#302 Karl Friston: How the Free Energy Principle Could Rewrite AI

This episode is sponsored by AGNTCY. Unlock agents at scale with an open Internet of Agents. Visit https://agntcy.org/ and add your support. How could Karl Friston's Free Energy Principle become a blueprint for the future of AI? In this episode of Eye on AI, host Craig Smith sits down with Karl Friston, the neuroscientist behind the Free Energy Principle and advisor to Verses AI, to explore how active inference and brain inspired generative models might move us beyond transformer based systems. They unpack how Axiom, Verses' new architecture, uses probabilistic beliefs and message passing to build agents that learn like brains instead of just predicting the next token. We look at why transformers face scaling and reliability limits, how Free Energy unifies prediction, perception, and action, and what it means for an AI system to carry explicit uncertainty instead of overconfident guesses. Learn how active inference supports continual learning without catastrophic forgetting, how structure learning lets models grow and prune themselves, and why embodiment and interaction with the real world are essential for grounding language and meaning. You will also hear how Axiom can sit beside or beneath large language models, how explicit uncertainty can reduce hallucinations in high stakes workflows, and where these ideas are already being tested in areas like logistics, robotics, and autonomous agents. By the end of the episode, you will have a clearer picture of how Karl Friston's Free Energy blueprint could reshape AI architectures, from enterprise planning systems to embodied agents that understand and act in the world. Stay Updated: Craig Smith on X: https://x.com/craigss Eye on A. I. on X: https://x.com/EyeOn_AI

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  We change our mind hundred milliseconds by hundred milliseconds as we engage with the world and we continually making sense, doing our sense-making with the world. And that's a process of inference. It's not necessarily learning, which will be a much slower process. And if the way that our brains work can be cast as inference, that means if our brains don't work properly, as in mental disorders, then we can understand that as false inference. The brain is in the game of basically trying to predict what it's sensing. So if there's a prediction error, that's newsworthy. So it can now use the prediction error to update, to revise its explanation, its representations of the latent causes. So in the moment, I literally change my mind.

• 0:43Speaker 00

  I literally change my mind on the basis of the prediction errors that have been elicited by comparing what I predict and what I actually saw.

• 0:53Craig S. Smith?

  Hi. The transformer algorithm created by a team at Google in 2017 has been the core of generative AI. But scaling transformers has hit practical theoretical limits. New releases are providing only incremental improvement. And as a result, many researchers are looking into new architectures aiming for better memory use, better generalization, and longer context reasoning. Among those researchers is the team at Versus, a startup implementing the ideas of Carl Friston, who is the most highly cited neuroscientist globally

• 1:36Craig S. Smith?

  and one of the most cited living scientists overall. Friston is the creator of the free energy principle, sweeping attempt to explain how all living systems, including brains, maintain order in a chaotic world. His work has inspired entire branches of computational neuroscience and new directions in artificial intelligence. Versus where he serves as chief scientist was started by some of his students. Its core project is called Axiom, a groundbreaking AI architecture built on the principles of Friston's work.

• 2:18Craig S. Smith?

  Taking inspiration from how the human brain predicts and adapts to its environment, not by memorizing data like today's models, but by constantly minimizing what Friston called surprise, which is a measure of the difference between prediction and reality. It's an ambitious attempt to create machines that learn and reason the way living systems do, grounding intelligence in physics rather than statistics. In fact, this is my second conversation with Friston because I was completely lost the first time around, and he says himself that he has trouble speaking in the vernacular.

• 3:02Craig S. Smith?

  So while some listeners will be able to understand what he's saying, most are better off letting the jargon wash over them, focusing instead on the higher concepts and how they're implemented in an AI model. You may not understand everything, but you'll come away with a picture of how and why post-transformer architectures are being developed. I hope you find the conversation as fascinating as I did. But first, I want to give a shout out to our sponsor. Build the future of multi-agent software with Agency. That's A-G-N-T-C-Y. Now an open source Linux Foundation project,

• 3:46Craig S. Smith?

  Agency is building the Internet of Agents, a collaborative layer where AI agents can discover, connect, and work across any framework. All the pieces engineers need to deploy multi-agent systems now belong to everyone who builds on Agency, including robust identity and access management that ensures every agent is authenticated and trusted before interacting.

• 4:18Craig S. Smith?

  Agency also provides open standardized tools for agent discovery, seamless protocols for agent to agent communication, and modular components for scalable workflows. Collaborate with developers from Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and more than 75 other supporting companies to build next-generation AI infrastructure together. Agency is dropping code, specs and services, no strings attached. Visit agntcy.org to contribute.

• 5:01Craig S. Smith?

  That's A-G-N-T-C-Y dot O-R-G.

• 5:06Speaker 00

  My name is Carl Friston. I'm a professor of neuroscience at University College London and Chief Scientific Officer of Versus, a cognitive computing company in Canada and America.

• 5:17Craig S. Smith?

  Versus is using your free energy principle in its models. And to talk about free energy, I thought we would start with physics, move on to neuroscience, and then discuss how it's implemented in or the free energy principle is implemented in in machine learning. My understanding is that free energy is a measure of the usable energy in a system. Can you can you talk about that? It's a very confusing concept.

• 5:57Craig S. Smith?

  It gets more confusing when it moves, when you take that principle to neuroscience. But can you talk, first of all, in layman's terms, what free energy is in physics?

• 6:09Speaker 00

  Yes, I can. Or perhaps I should qualify that. I'm not sure I can do that in layman's terms. I have a reputation for nothing. So you're going to have to hold my hand in in trying to make this accessible. So the free energy we're talking about here is closely related to the thermodynamic free energy that you're talking about, which is the amount of available work in a system that is not, if you like,

• 6:37Speaker 00

  Entailed or trapped or locked in because of the entropic or the disorder of the system. However, the particular mathematical free energy, the variation free energy we're talking about is a purely information theoretic construct. It actually inherits from Richard Feynman's work on quantum electrodynamics. And he was trying to solve a problem, which in fact, Versus is also trying to solve in one sense, which is finding the most efficient paths, the most efficient courses. In his context, these were the sort of paths of small particles, for example. So trying to find the, a probabilistic description of the paths of least action.

• 7:22Speaker 00

  And he solved that in almost intractable problem by turning an impossible marginalization or integration problem into an optimization problem. And what he did to do that was introduce this notion of variational free energy. And then that was taken up decades later by people like Geoffrey Hinton and David McKay and the like. And they were, they noticed that this information theoretic quantities, mathematical information theory quantity that has exactly the same components as a thermodynamic free energy, as an expected energy that's supplied by some world model or some probabilistic specification of

• 8:05Speaker 00

  the thing that you are trying to model minus the entropy was exactly the same that you would need to optimize things from the point of view of

• 8:16Speaker 00

  cryptic analysis and with a nod to the Russian notion of algorithmic complexity and that led through to universal computation.

• 8:29Speaker 00

  In terms of cryptanalysis. And also was exactly what you needed if you wanted to make something like a Helm-Holtz machine work. So if we go right back to sort of the 1990s, you'll find sort of prototypical variational free energies at the heart of the things like the Helm-Holtz machine, like the sleep-wake algorithm. And indeed, you can track that right through to things like variational auto-encoders. So this is like a universal objective function, and it has two bits to it. One bit is the entropy term, and the other bit is this constraint or expected energy,

• 9:14Speaker 00

  which is this basically something that enforces an accurate account. If you're trying to model something, say you're taking a sort of a neural network that was trying to predict the next word, or trying to compress information in an image to generate, to predict the image and thereby generate new kinds of images. Then you're going to want to do that as accurately as possible. But by including the entropy term into the free energy, then you're effectively applying a kind of Occam's razor. You're sort of not committing to a particular explanation, but you're including, or you're

• 9:58Speaker 00

  trying to not overfit the data because you're trying to maximize the dispersion or the uncertainty about your particular explanation for this content or for the, for these data. I slipped that in. Again, this is not vernacular, and I apologize, but there's a beautiful link between minimizing free energy, which is just this expected energy minus entropy, which means that you're implicitly, under constraints, maximizing the entropy, which of course is Jane's maximatory principle, which is a cornerstone of measurement physics. And, you know, it is that sort of, if you like, built into your cost function.

• 10:45Speaker 00

  It is that sort of uncertainty aspect that means that you preclude overfitting and elude many of the problems that we see in current, say, deep neural networks optimized using reinforcement learning. So this free energy principle is a first principle account of how to make neural networks or neural networks used as agents or as classifiers or inference machines as efficient as possible, going right back to the path integral formulation of Richard Feynman in terms of finding the paths of least action, the paths of maximum

• 11:26Speaker 00

  efficiency, the paths of least effort that include, if you like, this cost of overfitting, technically a sort of a complexity cost. You know, so I don't have to change my mind too much in order to account for these data, which means that I don't want to commit to a particular, I don't want to commit to a particular explanation for these data. Does that make sense?

• 11:51Craig S. Smith?

  It does, only because I've done a lot of reading, I'm thinking for listeners, one of the confusions, confusing things for me when we first spoke was the word energy when you apply it to machine learning. What you're, in machine learning, what you, energy is, is the, now, tell me if I'm wrong, but it's the, it's the difference between an input and the prediction is it's, it's, it's in effect an error function. Is that right?

• 12:29Speaker 00

  Yeah, no, absolutely. And indeed, you can sort of look at this first principle approach to optimization in terms of, you know, error correction or error minimization. Indeed, you know, in some, in some parts of the life sciences, you get things like predictive coding, which is basically sort of scores the accuracy of your machine.

• 12:54Speaker 00

  In terms of the prediction error and that's an important part of this free energy and indeed you could actually read free energy in the context in the in the setting of predictive coding as a precision weighted prediction error, just the mismatch. So that's absolutely right. So just just to address that, sort of unpack the notion of energy in this

• 13:18Speaker 00

  information theory context, and energy is just a potential and a potential is just the sort of negative log probability. So it's just a measure of the implausibility of something. So if something is very implausible or very surprising or very unlikely or very uncharacteristic, then that has a very high potential, it has a high energy and you really want to minimize that, you know, that energy in order to produce an unsurprising account or render this content unsurprising, given what you might have predicted if you knew the causes of this particular content.

• 13:56Craig S. Smith?

  Yeah. And and in neuroscience, I mean, you're a neuroscientist. You saw this principle as

• 14:06Craig S. Smith?

  as, you know, sensory input coming into the brain and and the brain trying to fit that in in its model of the world. And if it's surprising, there's a lot of energy there,

• 14:23Craig S. Smith?

  free energy and and the the brain then has to update its model of the world to to match

• 14:33Craig S. Smith?

  the sensory input that's coming in. Is that right?

• 14:35Speaker 00

  No, that's absolutely right. That's a perfect description of predictive coding formulations of this kind of free energy minimization, that the brain is in the game of basically trying to predict what it's sensing.

• 14:50Speaker 00

  Yeah, the sensory content. And if it's got the perfect prediction, clearly the free energy will be zero and there will be no prediction errors. So if there's a prediction error, that's newsworthy. So it can now use the prediction error exactly to say to update, to revise its explanation, its representations of the latent causes of these data. So one can describe that in terms, if you're a statistician, that would be called Bayesian belief updating. You were, if you're a philosopher, it would be sort of resolving prediction errors. It does really emphasize that sort of key point about, you know, this biomimetic principle of how we make sense of things and indeed act upon the world.

• 15:34Speaker 00

  We do so in a very sort of inside out ways, you know, we generate predictions and high based upon hypotheses about what could have caused this content, what could have caused this sensory input. And just use the mismatch, the ensuing free energy, if like the implausibility in relation to your predictions of your actual sensory content to update online and over time. So now sort of introducing a distinction between inference. So in the moment, I literally change my mind. I literally changed my mind on the basis of the prediction errors that have been elicited by comparing what I predict and what I actually saw.

• 16:17Speaker 00

  But also over time, in a much longer time scale, I can learn in the spirit of machine learning to be a better predictor. Given this, this, this kind of world.

• 16:27Craig S. Smith?

  Couple of things about the neuroscientific view that I've found fascinating is your view is that this is an instantaneous continual loop, that the brain is, is constantly doing this. And that some mental illnesses may be caused by an error or an inability of the brain to update correctly or to minimize the free energy principle.

• 16:56Speaker 00

  Absolutely. And in part, you know, that it was exactly that observation that motivated much of this work in the context of neuroscience and in particular computational psychiatry. And, you know, you could even go further and say that all mental illness. Could be construed as some kind of false inference. So you you you made the point that this process is ongoing, it's continuous. So, you know, it is something that we change our mind millisecond by millisecond, or at least 100 milliseconds by 100 milliseconds as we engage with the world and we continually making sense, doing our sense-making of the world. And that's a process of inference.

• 17:40Speaker 00

  It's not necessarily learning, which will be a much slower process. And that if if the way that our brains work can be cast as inference, then that means if our brains don't work properly, as in mental disorders, then we can understand that as false inference. And that makes a lot of sense in this. You know, and by false inference, I mean exactly what you would have done at school when, I don't know if you remember doing type one and type two errors when doing T-tests. You know, you can make a type one error and infer something is there when it's not. Of course, that's a very apt description of an hallucination or a delusion. But you can also make type two errors, type two false inferences, infer something is not there when it is.

• 18:25Speaker 00

  And of course, that's a precise explanation for many neglect syndromes in psychiatry, dissociative syndromes, what used to be called hysterical syndromes. So nearly everything in in mental health and mental disorders can be construed as inferring properly or some failure of inference. And then getting into the weeds of the actual mechanisms of this predictive coding is really quite informative and sort of pointing to the particular failures of message passing entailed by this view of the brain. So I'm generating predictions in an inside sort of way and then using the outside in prediction errors to correct and update and perform this kind of inference.

• 19:10Craig S. Smith?

  And another thing that fascinated me from what I read, that this cycle or this process matches the anatomy of the brain. I mean, that that there are layers, like in the visual cortex, and information is is passed both back and forth, up and down, which is one of the reasons why back propagation has never been. Validated as a process that happens in the brain because there isn't that kind of pathway for updating weights going down. Am I wrong on that?

• 19:49Speaker 00

  No, no, I think you've hit a really central argument and focus of both the machine learning community and people invested in predictive coding formulations. So just to sort of highlight the importance of that question, it is certainly the case that the functional anatomy, the computational anatomy of the brain has a hierarchical structure. And indeed, you could argue this is where deep learning comes from. It just means that the world model or the generative model that is entailed by the structure of our brains has a hierarchical depth. And of course, then that naturally translates into

• 20:32Speaker 00

  deep neural networks and deep RL, for example. However, there's a fundamental distinction between backpropagation on a deep neural network and the kind of optimization and message passing between the hierarchical levels in the brain. And exactly as you say, it is all local. So as you move from one hierarchical or say layer in the neuronal brain architecture, you are in exactly the way you described, sending messages to the layer below, which are the predictions, and then receiving messages from the layer below, which are the prediction errors that revise the representations equipped with

• 21:12Speaker 00

  uncertainty that then generate the predictions. So there's lots of local message passing. Crucially, the objective function that describes the dynamics implied by that message passing is local. So now you get into a message passing scheme, which is much more sympathetic. Well, the message passing scheme under predictive coding has this locality and it now acquires a biological causability, but also a massive increase in efficiency, which you don't get with backpropagation in terms of sending all the way to the top and then all the way back again, which is a very non-local sort of violation, not just of the principles

• 21:55Speaker 00

  of biological brain function, but also the principles of least action that we started with. Because if you can do it locally, you're doing it very, very efficiently. You know, interestingly, in the past five years, I would imagine, people in machine learning have been looking towards predictive coding as an alternative to backpropagation of errors. And certainly establishing that in terms of the efficiency, it is at least as good as, if not more efficient than backpropagation. There are issues about scaling it up to the enormous dimensionality of, you know, deep RL. But from a mathematical perspective, this is the right way to do it because it is just much more efficient.

• 22:35Speaker 00

  It is just much, and it just relies upon local, you know, local optimization. Well, beliefs in this context are used in a sort of technical sense of Bayesian beliefs. So just to try and, you know, not make that vernacular, but certainly qualify the notion of belief updating and the sense of Bayesian belief updating. So beliefs are read in my world, or in the world of active inference and the free energy principle, as conditional probability distributions. So all their sufficient statistics. So they're not sort of propositional beliefs, not something you can talk about.

• 23:20Speaker 00

  These are just sort of mathematical descriptions of a representation that is equipped with uncertainty. So there's content and the confidence that you have about this. So that's another bright line between the kind of networks that we have in our brain and indeed that are employed in active inference, for example, versus Axiom, where the nodes actually represent beliefs in terms of both the content and the uncertainty or the inverse uncertainty, which would be, which would be the precision as opposed to a neural network, which just has a value, it just has a cascade of functions that you couldn't interpret

• 23:58Speaker 00

  as representing content, but it is not equipped, not equipped with, with uncertainty. So once you've got a, one of these, I repeat, binometric kind of neuronal networks in play, of the kind that you Might find in predictive coding. Then you can talk about the changes in the values of the neural network in terms of belief updating. And of course, beliefs about what? Well, beliefs about the causes of your content, beliefs about the causes of what is observable. And so now there's a there's a distinction between the latent causes, the unobservable causes

• 24:41Speaker 00

  of observable data, sensations, images, you know, whatever is can can actually be explicitly measured or observed. So when we talk about belief updating, what we're talking about now is a Bayes optimal way of updating your beliefs about the causes, the observable causes of your sensations. Of course, we can only sense a very small part of the entire universe. So we have to, on the basis of, you know, very sparse sampling, build beliefs and update those beliefs continually about the states of affairs out there beyond the brain bound skull, or indeed beyond my deep neural, uh, deep neural neural network.

• 25:26Speaker 00

  So inference just is a description of the process of Bayesian belief updating. And Bayesian belief updating is just mathematically something that can be described as minimizing your uncertainty, minimizing your prediction error, minimizing your free energy in the right kind of way that includes this measure of uncertainty.

• 25:50Craig S. Smith?

  Another, when you move into machine learning, um, this Bayesian active inference, it's happening continually. You're working with probability distributions, not weights in a, at at least in, I mean, you're working with versus, and at least in versus is. Implementation, you're working with the parameters in the, in the model are probability distributions, not weights of individual nodes or neurons, uh, the, the, the influence the outcome more or less. It is, am I off there?

• 26:33Speaker 00

  No, no, no, I don't know. You're spot on. Yeah. So just to sort of contrast, um, the, you know, so a, um, a deep RL or a standard, um, uh, neural network, you know, including things like transform architectures that, um, underwrite large language models or most generative AI. These are basically, uh, just, um, universal function approximators. They just map through a series, a cascade, um, um, of, um, functions, values, um, afforded by the input, uh, to generate some transformed values, which are the output. Now, there's,

• 27:13Speaker 00

  there's could be predictions of the next token. There could be some, uh, action. Um, that is not how the, the brain works. It is not how, um, um, active influence works. And it's not, as instantiate say in axiom, um, um, which was a sort of used to demonstrate the, in the, you know, the, the fundamental increases in efficiency, if you do it using belief updating. So the equivalent architecture is certainly in place. There is a deep structure to it. And I, you know, I mentioned sort of world models, generative models before, the structure of the network, you know, determines the structure of the, your model and the implicit conditional dependencies and all the contingencies in there. But as

• 27:57Speaker 00

  you say, each node now stands in for a probability distribution, a belief. And therefore what you're passing around, what your, um, the influence of one node on another node is, um, determined by, um, driven by the, um, the optimization of your probabilistic beliefs or representation. And that's really important because, um, for example,

• 28:26Speaker 00

  If I don't have a representation of uncertainty, then I don't know what I don't know. And more importantly, I don't have any way of judging or evaluating the goodness of an action in terms of its ability to reduce my uncertainty, to reduce my prediction errors, to reduce my expected free energy. But if I now got a representation at each and every level of a deep generative model, a deep world model, I can work out, if I went and solicited or looked over there.

• 29:08Speaker 00

  So for example, you and I are extremely skilled at this. Every 250 milliseconds, every quarter of a second, we deploy our visual apparatus by looking over here, phobiotic, phobiating, phobiating over here. That's an incredibly skillful act, simply because we are choosing to look at the parts of the visual scene that will have the greatest information gain or resolve the greatest amount of uncertainty, given what we believe at the moment in this sub-personal Bayesian sense. So what we're talking about now is a mechanics of optimal smart data mining,

• 29:50Speaker 00

  knowing where to go and get the right kind of information that enables you to resolve your uncertainty so that you can now have a better model of the world in this probabilistic sense. And you know, Axiom sort of showcased that in the sense that it spoke to, you know, what could be construed as sort of two of the main problems with, you know, with the current tech on offer with generative AI, which is basically the inefficiency. And the lack of reliability. So the inefficiency can be resolved by appeal to these first principle accounts. And you know, as exemplified by things like predictive coding or

• 30:35Speaker 00

  other implementations of of active inference. Because now you've got this ability to resolve uncertainty in the the most efficient way in accord with these principles of least action. You can now not only outperform Deep RL on benchmarks as, you know, I think there's a 6% improvement in performance demonstrated by Axiom. But more importantly, you do it much more efficiently. So I think there's a, you know, using 3% of the compute. And that translates into efficiency in many, many different ways. So not only is it informationally more efficient or statistically more efficient,

• 31:18Speaker 00

  it also is thermodynamically more efficient. So you're losing, using less power. And it's also sample efficient. So you need a fraction of the data. So, you know, this is, if you like, by appealing to these biomimetic principles of, you know, from the physicist's point of view, principles of least action. The sort of underwrite the way that we deal with our world and make sense of our world. You you can resolve the efficiency problems that, you know, that are currently plaguing or preventing the right kind of enterprise update of of generative AI, for example, where you'd expect it to be deployed. It's not being deployed.

• 32:01Speaker 00

  And of course, reliability is, you know, of, you can't get hallucinations if if you if you've got an explicit uncertainty quantification.

• 32:12Speaker 00

  If you're not sure, if you're not sure, you're uncertain about a particular recommendation or a particular classification, then because you quantified it, then you inure yourself or you protect yourself against making overconfident, overreliant, overconfident predictions or classifications and the like.

• 32:36Craig S. Smith?

  Build the future of multi-agent software with AGENCY. That's A-G-N-T-C-Y. Now an open source Linux Foundation project, AGENCY is building the Internet of Agents, a collaborative layer where AI agents can discover, connect, and work across any framework. All the pieces engineers need to deploy multi-agent systems now belong to everyone who builds on AGENCY, including robust identity and access management that ensures every agent is authenticated and trusted before interacting.

• 33:20Craig S. Smith?

  AGENCY also provides open standardized tools for agent discovery, seamless protocols for agent-to-agent communication, and modular components for scalable workflows. Collaborate with developers from Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and more than 75 other supporting companies to build next-generation AI infrastructure together. AGENCY is dropping code, specs and services, no strings attached.

• 33:58Craig S. Smith?

  Visit agntcy.org to contribute. That's A-G-N-T-C-Y dot O R G. I should have said at the beginning, Axiom, that you've mentioned a couple of times, is versus is the startup that you're working with, and Axiom is the model implementing the free energy principle and Bayesian active inference and all of that. Is that right?

• 34:28Speaker 00

  Yeah, yes. It's a sort of...

• 34:32Speaker 00

  A minimal implementation just to demonstrate what you could achieve and the problems you can solve if you commit to these biomimetic principles. And there are many aspects that we could talk about. This one of them, interesting, which people sort of latch on to. And indeed speaks to one of your other questions about the architecture of the brain. I mean, the brain is not just a deep neural network. It has a lot of factorial structure. It has a lot of modularity. Some parts do sort of perception, other parts do planning, other parts do memory.

• 35:16Speaker 00

  So all of these different aspects of inference and sense making and decision making all imply and require a certain kind of structure. So your Axiom was built to show that if you put this kind of structure, a genetic kind of structure into a world model, into a generative model, and then you apply these variational principles of least action to the local updating, the local optimization, then you can get this kind of performance enhancement. And to my mind, part of the more important aspect is this efficiency. You could have guessed that from the meme that you and I can drive a car on 20 watts.

• 36:00Speaker 00

  But if you have to train a large language model, you may need a nuclear power station. So this is the...

• 36:10Speaker 00

  To my mind, this is the important part of the solutions to the current problems that are inherent with the current direction of travel with generative AI.

• 36:23Craig S. Smith?

  Yeah, and a couple of things that fascinated me is this is a dynamic model. It expands and contracts depending on the data that it's looking at, is that right? So and it learns continuously because the probability distributions are being refined, but they're not being overwritten. And that's a problem with neural networks, that new data coming in can... That's why they can't achieve continuous learning.

• 37:01Craig S. Smith?

  Because of catastrophic forgetting. And can you talk about that, the promise of continual learning in this kind of a system?

• 37:10Speaker 00

  Yes, I can. You've already summarized the bullet points. I'll just speak to those. No, absolutely. So, you know, we do not suffer from catastrophic learning. We will grow our models, our world models, in a way that is sufficient to account and optimally explain what's going on. And in growing them, we're now talking about something you implicitly implied, that the Jotin model in and of itself has an optimal size. For example, it could have, if you were just using a simple layered deep neural neural neural network, how many

• 37:54Speaker 00

  layers do I have? And in each layer, how many hidden units? And in my world, how would you factorize or modularize within each layer? So all these architectural aspects determine the structure of the model, which just is a specification of the cause effect structure statistically in the world you're trying to model. So if you remember where we started with the free energy and the importance of that, putting entropy into the objective function to implement sort of Occam's razor. What that means is there is an optimum complexity of the structure for any given set of content or world in which you are navigating, predicting,

• 38:34Speaker 00

  classifying, acting in. So you need to find the right sort of complexity of the structure, the right depth of the model, the right, you know, the right sort of factorial structure that is sufficiently complex to provide an accurate account, but not too complex. So this is exactly Einstein's, you know, keep everything as simple as possible, but no simpler. So there's a sweet spot. And in principle, and indeed in practice, one, indeed, axiom actually demonstrated this under the hood, though it was never, it wasn't sold as one of the key issues, but to my mind, it was a really interesting aspect, that

• 39:15Speaker 00

  it grew itself. And just to the right level of complexity before, if you like, opening up new slots or new parts of the model for objects that it had never, you know, it had never seen. So you've got this notion now that you can increase or decrease the structural complexity of the model in exactly the right way. So you don't overfit. And this contrasts with deep RL. Because in deep RL, you just start with something that's too complicated. You start with billions of parameters and then fondly hope you can reduce the number of parameters, say, connection strengths or weights by some kind of, say, pruning

• 39:57Speaker 00

  or dropout or putting noise in and doing mini batching, you know, in a way that eludes those sharp minima that you get from the overfitting, simply because you've got too many parameters, too many connection weights in your model. So conventional approaches use this sort of top-down approach. They start off with an overly expressive model with too many parameters, ninety-nine point nine percent of which will be redundant, and then try and find some engineering heuristics to eliminate the redundant, to get, to reduce that complexity in a way that you get for free if you'd use the free energy. But to use the free energy, you have to encode the uncertainty. And of course, deep RL doesn't

• 40:42Speaker 00

  have that. It doesn't have the capacity to represent the uncertainty unless you're dealing with a variational autoencoder. So this is another motivation for using the free energy as your ultimate objective function, because it now gives you a way of scoring not just the quality of your current beliefs in the moment, or the parameters of your objective model, the connection strengths as you assume, as simulation accumulate data and build better and better models, but also it scores the quality of the structure of your model. And that speaks to the possibility of growing models. So growing models from

• 41:24Speaker 00

  scratch, just so you can get to the right level of complexity, apt for the kind of content that you're dealing with. And this is, you know, it is a fascinating and I think really important, possibly vexed. Problem at the moment, which in the life sciences would come under the rubic of structure learning, learning the right kind of structure automatically in the face of data by optimizing your model with respect to this free energy score that can also be, well, provides a band. In machine learning, I should say, this is the elbow, this is the evidence lower band. In fact, it's an upper band because it's a negative in physics.

• 42:10Speaker 00

  And therefore, you can understand this sort of model growing and shrinking and adapting to new worlds and new contingencies

• 42:22Speaker 00

  as maximizing model evidence. Or if you treat your model as an agent, then the agent looks as if it is gathering evidence for its own model. And philosophers sometimes call this self-evidencing, a little twist on the philosophical, simply because you're trying to optimize the model evidence that is bound by the elbow or the variation, uh, variation of free energy. So this notion of gathering evidence can be played right through not just to beliefs in the moment, beliefs about states of affairs in the world or hidden states, but also the parameters of your model that would be optimized through learning, but also the very structure.

• 43:07Speaker 00

  I mean, statistics, we call that Bayesian model selection. So it's selecting the model, the structure that has the highest evidence in the face of the content or the data which is trying to explain. And if you pursue that notion, what you've now got is a mathematical image of natural selection. So you can now look at natural selection. As nature's way of doing Bayesian model selection, where you and I are the hypotheses and the structures that provide evidence that this is the kind of thing that can live in this in this eco niche. So again, we come back to sort of biological principles.

• 43:48Speaker 00

  You know, you can, you know, nature's already done that. You know, data's already done its Bayesian model selection, free energy minimization to create you and me as existence proofs of generalized intelligence.

• 44:01Craig S. Smith?

  Right now, in its implementation in Axiom, it's not a language model. I mean, it doesn't, the data that's coming in is primarily frame by frame video data, or it's

• 44:18Craig S. Smith?

  something like that sequential data. The inferences are decisions, and I understand it can be applied to robotics, for example. Is, could this also

• 44:32Craig S. Smith?

  work on language? And I guess two questions. If if you have this model, I mean, this is, you know, Yama Kuhn's theory that that if you have a model that has direct experience of the world, it'll eventually, and if it can learn continuously, it'll eventually catch up with human knowledge. But so much of human knowledge is embedded in in text. That's what makes language models so intelligent,

• 45:12Craig S. Smith?

  is that they, they, as inefficient as it is, they have that knowledge base to work from. So it is, can can this be applied to language? And if if not, or or even if it can, if you ran a model like this theoretically, long enough, would it eventually learn all of the the the knowledge that that humans have accumulated today?

• 45:46Speaker 00

  That's a challenging question, and which I could take in many directions because it's such an inviting question. And First of all, it would only learn language if it was exposed to people speaking that language. So there's a certain aspect of active inference that rests upon a coupling between the agent and the world that it is trying to model and gather evidence for. So if that world includes other agents like itself, then there will be emergent language.

• 46:25Speaker 00

  And certainly exposing this agent to communication with other agents like itself, it would eventually learn a language, just like you and I do, and just like our children do. Would it do that just by being exposed to large cohorts of textual language of the kind that are used to train large language models? I suspect not. I agree entirely with Jan. I think he's on exactly the right track here, with one small exception.

• 47:00Speaker 00

  And I think he would argue, and I would certainly argue, as would most of my colleagues in the life sciences and philosophy, that to build the right kind of generative model where the words have meaning, you have to be embodied and situated. So that you have to experience the world as part of that world in order to endow meanings to the actual language. So we're moving beyond just the statistical structure of language and the predictability of the next word

• 47:41Speaker 00

  to now read the words as labels or some representation of some latent causes or states that are part of your generative model. And in that sense, if you exposed an active inference agent to the world as originally encountered by people like you and me, or things like you and me. Then yes, it might indeed learn language. But in no way different from the way that your children learn language or my children learn language, it would have to be exposed to the lived world so that the meaning is all in the interactions, not just with the world, but also with other things in that world that are like me.

• 48:25Speaker 00

  I keep emphasizing the like me because for language or communication in general to have any utility, there has to be a common ground, a shared narrative, a shared reference frame. So from the point of view of the free energy principle, active inference, or indeed artificial intelligence research, that basically means we're talking about artifacts or agents or artificial intelligences that, and that ones, that have a shared world model. And if you have a shared world model, then you can share your beliefs through some kind of communication. You can only do that with a shared reference frame.

• 49:04Speaker 00

  So, you know, I think this interactive aspect becomes incredibly important when understanding the utility of language. And you could turn your question on its head and ask, why are large language models so potent? They are beautiful things, you know, the most beautiful invention of this century.

• 49:26Speaker 00

  Why are they so alluring? And I think it is just because they generate content that resonates with our shared narrative and that we can understand and we can anthropomorphize them. So, you know, but is that enough to, is that enough to build a world model? I don't think it is. I'd be much more with Yan and a lot of Ari Peatels and other colleagues, you know, with machine learning, that you actually have to have something that's embodied and situated, embedded and dynamically coupled to and engaging in its environment.

• 50:09Craig S. Smith?

  Yeah, although the two could be combined at some point.

• 50:14Speaker 00

  That was the other direction I wanted to take your question. Yeah, absolutely. And so, you know, does that mean that there is no role for all this beautiful engineering that we have, you know, access to, especially over the past few years? And no, if you, if you can, that in this, you know, in the context of active inference and building the right kind of generative models, you can certainly install transformer architectures and deep neural networks. And the way that you would do that, and in my world, that's called deep active inference, is literally putting a a standard machine learning neural network

• 50:59Speaker 00

  into one of these, technically what they're called factor graphs, but

• 51:05Speaker 00

  say neuronal networks with message passing and belief updating. And the way that you can massively accelerate the efficiency of this kind of predictive coding or belief updating or active inference is if you can map directly from content to the beliefs about the causes of the content, then you, then you can very much accelerate the efficiency of these continual learning and continual inference machines. But you can only do that if it's learnable. So on the, you know, as it says in machine learning, on the tin, it has to be learnable, which means that it has to be context insensitive. So if you can find some mappings between content

• 51:51Speaker 00

  and probabilistic beliefs, very much like a, you know, a variational auto-accorder goes from some data at the bottom to some posterior, some variational density with expectations and variances on it, you know, at the top.

• 52:07Speaker 00

  If you can learn that, because it's learnable, context sensitive, then you you'll you you would actually have a hybrid sort of, or you'd be using conventional technology, machine learning, to augment active inference schemes. But notice that you can only do that when it when it's learnable, which means it has to be context insensitive, it has to be, you know, has to have a certain symmetry and be exactly the same mapping over time. So you can look at this as basically using machine learning to learn how to infer.

• 52:41Speaker 00

  And and leverage that by, I repeat, putting that inside these message passing schemes. The other way that would be really interesting is to map from the actual posterior basing beliefs of your machine, beliefs about not only states of affairs in the world, you know, what did you cause this particular image, but also about its intentions and actions. So if it has if it has agency, which simply means that the world model, the Jeritan model

• 53:20Speaker 00

  entails, or part of the Jeritan model includes the consequences of my actions in the future, which in turn implies that I can plan by choosing a particular course of action, which in turn means that I now got true intentionality. I have intended counterfactual states in the future that are part of my Jeritan model. You get a large language model to tell you not only what this agent thinks about the current states of affairs, but what it intends to do and why it's doing this as opposed to doing that. I mean, that would that would be wonderful if one could sort of, you know, put a large language model to interrogate and to render, you know, what is private to the active inference

• 54:04Speaker 00

  neural network now publicly available, because it can now broadcast to the user or to other artificial intelligences. Not only its beliefs about what's going on, but also what it intends to do and why it intends to do that.

• 54:21Craig S. Smith?

  Yeah. And which direction do you think, do you think the research will go both directions? Or, I mean, which direction are you working on?

• 54:34Speaker 00

  And I, well, versus at the moment is, is really sort of focused on

• 54:42Speaker 00

  establishing, socializing this approach and also

• 54:47Speaker 00

  serving sort of customer needs, you know, there are a few sort of lighthouse customers. And demonstrating the improvements or the solutions for the problems of inefficiency and unreliability in alternative current options. You know, both in the, in the context of complex system modeling. So, you know, one example with the work with analog and, you know, trying to optimize the, you know, the placement of cab drivers. You know, so imagine that you're waiting for a cab and there's some scheme that is deploying cab drivers to your location.

• 55:33Speaker 00

  And yet this is, you know, an incredibly difficult problem to optimize because it depends upon the traffic flow. It depends upon whether Taylor Swift is in town, depends upon where your drivers are. So this is a complex system that has a cause effect structure that can now be learned on the basis of historical data and ongoing learning. If you put one of these active inference machines, trying to explain what the cause effect structure of traffic flows and the deployment of your drivers. And we, you know, we applied this, this sort of first principle approach. And in simulation, we're able to get at least 30% more rides

• 56:14Speaker 00

  in simulations for, you know, for, for the, for the cab drivers. So that would be one example. The other direction of travel, I think, speaks to your, you know, what we were talking about in terms of

• 56:29Speaker 00

  imbuing AI with the authentic kind of agency that rests upon having intentions and the ability to plan. I think this is probably best demonstrated in Versus' work in robotics, for example, seeing what they're doing, looking at sort of benchmarks in the sort of habitat setup.

• 56:53Speaker 00

  Where, you know, currently

• 56:59Speaker 00

  robots can now understand the, and infer the scene around them in terms of, for example, very simple notions like there are objects. And have in mind as part of their gerative model the consequences of reaching out or moving objects to the extent that if you can build a sufficiently deep gerative model with a separation temple scales and planning, you can simulate and indeed realise practically, and indeed the robotics team at Versus R and D have done this, you can get robots to sort of clean up a room or go and get something from the fridge

• 57:44Speaker 00

  and do it with no, to do it instantaneously, because this is, this is inference, you don't learn to take something from the fridge. You infer, what do I need, what am I going to do if I'm the kind of thing that needs to retrieve a bottle of milk? Well, where is a bottle of milk? Well, it's likely to be in the fridge. How do I resolve my uncertainty that the bottle of milk is in the fridge? Well, I have to open the door. Well, how do I open the door? Well, if I, if I'm that kind of thing, so opening the door, I should expect to feel my, my actuators, you know, sending these IMU signals. And then you just get this predictive coding sort of technology just to make all that happen. So that, you know, that will be another direction of travel, probably more apt, less for work with people

• 58:29Speaker 00

  doing complex system, you know, dealing with complex systems. Such as, you know, traffic flows or our work with um um company investment, um, you know, people responsible for investing, uh, you know, um, for for pensions who want to sort of maximize their return but minimize their volatility. That that kind of um application, uh, I think speaks to the efficiency of being able to have good world models of the system that your customer is dealing with. Um, as distinct from building agents that have a situation awareness and can use active inference to actually do stuff in the moment, which would obviously speak to things

• 59:10Speaker 00

  like robotics and uh and autonomous vehicles. So those are sort of two, to my mind anyway, two sort of uh ways of leveraging and exploiting these first principle biometric approaches.

• 59:26Craig S. Smith?

  I mean, it seems uh that there's so much that you can do with this. Sorry, how uh widespread is research on on this active inference or first energy principle in uh in modeling? I mean, is it is it, uh, for example, that uh with your academic hat on, uh, do you have a team of PhD students working on this? Or are there other uh institution or labs working on it?

• 1:00:03Speaker 00

  Right. Uh, that's a nice question. Yeah. So in the life sciences, and particularly in the in the in the neurosciences, um, active inference as basically a sort of the technical version of something called predictive processing is now the main paradigm and has been since the turn of the century. So, you know, last century, we had some sort of behaviorism and behavioral psychology that gave rise to sort of reinforcement learning in behavioral psychology that then was inspiration for much of, you know, say, Sutton and Barto and then subsequently, you know, Q learning and reinforcement learning. There was a move in the latter half in the 20th century to now away from behaviourism,

• 1:00:45Speaker 00

  to sort of, you know, thinking about cognitive processes, belief updating from, you know, from the point of view of this conversation. And then at the turn of the century, there was this inactive movement where people were emphasising the embeddedness, the inactivity, the extended aspects of cognition in exchange with the environment. At that point, in academia, certainly the cognitive neurosciences and in philosophy and the neurosciences, predictive processing, you know, became the major theme, active influences or so, if you like, the technical aspect of that.

• 1:01:27Speaker 00

  So in that world, it's now the standard model. In industry, company and applications of a commercial sort, I think we're just at the beginning. There are, there are now institutes, there's the Active Inference Institute based in California that's now a sort of not-for-profit registered charity, I think in America. We're just about to have our sixth international workshop on active inference, but these are little baby conferences, you know, they're dwarfed by NeurIPS, for example.

• 1:02:07Speaker 00

  So this is a very embryonic community with lots of bright young things, lots of promise, but, you know, from a, you know, a translational perspective into industry. And, you know, I think we're at very early days and, you know, for my part, it'll be a very exciting time. Just to answer your final question, I'm afraid all my PhD students grew up and left me. So.

• 1:02:34Speaker 00

  Seriously, I actually started my flexible retirement from university last year. So you know, I have sort of 80% full-time employment now, but I have no students left. They all grew up and what happened to them, they ended up in Versus, which is why I, one reason why I'm so committed to Versus. To keep my eye on all my, all my bright, favorite young things.

• 1:02:57Craig S. Smith?

  That's it for this week's podcast. I hope you learned something and I'll be doing more interviews with researchers working on post-transformer architectures because I think that's the future. See you next time.