Living in Flow with the most famous man in neuroscience

TRANSCRIPT

In Holiday Flow with Karl Friston on Love and Philosophy: A winter walk through neuroscience

Karl Friston: ​[00:00:00] Doing what you're told is not always the best idea,

 make the right choices on the basis of what you're being asked to do at each point. And as long as those choices are contextualized with an overall direction of travel, but not too worried about the particular choice at hand, then you do quite naturally flow along so I think that's, for me, flow and in the service of reaching a conclusion is just a description of existence. And it would look like things are measuring things, it would look like things are, observing things. And crucially, it's not that we're just passive ingesters and assimilators.

of the flow of sensory input, we can now act upon the world to sample and direct that flow. So you've now got this interesting circular perception action cycle, the circular causality between the inside and the outside, where I'm trying to make sense Through a process of [00:01:00] inference

if I can do that gracefully for a long period of time, I am basically in tune with my environment, with my world. keeping your mind open

and finding the simplest explanations that's the best way to, reach an understanding or explanation, for the world.

Andrea Hiott: It's too easy to mistake the model for the reality and you think you're fighting about the reality But you're fighting about the model

Karl Friston: these would not be issues if it was only me on Mars. I wouldn't need a theory of mind.

But as soon as there's something like me, like you, I now need a theory of mind.

 Because we're not talking about one particular state of being.

The world is always moving and as you say, it's not about, Finding the perfect, state of being it's finding the best fit to a constantly changing world and it's always constantly changing because you're changing it for me and I'm changing it for you.

So as soon as we put two things like us together, we guarantee that we're going to be learning from each other and constantly changing.

we're always [00:02:00] chasing away uncertainty, always chasing away expected surprise, always chasing away epistemic entropy, always on the move.

So this is what I meant by flow. Flowing on this attracting set or attracting manifold, where that manifold or that attracting set is all consistent with the kind of thing that I am. And it's incredibly itinerant. Technically, space filling that has low volume in the sense there's only a very small number of all the states I could be in that are consistent with me living and being alive.

Andrea Hiott: Hello, everyone. Welcome to love and philosophy beyond dichotomy. Where we try to look past traditional divides and divisions. And notice the pattern that connects. Today is a holiday episode, or even Karl called it the birthday episode, because this [00:03:00] has been one year since I've been putting these conversations online. And it's also the end of the year. 2025 is almost here.

It's holiday season. So this is holiday flow with Karl because this word flow as you're here is very important in this conversation comes up from the beginning and goes through to the end. And connects. Some of the terms that we're using also this tricky way that a lot of these terms are everyday terms, but mean something a bit different in mathematics and the neuroscience and

the way that Karl uses them in his work. Words like energy value, entropy. Belief. Connection surprise flow. We talk about all of those from different perspectives here. And I try to think about them. In a way that can be understood from whatever side you might be coming to, whether you're thinking of those in terms of mathematics. Or just in everyday life. Karl [00:04:00] first-in is maybe the most famous neuroscientist in the world.

 That's how he's often described. We talk about that and why that is as it is some of the inventions that he's come up with over his life but in the YouTube world or the broadcast world, he's most known for this idea of active inference or free energy principle. Which is a first principles account of sentient behavior, I guess is the short definition of it. He's trying to understand. What sentient behavior is and how we can go to the beginning of it. So to speak. Although we talk here about maybe not thinking of this in a linear way, but what connects life and Cynthia and behavior. And how can we understand that? As you know, this podcast is a lot about patterns that connect.

And I actually see a lot of the math, which is very difficult by the way. And we don't talk about math here too much. Certainly we don't talk about it in any way that you need to understand math to listen, but his work Karl's work and active inference in the free energy principle are very mathematical. [00:05:00] But I see this as a sort of maths of how we might notice the patterns that connect at different multi-scaledareas of behavior or cognition or life, depending on how you want to use those different words. And by that, I mean that it's a way we can externally represent these processes and activities so that we can see the patterns that connect them and what they have in common. What's different about life and non-life for example, or how we connect something like the way we move through the world to the way that we think through the world. I try to bring in some of those ideas here and ask him about them? And as you'll hear, we get to some really interesting places because. There's actually a lot of paradox being held in his work and in the math that he's doing. There's this kind of math got Bayesean statistics, which I think within itself is able to hold paradox and part of how we can now maybe have mathematical models [00:06:00] that do this, that help us understand the patterns that connect.

So we talk about that here. He explains in a way that I think is very understandable. This almost contradiction at the heart of the free energy principle and active inference and a lot of his ideas, which is that we're both minimizing and maximizing energy or entropy, I'm going to let you listen to him and to the conversation to understand that, but I just want to note it here that it's one of those patterns that I'm talking about and that he explains very well even though you might have to listen to it a few times or read some of his other papers before it starts to really fully open. But even if you don't want to read any of these hard papers or think about these difficult issues. You can listen to this conversation not needing to know all of this stuff, because. What we talk about is what that model is showing and it's this idea of flow of how does one stay in the sweet spot, which is to say, stay alive. And that is through this minimizing and maximizing [00:07:00] of energy and entropy at the same time. And at multiple scales and that has a lot to do with how we think about the body. What we think of is life. That's why this can seem like a theory of everything almost like you just can't handle to think about it because it's so big, but in this conversation, we walk through these ideas in a way that's very understandable. There's so much more here than I could really get to. For example, he mentioned Geoffrey Hinton and Peter Dayan and Jerry Edelman. All of those are figures.

I would love to have talked for an hour with Karl about because there's a lot of history. I'm going to link to some things that you could read. If you want to know more about that history, we do touch on quite a lot of it here. I think one of the most interesting parts for me is when we talk about what philosophers will recognize as a common problem or issue that is often talked about in terms of what is alive or not alive. Why isn't a thermostat or a Watts [00:08:00] governor, or some sort of machine that's functioning in an organized way, not able to be called live. In my work, I link it to affordances in the thesis I'm writing. Affordances are the way we can distinguish these things and solve some of these bigger problems. I bring that up here with Karl

 He

seems to agree with that, or it seems to fit very well with his work. Even though it takes a bit of time for us, for Karl and I to become clear about exactly what these words are, meaning as we're using them. He also says at some point that inference can be exchanged with computation or measurement the word itself because I push him a little bit on the idea of unconscious inference, which is an important word in his work. And it's a term that could also mean unconscious conclusion, for example. And Karl says that inference could be exchanged with computation or measurement or any of these words. And this points to another theme that we've talked about a lot here on the show, which is what's the difference between the [00:09:00] model and the process that's being modeled. Especially when we are the process doing the modeling of our own process. This can become very meta, confused, hard to understand. How am I as a person able to model my own patterns or habits to myself. And within that entanglement what's the model and what's not the model. And all of this has important repercussions in real life, because it starts to matter how we think about what the brain is what the body is. If it's a machine. If it's not what's objective. What's subjective. Which one of those sides should be taken seriously. All of these are very important themes for holding paradox. And we talk about those here in a very particular way. There's a lot of answers here to a lot of big questions if one is willing to be patient. Try to understand where Karl's coming from and uh, understand this line of questioning about how I'm trying to understand what the space is holding binaries. We also talk about listening and Karl says that the secret to success is doing what you're told that's another [00:10:00] thread that runs through this conversation that's very interesting. And they will become a little bit like holding the paradox by the end. Because being told what to do as Karl, eventually unpacks is actually being able to listen to what other people are saying to you. Relative to where you are to your perspective. This conversation connects to the last one with affordances, with Harry Haft and to the one coming next, which is with Evan Thompson. Were we discuss in a very pointed way a lot of the ideas that I raise here with Karl about subjective and objective. And the difference between the process and the science that's modeling or measuring that process. But I just want to say from the beginning, these are hard subjects and a lot of people are trying to figure them out in the best way possible. Karl has done so much work on so many levels. You'll hear about a lot of it here, but there's even so much more. He's really given a lot in all sorts of fields and disciplines. He's his dedicated his life to trying to articulate what I think of as the pattern that [00:11:00] connects. The work or the writing about the free energy principle can get criticized. For what I was just saying that it's taking the model for the territory which by the way Karl is definitely not doing, he's always saying this is just a model of the process. But because our language is so tied up with these binaries, that can be very hard to even talk about it without lapsing, as he says, back into that language and so it does get very confusing, but Karl himself is clear about it.

And some of the things I push him on here are just to clarify that and to make it understood for myself and for listeners. Because there's a lot of insight here that can be a gift if we are able to understand exactly what these terms are doing the work they're doing. And understand it as a way that we can see how these different patterns of life and cognition and consciousness connect at many different scales within our own lives, within our own bodies. But also within the wider [00:12:00] ecological world. So thanks for being here and I wish you a very beautiful holiday season. Wherever you may be. And if it's not a holiday season for you, I wish you a wonderful couple of weeks until I see you or hear you, or you hear me or see me again thanks for all that you gave this year. Thanks for becoming part of this endeavor and I send you lots of love and warm greetings.

Hi, Karl. We already said hello, but thank you so much for being on Love and Philosophy today.

Karl Friston: Oh, thank you very much for having me.

Andrea Hiott: To get started, I thought I'd ask, what's been on your mind over all these years?

You're humble and you often say, no, I'm not. But many people have told me or say that you're the most successful scientists of our times and so on. And you've worked on a lot of different things. But before we get into that, I wonder if There's been any one problem or idea that stands out for you as something you Had at the beginning of your career that you [00:13:00] still have

Karl Friston: that's a nice question to start with.

I think the secret of Success and certainly my success is to do what you're told So the very fact that I have got involved in so many things. It's because Most of my work is in response to collaborations or interactions, usually with sort of younger people, with students, uh, to whom I am responsible.

So, and they go off in very, very many different directions and one has to follow, uh, to a certain extent. Um, has there been one overriding ambition or question? Yeah, there certainly has. Um, you know, I started my career. Um, or designed my career really, or made the choices that enabled me to study, um, arrogantly is, um, Probably, uh, the most interesting thing in the universe, which is me, or at least my brain.

Uh, so, from a very [00:14:00] early age, uh, I was looking forward to a time in life where I could, I could just focus,, scientifically, um, well, on many levels, um, on how, uh, I work or how one's brain work. And of course, that necessarily entails an understanding of how other people's brains work. Um, so that had implications in terms of, um, my journey, specifically in terms of, um, But doing both, mathematics and physics to formalize that research and that understanding, but at the same time, also going into medicine and psychiatry and qualifying as a psychiatrist.

So, um, in later life was able to combine these dual routes to, to, um, self understanding and probably more importantly, other understanding. I'm glad you

Andrea Hiott: brought up those different roads that you've met or have been part of your life, but as you were talking too, two things came up. I was wondering if you ever [00:15:00] say no to things because I know a lot of people you help and work with.

So one wonders how you can possibly do it all. You seem to say yes and somehow you're able to do it. Connected to that, I seem to remember That's how you got into psychiatry in a way. Someone told you that you needed to have all of that in order to pursue the more physics or mathematics based interest that you had? Was it that also a matter of doing what someone wanted from you?

Karl Friston: Well, doing what you're told, yes. Yeah, doing

Andrea Hiott: what you're told.

Karl Friston: Yes, well, it's nice you heard that story, which is absolutely true. Yes. So I was, uh, in my late teens, deciding what kind of A levels or baccalaureates and then what kind of course I'd like to do at university. And there were two funny stories. One funny story was that this was a time that computers were being invented and used, and one of the computers was deployed to try and predict the perfect career for school [00:16:00] children. And I was a guinea pig, and I remember actually being televised by the local, um, television, uh, station.

Um, and I putting in all my preferences for the perfect life into this computer and then waiting. I was about 15 years of age, 16 years of age. And, um, but it was in those days, there was no email. You had to sort of print out things on cards and send them to Liverpool University and wait for the response.

So a week later, the envelope arrived, my perfect career. Yeah. Having stated, um, my enthusiasm for nature and working outside, but I also like, sort of, physics and engineering and electrical things. So, my perfect career, apparently, should have been a television aerial erector. Which means I could

Andrea Hiott: What is that?

Sorry, a television aerial erector? Erector, yes. I've never heard of such a thing.

Karl Friston: Well, well, yes, I suppose nowadays with sky, uh, [00:17:00] with dishes and, and cable, you wouldn't need them. In those days, you had to have an aerial on the top of your house in order to receive the radio waves to make them come in. Oh,

Andrea Hiott: like an antenna.

Karl Friston: An antenna. Yes, a television antenna. Yeah.

Andrea Hiott: Okay.

Karl Friston: Um,

Andrea Hiott: Wow. I,

Karl Friston: I took this to my, um, career advisor and said, well, actually what I'd really like to do is, um, the mathematics or the physics of psychology. And he said, Oh, if you want to be a psychologist, you have to be a doctor first. And what he thought what he meant was he thought I wanted to be a psychiatrist.

as opposed to a psychologist. And of course in the UK you have to become a doctor before you can become a psychiatrist. So neither of us knew the difference between psychology and psychiatry. So I dutifully applied to do medicine and then went to university to do medicine and then discovered afterwards that that wasn't [00:18:00] quite, I didn't quite have the mathematics or the physics.

So I was able to put that into my early undergraduate training by mixing and matching courses. Yeah, so you're absolutely right. Doing what you're told is not always the best idea, if what you're being told is not quite what you had in mind.

Andrea Hiott: In this case though, maybe, I mean, considering where you are now, something good happened.

Yeah, it

Karl Friston: was very fortuitous, absolutely, yeah. The quest for, um, understanding. others and self, um, clearly, um, was made much more poignant and, pertinent by the fact that, you know, I ended up trying to learn about, diagnose, treat, help people with, um, various psychiatric conditions and particularly schizophrenia.

Andrea Hiott: I wonder where, when, if there's a memory of the mathematics becoming such a fascinating puzzle for you. Also, the [00:19:00] psychological, I've heard you talk something about maybe you wanted to do psychological mathematics or mathematical psychology, but that's not, I mean, that's becoming a more normal thing to put those together, but I don't think back in the days when you were being told to become an antenna erector, those were really commonly put together.

Maybe I'm wrong, but I wonder if you remember trying to put those together already early on.

Karl Friston: Yes, I mean, very explicitly, you know, I applied to a university, Cambridge University in the UK, specifically because they allowed you to do, um, and still do, a natural sciences tripos that allowed you to do a mixture of psychology and physics, so I was able to study quantum physics and probability theory at the same time as studying psychology in service of acquiring a medical degree.

That decision, I think, was, Probably much simplified, but um, now looking back, I think it was um, in deference to the dual interests [00:20:00] of my parents. So my mother loved psychology and she used to get all the popular psychology books, you know, thinking laterally and all the books that one would read in the 1960s to, you know, gain insights into, into yourself.

So I had to read all of those. Then my father was an engineer. And he, had a passion for understanding things, um, mathematically in terms of the physics and made me read, again, an instance of doing, as you're told, reading Sir Arthur Eddington's Space, Time and Gravitation. So there were two, these two things.

So it was quite natural for me to, you know, to try and relate one to the other. And to build a career that allows you to do that. And you're absolutely right. I mean, um, you know, now that is that that is, um, almost the, the norm, isn't it? If you think about computational neuroscience, theoretical neurobiology, or indeed nowadays, machine learning, artificial intelligence research, you know, to make a difference practically play.

Um, in, you know, in the 21st century, you know, you have to be at least, have some fluency with computer [00:21:00] science and with programming, with data science, um, and the underlying maths of that. And of course, the most interesting thing to apply that to, uh, is other people and other users.

Andrea Hiott: Yes, exactly. There's computational everything now, psychiatry, neuropsychiatry.

 But you mentioned schizophrenia and that's an important part of your early world coming into the, this neuroscience, was it a residency or I'm not sure what it was, but I feel you spent some years with Living with patients, people who have what we call schizophrenia, is that correct?

Karl Friston: Yes, absolutely. Yes, it was exactly, um, a residency, um, it was a training program in psychiatry, a long program, you know, lasting for many years before, one became a member of the Royal College of Psychiatrists in the UK. And two of those years, um, involved, partaking or being part of a therapeutic community for people with chronic [00:22:00] schizophrenia.

Absolutely. Yeah,

Andrea Hiott: that's a very strong, I mean, for me to try to imagine being in that environment for so long. Would you say that influenced all that was to come? Or was it just a sort of interlude? I'm wondering when you're in that environment with people who are operating or living in a way that's, different from the so called norm, whatever that might be. If you, if that influenced the way you think of, you thought of things like energy and entropy and the way they can be different for different people in different circumstances, because later on you would deal with all of that in different ways, but.

Karl Friston: Yes, I think at the time, um, what I recall, was the de skilling that is required to be open to understanding and living with people who, you know, in this instance had, you know, had psychotic, [00:23:00] chronic psychotic disorders. So that really requires you, um, to put down and to de reify everything you thought you knew in order to be receptive to, um, a different perspective and a different set of experiences.

So in fact, the uh, formalization or the naturalization of the way that people made sense of things and made decisions, um, that I was dealing with as a young psychiatrist came after the event, after many, many years. So it was, you know, what needed to be understood. Um, was the gift of that kind of, uh, experience, early clinical experience.

So at the time, you were not thinking about free energy or entropy. You were basically trying to understand why this person was distressed or enjoying their, your neologisms and their lateral thinking. Um, but [00:24:00] then, as time went on, you got more and more into, uh biological psychiatry and neuropharmacology and the biophysical mechanisms that underwrote this kind of sentience and decision making.

Um, then you started to apply what you had learned as, as a student in terms of modeling and the physics, both in application like MRI physics in your imaging, but also in terms of inference and modeling, um, that inherits from a knowledge of probability theory and statistics.

Andrea Hiott: So you wrote this paper with Chris Frith that it's pretty much was and still is a different notion of how to deal with something like schizophrenia, the disconnection hypothesis.

In there, I think you talk about value and belief or the, or I read somewhere maybe that was part of what inspired that, that hypothesis that the disconnection hypothesis, that different, that cells aren't. [00:25:00] Connecting properly, you can explain it better than me, might be, maybe in the prefrontal and temporal lobe, have to do with some of the manifestations of schizophrenia. But I feel like that was already connected to words like belief and value and so on. Is that right or was that later too?

Karl Friston: No, that's absolutely right. Um, and you phrased it very nicely that, um, so in getting into research, I was luckily, um, I did as I was told, I was focusing on brain imaging, by, uh, people like Peter Little, um, who was at that time, you know, one of, one of the young, good and great of schizophrenia researchers who had seen the potential of brain imaging, and the people in charge of assembling an expert team to deploy brain imaging to understand schizophrenia, um, included people from Chris Frith, who was at that time at Northwick Park.

That was the scene setting, um, in the UK at the inception of modern human brain mapping. [00:26:00] But I did a, um, a sabbatical with Jerry Edelman at the Neurosciences Institute at the and subsequently, um, Scripps Institute or in California. And, um, our focus at that time, and I say our, uh, including people like Reid Montagu and, and Giulio Tononi, Olaf Sporns, the, uh, sort of key young men that, um, Edelman was mentoring, was on value dependent learning.

So I think that's where the value gets in. And you're absolutely right. Now, to understand how we wire ourselves to understand the fundaments of how we make associations and how that is, those associations are installed in terms of synaptic connections requires a very mechanistic understanding of when and where you should consolidate those synaptic connections.

So it was a natural step then to think, well, if people who have positive or negative symptoms of schizophrenia in some way, um, [00:27:00] are suffering from a failure of this kind of associative Um, plasticity, this, this, um, what subsequently now could be regarded as a kind of synaptopathy, a pathology of synaptic connections, then this would be a natural explanation for the disintegration of the psyche that, uh, described by people like Bleuler, you know, a century previously.

So, um, What that pointed to was a failure of functional integration or a failure of connectivity in the sense that the connection, the connectivity dictates the coordination of neuronal networks, neural neural activity that underwrites our sensemaking and the way we behave. So on return to. England, um, and analyzing the brain imaging data from schizophrenia, we were looking for evidence of this pernicious kind of disconnections, [00:28:00] disintegration, not of the sort that you get with a stroke, uh, or a neurodegenerative disease where, uh, most certainly sort of, um, for example, leukodystrophies where you might get, um, a disconnection due to a cutting of the vise.

More. A functional disconnection that rested upon a failure of synaptic integration, um, at the sort of molecular, uh, level, and more specifically, the failure of, um, to modulate appropriately to contextualize when and where, um, and how you should consolidate this connection or modulate this, this connection.

So the value here was, um, a rather generic notion of value, the kind of, value that you find in utility theory and economics or, um, reward in behavioral reinforcement learning um, from Jerry Edelman's point of view, it was even more generic. It's just basically adaptive fitness. [00:29:00] How fit is this?

pattern of, um, connections in order to, um, ensure the persistence or longevity of these neuronal groups in response to this sensory input. So he read value, um, very much as a theoretical biologist might read, uh, adaptive fitness. You know, simply, how is it that certain connections and certain, um, on, uh, cell assemblies, neuronal groups come to, into existence in the same spirit of, you know, how is it that phenotypes such as you and me, come into existence on an evolutionary timescale and then looked at the same maths to try and explain things in terms of this value dependent or fitness dependent, um, plasticity.

Andrea Hiott: I'm trying to imagine the time period because it feels like to me, was this the beginning of, something like fMRI, [00:30:00] which was was that was like the 90s, right? When that first became. That's right. Okay, so this was really in the moment of the first time a lot of this was happening.

And when, at what you just described, was that an application of psychology to brain imaging? Something like, I think it's called Schwartz theory of value or these, it makes me think of circular causality and loops because there's all these old ways of thinking in psychology. There's sort of these loops and circles of usually binary sort of categorizations of what's called value and values are considered to be beliefs and so on. So is that a jump? Was it literally being applied to the mathematics and the mechanics, those at that time in those early years, or was it already removed from those sources?

Karl Friston: No, no, I, I think the, I, I think there was, um, uh, an intimate and deep [00:31:00] link, um, and you're absolutely right that these, uh, this early work. Um, was coincided with the inception of things like functional magnetic resonance imaging. In fact, the very early, the empirical evidence for the disconnection hypothesis in schizophrenia actually came from positron emission tomography, which was, had been going for, um, about half a decade before fMRI started, you know, to become available routinely in the early nineties.

So slightly before fMRI, uh, were the first evidence, um, not just from our group, um, but, a number of groups who for the first time could image the entire brain. And that's important because if you want to make a comment about or infer a failure. of dynamic coordination or functional integration that involves all of the brain.

You have to be able to measure all of the brain at the same time. So brain imaging was very important to be able to do that. It was the first time ever that [00:32:00] people had a window on the function integration over all of the brain and were able to make comments about the Coupling between brain areas, the connect, the functional connectivity, as it's currently known, um, because you were able to measure that at the same, at the same time.

The maths in terms of the, the beliefs and the value, um, so you could certainly read it from a sort of psychological perspective. Um, but you can also read it from the point of view of behavioral psychology in terms of reward learning. So the kind of value that Gerry Adelman was talking about and that we were talking about transpired to be easily translated in terms of reward, for example, it's not, you know, there are certain.

Um, it subsequently transpired that that was not the complete story, but certainly you could talk very comfortably to, um, with people in behavioral [00:33:00] psychology, reinforcement learning, and the early machine learning and computational, uh, or neural network people who had been working on similar, um, you know, what was the right objective function, uh, to, in order to learn the right weights in a neural as opposed to a neuronal, uh, network.

I think the notion of value that you are referring to. And the circle of causality, was, um, came to the fore a little bit later in my life, um, in terms of understanding the more generic principles of the self organizing brain. And those, I think, were inspired by, um, work on predictive coding and certainly, um, the work of, um, My next door neighbor was Geoffrey Hinton, who was, you know, at that time he'd come from the Americas and was working next door in Queens [00:34:00] Square at the Gatsby Computational Neuroscience Unit.

So his work was inspirational, along with his younger colleague, Peter Diane, in terms of taking Helmholtzian ideas about the unconscious inference, and also the mechanics of how an artifact, you know, a neural network, for example, would comply with those basic principles of a neural network. modeling and understanding that could be read in terms of a Helmholtzian notion of unconscious inference.

And at that point, I think the kind of value that you're talking about, um, became center stage in terms of what is valuable to me. What are my prior preferences? Um, you know, how would I write that down in terms of a physical principle of least action? that which is just a biology principle, um, and how would I demonstrate and reach out to cognitive neuroscience and, um, philosophy, sorry, [00:35:00] not philosophy, psychology more generally, and the philosophy of mind and psychology.

Andrea Hiott: It's fascinating. I think a lot about dichotomies and this beyond dichotomy and binaries. And of course, in everything that we've already brought up, there's a lot of binary built in, whether it's zero one, or whether it's these models that are usually either or or versus, I think that Schwarz model, if I remember correctly, is something versus something always in psychology. And I, I'm wondering if it was sort of co creating, you know, the technology and the way that we approach. Mind body or, I mean, there's all these not dualisms, but just binaries of the way that we've come to understand the body and the brain and all the Cognition and you know, we see we seem to have used binaries to progress up to this point and it's interesting to see how the technology and the ideas were sort of co evolving, what was influencing, what, at that time, it's very interesting that you [00:36:00] were with Hinton, and you also brought up unconscious inference, which I want to ask you about, but first, um I should just say, I remember once, I think the first time I was in, with you in the same room or something was in Berlin at the technical university when you came, and that was a while ago. Maybe 2019. And, I think I remember someone saying, here's Karl Friston, basically every image you see of the brain depends on him or, or you're looking at his work or something because, because of the software you invented, which I can't remember all of it.

I know there's a lot like VBM and SPM and, um, the, the dynamical. DCM. DCM. Yeah.

Karl Friston: Okay. But the

Andrea Hiott: point for people who are listening, in case they don't know, they only know that all this other stuff you're famous for, you actually have invented, or I guess that's the word, invented a lot of software or programs that have really become so important to the way we image the brain.

You brought up this amazing moment of being able to actually look at the brain. As you were [00:37:00] describing it, I was thinking of the first time we could take a picture of the earth from the moon or something. This kind of moment where things change, where you get a different perspective and you can see in a whole way.

I wonder if you felt that in the community at that time of a perspective shift. And then I I wonder why you wanted to make this new software that has become so important to everyone who's looking at that bigger perspective. In a, as general as possible way, Maybe you could walk me through that part of this part of your journey.

Karl Friston: Right. Um, that's a good question. And I have to say because, um, I was new to the game. For me, it wasn't a, um, drop the mic moment or, you know, it was, for me, this seemed to be the natural way to study the brain to basically image, uh, the brain in action. Um, so, And, um, well, let me come back to, you know, answering that question, but just briefly, um, [00:38:00] you know, for your, for your listeners, just, um explain the nature of that software and its importance at that time.

So we have lots and lots of data. So this was the first instance of big data. So this is before the human genome. So this is, this was the first time that people had to deal, um, with very large time series with, you know, thousands, millions of, of, um, pixel elements or volume elements called voxels that constituted a brain image.

Not just at any one point in time, but a sequence of these things that could last for minutes, if not hours. So to handle that, you basically had to solve, um, A relatively difficult statistical problem. which is inferring where your interventions during these successive acquisitions of brain images were causing a difference in brain activity [00:39:00] and neuronal response of neurophysiological or hemodynamic response based upon the blood flow changes.

Um, the difficulty was um, but there was a certain smoothness imposed by the, um, the brain imaging, which meant that you had to adjust for an account for, um, that statistically. So basically what, uh, we did was invent a, um, a statistical procedure that effectively gave you x-rays of brain activation. Um, so, you know, when people point to, uh, these, um. heat maps of brain responses. They are almost universally images of statistics, or not images of activation. They're the images of how significant is an activation, understanding carefully all the correlations and all the confounds that could have caused that. So that was statistical parametric mapping.

Andrea Hiott: Just to stop [00:40:00] you for a second, for people who are listening, so almost everyone in some science magazine or just anywhere on media, social media, you see these images now, right, of, I almost sometimes think we need this little, caption underneath like c'est n'a pas, this is not a pipe, this this is not a brain, because those images are almost the way we think of brains now, as those are brains with like little parts that are lighted up, but that, we're basically looking at your work in a way, right?

Karl Friston: Well, yeah, I think that's slightly understated. It's

Andrea Hiott: a little grandiose, but that's what it is. Everyone has seen it, probably, even if you've never studied neuroscience.

Karl Friston: But you, you, you've said something very important there about this sort of, you know, understanding what you are looking at in terms of lots of little activations, lots of little, uh, bits of the brain, uh, lighting up.

Of course, that's not the complete story. Um, and perhaps unkindly per perhaps, um, in a toyful way, people used to refer [00:41:00] to this as cartography or neocartography, map making in terms of, you know, this is the part of the brain, the city that does color. This is the part that does sound. This is a part that does edges. This is a part that does love, you know, just building little maps of the brain. Of course, that's not Um, an understanding does not convey an understanding, of the functional architecture of the brain. It does to a certain extent in the sense that you can now confirm the principle of functional segregation, by which I mean that the, the brain, is organized according to a principle where different functions are anatomically segregated, specialized in anatomically segregated regions.

And that's basically what you're looking at with these little heat maps. The other side of the coin, though, is a functional integration. It's the coordination that we were talking about before, the connectivity. So these dual principles of functional [00:42:00] specialization and functional integration You can even say functional differentiation or functional integration, um, were, um, part of the journey of developing the right kind of maths and software in order for people to, for people to answer questions about not just where in the brain their favorite cognitive process or their favorite decision making or their favorite, um, perceptual synthesis was occurring, um, but, um, But how that processing was distributed and coordinated, um, in terms of the coupling between the, how it was integrated in terms of the coupling.

And that's where the dynamic causal modeling came in. So that was the next wave. It's moving on from cartography to connectivity, um, and that, um, uh, repeat, reflects this sort of, um, distinction between specialization and integration as these dual principles of functional brain, functional brain [00:43:00] architectures.,

Andrea Hiott: then there was VBM. Well,

Karl Friston: VBM. Um, yeah. VBM is just, is just a, um. application of the same, statistical procedures, uh, statistical parametric mapping, not to functional images that are acquired over time, but structural images of anatomy.

that are acquired over different people. So that allows you to pick out if there has been say a particular deficit in the anatomical structure in this brain region when I compare a whole bunch of people who are neurotypical with, say, people who have Alzheimer's disease. And you can track that longitudinally as well uh, voxel based morphometry is just an application of SPM to structural data. Um, so, you know, the bright line really is the move from, um, these whole brain cystical analyses to the connectivity and the, and the [00:44:00] dynamics, uh, and the cause and causal. Yeah. what causes what to come back to your circular causality, uh, which is important for the brain because there are no, well, with the exception of some, um, uh, corticobasal ganglia connections, there are no connections in the brain that are not recurrent.

So everything All sort of organization is inherently circular, um, but to come back to your, your, well, was this, um did people realize that this was a moonshot moment or, um, there were people, so there were people, for example, people like Semir Zeki who was one of, you know, the UK's, um, um, great and good in terms of understanding Functional anatomy, but all of his life and all of his intellectual forefathers could only have very limited views of what, how the brain was put together, either through [00:45:00] neuropsychology, looking at the lesion deficit.

model in the sense that if I leave in this little part of the brain, what function is, is lost or using, um, in, electrophysiology just to measure one little part of the brain at a time, or indeed, even, um, possibly even one neuron and neural population at a time. So no one had actually looked at the entire brain in action.

Um, in a context sensitive way previously. So I think for, for neuroanatomists and neurophysiologists who had spent their life either doing neuropsychology or neuroanatomy or doing neurophysiology in animal models that this was a remarkable moment. I mean, there was a degree of, um, not physics envy, but sort of imaging envy from these people who couldn't actually get access to it.

So it wasn't all, uh, joy and, uh, and, um, back slapping [00:46:00] a lot of, because, um um, yeah, what I do remember quite acutely is, is the monkey electrophysiologists got very cross with the, um, the young. Turks that were doing all this brain imaging simply because they could acquire their data in one week, take a week to analyze the data and then write a paper and it would always get into nature because it was such an exciting new technology.

Whereas the poor monkey electrophysiologists would spend two years training their monkey, acquiring the data carefully in a crafted way for six months, spend a year. arguing with reviewers before they get their paper published. And, you know, they just could not understand why you could short circuit all of this hard work.

And crucially, of course, you were brain imaging was attracting all the talent, all the young PhD students now wanted to do brain imaging. So there was a lot of tension in that paradigm shift that basically [00:47:00] took cognitive science into the area of cognitive neuroscience. I think you could really Um,

say that neuro brain imaging of this kind, neuroimaging this kind, um, sort of made cognitive science into a neurobiological science, simply because now you have an empirical measurement of the cognitive mechanics that people have been studying up until that time. But I didn't know that. Remember I came from psychiatry and theoretical neurology and I had no experience with monkey electrophysiology or, or neuropsychology.

Uh, so for me it was just quite just the thing to do

Andrea Hiott: You just fit sort of perfectly in the time where that's what you learned. Absolutely. As your skillset. Right?

Karl Friston: Very, very lucky. But again, you just wrap upon doing what you're told. This thing needs to be done with you.

Andrea Hiott: I'm not sure I believe you always just do what you're told, but I mean, was was it really so, or maybe it's, it, there [00:48:00] is a kind of, um, what's, I know, I don't think you meditate or anything, but there's a kind of.

meditative sense to you and your work that you do sort of stay in a flow. Has that just been natural for you since those early years, that you kind of found a way to flow with life? I know I'm being, I'm simplifying it, but Is that kind of what you mean by doing what you're told or was it really just Yes,

Karl Friston: yes, I think that would be a nice way of describing it.

So you just have to make, um, listen to the right people, make the right choices on the basis of what you're being asked to do at each point. And, you know, as long as those choices are contextualized with an overall direction of travel, but not too worried about the particular choice at hand, um, then you do quite naturally flow along.

Absolutely, I think it's a very nice way of describing it.

Andrea Hiott: And there's something about also, getting a perspective and an overview because you have to choose which person you're going to listen to because a lot of people will be telling you what to do, or could [00:49:00] be telling you what to

Karl Friston: do. Yes.

Um, yeah, well, um, as you get older, there are more people who want to tell you what to do. Oh,

Andrea Hiott: really? As you get older?

Karl Friston: Oh yes, yes, because you get more famous and more people start to think that you should be doing this and doing that. When you're young you're just grateful for the opportunity to do something, so it wasn't quite as difficult to make those choices as it would be nowadays.

Andrea Hiott: So this is one part of why you're so famous and everyone wants to tell you what to do, is because you created these amazing software things that now everyone uses, and you're, I think you're double in the h index and quoted more than anyone a lot because of that, because everybody's using it.

But you mentioned this unconscious inference from Helmholz, and that moves into another thing that you're famous for, which is this active inference free energy principle. There's a lot of connections there, but I want to just, this word bothers me, because So I wonder about Helmholz, unbewusster schluss, unbewusster [00:50:00] schluss. I know German, but I sound terrible saying it.

Sorry, everyone. But, as I understand that, that's an unconscious Conclusion or statement or something like this. So why is it translated as inference? Because this gets to some of the confusion I have personally about for example, the image of the brain with the brain itself or, Life flowing along in a way with metacognition, life knowing that it's flowing along. Could you just maybe tell me how you came across Helmholz and what this word means for you, the unconscious inference?

Karl Friston: Right, yeah, well, um again, excellent question, which has so many different meanings. Um, answers, um, that present themselves to, to sort of foreground the points you're making. Um, first of all, um, inference is read here in a very generic sense, um, and it's just read in the sense of, indeed, reaching a conclusion on the basis of some [00:51:00] evidence.

It just so happens that in, um, Helmholtz's unconscious, um, inference, um, the evidence at hand is his sensory input. But, you know, the notion of inference here is just basically explanation to the, you know, an explanation predicated upon evidence, the evidence, what the evidence, what is the thing that has the attribute of evidence?

Well, it's a conclusion, it's a hypothesis, it's a model. Um, so that conclusion of alternative conclusions or hypotheses that has the greatest evidence is the one that you normally commit to and that committing is a process and it's a process that can be described as inference, but it's also a process that can be described in terms of, um, dynamics and physics.

So you could argue, um, that everything is, or can be read as, or [00:52:00] described as a process of. inference, if it looks as if it is measuring something or observing something. So I'm quite happy, happy, treating synonymously inference, um, measurement, observation, computation, all of these things are just different words for the same, um, underlying process.

That entails some optimization of the evidence for your model of either your data and that you're in dynamic causal modeling, for example, or, um, of your sensorium and notice because it's a process. It is a flow, it is literally a flow, and indeed the free energy principle, or the classical physics derivation of the free energy principle, starts with a flow equation.

It just says, if [00:53:00] the world can be written down, or if some universe can be written down, in terms of the way that my states, or any state in the universe, flows as a function of where you are in state space, it is a function of where you are in state space. That's it. That's, that's what the starting point of the free energy principle.

And all you're saying is that this particular flow flows on a gradient or in a direction that maximizes the evidence for an implicit model. And then you have to get into the game of a model of what, well, the model of everything that's causing my sensory inputs or my data, and then you have to carefully define the boundary that defines that, you know, the sensory inputs from the world out there.

from the thing that's doing the sense making. So for me, it's quite

Andrea Hiott: The mark of blanket sort of Absolutely.

Karl Friston: Yeah, that is exactly the Markov blanket or the boundary that separates you from everything else. And the universe impresses itself very much in [00:54:00] the spirit of Helmholtz's notion. of sensations impressing themselves on your sensory veil, on your sense, sensory epithelia, that would just be the sensory sector of your Markov boundary or your Markov blanket.

And then that induces a flow on the inside and that flow is inevitably or necessarily under the free energy principle, um, describe it, you can describe it as inference simply because it's maximizing the, the basic model evidence for the underlying, um, or the implicit generative model of, of what's going on.

So I think that's, you know, for me, flow and in the service of reaching a conclusion. is just a description of, of of existence. And it would look like things are measuring things, it would look like things are, um, observing things. And crucially, just to bring in your circular causality again, of course, uh, [00:55:00] it's not that we're just passive sessile, um, ingesters and assimilators.

of the flow of sensory input, we can now act upon the world to sample and direct that flow. So, you've now got this interesting circular perception action cycle, the circular causality between the inside and the outside, where I'm trying to make sense Through a process of inference, um, of the data that I am myself creating through action.

Uh, and if I can do that gracefully for a long period of time, I am basically in tune with my environment, with my world. I have this adaptive fitness in the sense of Edelman's value, this generic notion of value to come back to, uh, our earlier conversation.

Andrea Hiott: Yeah, that's actually a good description of what I was trying to describe is what you you meant by doing what people told you to do But Karl this part is really hard for me, and I [00:56:00] hope it's okay We already time is flying by but I really hope we can dig into this just a minute because I find it So confusing and also so inspiring that what you just said that there you start with the flow that flow that is the first equation, and I feel like what, what a lot of your work with the free energy principle and active inference does is give a model to that ongoing process that can't ever be fully modeled, and it does it in a way that such that the model itself can continue to change and evolve over time.

Because the process has to continue and to change and evolve, which is a very different kind of math and a very different kind of statistics. And I think related to why Bayes, Bayesian statistics is involved. But just this unconscious inference versus unconscious conclusion. I wonder what you think. For me, it seems like what Helmholz was talking about , or what you even are just describing is that life process that can't be separated. It it's not inferring [00:57:00] in the way, at least not in the everyday use of that term, in the sense that something is setting against itself and making a decision. it feels more like that flow or that alignment such that, for example, my body comes in and every single thing that it's encountering, it has to not be molded by, but in a sense be molded by, fit to, I guess would be a word some people use, fitting to the world and the brain part of that, a very big part of that and then, out of that same continuity and that same flow we can begin to recognize the ways that we've aligned and start to make something like inferences. That's how it seems for me, so then. when I, I don't hear those things distinguished. Maybe they don't need to be distinguished, or maybe I just misunderstand it, but does that make sense to you at all?

Karl Friston: No, it does absolutely not. I think, I think they certainly need to be distinguished. Um, so let me, um, [00:58:00] um, be very clear about the way that, uh, people like me use the word inference. So notice that unconscious inference immediately tells you this is not about a propositional, declarative, Subjective, explicit act.

It's not a mental act at all. This is something Because of the

Andrea Hiott: unconscious.

Karl Friston: Exactly. Or Yeah. Exactly. So I think what you're saying is, um, that if we now want to have, um, conclusions and inferring things in a propositional way that I could actually articulate, not just to somebody else, but to myself, um, then you'd certainly need to have now a meta reference, um, a metacognition that's looking at your unconscious, very low level processes of the kind of thermostat could be accused of having.

You could argue that a thermostat is an instance, a very, very simple one, but it is an instance of [00:59:00] unconscious inference in the sense that it has a prior Bayesian belief. And again, I'm using belief here, not in the vernacular. It's just a description of a Bayesian probability distribution. But you could argue that the, um, the thermostat has a Bayesian belief that the temperature of this room should be at a fixed set point.

And everything that it does. Uh, all the way that it acts upon the environment is to make that belief come true. So it's basically doing active inference. It's just inferring what the cause of its thermoreceptors, uh, is, and if it deviates from its prior Bayesian belief, It will act upon, uh, on the world to, um, to, um, realize it's, uh, it's priority.

But at no point would you say that the, the thermostat is thinking or has any awareness or is making any decisions or reaching any conclusions. I think you then need a layer on top of [01:00:00] a myriad. of coupled thermostats, um, for, for you and me, of course, these would be autonomic reflexes and motor reflexes, you know, truly embodied the brain and the body perspective, um, and you'd be looking at that, and you may have to do that several times before you've got to, this is not a pipe.

Uh, to know that you are perceiving, to know that you are, um, engaging in unconscious inference consciously requires an enormously elaborate kind of generative model. And interestingly, you know, um, from the point of view of writing down the kind of models that you need to be able to do this kind of metacognition and have self awareness, they have a hierarchical depth.

So you're now very much in the same sort of architectural space. as deep learning. So the deep in deep learning just says you've got hierarchical depth. So I think that's what you're talking about. And it's really, you know, you mentioned, you mentioned a couple of [01:01:00] times, you know, this, um, the way that we understand the world.

Um, very often can be cast in terms of a dichotomy or a binary, um, um, kind of representational state space. I think that's very insightful. I would read that as basically, um, carving nature at its joints in order to understand it. Um, and certainly that's been my experience in computational modeling.

That's been a move. Away from these continuous, um, flow like generative models up the periphery near our sensations. As we get deeper and deeper and deeper into these models, they'll become very categorical. You know, for example, words, you know, there's no, um, you know, there's no sense in which a particular word, say the definite article, has any continuous representation.

It's just what it is. Uh, as opposed to another word, so I, I think that the, the, [01:02:00] the, the way that we carve nature at its joints as installed in these generative models does speak to this simple binary categorization at some point that we could articulate and actually talk to ourselves about and, and, um, you know, and communicate not just with other people, but also, you know, with ourselves and that.

Uh, but that's a very, very deep and abstract in kind of inference, which now may encroach upon conscious inference and awareness and experience in the way that I've carved my, uh, world at its joints and my world is largely all about me and my body and my interactions with other things that have bodies like me and, and I infer minds like me.

So, you know, I often get told off as talking about sort of beliefs. Um, in my world, um, a belief is just a probability distribution. So you talk quite freely about Bayesian belief [01:03:00] updating. So a thermostat can do Bayesian belief updating. Of course, it's a very trivial kind, but it does it. Um, um, and so, you know, the Bayesian aspect means that it is not the vernacular kind of belief.

So I'm assuming When you talk about conclusion, you actually meant a conclusion that I was aware of, that I'm committed to this as opposed to that. And that would require a very special kind of artifact or phenotype or system, um, you know, beyond, uh, I would imagine, um, insects or even some small, uh, small mammals.

It would require something like you and me, you know, as if it's by communication language, for example.

Andrea Hiott: Yeah, thank you. There's so many things I want to say in that. I think maybe the thermostat is Something that could help me unpack this. I think what I meant by conclusion wasn't, again, I know you said like languages can be distinguished, but even that we're coming from a [01:04:00] perspective and oftentimes we use the same words in different ways and that's part of what gets so confusing here and why I'm looking deeply into this word inference and conclusion, but I really meant more something kind of at one with itself, like schluss, like it's closed together or something, but that doesn't necessarily mean static. Um, but let's, let me think about the thermostat because this is a big philosophical problem. There's all kinds of people writing about whether, you know, , can, thermostats, consciousness and so on, but. I think it's very different, that flow that we started with and what the equations are talking about being, it's, it's dynamic and it has its own trajectory. You're often talking about Um, things referring back to their own past, um, like sampling their own, going back to where, to their own habits or patterns in a way this, this dynamic flow is, it's very hard to talk about, but that, that is itself already a [01:05:00] trajectory.

It's always changing, but that we can sort of go back to and revisit. I mean, that's what memory is in a way. And I don't think a thermostat has that, um, that continuity or it doesn't have affordances that apply only to its own trajectory. That's how I would say it. It's almost a distinction between living and non living, right? So, um, the affordances of the thermostat are our affordances, and they apply to our ongoing trajectory. So, I guess what I'm saying is exactly that's the difference for me, that the thermostat is going to just be doing what it's doing mechanically. And it's built on a binary way that we've understood ourselves and so we can create sort of a little model or representation of what a body does in the thermostat to help control the room.

And like that sort of seems the direction of that technology in relation to the flow for me. But what I'm talking about is like actual life [01:06:00] and how we

Karl Friston: We

Andrea Hiott: are a bit like life becoming aware of itself as life, and I think your work helps us understand that. It's putting a kind of mathematics on letting us glimpse at that process, but I often hear it described more like the way that was describing the thermostat, which confuses me because it seems I feel like we can see beyond the binaries a bit, even though they've been so helpful. And this might be a way to do that, but we keep trying to fit everything back into those binaries.

Karl Friston: Yeah, I think that that was an excellent deconstruction of a key distinction between the thermostat and you and me. I think that everything rests upon what you said, um, in relation to where is the affordance.

So if you wanted, and I am flattered and pleased that you think that the free energy principle might afford a mathematical image [01:07:00] of the, you know, the distinction between life as we know it, or life as, as if in spite things like you and you and me, um, uh, that is not a characteristic of, uh, the thermostat.

I think the bright line between. Um, thermostats, or what's governors, or viruses, um, and the kind of life that you're talking about. I say prions, not viruses, viruses that are alive. Um, is, um, all rests upon this notion of affordance. So just think what affordance is an attribute of. It is an attribute of what I can do.

And as soon as you say I and do, you've brought two really big things to the table that the thermostat doesn't have. First of all, the thermostat has no notion of, um, itself and crucially it doesn't know what it's doing. Whereas as to have a [01:08:00] affordance, you have to have. A generative model that can generate the consequences of your own action.

And if those consequences are apt for the kind of thing that you are and conform to your preferences that define that attracting set you were talking about. You're returning to, uh, time and time again to these characteristic states that constitute the, the preferred states of being or the technically, uh, the attracting, the attracting states that define the kind of thing that you are.

If you can do, if you can run out into the future and evaluate the consequences of this action or that action, then you can evaluate the affords. So now the game is to try and, if you like, mathematicalize that. How would you write that down in terms of a principle of least action? Um, and it's actually quite simple.

I mean, you're, you could argue that people like Richard Feynman were doing this during his PhD thesis, you know, you know, [01:09:00] uh, excuse me, um, you know, many decades ago. Um, so all you need to do is to, under your model of the consequence of your action in the future, just roll out and measure the, um, the value, for me, that would be the free energy you would expect if you committed to this course of action in the future, and then just add up the value.

all the values and that can be cast as a path integral and the addition of all the energies is known as an action in physics and then you find the path of least action. So the path of least action into the future under this plan or that plan, the things that I do, um, now Uh, describes the best kinds of action, and then you can simulate that, and now you have, I think, a much more expressive kind of physics that does not apply to a thermostat, but would apply to you if you made choices [01:10:00] about the things that you could do in the world, and you could respond to affordances, and not just affordances of the kind, oh, I can sit in that chair.

We're talking about a much more generic notion of affordance, which is that which, um, minimizes your uncertainty about the world, which can also be read as an epistemic affordance, which basically makes us curious creatures. So I think that's the difference between a thermostat, and what could even argue a large language model, and you and me, is that we have epistemic affordances built into our beliefs about the kind of thing that we are, and then we just select the path of least action, the path of least expected free energy, uh, to commit to, because that's the kind of thing that we do.

We end up watching television shows, reading books, doing podcasts, asking questions of each other, [01:11:00] simply because that's the kind of thing we are. This is, these are the attracting paths, these are the attracting states. And notice that in so doing now, We've again got another bright line between the thermostat and you and me because we're not talking about one particular state of being.

We're talking about a very itinerant and attractive set that is constantly evolving and we're always chasing away uncertainty, always chasing away expected surprise, always chasing away epistemic entropy, always on the move. The world is always moving and as you say, it's not about, um, Finding the perfect, um, state of being it's finding the best fit to a constantly changing world and it's always constantly changing because you're changing it for me and I'm changing it for you.

So as soon as we put two things like us together, we guarantee that we're going to be learning from each other and constantly changing. The thermostat on the other hand has an attracting [01:12:00] set which is technically a fixed point attractor. So all roads lead to one point. It's too hot, it'll cool things down.

It's too cold, It'll heat things up. That's not the kind of attracting set that characterizes life, and certainly not you and me. So the attracting set that we're talking about is enormously space filling. I could be in many, many different states and still be me, but I have a lawful flow. from one of these attracting states to another one.

So this is what I meant by flow. Flowing on this attracting set or, or attracting manifold, where that manifold or that attracting set is all consistent with the kind of thing that I am. And it's incredibly itinerant. Technically, space filling that has low volume in the sense there's only a very small number of all the states I could be in that are consistent with me living and being alive.

So, again, I come back to your notion of flow because, you know, mathematically, [01:13:00] this is just a flow on an attracting set, and those flows trace out paths of least action. And the action is the expected free energy or uncertainty or surprise. You could write it another way, um, instead of trying to minimize expected free energy, you're trying to maximize the expected evidence for your model of the world.

that dictates the direction of flow on this manifold. So on that view, life would really be defined in terms of the space filling attracting set. And as soon as you have something that's not a point attractor like a thermostat, what that tells you immediately is that it's going to attractor. This, you, you know, you, you, you were quite explicit about, um, returning to various states of being.

And what would that look like? Well, it would look like very complicated oscillations. So for example, at every scale. So, [01:14:00] uh, it could be the kind of oscillations that, uh, Gyuri Buzsaki, uh, my friend Gyuri loves to talk about, you know, in terms of fast gamma oscillations or, uh, other kinds of rhythms in the brain.

But also at different scales and different parts of me, for example, my cardiac cycle has exactly this itinerant quasi periodic orbit. You can go beyond that. You can go right through. to fluctuations over the month, over the year. And indeed, at an evolutionary time scale, this reproduction itself is this life cycle.

So wherever you look, there's this characteristic dynamic that speaks to cycles and rhythms, and oscillations that are not part of the thermostat subtractive set, which is just sitting at one point. And what that means for, um, basically if you do have a point attractor, that means you're dead. And that basically means, um, you [01:15:00] have attained an equilibrium steady state.

Um, where you are now subject to the laws of, uh, your schoolboy physics laws and you start to dissipate and decay as your free energy minimizes. So just keeping yourself away from this fixed point attractor in the flow, literally all the time on this attracting set is exactly what the free energy principle is trying is trying to describe.

Andrea Hiott: Yes, and I think that's why affordance is so helpful here, or why it even opens up affordance in a new way. I think of J. J. Gibson talking about wanting to overcome subject object dichotomy, just to get back to that theme of dichotomy, and binary, and how I see this as, It's pushing out of that, because to have an affordance, it requires this looping relationship depending on, from wherever you're going to try to measure it, there's always going to be, and we keep talking about circles and that's another thing that bothers me a bit because a circle seems to always come back exactly to where it started.

But we're sort of coming back to where we started, but it's a new [01:16:00] place always. It's more of a. Yes. a spiral. I mean, we're never coming back exactly to where we are, were, but the regularities and the pattern is enough such that it's the same sort of place that we're sampling. Is that, does that make sense?

Karl Friston: No, absolutely. Technically that's what I would describe as stochastic chaos. So when I'm writing carefully, for an informed reader, I will say, to the neighborhood of where you once were. You never, ever. No man steps in the same river twice, but you can be in the same locale. So these attracting sets are not circles and they're not points.

Um, you know, if you were dead, um, they are, they are regimes. Uh, so you have to be in the neighborhood and you're always being pushed off this attracting set by random fluctuations. Um, in the physics version of the free energy principle, but I'll draw back through this flow. So this flow just keeps you close to, so yeah, every day you'll go through the same [01:17:00] routine, but you'll never be in exactly the same state twice, but you'd be pretty close.

And that defines it.

Andrea Hiott: Even if everything changes, it could be the regularities are similar.

Karl Friston: They're with some dimensions that are replicated, absolutely. And that's an important point that, you know, many of these, um, dimensions of these state spaces are very forgiving. You know, I can be me in Melbourne, and I can be me in London.

But I can only be me between 36. 2 and 37. 8 degrees centigrade. So some of the things that characterize me, and indeed sometimes. One talks about the characteristic set of states that characterize a particular entity, for example, me, um, as being the members of this attracting set. Some of them are very, very precise, but some of them are very forgiving.

So, you know, depending upon what, what attribute you're looking at, you know, things could be vastly different, but from another perspective, yeah, you're exactly, you know, this temperature or you're, [01:18:00] you're having your evening meal at eight o'clock PM, just like you were yesterday.

Andrea Hiott: There are these sort of sweet spots, but I think that, that reminds me of when you were talking about measurement and that requires that we understand we're always measuring from a particular point or perspective.

So if we're going to measure from the cell or we're going to measure from the whole human body or we're going to measure from, then those parameters are going to look a bit different fracticality or something that I wanted to talk to you about in terms of getting through the binary. But, To get into that, so a free energy, the free energy principle basically states, you could say it better, but systems change so as to decrease their free, free energy.

 Some people say to, to sort of avoid entropy or there's confusion here for me between free energy and the entropy of the free energy or how free energy and entropy might be different. Maybe you could help me or we could unpack that just a minute just to link it to that practicality I'm [01:19:00] talking about, it seems to me from different perspectives, what's free energy, or what's entropy, or those parameters are also shifting, like maybe some system's entropy is another's energy, or vice versa, but I wonder what you think about that.

Karl Friston: Yes, I think that's absolutely right. Well, you've said three key things there. First of all, the fractal, the fractional dimension, the fractality of these things. I tend to not to use that because fractal has a very specific meaning in the context of chaotic attractors. and the dimensionality. I said before that these, um the attracting sets of things like you and me are space filling but with low measure and effectively the fractal dimension measures the dimensionality of the actual, uh um, space that is actually occupiable by any given particle or person.

What I tend to prefer is scale free or [01:20:00] scale invariant. I think that's what you're talking about. So you can apply the free energy principle at all scales, all the way down and all the way up. So you could apply it to a single cell, And indeed it's been applied to gender reorganization. You can also apply it to entire behavior of the embodied brain.

You could even apply it in principle as people have them to communities, um, in exchanging beliefs or emails, or modeling the extended beliefs, um, and emails and messages. Um, so you can apply it to many, many different scales. And of course, each scale provides the context of the scale below. So I can't have a free energy minimizing.

Um, uh, almost sort of thermostat like single cell unless I have a particular organ or organelle. Um, and I can't have an organ unless I've got a body and I can't have a body unless I've got a family and I can't have a family unless I've got an institution, I can't have an institution unless I've got a [01:21:00] country, I can't have a country unless I've got a biopsy, I can't have a biopsy, and so on and so forth.

And every level, the way in which you measure things or describe things, Um, should in principle conform to the free energy principle. Um, and there's always a description of this in terms of some kind of inference. It could be really elemental like a thermostat or the moon going around the sun. But notice again the circular aspect that, you know, the, the rhythm.

Um, of the heavenly bodies at very large scales, um, or it could be much more intricate and space filling and itinerant of the kind that, you know, we talked about that would apply to our scale of, you know, say from multiple cells, multicellular organization through, through to, to say a, a, a, a body. So I think that scale Uh, invariance is vitally important, and you should always remember there's always a scale above that contextualizes the scale, both in space and time that you're applying the free energy principle to, and there's always a [01:22:00] scale below, there's always a microscopic scale below that, that is, if you like, feeding.

The, um, um, the key order parameters or macroscopic variables that you're dealing with in terms of applying the free energy principle. So I think, that, that was a, an important point. Remind me what else you, you, you said about three things, which were really quite important. Yeah, there were too many things in

Andrea Hiott: there.

I was talking about entropy and energy and whether the, what they, what's different from free energy in the entropy of the free energy or, or if one system's entropy can be another system's energy.

Karl Friston: That's right. That was a . And, um, and you know, on a sort of, um, fairly heuristic thermodynamic explanation, uh, or thermodynamic account of, uh, variation free energy minimization, that is absolutely true.

That, that I would require. free energy in order to do my belief updating, my inference, even if I was a thermostat. Um, and there are sort of physics [01:23:00] principles that say, uh, how that works. And indeed, um, Chris Fields, uh, a colleague and friend of mine, um, who, has pursued a quantum theoretic formulation.

The French principle has a very explicit set of arguments about the thermodynamics of this, often read in terms of, um, that we are, um, offloading energy into the environment in order to do, you know, to maintain our self organization, and there are lots of different ways of articulating that. However, you don't need to get into the thermodynamics, I think, to answer the question you're asking, which I think is quite crucial.

This is going to get a bit technical, but I think it's quite important to understand it. So entropy is just a description of a probability distribution, a little bit like the mean of something or the variance of something. So there is no one entropy, but whenever there's a probability distribution, there is an entropy.

So you have to say the entropy of what. So the, the entropy that you're trying to [01:24:00] minimize when you're invoking the free energy principle as a simple explanation or description of things that resist the second law of thermodynamics, should it apply. Um, then, uh, we're talking about the entropy of the Markov boundary states or the sensory states.

So what we're saying is that certain things in virtue of maintaining themselves, um, on their attracting manifold or within their attracting set or within their characteristic states, certain things that have this generalized homeostasis have a very low dispersion to the sensory states that they express.

And that has a low entropy. So the entropy, the sensory entropy is just the average of the self information or the surprise or, or the free energy. So the free energy scores in information theory, the self information often summarize in terms of the surprise or surprise or, um, [01:25:00] and you want to keep that small.

And the, the entropy of the, the sensory entropy is. the average of the free energy. So if you keep the free energy small all the time, basically avoid surprising sensory, um, samples from the world, um, then you are going to minimize your sensory entropy. Um, that's the entropy that has been minimized. And that means that you, um, don't, um, dissipate your sensory states.

Those observable states, uh, on your Markov blanket, uh, do not dissipate, you don't dissolve into the environment, you, you maintain your self organization through minimizing the average self, uh, information of the sensory states, um, namely the sensory entropy. Now, there's another entropy which is, um, which is actually maximized.

So remember, free energy is equal to energy, total energy, minus entropy. So if I want [01:26:00] to minimize my variational free energy, I now want to maximize my variational entropy. What is the variational entropy? It's the entropy, of my Bayesian beliefs about the causes of my sensory input. So you've got this completely opposite yin yang.

You want to explain things that minimize their sensory entropy, the entropy of over time of real deterministic sensory inputs, and in fact the action upon the world as well, but the mark, you know, everything that constitutes a Markov blanket. Whilst, and you do that by maximizing the entropy of your beliefs.

And this is very, very confusing for people. But it's a shame because once you realize it's important to maximize Your Bayesian, maximize the entropy of your Bayesian beliefs, the internal or the, um, if you like, representational entropy. You can now see [01:27:00] that the free energy principle is due to something which is very important in physics, which is James's maximum entropy principle.

So maximum entropy, uh, there is a maximum entropy principle, which is talks about not the thing that is being measured, but the inferences. of that measurement. So, if you want to infer how long something is, if you're measuring something, then you want a, um, a probability, a belief about the cause of your measurement.

in the outside world that is as non committal as possible to keep your options open. And this is very closely related to things like Occam's principle, finding the simplest explanation possible, uh, but no simpler, finding an accurate explanation that is minimally complex. And that's why you need to maximize the entropy that is put together with the accurate, with the, [01:28:00] um, the energy or the, um, the log likelihood.

Um, to constitute the variation free energy. So it can be very confusing, but it actually works very, very beautifully. Well, once you realize that there's no, there's no one entropy, you either talk about the entropy of your sensory exchanges with the world, or you can talk about the entropy of your inferred beliefs, Bayesian beliefs encoded by your internal states.

And one you can try to maximize in order to minimize the other one. Does that help? Wonderful. I

Andrea Hiott: love that because I I'm I'm always talking about trying to hold the paradox in a way because I feel like that's where you start to get exactly what you just did.

And by hold the paradox, I don't mean we squish together the parts or try to resolve them or reconcile them or that the paradox is bad, but that you just open the space for it in the way you just did, describing that. And then there's a kind of another view that becomes. And again, that relates to [01:29:00] this idea I'm trying to push a little bit of how to get unstuck from certain binary ways of being in the world. this gets to bigger issues, because a lot of this science and philosophy that we're talking about. reverberates and through these very same processes we're talking about into all other areas of life. So when we're discussing these, I think we're also discussing that at a different level.

It's why I brought up fractals and I'm very interested that you don't like that word and , I also use multiscale or, , scale free and maybe that's a better way of describing it, but I think people are attracted to the idea of fractal for the reason I just brought up, that we can start to understand that there are similar patterns maybe at these different scales and, which is what I think, again, correct me, but something like the maths, that you're providing show the patterns that connect at these different scales. I don't [01:30:00] know if you would agree with that, how you would see that or or if what you just said is an example of how there might be another way of understanding how these different kinds of systems can look different from different places, I guess another way of asking this question is there only one correct model or can there be many different models depending where we've set all of our experimental parameters and so forth?

Karl Friston: Right, yes, yes and yes. Um, I got um, so yeah, don't worry about using fractal.

Um, it's just technically, um, you get this, um, similar patterns at different scales. within the definition of scale invariance. And you're absolutely right that you're, what the free energy principle is committed to is that the pattern of self organization [01:31:00] read in terms of the dynamics or the flow on a Lagrangian or in this instance the free energy is conserved exactly at each level.

You have to renormalize the variables and the like, but the actual pattern of the flow is identical in its functional form. So that's exactly what, you know, what defines a Lagrangian. the scaling variance and certainly a, uh, an application of things like the renormalization group to the free energy principle, applying the free energy principle in the context of the renormalization group.

So, you know, the spirit of fractal is exactly what underwrites this application.

Andrea Hiott: The spirit of fractal that's a better way of saying it more than like a Mandelbrot set or something. It's the spirit of what fractal means

Karl Friston: I think you want to talk about self similarity and I think that's exactly

Andrea Hiott: what happened.

Karl Friston: Yeah Yes So forgive me when I when I said I do I do use the word fractal It's just that I read that as a fractional dimension. Whereas I think what [01:32:00] you mean is this fundamental aspect, which is a self similarity affording multi scale perspectives on the same thing at different levels of analysis or different scales of analysis, which is one way of, I think, sort of breaking, I'm now understanding more clearly what you meant by binary.

You mean a sort of commitment to just, it's either like this or it's not. Um, and of course that's exactly what you're trying to avoid. with, uh, the view of the free energy principle, uh, as a maximum entropy principle or a constrained maximum entropy principle. So I, I use the phrase before keeping your options open.

So as, as you try to form a belief about what's out there or an explanation about what, you know, what could have caused us because. This sensory, well, this sensory trajectory, you're trying to do it in a way that keeps your, to keep your options open, and there could be lots of possible explanations. So you're [01:33:00] not committing to one.

If you commit to one particular binary, um, hypothesis or explanation for these data, then you are basically limiting yourself in a very dangerous way because that particular explanation will not generalize. So you see this practically in machine learning, for example, where you have a very brittle explanation, too many parameters, you maximize the accuracy, but you haven't, you haven't done so in a way that keeps your options open by minimizing the complexity.

So, uh, you know, I think so the non binary where we're reading binary now is a very discreet, um, um, unitary explanation for things, that doesn't have much riddle room. That's exactly the wrong way to do things. You need to maximize the entropy and thereby minimize the free energy and maximize the, uh, effectively the, uh, the [01:34:00] evidence for your, uh, for the evidence in the data.

for your particular model of what's going on. So I think, I think there is a very comfortable argument you could certainly make there from first principles, you know, including not just the Theology Principle, but James's Maxwell Henry Principle that you can then relate back to Ockham's Principle that, um, you know, keeping your mind open.

and finding the simplest explanations. That's the way that things generalize, and that's the best way to, to to, um, reach an understanding or explanation, uh, explanation for, for the world. And then you, uh, The final point you made about the dangers of being binary.

Andrea Hiott: I think you addressed it, or I don't even remember my own thing now.

But I can't, unless you remember it, but there's something you just opened up there too. That was beautiful. And thank you. And I want to think about hierarchies a little bit. Are you still okay for just a little [01:35:00] while?

Karl Friston: Yes, of course.

Andrea Hiott: Okay. So the hierarchies is obviously a very important word in your work and in all of this. And I want to talk about it in two ways, one, how the difference between the model and the brain or the ongoing flow and process, I I keep bringing it up.

But first, This fracticality, that's the spirit of fracticality, or multi scale competency, I think, that's Levin, or, these different ways we're trying to open up beyond the binary in, and the math is opening up, I think Bayesian statistics opens it up because it implies the subjectivity, that people often accuse it of because you have to you can set your priors and that's gonna account for these different perspectives maybe in a way that we've been talking about. But to make it a little more grounded, when you're talking about something like hierarchy, you often talk about the low road or the high road or open or closed or, there's, the way we understand it is right now through these kind of binaries. Although [01:36:00] most, usually, I feel like the answer is that it's open and closed.

It's low and high. It's just that We're using these ways to talk about it. So I wonder what, if that sounds true to you. I wonder when you're really thinking about the brain itself and its hierarchical composition, if you see it as, something more along the lines of this nested multiscale, not necessarily linear from high to low, maybe it could be from left to right and in all directions in a more fractal or multiscale sense, if if that's more akin

to what you're meaning by hierarchy, because I think it often gets taken as almost like in a class system or something. It's, steps up the ladder. And I, I think that's also a confusion. At least it feels like one sometimes to me.

Karl Friston: Right. Well, again, that's absolutely right. Um, so, um, in virtue of the spirit of fractality and the self similarity out [01:37:00] there that we now have to make sense of and in fact implicitly model.

That immediately tells you that apt good models of the data generated by this multi scale self similar world must itself have a, um, a self similar structure. And that's what I mean by a hierarchy. So, um, it literally means, um, there is a depth to the Cause of explanations you bring to the table that is not binary in the sense and that was your third point in the sense that there are multiple levels of hierarchical abstraction.

So you can, each level constrains the other. Um, so there's no one privileged level that in your head, you literally have cortical hierarchies. Um that represent the same kinds of things, but different levels of, of obstruction and course grading over different time courses. Um, [01:38:00] so there's no one privilege le level of explanation.

There's no one privilege conclusion. Um, you know, it depends upon which part of the brain you're looking at in terms of the structure of these things. Um, I often think of this not in terms of a ladder. a subsumption hierarchy, uh, but much more in terms of a series of concentric spheres. So for me, what's on the outside is basically our sensory epithelia and our actuators.

And then as we get further and further in, monosynaptic neural connection by neural connection, you get further and further, further and you get deeper and deeper and deeper into, into inner layers. Um, and then, um, what that creates is now a view of the brain, for example, where this surface like, um, like a big onion.

Um, so in the same way that an onion has a layered structure and has a deep concentric structure, one can also envisage. So the connectivity of the [01:39:00] brain conforming to the same kind of architecture. And as you get deeper and deeper, things, um, in the center can see much more of the range of the surface.

So if I had the, this third of the onion was auditory. And this third was interoceptive, then at some point, there's going to be a deeper representation can see both. Now, these become now a modal representations that basically explain causes of my sensations in both an auditory and an interoceptive modality.

And what also tends to happen in the mechanics of this, um, is that your, um, Um, your belief structures, your, uh, your, the states, the latent states, sometimes called hidden states that are represented by neuronal dynamics, um, cover more extended periods of time. Again, in a self similar way, in the, in the way that you zoom out in a Malden Broat set, [01:40:00] you're also zooming out in time in this, in the flow, which means that as you get deeper and deeper and deeper.

You are representing things that can be read as the context that lasts longer than the content at the level below or the more superficial levels as you, you know, get right out to very fast reflexes at the, um, the periphery. So that's what I mean by hierarchy I'm talking about a very particular sparsity structure, sparse causal structure, where Um, that defines a hierarchy in the sense that, um, you only have connections or circular causality between sub adjacent or adjacent levels of a hierarchy, but it's a concentric one.

Uh, you know, it has a certain closure. It probably is a hierarchy. So there's, it probably isn't a nice orderly onion, uh, but, you know, it, it, you know,

Andrea Hiott: maybe a broccoli.

Karl Friston: A broccoli made of

Andrea Hiott: broccoli. [01:41:00] I love the onion thing, but it's, it's even feels too constraining for me because it seems so neat and that you have these layers and it's just, but it's more like there's, or I see it more as there's many onions within onions.

And depending on,

Karl Friston: Yes.

Andrea Hiott: Like you don't see all the other onions. You think you're in one onion, but if you kind of went to a different perspective, you'd be in a whole other onion and all those layers are sort of overlapping in different ways. Does that make sense?

Karl Friston: Yes, absolutely. I mean, there's a little literature on that in terms of Markov blankets of Markov blankets and you have bags of onions and then bags and bags of onions and it just goes on forever.

Yeah, absolutely. And it has to go on forever because of this self similarity. It has to go on to the biggest and, and, and to the smallest. But of course there is no biggest and there is no smallest 'cause that depends on

Andrea Hiott: your position, right? So there can be infinite kinds of big, bigger or smaller depending where you're gonna shift your position of measurement, which I think this is really like a hard paradox to hold [01:42:00] because we have trouble getting out of the position we're in, uh, to understand that.

But maybe that's what technology helps us do in math. In a way, yes,

Karl Friston: absolutely. Well, you know, uh, what comes to mind when you said that was sort of the use of microscopes and telescopes to try and reach up into larger scales and reach down into smaller scales. And of course, you could have mathematical microscopes if you're clever.

But I think there's something, you know, you were talking about holding the paradox before. I think you have to hold the fact that you cannot see. beyond the too many scales above simply because you don't last long enough. So remember, unlike the Mandelbrot set, there is also a self similarity in time.

Things last for shorter periods of time at any particular scale relative to the scale above, which means me, as an observer of the universe, cannot last long enough to observe one flow. or one part of the flow, the [01:43:00] scale above. Furthermore, when I look down, things change. They don't exist for long enough to be measurable.

And if you think about what that means, at a very small scale, something cannot be actually measured because it doesn't last long enough to be measured. That's quantum physics. So now all we can talk about is the probability distribution of things if I were able to measure it, but now I can just talk about the bag or the probability distribution of a measurement because it's changed so quickly, I can never actually measure it.

So, you know, there are, there are limits on the mic, the depth that you can look down over scales, you know, using, even using. Technological extensions such as microscopes at some point, you know, you can't do it. So we're all coming back to sort of dissolving the binary perspective and Celebrating, um, the perspectival or relational aspect of the way that we make sense of the universe.

You know, we are just [01:44:00] tied to our scale, and we can't see many more scales above or below. And that, you know, we can only sense make and exist within that particular scale. Or, you know, one or two scales above and below. I think people think it's six, in the sense that you can't fold a piece of paper more than six times.

I'm not sure. That's right. Oh, these scales you can look at.

Andrea Hiott: Yeah, but that's from our scale, so it's very hard to, to hold that, that there might not even be necessarily this huge conflict, for example, between these different kinds of physics and systems. It might just be that we're positioning ourselves in places where those different maths are going to work in the, from those positions and not from others.

And we could hold that paradox in a way. That's what I meant by, does it only have to be one model that fits to everything? Because the more we learn how to put ourselves in different positions, maybe we're going to have to learn to hold the idea that different models work in those different positions, which is also very hard.

Karl Friston: Well, I mean, a lovely example of that is, uh, the difference between [01:45:00] say, um, general relativity and quantum mechanics. I mean, general relativity is that that takes things. Um, and quantum mechanics is about very, very small things. I think scale really matters because of this temporal dilation or, you know, you get with this sort of renormalization perspective.

So it could be the same kind of thing, but they are very, very different models and yet they're completely consistent. They can be held together just because you can only apply it at the scales above, um, and you apply the quantum mechanics at the scale, scales below. I don't want to demean quantum information theory that there are quantum information theory is actually also scale free, but quantum mechanics of small particles where random fluctuations dominate that's limited to very, very small things.

This gets to the

Andrea Hiott: other, sorry.

Karl Friston: No, I was gonna say, if you were talking to Mike Levin, um, you, you, his way of articulating that things only get interesting when you move from one cell or [01:46:00] one mo molecule to multiple cells or macro molecules. You know, you when you're building up to higher scales. So things start to get interesting.

Andrea Hiott: Yeah, I think Mike and I have had, had some conversations, maybe three already, about different ways of thinking about scales in the biological sense, so it definitely applies. And I think it also touches on this difference that I've been trying to get at between the models and the process or the flow, um, and how those get confused because there's a lot of fighting about, you know, classical mechanics versus quantum, this is this versus that.

And it's hard to understand how they can be held in, uh, a space that we haven't explored yet, right? Um but also I think part of the confusion can be at least from my point of view that It's too easy to mistake the model for the reality and you think you're fighting about the reality But you're fighting about the model and sometimes I feel this in the literature around free energy and I wonder if it's [01:47:00] For example, when you say something like that you're using hierarchical, no, you, I think you say the brain uses hierarchical models, um, correct me, but it maybe uses them in different contexts.

It, but the, it's, the brain is talked about as if the brain is using the model. And for me, that's a little bit confusing because I feel like from our position, we've understood that the brain is working in a certain way and we've modeled it, which can be different from the brain using a model. Although the brain could perhaps be using a model, for example, in corollary discharge or efferent copies or something like that. Maybe that's the kind of way in which the brain is used, the body is using its own model. But I feel like these. are different things, and they get confused. Or maybe I'm just confusing them, but

Karl Friston: Yeah, I think that, um, that confusion is prevalent, um, in the philosophy literature, or the philo certainly in the philosophy that, um, likes to, um, [01:48:00] shout about the free energy principle, um, and sometimes articulate in terms of the map territory fallacy, and the fallacy of the map territory fallacy, and, Well,

Andrea Hiott: the math, you're definitely not saying that, but it can I can understand how it easily gets confused.

It sounds like that. Do you, do you know?

Karl Friston: Yes, I do. And when I, sometimes I give up and sometimes I, I lapse. But, um, often when I, um, hear people write down, the brain uses a generative model, I delete that and say, um, the, the generative model. Um, it entails a generative model, but that generative model is just implicit.

It's just a mathematical construct that allows you to understand the flow, literally the gradient flow on the, uh, variation free energy. Yeah. You say it

Andrea Hiott: all the time. You also, I've noticed you say, describe it as a model or something like this. But it, it gets collapsed together, I think.

Must

Karl Friston: be frustrating. Yes, you can do, uh, well, no, it's entertaining. As long as somebody doesn't shout at me or tell me [01:49:00] to do something that I can't do, um, uh, but, you know, I, I, I think it's useful just to remind people, um, that you'll never quite know what the brain is doing unless you break it, unless you do some invasive physiology, in which case it's not the brain anymore.

So all you can do is infer. and describe the, um, what you can see of me, uh, in terms, it looks as if I am doing belief updating under a generative model and therefore it looks as if my brain and all the neural, uh, message passing is being updated. entails a generative model. And I can try and reverse what kind of generative model would best explain my behavior.

And that's quite a practical application of the entropy principle when it comes to things like psychiatry. So reverse engineering the generative model that would have been used if I wanted to explain this behavior in light of these, uh, these, these data. So I think it's a [01:50:00] useful construct when it comes to simulation and explanation, because you'll never know.

Because you can't, you know, whatever's going on in my, behind my Markov blanket is private to me. Uh, by definition, is conditionally independent of anybody observing me. Uh, so, you know, to me, it's, it's, it's, um, you know, it's not a worrisome question, but it is a vexed philosophical question, I, I, I fully accept.

Yeah, and I,

Andrea Hiott: I know it must be annoying for you, but I, I think it has repercussions all the way up and into the side and in all directions in terms of the way humans treat one another and the way we treat the world around us, because when we confuse our models and the processes that they're modeling, something strange happens in in terms of, , the way we think about, Perspective like what I hear you saying is that we always need to understand we're modeling from a perspective and We're modeling a [01:51:00] part of a process But that gets lost completely and that seems like the important part that situatedness Which is why I'm just trying to bring it back up to people to bring the situatedness back into it

Karl Friston: yeah, and one important aspect of that situatedness, um, is the fact that I spend most of my life inferring things about other people. So I think that's, that's something which which often gets lost because once you recognize that your lived world is basically generated by other things like you. So you spend most of your generative modeling trying to infer, predict, understand other things like you.

And then the perspective taking, empathy, attribution of agency, um, the theory of mind, all of these things now become really, really important attributes of good generative models for a community of conspecifics that, uh, live together. Yeah, these would not be issues if it was only [01:52:00] me on Mars. I wouldn't need a theory of mind.

But as soon as there's something like me, like you, I now need a theory of mind. And of course, then the as if ness now becomes, from my point of view, uh, quite a useful hypothesis or fantasy. that provides a simple explanation for the way that you behave. I think you're conscious, I think that you are inferring, I think you are, and I might be like you, so perhaps I am conscious, you know, and then you get to these sort of meta levels.

Andrea Hiott: It reminds me of the connection and to try to bring it to love a little bit because this is love and philosophy. Uh, I wonder about, you know, when, when you're, the disconnection hypothesis, right, of that we started with way early in your career with the schizophrenia patients and the cells were not connecting.

Maybe they were connecting in a different way um, I wonder If this happens on larger levels, in what you were just talking about, because we're trying to understand each other all the time as ways of understanding ourselves, and all this is very much [01:53:00] entangled, and I wonder if there's also ways we get disconnected, so to speak, in this fractal way we were talking, or if that's too metaphorical and too, , going too far out, or if you see that maybe there is some similarity, self similarity to these sorts of things.

Karl Friston: No, I think there absolutely is. And, um, indeed there are simulations, you know, baby step simulations, but certainly it's a growing field, uh, applying the pre energy principle, specifically active inference in a, uh, ethological or social neuroscience context. Indeed, there are funded programs by the EU now to look at, to apply these ideas to look and say, um, voting behavior, population voting behavior.

There's some very nice simulations, uh, already out there. that exactly lead to, to what you were talking about, which is, you know, the segregation into, into in groups and out groups, um, that rests upon effectively me inferring whether you are [01:54:00] like me or not. And if I can infer that you are like me, then you now, uh, interactions with you acquire an epistemic affordance and then we can consolidate our beliefs.

But if I infer you're not like me, Um, then, um, I am not going to, um, associate interactions or communication with you with any epistemic forms. And I'm going to effectively, uh, at the same time as basically dehumanizing you because, um, I now do not see you as the same kind of thing as me, therefore you are not human.

Not only that, but I will not listen to you. I will, uh, or I will treat everything you say as fake news. So these are completely Bayes optimal, um, you know, sort of processes, processes of inference at a societal or interpersonal level beyond the dyadic and can be applied to sort of, um, group formation and you can simulate the different conditions.

Um, but all, as you say, rest upon the fact that you are, um, [01:55:00] you are taking a particular perspective and also. taking the other person's perspective if and only if they share with you the same implicit generative model. So if there's a shared generative model, a shared narrative, a shared set of conclusions, and I can infer, sorry, if I can infer that there is a shared narrative, that you are sufficiently like me to share the same kind of generative model, then we can talk, and you are part of my group.

If not, then, you know, you are not, not part of me, and that can lead to, um, you know, uh, uh, The kind of segregation you might see in sort of reaction diffusion systems, where you get this sort of separation, segregation, which again, I say is completely base optimal and slightly, um, paradoxical in the sense that you'd think that, you know, our job is to be mutually predictable, and therefore we should all, um, share, sing from the same hymn sheet, so we can all predict [01:56:00] ourselves.

That's not what happens when you break the symmetry. Um, you know, unless if you took everything away apart from all, you know, all human beings and made them speak the same language, then you might get that kind of, um, relatively unstable group dynamic. Uh, but in the absence of that, then that kind of segregation is going to be an emergent property.

Andrea Hiott: Yeah, it's almost like learning a new rhythm. This, we've touched on it with the holding the paradox and the way you've described things, which are quite difficult and challenging to, for one to be able to understand. But I do feel this is a case in which working it out mathematically and working it out in these difficult situations, that are very practical in terms of bodily and brain dynamics. I do think there's a pattern there in the way we were talking about regularities and that is not exactly the same parts, but the process and the patterns maybe can help us understand [01:57:00] how to not do that segregation or to understand that certain ways we're going to be segregated from others, but if If we move our position, those regularities are going to align differently, and there are going to be ways in which we do connect, with just about anyone. Uh, if we could figure out how to do that fractal dance, which I think is it's really quite, quite difficult. But I wonder if that rings true for you or if you've found in your own life that living in that state of flow is a matter of being able to understand how to shift perspective in that way a little bit.

Karl Friston: Yeah, I think it comes back to doing as in what you're told to do. To do what you're told means you have to hear people. To hear people means that you are not invoking, if I get a mathematical kind of sensory attenuation, Um, and I use the word sensory attenuation because, um, of your observation that one can [01:58:00] apply the maths and the mechanics of computational psychiatry to societal interactions.

And I think that the ability to ignore, uh, the ability to evaluate the epistemic affordance of interactions with that kind of person as very, very small means that I'm going to, ignore them effectively. I'm going to attenuate any truthiness or precision, um, in relation to their exchanges, their posts, their messages, their ideology.

Um, so, um, that basically means I don't hear them. if I don't hear them, you can't do what you're told. So I think you're absolutely right. Being in the flow in the, in the, in the, in the sort of, um, more poetic sense, um, really does require you to suspend sensory attenuation and to listen, um, and to, well, both to listen and to also evaluate the precision of, of, of, of what you're, of what you're hearing.

Um, and then there will be an opportunity to. re evaluate your prior [01:59:00] conviction that this person is or is not human. For example, you know, thinking about current, you know, political crises and wars and the like.

Andrea Hiott: Exactly. That's a very, very beautiful way to think about it, and it makes me wonder if that's been your way of being in love.

Of course we could all mean different things by love, but when you were talking, I was thinking about presence and awareness that it takes to listen. You have to, there's a kind of stance where you step out of your narrative or yourself.

but more into your full body, which means you're, you are immersed in all that you think is not you, that that flow that we talked about or that place that you're in. It, it. It is a bit sentimental to say, but you do need something of motivation to go through a life doing the science that you've been doing.

It's a lot of hard work. And what you just described, that's a way of, for me, that seems like letting the energy refresh or, it it shows me how [02:00:00] maybe you were, when you're in that flow, it aids, it helps that process, right?

Karl Friston: Yes. Um, again, for me, I, I,, I suspect for you as well, uh, I know no other way, um, but it was nice you used the word motivation because the motivation is just the affordances and the affordances are largely epistemic.

So that's just curiosity and to be curious means you have to be able to hear, to listen.

Andrea Hiott: To listen is a, some can be an act of, of love. To, to listen and to act, to perceive and to act in, in that way, I think can also be an act of love, in the most kind of banal sense, but the most important sense, well, Karl, I really appreciate your work and your perspective and everything that you, you do. And I, you know, That's going to be a new year when I'm publishing this. So is there any, Thing you're looking forward to or any, any thing you wanna minimize, any surprise you want minimized or, what,, what [02:01:00] what's your feeling about the coming year?

 Um,

Karl Friston: no, I'm afraid. Yes. Uh, my new New Year's resolution will be not to make any more new New Year resolutions. So as you, as you get older, you just want to, as you say, minimize your surprise and carry on and keep calm.

Andrea Hiott: That sounds good. Well, I wish you the best and again, thank you very much.

Is there anything that you want to make sure gets said?

Karl Friston: I think we've said all the important things. Thanks for spending this time. My pleasure. I've enjoyed speaking and speaking with you. Yes.

Andrea Hiott: Happy Christmas. Happy New Year and be well.

Karl Friston: Bye.