How Your Brain Predicts Reality | Flow, Intuition & Mental Flexibility

One of the most influential figures in modern brain science. He is one of the most highly cited living scientists in general and was ranked number 1 by Semantic Scholar in the list of top 10 most influential neuroscientists.

So just to have an out-of-body experience immediately tells you you're dealing with a kind of sense-making that entails the hypothesis that I am a self that is embodied.

Out of metabolism and schizophrenia comes the largest revolution in modern neuroscience.

Do you think that we will ultimately find a unified theory of brain dynamics?

Stuart Brand once said the only sustainable happiness is the satisfaction of a job well done.

And phenomenological research that flow correlates with well-being, lower levels of depression, lower levels of anxiety, higher quality of life.

The things that we've been talking about today are now the aspirations of tomorrow's artificial intelligence.

Hey everybody, welcome to Flow Radio. My name is Dr. Michael Mannino. I'm the Chief Science Officer of the Flow Research Collective, your host today, and I am joined by our Executive Director and founder of the Flow Research Collective, bestselling author and Flow researcher Stephen Kotler. And today I have the honor of introducing our distinguished guest, Dr. Karl Friston. Karl's a theoretical neuroscientist and one of the most influential figures in modern brain science. He is one of the most highly cited living scientists in general and was ranked number 1 by Semantic Scholar in the list of top 10 most influential neuroscientists. He's invented statistical parametric mapping, SPM, and voxel-based morphometry and dynamical causal modeling. These methods have been foundational to modern brain imaging. He's also developed a theoretical framework called the free energy principle, which we'll talk about today. It's a unifying framework for action and perception that's been called the closest thing we have to a general theory of intelligence. Across his career, Karl's earned nearly every major accolade in neuroscience. From the Minerva Golden Brain Award to the Glass Brain Award for lifetime achievement in human brain mapping. And he's also a fellow of the Royal Society, the Royal Society of Biology, and Academia Europea. So his nomination for the Royal Society reads: Karl Friston pioneered and developed the single most powerful technique for analyzing the results of brain imaging studies and unraveling the patterns of cortical activity and the relationship of different cortical areas to one another. Currently, over 90% of papers published in brain imaging use this method, and this approach is now finding more diverse applications in EEG and MEG, which we'll define soon. His method has revolutionized studies of the human brain and given us profound insights into its operations. None has had a major influence major influence as Friston on the development of human brain studies in the past 25 years. And if all of those things weren't impressive enough, he's also our collaborator here at the Flow Research Collective. So together we've published a couple of papers, two recent papers just last week in Nature Communications Biology called Pathfinding: A Neurodynamical Account of Intuition, and earlier this year in 2025 in Neuroscience of Consciousness on Flow and Intuition: A Systems Neuroscience Comparison. So let's get into our conversation. I am going to pass it off to Stephen Kotler for our first couple of questions. Dr. Friston, thank you for joining us. Stephen, go ahead, kick us off.

Thank you, Michael. It was a great introduction. I don't think I've ever heard a Royal Society nomination. And, I'm really now wondering, Karl, which is bigger, the golden brain or the glass brain? Did they give you brains? Which one's bigger?

The golden brain is actually physically bigger. Okay.

That was my first question. Second question, we're a peak performance organization and we're going to talk about your astounding career in science as we go along. But, I have a peak performance question I want to ask you because I've noticed because we've been collaborating for a little while, you're incredibly fast. Like, you can turn around a paper, it comes in with really learned comments, footnotes, references, and it like, and it's fast. So how did you learn to do that?

I think it's because I don't like worrying too much about the future. So whenever I have something to do, I like to do it. Usually it's quite enjoyable doing it. And I often reflect actually, You know, what gives me the most pleasure? And sometimes it's just after a hard day's work and doing something that matters. And for me, of course, going through people's papers and thinking about sort of the key issues at hand matters. And, you know, having done it, it does leave you with a sense of accomplishment, and I can then relax, smoke my pipe, and watch some television until the next paper comes along.

Stewart Brand once said, Stewart Brand once said, the only sustainable happiness is the satisfaction of a job well done. And that always sort of stuck with me. I think he's right. Like over and over, I really think he's right. And it certainly seems to be the structure of how the brain records memory and kind of adds up into things like meaning and satisfaction. This is a very off-topic statement. So let me go into my next question, which is a little bit historical. I like to start with a sort of a media historical question often because it just frames things up really well for our listeners. So, we're going to talk a lot about the free energy principle, which I'm going to ask you to define in a second, but it emerged out of two different things. It seemed very unrelated. You were studying metabolism and you were working on schizophrenia. And out of metabolism and schizophrenia comes the largest revolution in modern neuroscience. So start by walking us through how that happened. How did they come together? How did they birth the free energy principle? What is it?

As with most things, I think a happy coincidence. So as a young scientist, I was briefed with making sense of the first wave of human brain mapping data at the inception of imaging neuroscience with positron emission tomography. And that was my first research job. Prior to that, I'd been working at a therapeutic community with people with chronic schizophrenia. So I was tasked with trying to make sense of brain imaging scans from cohorts of people with schizophrenia to understand differences in brain activation using this new technology. So we were searching for the best ways of making sense of data, and after about a decade, we identified the most apt approaches, coded that up, and made that code available to other people. And that was the statistical parametric mapping that Michael mentioned in the introduction. At the same time, next door, Geoffrey Hinton arrived as the director of the Gatsby Computational Neuroscience Unit. Excuse me. The Gatsby Computational Neuroscience Unit. And was inspirational in the sense that it became obvious that the best way to make sense of brain imaging data could also be applied to the brain in and of itself, in the sense that the brain has to make sense of its visual data, its streams of interoceptive data, proprioceptive data. So, we applied exactly the same mechanics that we were using to analyze neuroimaging time series to the theoretical neurobiology of how the brain does its analysis. And there was a beautiful match in terms of the requisite computational architectures. From my point of view, mapping those computational architectures onto the functional anatomy was an obvious move. so that one could then understand message passing in the brain in terms of the kind of inference that we were doing when effectively creating X-rays of function, understanding the brain's functional anatomy in terms of identifying which parts of the brain were specialized for this or specialized for that, how that specialization was integrated by distributed processing and neuronal message passing, and then identifying the requisite structures, the requisite architectures that would enable us to understand the brain working in exactly the same way that we were working in terms of identifying the— making sense of that brain imaging data.

Oh, Stephen, can I piggyback on that real quick?

Yeah, no, I will. Over to you, Michael, it's your question.

Oh no, no, I just want to like follow up with this. And for our viewers, um, Carl, I introduced you as a theoretical neuroscientist and you just mentioned Hinton, who just won the Turing Award and the Nobel Prize for Physics, as a computational neuroscientist. So We've talked a little bit about this before on the podcast, but just for our viewers now, the idea of computational neuroscience, and this will lead to a question for you, Karl, the idea of computational neuroscience and theoretical neuroscience as coming up with a mathematical theoretical framework that describes and explains and makes predictions about the brain. And the fact that you can apply physics and math to the brain, which is flesh, right? And it's just so fascinating, just like you can apply, you know, physics is a hard science, right? And so computational neuroscience is this idea of mathematically modeling the brain. And the fact that you can do that is something that's really amazing. And you're, you know, one of the godfathers of this. And I'm just wondering, how, and so this is different from behavioral neuroscience, right? For example, or molecular neuroscience, although there's a lot of computation there, but you started with data and then you came up with a theory, or did you start with the theoretical framework first and then find data?

Ooh, interesting question.

Empirically support it, or was it sort of both? And this leads to a big question to start off right away. "Do you think that we will ultimately find a unified theory of brain dynamics, or do you think the free energy principle is that?" Gentlemen, could somebody define the free energy principle for our listeners before we go on? Yes, Karl, I will let you do that. You're the father of the theory, the creator of the theory.

Yes, I am, and as we were musing before we started, I'm the worst person to try and summarize clearly some of these issues. So the free energy principle is basically a method or a, a law which describes or allows you to describe sense-making and decision-making in the brain in terms of physics, in terms of mathematics, specifically the physics or mathematics of probability distributions or probability densities, which we fondly refer to as beliefs, Bayesian beliefs, or conditional probability distributions. And you can understand the free energy principle from two perspectives, sometimes referred to as the high road and the low road. So the low road is, I think, a more intuitive approach whereby the brain is viewed as a constructive organ, a statistical organ that's actively constructing explanations for its sensory data. On this view, perception is very much akin to the notion of unconscious inferences introduced by Helmholtz, namely that you have to have a fantasy a hypothesis, some explanation in mind in order to explain the sensorium. So this is very much, as Andy Clark would say, an inside-out sort of process, as opposed to a more Marian or 20th century view where sensory data impresses itself upon the brain and somehow information is extracted, a an outside-in process. Probably best, this sort of inside-out process, this constructive process, probably best understood in terms of predictive coding, which has now become one of the favorite metaphors, mathematical images of message passing and predictive processing in the brain. The idea here is that you predict what you would see if your particular hypothesis about the current states of affairs was correct. And then you use the sensory input to test your hypothesis. And if your hypothesis is good, then there should be no prediction error, where the prediction error is simply the difference between what you predicted in this sort of inside-out or top-down way minus the inputs, the outside-in or bottom-up inputs. So, the prediction error is just the mismatch between what I predict and what I'm actually sensing. And if I'm right, there won't be any prediction error. If I'm wrong, there will be a prediction error, but I can use that prediction error now to drive my beliefs, technically Bayesian belief updating, in order to supply better predictions that then effectively resolve or explain away the prediction errors. So, that will be the low road, and there's lots of, you know, wonderful philosophy from Kant, you know, that predated Helmholtz's notion of unconscious inference behind that, developed in many arenas such as analysis by synthesis. And indeed, Geoffrey Hinton and Peter Dayan actually created a Helmholtz machine shortly before coming to London that emulated this mathematically. The high road gives you the same destination but allows you to look at exactly the same mechanics and the same principles from the point of view of self-organization. So the free energy principle is more the sort of physics backstory to this predictive processing, this sort of constructivist understanding of perception and indeed action in terms of planning as inference. So, the free energy principle just starts off by saying, what are the characteristic things that something might have to express or manifest if it existed for a substantial period of time. And then you start to ask, well, what is a thing? And then you realize that a thing has to be separated from everything else. And you then think about the mathematical basis of that separation or that disambiguation, and you get to something called a Markov boundary or a Markov blanket. And then you just apply the laws of physics, usually a path integral of the kind that Richard Feynman introduced, in order to describe the dynamics of the inside of the thing in question that lies beneath the Markov boundary or the Markov brain in relation to the outside, the external world. And at that point, you notice that you can describe the dynamics of any internal states as effectively a gradient flow or a dynamics that tries to minimize exactly the same quantity that underwrites Bayesian belief updating. And that quantity is— it comes along in many different flavors. If you want to minimize it, it's known as self-information. Surprisal or surprise, hence the close connection with prediction errors. Prediction error just registers your surprise if you're wrong, and thereby if you minimize your surprise, you're going to be minimizing your prediction errors. The negative or the complement of this self-information or surprisal is called Bayesian model evidence in statistics. In machine learning, it would be called an evidence bound, an evidence, an ELBO, an evidence lower bound, also known as a marginal likelihood. All of these words are just ways of expressing the plausibility of your sensations given your model of those sensations. I like the model evidence, the Bayesian model evidence, because what that means is that just by existing, you are— or it looks as if your internal dynamics are always trying to maximize model evidence. So another way of expressing that is that you are effectively self-evidencing. You are securing or making sense of data in a way that provides or furnishes evidence for your model of the world. And that model is a generative model or a world model if you're in machine learning. So, this is a phrase that was introduced by Jakob Howie, the notion of self-evidencing, that just by existing in a way that enables me to be separate from my environment and yet part of it and exchanging, in open exchange with the environment, across my Markov boundary or my Markov blanket, I'm effectively garnering evidence for my own model of the world. In fact, I'm garnering evidence for me. I'll just make one more comment just to sort of speak to your question about dynamics. So what you have at hand basically is a description of sense-making and decision-making perception and action that provides the sensations that you're making sense of, which can be written down just as a particular kind of very simple dynamics. So in answer to your question, I think the free energy principle is— it is just a principle of least action that describes the dynamic, the internal dynamics of something that is in exchange with this world.

Yes.

I, I'm gonna ask a very— I've got a follow-up. I want to talk about intuition in a second, but as you were talking, this may be a really dumb, obvious statement question. You've worked on psychedelics. Um, I was wondering, like, is— if we're having an out-of-body experience, we're having a psychedelic experience, is that— that's a radical shift in the Markov blanket that surrounds the organism. If we're thinking about an out-of-body experience as a consciousness moving, or is that a totally dumb idea?

Is it—

am I anywhere close? I was just wondering it when I was— when you were talking, it was like, do we use— I know we use it to describe the sensorium and the boundary between the sensorium in and out, and all these things change consciousness. Does it direct impact on the Markov blanket? Is that how we're modeling it?

I think an out-of-body experience immediately begs a really deep question that you have as part of your repertoire, as part of your world model, as part of your generative model, the notion that you are a self. You actually have a body, which is quite a gift. You won't find this kind of very expressive generative model in a virus or a thermostat. So, just to have an out-of-body experience immediately tells you you're dealing with a kind of sense-making that entails the hypothesis that I am a self that is embodied, which is quite remarkable, you know. And then to furthermore be able to entertain the hypothesis that I am now outside my body is even more remarkable, in the sense it is a counterfactual that is very, very seldom experienced. You could even argue that that kind of hypothesis, that I can be me but divorced from my embodied self is the kind of hypothesis that philosophers entertain, in the sense of the philosophical zombie. Yeah, the very idea that you can have a being that does not experience itself is, if you like, in and of itself a quite remarkable feat. And again, it speaks to the fact that the way that we make sense of our world is that we put ourselves at the center of that world. So the Markov blanket, from the point of view of the physicist and the observer of the person having an out-of-body experience, will never change. But the very fact that from the first-person point of view, subjectively, they now realize that they were They are separate from the world, so they have a self. There's a distinction between self and non-self. And yet now there has been a dissociation between the sort of the physical embodiment and the concept of self.

So that dissociation doesn't shift the blanket per se. That was what I was trying to get at is, you know, or when Shahar Arzy's really crazy work in the University of Jerusalem on the doppelganger experience, which like he worked on with Moshe Adel, comes out of Kabbalah, where people actually see a projection. So, there's the inverse of the out-of-body experience, so that the sense of you leaving the body, you see a version of yourself. It still seems to be right in this temporal-parietal junction, the same part of our brain that does perspective-taking seems to be going haywire during these experiences, but it doesn't— the Markov blanket just governs the outside of the sensorium no matter how big or small or what we're looking at.

Yes, I think that's right. But I mean, you know, it'd be really nice to just pursue this line of thinking because, you know, in certainly sort of body swap illusions, for example, there's something quite profound about—

the hand, Peter Berger, right, in Switzerland, who was working on the hand, right, people would stab the hand.

Yeah, yes, and the rubber hand illusions and the like. So, these tell you something, I think, which, you know, quite profound in the sense that it is easy to induce these illusions that I am somewhere else or that this artificial or rubber hand is my hand. And that tells you that our everyday, almost unacknowledged belief that we are an embodied self is also an illusion. It's just the best explanation that we have for the sensory data at hand as we sort of live our daily lives in the lived world. And yet it it can be quite fragile, certainly with, you know, say psychedelics. It is just a hypothesis, it's a fantasy, and I use the word fantasy because that basically means that the brain is a fantastic organ. It is a purveyor and the generator of fantasies, including the fantasy that I am me and that I have a body, and it can Of course, with the right kind of experimental manipulation, usually, you know, of the kind that induces an illusion, you can so easily destroy it. But of course, you can also revisit or challenge these hypotheses or these illusions in many, many different ways. You know, meditation, psychedelics, or, you know, anything that will produce things like out-of-body experiences out-of-body experiences, or even the brain itself.

Like, there are diseases like somatoparaphrenia and alien hand syndrome that people think that their arm does not belong to them even though it's attached. And also the reverse with phantom limb syndrome as well. So it's— it is a fantastical hypothesis that the brain creates. Yeah, Carl, I want to piggyback on the psychedelics for a little bit real quick. The— so Free energy principle tries to minimize the surprise. The brain tries to minimize the uncertainty between its expected and what it's— the sensorium, what's coming in, and the model it has. So there's two ways that you could change and make the prediction error— and to minimize the prediction error, that is to minimize the discrepancy between the prediction and the expectation And so the two ways are you can change the models in your head and you could change the world. And that's where your notion of active inference comes into play. And I want to tie that back to psychedelics. We, we wrote a paper on flow in 2022 with Scott Kelso and Richard Husky on how the brain transitions into flow state. And in that paper, we talked about the free energy principle and Scott Kelso's theory of metastability, which we can get into later. I just want to seed the audience now that we're going to talk about that, and flow and psychedelics, and it's sort of a theoretical overlap between those. Your theory that you posited in 2019 called Rebus, the relaxed belief under psychedelics, I just want you to talk a little bit about that and how, how you came to that insight. But I think the general idea, if I'm not mistaken, is psychadelics relaxes the hold of the default mode network on how the brain interprets information coming in and its own models, so that you can change these assumptions and update your beliefs about yourself, maybe your past and your models, and thereby heal mentally with psychedelics. And we argued in that paper that flow also does that as well under some similar mechanisms and can help with PTSD and things like post-traumatic growth, which we're now actually doing research on with, with you, Carl, on mental health and flow using sort of the free energy principles of theoretical framework. So I'm wondering if you could speak to how you came to this idea and how psychedelics do that. They relax that one part, the hold of the models.

Yeah, I mean, you've introduced some quite fundamental issues and points there. The simple distinction between action and perception, I think, just needs to be rehearsed. But we're going to be talking about a very special kind of action, which is action on the inside, which is the attenuation or augmentation of attentional decision when picking out the kind of sensory signals, specifically predictive prediction errors in a predictive coding formalism, that are privileged in the sense that they can induce belief updating. But let me just sort of reiterate the importance of what you just said, just setting up the this kind of mental action. So the free energy principle talks about the distinction between self and non-self as in terms of a separation intervening, or an intervening boundary, and that boundary allows for a bi-directional openness to the world, to the environment. So the world influences you through the sensory sector of your Markov blanket, and in an exactly symmetrical way, you influence the world through the active sector of your Markov blanket. So there's a lovely circular causality, you know, often conceived of in terms of a perception-action cycle, which means that everything that I think and do is in the service of minimizing surprise, free energy, or prediction errors. And as you've just said, there are two ways that we can minimize prediction errors. I can just change my mind so that I now have a different understanding of the causes of my sensations, and that will resolve my prediction errors. Or I can just make my predictions come true. And I can do that in the motor domain just by moving or I can do it in the autonomic domain by invoking autonomic reflexes, so that now predictions about my homeostasis and my interoception are fulfilled. And in that way, we construct or create a world that is sampled in such a way that I can always predict what's going to happen. And this, you know, if you like, closes the loop, and it enables me to survive in a capricious and changing world. What I think you want to talk about, and what I would like to talk about, is a very special kind of action though, and this is not actually moving with a muscle or indeed moving with a secretory organ, engaging an autonomic reflex. This is action on the inside, of the brain in and of itself. And I think it's probably best understood using the predictive coding formalism. So, we've said before that the Bayesian belief updating, the sense-making, the dynamics that enable us to infer the causes of our sensations are driven by these bottom-up or ascending prediction errors that have yet— the newsworthy information from the senses that has yet to be explained. But I have to select which news channel to attend to. I'm bombarded with a very high-dimensional sensory input, and I have to actively select. So this is another important aspect of active inference. I have to select which prediction errors to listen to. So how do I do that? Well, if as a psychologist, I'm going to attend to this kind of information or this modality or listen to this social media or this news channel, if I was a statistician, statistician, what I'm basically saying is I'm placing more confidence or trust or precision in these prediction errors relative to those prediction errors. And if I was a pharmacologist or a neurophysiologist, what would that mean? Well, it means that I have to now equip the neurons reporting these prediction errors with more potency, more gain, literally, the postsynaptic gain, so that would be manifest as an attentional gain, so that now these particular prediction errors are selected and have privileged access to belief updating, usually in a hierarchical generative model. So, now we have this notion of, well, to what extent then do I allow my prediction errors to revise my beliefs? Because I can just ignore them. This would be called something like sensory attenuation, and therefore my prior beliefs before I looked at the evidence at hand will remain unchanged.

But would that also be cognitive bias? Is that—

Yes, yeah, absolutely, yeah, yeah. You're relying upon my prior beliefs Just our biases, yes, and they're good biases, the biases that make me work and the things that I've inherited through my experience-dependent learning, and you know, also probably some of these endowed by evolution. So we've got this sort of balance, as with any kind of inference, and particularly Bayesian inference, between how much we let the sensory evidence update our prior beliefs. If the data are really imprecise and really noisy, then I'm not going to change my mind very much, which means that the precision I'm assigning to my prior beliefs, high and/or deep in any cortical or subcortical hierarchy, outweighs the precision I'm affording the sensory evidence. Conversely, if I now relax the priors of my deeply held convictions, I'm now empowering the sensory prediction errors to change my mind. [Speaker] Right. [Speaker] So, and we have to all the time, moment by moment, we have to sort of dynamically adjust and get this balance right. And this is what I meant by the sort of the internal action with lots of Markov blankets on the inside. Which is, I say that for Stephen because there is a way in which you could have your Markov blankets do change because in a sense if I am in control of the precision afforded prediction errors at any level in the hierarchy.

Oh yeah, of course.

Each level is a Markov blanket. I can now dynamically reconfigure my internal Markov blankets in a way which I think fits very comfortably with your introduction of the rebus hypothesis with Robin Carhart-Harris. So this was the notion that psychedelics and many other modulatory drugs that act upon modulatory receptors that have neuromodulatory effects have a selective effect on the gain or the excitability or the precision of higher or prior beliefs, thereby, relatively speaking, equipping the ascending prediction errors with more potency in terms of them being able to revise or update my beliefs. So, dissolving those cognitive biases, those prior beliefs, and sensitizing them to the sensory input. And the, you know, it's almost, if you like, a natural conclusion if one just thinks about the computational and physiological anatomy of predictive coding in the brain. That it is almost obvious that if you want more peripheral, closer to the sensory, more, if you like, shallow levels to have more influences on higher levels, then you just increase the excitability of the lower levels that are the source of these ascending messages, the ascending prediction errors. So that's in a nutshell where the rebus hypothesis came from. And as you say, you're basically liberating yourself, if you can, from these cognitive biases that just are your priors. And sometimes those priors will include the illusions we were just talking about, the prior belief that I am a self and I am embodied.

Yes.

These are just illusions. And if I reduce the precision of these so that I can entertain other hypotheses— so perhaps just to put a bit of flesh on that, the precision effectively can also be understood as the dispersion. So if I have a number of hypotheses hypotheses about states of being at the moment. A very precise belief would be one where the probability of being in this state is 1 and 0 everywhere else. Got it. I may have alternative hypotheses about states of being, and, and if I reduce the precision, that basically means all are possibly equally viable. and then the probability distribution spreads. So, when I talk about sort of relaxing the precision, what I'm effectively saying, or what Robin and in fact you are saying, is that you're now allowing other hypotheses, other fantasies, to be part of the game in trying to explain the sensorium, and implicitly, just by relaxing the precision over alternative hypotheses of states of being, we are now sensitizing ourselves to all sorts of sensory information. Because if I don't know what's my state of mind or my state of being, then the only way I can resolve my uncertainty is by attending to the evidence. What does the sensorium say? And of course, —then you can understand the psychedelic experience in terms of now focusing on or attending to the elemental sensory evidence, you know, that will hopefully adjudicate or will now be potentially explained by these relaxed priors in the sense of rebus. Is that what you were driving at?

Yes, yes, that's a great answer to your question. Yeah, yeah, with the nature of the rebus model and how it works. No, so I don't know, Stephen, maybe we could talk about the, the idea of flow here as well. We argue also, just piggybacking on that a little bit, that, that flow, um, the flow state, which is neuroprotective— a very recent study actually came out showing, um, just last year actually, showing it was a huge study on 9,000 the genetic study on 9,000, 10,000, 10,000. Yep. Set of twins showed that flow is neuroprotective against things like, and maybe like psychologically protective, but I use the word neuroprotective, anxiety, cardiovascular disease, depression, schizophrenia, bipolar disorder, bipolar disorder and schizophrenia. Yeah. So, there must be some, and we argue, we're working on this, right? We're researching this. There must be some argument that flow has possibly some similar effects. And we actually started that argument in this 2022 paper.

We also saw, you know, one of the things with psychedelics is you sort of get a bigger brain effect, right? Prior, get relaxed sort of all over the place with flow. It's the same thing, but it's very constrained to the task at hand. You get very creative about what you're doing, but it's bounded. Whereas with psychedelic experiences, it's not, except I was thinking of this as you were talking, Carl, you don't know this. My original work was with Dr. Andrew Newberg and I came in right when he was looking, doing SPECT imaging of Tibetan Buddhists and Franciscan nuns experiencing cosmic unity, that feeling of one with everything. And, I met Andrew because I was studying flow and this was happening all the time in my athletes and nobody wanted to talk about it because they thought they'd sound like they were crazy. They didn't want to actually get kicked off their teams for having mental illness. And I was like, no, no, no, I think there's an explanation for this feeling of oneness with everything. And I called up Andrew and that's sort of where this work started. But it was that. So that's essentially, but to put it in the language that we've just been using, you've relaxed a lot of priors about where your body ends and where the rest of the world begins. It happens because of something that goes on in the temporal parietal junction, brain structure we talked about earlier, among many other structures, I would guess. But that— so that's— there is definitely overlap there, and there's definitely overlap with the creativity that you see in psychedelics. It's just more specific.

I mean, it may be useful just to sort of celebrate the very close connection between psychedelics, non-dual states, meditation. I hesitate to say mindfulness, but I think mindfulness may have a similar kind of mechanism. All of these things are, I think, from the point of view of active inference, that all of these things are really different states of the balance of the precision in this hierarchical context. And further to that, many of these are skills. Now notice that you can be a skilled meditator or you have to be a skilled mountain climber to be in the flow. Notice that these are all things that you do. This is not sort of unconscious inference, it's not sense-making, this is basically a reflection of your skill in terms of doing things. And this instance, it's having some volitional control over the deployment of your attention. So not all precision weighting, not all sensory attenuation, which would be the reduction in the precision, the ability to ignore stuff that is not relevant, or selective attention, which would be the, you know, the complement, increasing the precision. Not all of this is under volitional control. So my favorite example is saccadic suppression. So every 250 milliseconds, we move our eyes through a saccadic eye movement to visually palpate one part of the visual scene or another part of the visual scene, gathering evidence to resolve our uncertainty about, you know, what's causing my sensations at the moment. In this instance, you know, looking at your facial expressions just to make sure that, you know, the intended state of the world is pursuing my, or the state of the world is pursuing what I predicted and thereby intended. That ability is not always under volitional control. So, for saccadic eye movements, when we're actually moving our eye, that produces optic flow. You know, it's as if the whole world was moving, and yet we don't see that. We don't see the world move. So what happens, and the reason we don't, is we, for about 50 milliseconds, reduce the precision. We attenuate the precision. So that's what I meant by sensory attenuation. We attenuate the precision in order— because that's not useful information. I cause that information. I cause that sensory flux or that optical flow, therefore there's no useful information. And in fact, if I did attend to it, I wouldn't be able to move my eyes because the prediction errors would be saying, 'Oh, your eyes are not moving,' so I have to attenuate in order to move. But notice you can't intervene on that volitionally. So there's, I think, a great challenge, certainly in terms of sort of mental well-being and mental health, or an opportunity, which is to render what sometimes is automatic and not under volitional control controllable by will, or by practice, or by some kind of skilled behavior. So that's why I was emphasizing the fact that we are talking about very often skilled behavior. So to be a skilled meditator is to be able to deploy your attention to, say, your breathing at the expense of everything else in a way that, you know, would normally not be accessible to conscious manipulation. So, you know, the ability to gain some control over the deployment of precision is, to my mind, one of the key aims of many therapeutic interventions. And just to sort of link that back to your observation about how flow has a neuroprotective aspect, I think that fits beautifully with the computational psychiatry approach to disorders of mind, which are invariably describable as some kind of false inference. So if you remember back to your school days, there are two types of false inference: Type 1 errors and Type 2 errors. So Type 1 is inferring something is there when it's not, so that would be like having an hallucination or a delusion. And then Type 2 errors are inferring something is not there when it is, and of course that's a very apt description of things like dissociation syndromes, neglect syndromes, and you mentioned some, you know, sort of phantom limb, some beautiful examples of false inference. Now, how do you get false inference? How do you break inference? The simplest way to break it is to get this precision wrong. So imagine Well, let's just take a statistician doing a t-test or trying to work out an effect size. So what she would do is she would estimate the effect of an intervention, and then to get the normalized effect size, she'd have to actually now estimate the standard error or the standard deviation or the variability, or its inverse, the precision. So in order to make an inference, you have to be able to know, represent, and modulate in the right kind of way the precision. So if you are, if you are in a position and you have become skilled in terms of manipulating your, volitionally controlling your precision, I would imagine that that would certainly, from a sort of statistical or computational perspective protect you against the kind of false inference that characterizes many of the psychiatric syndromes that you just listed.

We, in the work that we did together, one of the things we did was we looked at intuition through the lens of the free energy principle and metastability, Kelso's theory. I was wondering, this is clearly not the first time you were thinking about intuition. I know you've worked on meaning for for a really long time and things along those lines. And I know that in our most recent paper that we probably shouldn't talk about much because it's not published at all, but we're looking at purpose as it relates also. And it seems like all of these things, much like your description of kind of the altered states, it seems like obviously if we're looking at a unifying theory in brain science, which I think we are as well with the free energy principle, But when did you start thinking about intuition? Where'd that come from? I know we dragged you into it in our approach, but I know that wasn't the beginning for you.

Well, in terms of intuition, it probably was actually. Oh. But there are related notions succumb to an analysis under active inference in the free energy principle. And they basically, again, appeal to this notion of a repertoire of hypotheses, a repertoire of fantasies about the way things could be, and being able to very efficiently disambiguate or adjudicate among these different structures or these different models or different explanations for what could best explain that. So, what I'm talking about is a kind of Bayesian inference that's not in the moment in terms of inferring states, but much more about the very structure of the model in and of itself. And some people refer to this as structure learning, and thinking of people like Josh Tenenbaum, where the structure has the symmetry, the rule, the invariance that, if you like, forms the basis of an intuition. That, you know, I have an intuition that this is— it could be work like this, or it could work like that. And as soon as you've got this space of hypotheses, these spaces, space of laws, or space of symmetries, then you can go and just look at all the evidence that you have accrued through experience-dependent plasticity, and ask, well, which is the best explanation? Which intuition is best served by my experience? But of course, the challenge here is to actually generate this space of structures that you can now decide among. To identify what is the best explanation. And that certainly would require a degree of creativity because what you're saying is you're now bringing to the table structures and generative models that have not previously been entertained. And that clearly would require— to refer back to our early conversation— a relaxation of our cognitive biases or priors in order to just be able to generate explanations in a creative way that have never been previously entertained, so that you can now test one against each other in relation to the accumulated or accrued experience to say, 'Ah yes, that's how it works.' So what I'm thinking of here is sort simulations of aha moments where your intuition is verified, and that basically, you know, is suddenly a collapse of your uncertainty about it. Could have been like this, this, this, or this. Ah, it's definitely one of those things, and you, you, you, you, you resolve your uncertainty about the way things were. And usually it's, it's finding the same explanation for multiple instances. It's, oh, it's just one of those things. So, it's usually finding a simpler explanation for things that go. But I'd also like to turn to the, you know, the notion of metastability, because I think that's quite important as well in terms of the mechanisms that allow you to explore alternative intuitions or explanations. If you've If you are overly precise, then that effectively means you're stuck in a rut. Yeah, so you can imagine the, um, a Waddington landscape of hypotheses or ways of being or ways of acting or ways of getting, um, dealing with a particular situation. And if you're, if you have overly precise prior beliefs about you and how you are responding in this situation. You can imagine that as effectively a landscape where the negative probability or the sort of the, you know, the surprise is very, very narrow. So if you want to minimize surprise, you're at the bottom of a very, very a narrow ravine, or a rut, or a canal. So, the relaxing of the priors basically just opens things up for you, allowing you to escape that rut or that canal. And in a sense, metastability is basically a statement that you can jump from one rut to another rut and start to explore different explanations, different local minima, in this Waddington landscape, which, you know, for me would be a free energy landscape. So by relaxing the prize, you are now revealing the potential for metastability and to be able to explore different states of being. So I love the notion of metastability, which almost comes for free with the free energy principle, in the sense that there is one sort of paradoxical aspect of sense-making in this Bayesian style, or under the free energy principle, that actually inherits from counterpart in statistical— in the physics of measurement, which is called Jaynes' maximum entropy principle. So in making sense of things, almost paradoxically, what you are trying to do is to keep an open mind as possible and not commit to a very particular overfitted explanation. And that, from the point of view of Jaynes, would be maximizing the entropy, hence the maximum entropy principle. What it means from the point of view of the Worthington landscape is you're always trying to find an explanation that doesn't lock you in too much. So you're always trying to find an explanation where you've got a bit of latitude to move around. Technically, this basically means that the curvature of this landscape is relatively small. So by definition, the free energy minima are always quite broad ruts or canals or valleys, if you like, providing therefore the opportunity to jump to another meta, you know, in the sense of metastability. All of this, um, you know, depends upon having a sufficient relaxed prior precision. So I repeat, the precision basically determines the curvature or the depth of these landscapes that can sometimes keep you trapped with one cognitive bias, if you like, that you cannot escape simply because you've endowed it with too much precision.

Metastability, by the way, oh, I was gonna define it for our people. Yeah, it's metastability. The two things that you said that I wanna double-click on, 'cause first of all, intuition is a reduction of surprise, right? It's part of this, everything we've been talking about, that's what we're talking about. We talk about intuition as a pathfinding system. So it's kind of a bounded reduction of surprise within a certain set of parameters. And metastability is the ability for populations of neurons to have individual behavior and have group behavior. And they can be separate. They can be different things. It's a tendency for oscillating networks to almost come together and lock into a pattern, which would be that really steep valley that he was talking about. And right there, they burst free and go elsewhere. If we were to talk about this psychologically, I always think about it psychologically as adaptive flexibility. Metastability is like the internal version of adaptive flexibility. If you guys remember our conversation with George Bonanno, adaptive flexibility is the heart of resilience. So, another thing to think about is that metastability is at the heart of resilience and some of the work that we've been doing recently that Michael alluded to earlier on PTSD and depression, those kinds of things. One of the things that we're going after, those are— in all those conditions, brain networks become very inflexible. Default mode network in depression just keeps spiraling, right? In PTSD, it's the amygdala and the hippocampus, and they can't stop firing. Addiction is similar, but it's the goal-directed network, etc. So all of these disorders are network inflexibility at a root cause, and by restoring metastability, things get flexible again. Okay, I'll stop there. Ryan, you were going to jump in and say something?

So in terms of flow and these valleys that we're talking about, is that, that's basically what you're asking, Carl? And metastability. And metastability, yeah. I agree with Stephen in the conversation about cognitive flexibility and metastability being Carl, I like the marriage that you talk about in terms of free energy principle, metastability, being a mechanism for how the brain moves around in this landscape of free energy, this attractor landscape as it's called. And I think flow is very interesting for several reasons and fits into this for several reasons. One is when we're skilled at a task, that is governed by the challenge-skills balance, which is a flow trigger, I think this focuses the brain for higher levels of cognitive control. And so we argue, of course, the brain is a self-organizing mechanism. It's a dynamical system, and metastability allows for that. So metastability really allows for this mechanism of being, you know, so— like, that's the phenomenological experience. I think metastability accounts for both the phenomenological experience and the underlying neurobiological mechanisms for how the brain switches gears into a higher level of cognitive control, which is really kind of what we're arguing about flow theoretically. Empirically, we actually just submitted a paper which is being published in NeuroImage where we looked at some data, some fMRI data of folks in flow playing a video game in this case, and found higher levels of metastability with specific dynamical complexity metrics while they were in flow as compared to boredom. So usually when you're looking at flow at the brain, you're comparing different conditions. That is the boredom condition where the challenge is much, much lower than the skill. The flow condition, where the challenge-skill balance is there and the challenge is a little bit greater than the skill of the task that you're doing. And then the frustration condition, where the challenge is much greater than the skill. And we see differences in the brain in flow conditions as compared to the boredom condition and the frustration condition. And we see, we see basically looking at complexity metrics of neural time series analysis, that there is some, there is indication that metastability is at work as an opera— so when you operationalize metastability, that is with these mathematical metrics, we see that those can actually capture and track these three conditions. The flow condition, the boredom condition, and the frustration condition. And so we argue that that's— we found empirically that that's the case. We've argued theoretically that that's the case. And then we combine that with the free energy principle as a theoretical explanatory framework for this mechanism. So that's for— and that we're seeing, we're seeing that in the brain. More research needs to be done, but it's a computational approach to neural activity when you're in flow as compared to not. And so that's very interesting and there's some really good early empirical results justifying that. Does that answer your question? Yeah, no, that's—

Talk about the edge of criticality a little bit for folks too, Michael, and what metastability does there and how that influences flow's creativity.

Yeah, so the brain sits at the edge of chaos. That's what the theory— that's what the words people use, Carl. I see the smile on your face because it's a naive— I don't want to naively oversimplify this, but meaning, so in fact, the brain, when we talk about chaos theory, the brain is sort of chaotic in that it's Areas, as Stephen, you were talking about earlier, synchronize and then they asynchronize and then they come in synchrony and then they go out of synchrony. And it's this balance between going in and out that allows for metastability and functioning of the brain. And so in fact, when people have seizures, a large swath of the brain start to really synchronize. And so we say the brain is at the edge of criticality. There's this self-organized criticality, and the metaphor in the literature coming from some scientists, early neuroscientists like Peter Bach, is that you have a grain of sand on a pile of sand, and the grain falls and causes an avalanche, and that's, you know, sensitively dependent on how the structure, the initial conditions of that pile of sand where the avalanche happens. And so that's really how the brain operates at the edge of criticality.

Hopefully that was— Michael, you just provided me— this is totally off topic, but that damn Sam metaphor, when I was trying to learn this stuff early on, I was so confused. I was like, wait, huh? What? It just— the Sam metaphor, you're taking me back like 20 years.

Sorry if that's too crude.

So I, yeah, I think, um, yeah, I'm really excited about the research that we're doing for flow and, um, using metastability and free energy principles, backdrops, and explanatory frameworks for how flow works.

It's very promising. Um, that notion of self-organized criticality, I think, fits in really in a very telling and, and compelling way here. Um, I, I likewise smile when I hear, when I remember the pile of sand. I had totally forgotten about that until you brought it up. Yeah. From a maths perspective, critical slowing would be another way of inducing, and I imagine that your complexity measures were basically measuring scale-free or long-term correlations. That's right. The system has, if you like, a much more enduring memory, so its time constants are now distributed far into the future and far into the past. And I think that's really interesting from the point of view of the inference mechanics, in the sense that what you could argue I would argue that when the skill balance, challenge-skill balance, is apt, there is a future, deep into the future, that I can imagine myself doing this for quite a long time. I don't know exactly what I'm going to be doing, but I can imagine myself committing to this and pursuing this course of action into the long-term future, which, from the point of view of the neuronal interpretation of neural dynamics, would require a certain kind of critical slowing that would look, you know, that would have the hallmarks of self-organized criticality and implicitly metastability. Whereas, if the challenge is, if I'm bored, then, you know, I don't know what I'm going to be doing in the future, because I may not engage with this. And if the challenge is too high, I'm just dealing with basically white noise, then again, there is no clear future, so that my time constants in terms of my generative model— Yes. —or the consequences of my action really shrinks down. Whereas if I've now got, if I'm in the flow, then literally a dynamical flow takes me to another regime of my attracting set or attracting manifold. In a reliable way into the future, which would then require a critical slowing and effectively evince a self-organized criticality. Is that what you were hinting at with the complexity?

I think, and going back to our earlier conversation about rebus and psychedelics, this is the sort of the argument here is that we know from a lot of research, psychological research and phenomenological research that flow correlates with well-being, lower levels of depression, lower levels of anxiety, higher quality of life, life satisfaction in general. And more flow in your life correlates with higher levels of happiness and those things. And so if psychedelics, you know, help with that, the mechanisms must be similar. Something similar. Even though flow is you're performing a task, you're actually doing something in the world, as you said, versus letting the world sort of happen to you in the psychedelic state. And so I think some of the mechanisms of free energy principle and metastability also explain the reason why flow is correlated with higher levels of mental well-being and can alleviate mental health issues, I should say, as well.

Yeah. Can I ask a related question?

And it's psychedelics or flow because they have similar impacts. And I know we just touched on this a little bit in our purpose paper, but one of the questions, if both, if these experiences are starting with sort of very relaxed priors, one of the things that happens on the back end of both is our beliefs get updated. Of course, we have a really strong attachment to those updates, right? For example, flow, we tend to remember flow experiences. They lock into memory, right? You know, one of the chances are if you have really strong, good, positive memories of the past, if they're not pegged to a time, a date— I graduated, I got married, that sort of thing— you're often remembering flow states. So we're going with very, very relaxed priors, but is there a mechanism on the back end of these that locks them in and tightens them down? They become more rigorous on the back end? Rigorous is the wrong word here. Like, they become really locked posteriors? Is that the right way of saying this?

Yes. And you should use the precise word. Yes. You are now rendering, in this context, your belief structure and your generative model much more precise. The precision here is is not so much about the flux of states in inference, but your posteriors over the parameters of the generative model, the coupling constants, the parameters of the equations of motion of the dynamics, as it were. And certainly, that would just be learning. It would just be associative learning, but in a particular context. I guess, I'm just wondering, and perhaps I should turn this into a question, you know, it strikes me that to actually engage and to be sufficiently skilled to be in a flow state equips you with another volitional way to change the precision of either at the sensory level to what you're attending to, or 'Um, the kind of person I am,' or you can take your mind off it and relax in a rebirth sense. It doesn't matter what kind of person I am, I'm playing this video game, or I'm playing this musical instrument, or I'm climbing this mountain. Um, so, but crucially, you've got control over it, right? Unlike many, um, aberrant precision control mechanisms that underpin psychiatry, if you can get control over it, you've basically, you know, you've got a therapeutic handle built in to the way that you lead your life, which I think is very empowering.

To your point, there's some research, Carl, that shows higher levels of cognitive control are associated with lower levels of PTSD. And so there's an inverse relationship there. And people with PTSD who start to do techniques to elevate their level of cognitive control, it helps. And it helps them focus more. So there's definitely some, some mechanism at play there. So Carl, can I just— Stephen, can I finish with one more question?

Or do you want to? Yeah, you get the last question. Over to you, Michael.

All right. All right. We got to bring AI into the conversation. And so, Carl, what do you— you've been— you talked about Geoffrey Hinton earlier in the conversation. He's been called also the godfather of neural networks and modern AI, especially deep learning and generative AI. Very specific question. Do you think the free energy principle or theoretical frameworks like metastability or the free energy principle are really needed to advance and move toward the future of AI and to advance AI?

Um, I, I think they do. That's not to decry or distract from the remarkable engineering achievements of generative AI and large language models in the past, uh, in the past few years, but I think the next move people fondly refer to as AGI, or I would prefer to think of it as a move towards natural intelligence as opposed to generalized intelligence. I think that you will need to recourse to principles, and the free energy principle for me is the obvious one. It's just a path of least action, and I use the phrase path of least action because the path is the dynamics literally it is the flow, not quite in the sense that you mean it, but it is the flow that underwrites our very existence. So one's got to apply the right kinds of principles, and I can see that now becoming a thing at many levels. So for example, if you take the ARC Challenge that François Chalet has So that's now moving exactly in the direction of reasoning, planning, and intuitions. It is basically a challenge that forces the machine learning or the artificial intelligence community to think about the principles and the schemes that can have intuitions and act upon them. And of course, this is exactly the provenance of active inference when it comes to thinking about the future and me as an agent in that future, and having exactly the kind of intuitions and hypotheses that we were talking about before, So I think that, you know, the physics of intuition and the physics of purpose and the physics of intention, read from my point of view as active inference, is going to become increasingly important as we move, you know, as we take the next steps from generative AI without generative models, or more particularly generative AI without agency, I think first of all, we have to put agency in. You can't have intuitions or intentions or purpose if you're not an agent. So you have to have a generative model of the future. And of course, you know, our past conversation about self-organized criticality and how far into the future you can get means that your agency is going to, if you like, it is the definitive aspect aspect of artificial intelligence that I think will take you to AGI in the next step. So I think the things that we've been talking about today are now the aspirations of tomorrow's artificial intelligence. Excellent. Can I ask a weird question?

Whenever I think about agency, I always think about, and I think we wrote about this in one of our papers somewhere, I always think about Stuart Kauffman. Complexity biology and his notion that for agency, you really just need 3 things. You need a system that remembers what happened. You need a way to make this— you need a decision-making apparatus, right? And then you need to sample the environment, right? To get information. And if you look at it from his lens, don't like— even if you're looking at something like Facebook, or even Google, isn't there agency baked in there already? Am I making a thought error somewhere along the way? I've been asking that question a lot today.

I think that's a deep question at many levels. So I would ask you, is a thermostat an agent? It has goals, it has, the capacity to sample the environment. Sample the environment. Yeah, and, you know, it wants to keep the room at that particular temperature. So, I don't have an answer, and I've indeed wrestled with this, trying to define agency mathematically. What I tend to do, though, is to say that authentic agency of the kind that we were implying, that we have comes equipped with intentions. So I don't think a thermostat has intentions. Intentions are read as future states. So it all comes back to the future, and it all comes back to the long-term memory in the metastability or self-organized criticality. A thermostat, you could say, intends to keep the temperature of the room within a particular range. But by intentions, what I mean is I intend to be in this state in the future. And now it's the path to this state. And of course, that path could be a flow state. Right.

But this also means this is like, you know, cells have intentions. Thermostats don't, right? Because the cell can swim up a chemical gradient to food. So it starts to get really— the line gets thinner and thinner and thinner, right? As these are such great questions because they really just warp your brain a little bit when you think about them.

Well, we could come back to the pile of sand in the sense that if you're a philosopher, then to have vague— perhaps agency is vague in the sense that how many grains of sand constitute pile. Of course, there is no definitive— so, you know, is a thermostat an agent? Well, a very simple kind. What about the chemotactic cell? You know, does it dream about tomorrow? Probably not, but it's a little bit more intentional in its behavior. Whereas, you know, you and I, or our children worrying about whether to get married or which university to go to, they look deep into the future. So perhaps agency is a vague concept.

Well, that's it. But I like it as a vague concept. It's clearly a spectrum, right? Like most things, it's clearly a spectrum. But you did something really interesting there, because it's a spectrum defined, at least partially, one variable is how far can that spectrum predict into the future, see into the future, act into the future, right? That's interesting.

And it fits nicely with this notion of self-organized criticality that accompanies metastability, because mathematically it's critical slowing. You know, it's the accumulation of long-term correlations over time, so it's all about the future. So I think that, you know, the depth into the future that your world models can take you, or the depth into the future that your state of mind, for example, being in a flow state, can take you is a really key, if you like, dimensional access. In another dimension, electric sheep.

Nice Philip K. Dick reference.

Well done. All right. Thank you. Thank you. I see you, Carl. Thank you so much for this.

Thank you. This was lovely. Deep conversation, very helpful.

And our research with you has been incredible, and onwards and upwards with our collaboration, hopefully.

Really fun. And we're gonna have to have you back because there's about a million and a half other questions that we didn't have time to talk about. We're gonna do this again soon, I hope.

No, I really enjoyed that. And I've learned a little bit about bit more about flow and intuition. So thank you very much indeed.

This is welcome.