzzled by that because to me flipping on the lights is kind of a metaphor for acting on the world to make it less predictable right and so if you're if your system that seeking a predictable world or predictable environment. And he even said, you know, natural selection is basian model selection, which I think was a really clear articulation of it. Um, if you're acting on the world to reduce uncertainty, there's still a sort of counterfactual there. There's a there's a um imagination of what might be possible if the lights were on. You know that that seems that variance if you leave the lights off you'll never discover.
But you'll also you'll you'll live in a more predictable world. So that leaves me that in some ways it leaves me with a deeper question than I had before the meeting. I think.
Anyway, oh guys, we're almost to the hour. Um, so I just want to thank you guys and um I'll go ahead and turn off the recording if unless anyone else has a an an um a burning comment or question. Alicia.
Oh, you're you're muted. Sorry.
You're muted, Alicia. I'm sorry.
You're muted.
Let's see.
>> I'm just saying thank you. I need to leave. I'm sorry. Thank you. Thank you so much. See you next time. Thank you.
>> And thank you, Lisa, for joining again.
Yeah. Great. Okay, guys.
Bye. Bye.
>> Thanks, Charlie.
>> Let's see there.
Ähnliche Videos
Recovery pronouns. Neuroplasticity & practical neuroscience tips to help recover from pain & fatigue
Fantasticneuroplastic
907 views•2026-05-31
No Eyes, No Darkness? 👀😱
Huwatif
630 views•2026-06-02
I Saw the Thing Crash. Then I Lost Hours | Beyond Black Budget
BeyondBlackBudget
148 views•2026-05-30
Physical vs. Computational Causation Explained #shorts
PhilosophiaVL
641 views•2026-05-30
Your Brain Is Actively Deleting Your Childhood Memories! 🧠🗑️ #Shorts #Anatomy #DidYouKnow
voiceless2345
225 views•2026-06-01
What are you looking at
SuperStaticPro
1K views•2026-05-31
Deep Pressure & Anxiety Explained
OccupationalTherapyForChildren
145 views•2026-06-01
Size Illusion
WTFactt_t
1K views•2026-06-03
Trends
Why Batman Lets The Joker Live 🤨
zackdfilms
9222K views•2026-05-30
They're Complete Trash
penguinz0
558K views•2026-06-04
The Murder of Deputy Caleb Conley
MidwestSafety
810K views•2026-06-04
I Bought FAKE HopeScope Merch (and paid a subscriber to give it a makeover) | Hopeful Hauls
HangWithHopescope
158K views•2026-06-04