Dr. MAXWELL RAMSTEAD - The Physics of Survival

we're here to talk about the free energy principle folks and no this is not a scheme to reduce your electricity bill now today we granted the indubitable pleasure of hosting a truly multifarious intellectual a paragon Maxwell ramstead a philosopher mathematician and leading prognosticator on the free energy principle now from his earliest years Maxwell exhibited a precocious interest in mathematics and science finding solace in Realms where rigorous axiomatic thinking could elicit precise answers to profound questions yet his intellectual Journey was not confined to these disciplines he was equally captivated by the labyrinthine intricacies of philosophy seeking Clarity amidst the nebulous mysteries of existence in a testament to his indefacable intellectual curiosity Maxwell secured a PhD in philosophy and cognitive science as he descended deeper down the academic Rabbit Hole he developed a Fascination for complex systems theory immersing himself in the study of dynamical systems ecological psychology and other methods of formal cognition modeling um so there there's an argument to be made that really this is really a theory like a a formal theory of uh cool I mean the the fep itself is not a theory as we discussed it's a principle but it's a formal approach to that begins to give you a grip on how to model nested systems and systems of systems of systems uh the cool thing is that the whole stack operates on the same quantity basically so you have one objective function it's the same pattern at every scale it's the now in the years of 2014 and 15 Maxwell underwent a significant intellectual metamorphosis he chanced upon the free energy principle an adventitious Discovery originating from the esteemed neuroscientist Professor Carl friston this principle manifested like a Rosetta Stone deciphering the autodidactic yearnings of Maxwell's mind providing a unified theory encompassing mathematics physics philosophy complex systems and cognition so what Transcendent concept is this free energy principle well it's an audacious inversion of the traditional survival hypotheses so Kristen is an absolute honor to meet you right it's lovely to see you um tell me a little bit about where where are we you're in my office um post pandemic so this is a place that I very infrequently visited in the past few years it's a pleasure to be back a bit Nostalgia got you wonderful and and we're we're here in central London yeah so we're in the Bloomsbury area we're at Queen's Square this is the welcome uh Center for human neuroimaging uh across the road you have the queen Square National Hospital for um um neurology and neurosurgery um and The Institute of Neurology so this is if you like part of um Bloomsbury Academia part of UCL with a bit of clinical and medical uh with a special focus on neurology expertise or into the mix and how many years have you been basically I can't remember decades so I came here uh I came we came back or I came back from America uh working in California uh part of a reverse brain drain if I remember correctly with John Bell who has done very well subsequently um and this was in the sort of late 90s and I've been here ever since instead of querying which behaviors promote survival it poses a seemingly prosaic but ultimately profound question given that creatures exist what must they do the principle posits that in order to safeguard their existence systems must minimize free energy this requires systems to be vigilantly phlegmatic maintaining their internal States as close as possible to anticipated States derived from a world model failure to do so leads inexorably to chaos the system would vitiates dissolving into disorder thus the free energy principle offers an Innovative explanation for why living beings are seemingly engaged in a Relentless struggle to model predict and comprehend their surroundings it's a survival imperative and necessary counterbalance to the entropic forces constantly threatening their very existence so essentially the free energy principle is the merry condo of the Mind constantly tidying up always trying to match expectations with reality the utility and relevance of the free energy principle are not categorically delimited it ensconces itself across disparate scales from solitary cells to neural networks organisms to vast societies its influence is pervasive and omnipresent wherever order and Longevity are found the Invisible Hand of the free energy principle can be discerned silently coaxing systems towards equilibrium now whether yearning for deeper understanding still burning bright Maxwell embarked on a journey to the hallowed Labs of Professor Carl freston at University College London now this self-imposed intellectual furlough allowed Maxwell to delve headfirst into this Paradigm understanding its native language if you like the language of mathematics and physics now his work didn't go unnoticed Maxwell was now considered the leading figure or one of the leading figures in the field now Maxwell views the free energy principle not just as an academic curiosity but as a comprehensive physics of the mind on par with such seminal intellectual epiphanies as Newton's law of motion now just as Newton's Laws reconciled the celestial and the terrestrial Realms under the unifying Banner of classical mechanics the free energy principle extends its intellectual dominion over life cognition and intelligence bringing them together in harmony with the laws of physics now it was originally proposed by Professor Carl friston and it is essentially a mathematical unifying Theory of Everything it explains how living things from single cells to galaxies maintain their existence in a sea of disorder by minimizing free energy so join us as we feel our way around the idiosyncratic free energy principle and Ponder its staggering implications how does this principle elucidate the Enigma of cognition what can it reveal about the fundamental laws of existence and how could it precipitate seismic shifts in areas as diverse as Neuroscience artificial intelligence and philosophy our conversation today will oscillate between the technical and the existential from the esoteric to the Germain Maxwell with his unique blend of mathematical incisiveness and philosophical sagacity is the quintessential intellectual guide for this Expedition so prepare to submerge yourself in his articular incogent thoughts they promise to be an edifying experience I can tell you that for now and they will cast light on an unexplored corner of our understanding and hopefully Kindle mental synapses with fresh insights enjoy well it's it's a thrill and an honor to be here with you guys like frankly I'm just so pleased this is really nice Maxwell it's an absolute pleasure uh to have you here today and I want to thank you for coming on the show because you're one of the uncommon type of combinations that I that I love talking to and find you know so enlightening which is a philosopher who's also fluent in mathematics and I think that brings an interesting rigor you know to the philosophy and then there's actually like a really cool interplay between philosophy and Mathematics and Science so I think I think you gain a lot as a philosopher the more science and math you know and vice versa you gain a lot as a scientist and a mathematician the more philosophy you know well thanks very much for your kind words um I'm inclined to uh agree that philosophy and formal training go uh hand in hand and you know that it is something of a strange historical circumstance that we we don't think of philosophy as a very formal discipline anymore uh you know if if you went back a hundred years uh some of them there was no separation there was no distinction between science and philosophy right absolutely think back to the principia Mathematica right uh you know uh Russell and Whitehead uh philosopher mathematicians uh you know the it was very common back in the day and uh less so now but you know let's hope that uh the trend is reversing uh I I I'm definitely optimistic about this I think uh I I increasingly uh find myself surrounded by a very technical uh colleagues who also have uh strong philosophical backgrounds uh so uh yeah hopefully things are changing for the better yeah and you know it's interesting I'm going to take the time to ask you since you are you know a philosopher like by by education and I mean no offense folks to philosophers whatsoever because I think science has lost out on losing losing philosophy and philosophy is lost out on losing science and like the the rigor that you know confirming your thoughts against reality uh can have I mean do you find that to be the case which is that I find a lot of philosophers you know without that grounding kind of then in Science and Mathematics they spend a lot of time pontificating really about things that just amount to the to the vagaries and ambiguity of like natural language rather than than being able to map it to something symbolic something concrete you know well um so I think uh I think it depends what kind of philosophy we're talking about um I mean in in general I think uh the ultimate problem is that being a philosopher of science is difficult because you have to be a philosopher and a scientist um and usually when you're training you can only do one PhD so I mean I think to be like a proper philosopher of science you know you would have to be familiar with the the history of philosophy and contemporary philosophy of science and also uh to get some proper training in the discipline upon upon which you are you know uh directing your Reflections um and so you know in my case uh I had to do a lot of um kind of training on my own so my my uh most of my coursework uh was in philosophy and cognitive science and so the the the more formal stuff that I wheeled uh I've had to teach myself as a kind of side job so like when I was doing you know my masters and my PhD uh in philosophy and cockside I would also be teaching myself math on the side and I think that you know to to to to to uh you know exist effectively at these intersections you can't help but uh get some multi-disciplinary training and you know I think a good philosopher of science is necessarily A polymath in some sense like you you need oh that's absolutely right and I think that's that is what makes it that is what makes it kind of a rare breed right like uh you know like one of my favorite philosophers of mathematics is Eric curiel and he has so much great you know content on on general relativity and Mathematics and philosophy in general but like you say there are these rare these rare combinations and or it takes a lot of self-study you know to develop that that's right yeah um and I mean you know topics like the free energy principle and active inference are inherently multi-disciplinary so it's it's beyond just like the philosophy and the math uh you know what if if you're in some sense if you're doing it well then uh insights from every discipline will wind up making contributions to the way that you're thinking uh Feb theoretically so um yeah uh but I um this is one of the things that I love the most about uh our intellectual Community uh around the fep and active inferences the multi-disciplinary the multi-disciplinary nature of it and that's sort of unavoidable right um so it you know the free energy principle uh which we'll be discussing uh today is a an explanatory principle that that its proponents at least purport applies to uh every scale at which physical systems self-organize and therefore uh you know insights from all of the different disciplines working on on all the different kinds of things uh become relevant and the fep then acts as a kind of like meta theoretical architecture to fit these claims together uh in a way that's tractable uh from a physics and from a mathematics perspective well one of my earliest childhood memories is uh understanding how Cartesian coordinates work when I was uh I guess uh uh yeah 11 or 12 years old uh kind of really like understanding the mechanics of that and going like wow like there are disciplines where there are exact answers to the questions that we have uh and so yeah I was always uh very into the sciences and math but I I've also always been very interested in the big questions um and by the time I was done with High School uh I would I I was either going into um uh philosophy uh physics or film and so uh I ended up deciding to go into philosophy and I uh I remained a physics and math geek throughout so I'm self-taught in a lot of ways uh and uh yeah I ended up specializing in philosophy and cognitive science the philosophy of cognitive science and formal approaches in cognitive science in particular you know when I was doing my undergrad uh the embodied and active embedded extended traditions and cognitive science were very hot and so I I learned some dynamical systems theory uh and I was very into ecological psychology which is a branch of psychology that attempts to apply principles from physics to understand action perception Loops so I was very into all of this cluster of ideas and um around late 2014 early 2015 I was exposed to the free energy principle and the ideas of Carl friston um in part uh by reading uh this this really wonderful and by now seminal paper of Andy Clark's called whatever next um which was his uh famous BBS paper where he kind of reintroduced uh the the predictive processing framework um and uh you know active inference along with the free energy principle um so that essentially uh combined everything that I had ever found interesting in some sense and I had a real kind of conversion experience if you want to call it that um and so uh yeah I I had the pleasure and privilege of meeting Carl In the Flesh in May 2016. um at a decision-making conference in Montreal and uh you know you've spoken to Carl a few times he was uh very typical of Carl he was very extremely generous with his time and friendly and insightful um and at the time I was working on socio-cultural Dynamics and I was wondering whether it was apt to apply the free energy principle to explain this kind of you know Ensemble Dynamics and you know Carl uh responded with an affirmative answer uh and encouraged me to uh you know take this seriously and exchange with a few other people um and what began was a very extensive email correspondence that turned into some papers uh Carl then became my PhD supervisor and I spent about yeah the better part of two years in London uh with his group at UCL learning the ins and outs of the free energy principle um yeah and so um so I guess uh that's that's sort of my story and how I met Carl uh that's interesting you know and the free energy principle so um you learned about it much earlier than than I did I wish I had known about it earlier but I didn't come across it until I guess a couple years ago whenever the first you know the First mlst episodes that that we did on it but I had had quite a bit of Bayesian you know background up in that point so when I first saw it I thought wow this is It's really ingenious you know I like I'm fascinated by it and we wanted to learn more and so we've had we've had Professor ferson on the show a couple of times and like you said tremendously insightful uh you know one of the most brilliant people I've ever spoken to it's such a joy to talk to them and I think even though we he's been on the show a few times and talked about we should probably frame up a little bit about what the free energy principle is absolutely it'll play a big role in our communication going forward here so let me take a stab at it absolutely that's because you're the expert like well you're usually really good at summarizing well you know what fascinates me about I think is and I want to get to the the Crux of this this beautiful statement that was made about it being the ultimate existential question right because usually we think um what does what does a life form or a thing even for that matter what does it have to do to survive right like this is kind of the emphasis of you know are you fit the darwinian sense you know what kind of Fitness does it take to survive but this the free energy principle completely inverts that and it says okay if things exist if things survive what must they do right and then it turns it on the head in this way which is let's assume that there is a thing you know and it and it continues to survive it continues to exist just by knowing that what must it do you know what what Dynamics what Behavior must it have is that a fair Framing and what is what are those behaviors that things that exist must do um so uh yeah no I think uh the way that you uh describe things is uh accurate and an insightful way of putting things uh the free energy principle uh is not just a uh basically a theory uh According to which uh things that exist uh must be doing this or that as in it's not it's not trying to um to tell you uh here's something that things do in order to exist what it's what it's telling you is we we observe that things exist in the sense that uh there are uh uh there are systems or uh particles or things uh that can be reliably re-identified that are separated from but coupled to their environment and given that we observe these things that exist uh what must be true of them so it's a kind of inversion of the explanation moving to like a kind of uh first principles account of what what what must necessarily be the case uh if you exist and essentially with the free energy principle tells you is that if you exist in the sense that you are separate from but coupled to an embedding environment then it will look as if you're tracking the statistical structure of your environment or more precisely it'll look as if the states or paths that are internal to uh a given to your boundary of a given thing track things that are external to that boundary so in some sense it it explains why uh or it provides a a principle allowing us to uh uh yeah explain why uh it it looks as if everything that exists is uh yeah tracking or representing depending on how you want to think about it uh features that are external to it um and this this tracking or representing relation is um it is rather weak in some sense but we're not talking about like necessarily contentful representations little images in my head what we're talking about is uh something I think more fundamental or existential so for example uh I can see from uh your your camera uh feed that you're wearing a uh a button shirt uh that tells me something about your environment uh so I I can read off you from the fact that you know you're wearing a button shirt that you're probably not in the Arctic somewhere I mean I can see your background also that that's kind of cheating uh the basically we um we can read off anything uh properties of it's in of its embedding environment by virtue of the fact that it exists right right so yeah I think there's an interesting point in there which is that it isn't this exact one-to-one correspondence and I mean really how could it be right like how could a how could say a subset of the system precisely represent the entire system but instead it in a sense it's representing an abstraction absolute system you know and I mean I think there are even good energetic arguments why that would be the case because you know you since you can't maintain total information you maintain an abstraction of it you maintain enough in order that you can predict and track you know behaviors because you exist you must continue to exist but you don't require complete information only sufficient sufficient tracking absolutely yeah heuristically you can think of the the free energy principle and this is metaphorical of course but as a a map of that part of the territory that behaves like a map so uh it is a a scientific uh principle that we can use to construct models of systems that appear as if they are in turn modeling or tracking aspects of their environment um and I think thinking about modeling for a second is useful here so if you had a one one Scale map is it Borgeous who who uh presents that in a in a story at some point uh yeah I think so yeah so a one-to-one scale map would be completely useless right I mean just imagine the quantity of paper that you would need to have a map of say la right a one-to-one scale map of La it wouldn't be useful in the generation of your actions uh so the map is necessarily simpler than the territory um but that's okay in some sense right because the implication of that is that using the map you're going to encounter errors you're going to generate errors but uh in the active inference and the free energy principled approach those errors are the relevant signal basically so all all you have to do is have a good enough map and act in such a way that you are informed by what your map contains in terms of information and to in real time course correct based on the errors that you're generating um so it's the these errors in the kind of oversimplified character of the model or features rather than bugs uh you you would need that in order to have a signal at all in some sense yeah and there there is this iterative nature to it that I think is is sometimes forgotten because uh you know there are these two components in the free energy principle um one is Fidelity how accurately does it kind of map to uh the environment like so so if you if you think of the the the entity that's surviving the thing that's surviving you know has to have a model of the environment it has to be have some degree of fidelity because if it doesn't it's not accurate enough to maintain its survival but at the same time it also has to have adaptability right because the information is never complete there's always new uh you know phenomenon occurring to it the the environment is changing or whatever so the model has to maintain a degree of flexibility right and that's what this this kind of entropy term in there is it's saying precisely um you know you need to maintain a certain amount of entropy because that is a form of flexibility is that correct I mean that's absolutely correct and you can think of the entropy in a few different ways um and it the the entropy term in previous discussions with uh Carl friston you have highlighted that it's technical importance um I mean basically what we are what we're trying to do when we minimize free energy is to increase the predictive accuracy uh of our model uh so that is to have a model that generates predictions that are as close as possible to the real data that I'm ingesting um but the free energy principle uh allows us to in a in a principled manner penalize the complexity of the model right because you don't just want an arbitrary explanation as you know you can construct an arbitrary uh explanation for any uh data set and uh you know it's looking to be deleterious if you have incomplete information and you model it too accurately accurately in a loose sense then actually you're you're just you're you're memorizing uh spurious information that doesn't generalize absolutely yeah absolutely and so um the free energy principle uh when you're when you're applying it and you're saying that systems that exist minimize this this quantity variational free energy uh the the variational free energy can be decomposed uh roughly speaking into predictive accuracy minus complexity and so what you're doing is you're you're penalizing your gains in predictive accuracy against uh the the complexity cost of your model basically penalizing every new degree of Freedom that you need to introduce into your model to explain the data so in some sense uh the free energy principle uh is sort of like you can think of it as a kind of like uh statistical predictive accuracy uh but also Occam's razor right right right so yeah no no it makes sense I mean and that's it's this interesting balance and the free energy principle encodes that balance and I guess the simplest the simplest way possible this is a sense in which the free energy free energy principle applies to itself because it's that's right it's almost the simplest formulation of that of that balance right yeah yeah I mean um there are all sorts of ways to control for complexity that have been introduced in machine learning um but uh those all uh might seem ad hoc in some sense what we try to do is to build in this complexity control uh right into the objective function that we're using right so it's it's at the core of the architecture it's that rather than just use reward uh which you could probably cast in terms of like predictive accuracy so I think we we have one more foundational concept that I'd like to kind of get on the table and that is just the concept of a thing you know because because a thing the idea of a thing is is defined in a very certain way in the context of the the free energy principle and it's all about this this Markov blanket and and we've we've talked about markup blankets a few times on the show they tend to be confusing you know I I tend to I I almost visualize them usually as uh cells like in the human body so like you know you have a cell and it has stuff inside and then it has the cell membrane and then there's the stuff outside which is kind of the environment and that membrane in a sense is the Markov blank and it's the set of stuff and States between you know what's inside and outside but can you talk a bit about you know why Markov blankets are important what they are you know how they're defined um and maybe some of the edge cases like there's an eternal flame have a Markov blanket why or why not that's those that's a great set of questions um so uh you can think of the free energy Principle as um a the same kind of thing as the principle of least action in the sense that it's it's a principle that allows you to write down uh mechanical theories or mechanics right uh so uh classical mechanics has the principle of least action and the principle of least action is basically a way of uh specifying the uh the the conservation laws that we want to see our systems conform to in particular the the balance of the potential and kinetic energies uh are zero and the real trajectory of systems physical systems are those for which that balance uh holds and uh they they are maybe I'll just quickly interject here it's not so much that we want them to conform to that it's just that they do yeah and that's actually key to the free energy principle too because it's not like it's not like we want things to obey the free energy principles they must obey the freedoms in principle if if they you know exist right absolutely I was speaking in in the manner of a mathematician oh yeah no totally totally get it I just wanted to point this out to our readers is that because as we go along like you are almost these are principles that in the case of classical mechanics just are they're just the way things behave right and in the case of the free energy principle if you survive if you exist you're inexorably drawn to this set of Dynamics to the set of mechanics otherwise you don't exist well to get technical for a second I think there are two issues uh that that are both striking uh and then speak to what you just said so things like the principle of least action and the free energy principle and the principle of Maximum entropy they are in some sense true apriori or mathematically they are mathematical truths so you wouldn't try to falsify the principle of least action empirically any more than you would say uh try to falsify calculus or probability Theory by coming up with an empirical counter example so there's a sense in which like the the truth of these statements is robust and mathematical um having said that it is a striking empirical fact that the physical Universe does in fact seem to conform to these mathematical regularities so it's sort of a one-two punch uh in some sense and uh right right it's the unreasonable effectiveness of absolutely yeah uh and it's getting even more unreasonable now with the free energy principle coming into play um so the way I like to think about the free energy principle is like I was saying as introducing a new family of mechanical theories or mechanics so uh you have classical mechanics uh which follows from the principle least action and you have quantum mechanics and you have statistical mechanics so basically the idea behind the free energy principle is let's get the rest of physics working in the background so you get you get your classical statistical and quantum mechanics working with all of those uh yeah all of that mechanics is in play and then you ask a simple question what does it mean to be a thing in this context I.E what could it possibly mean to uh re-identify some status the system is being the same state over time um and so yeah unpacking that question leads to a Bayesian mechanics uh so Bayesian mechanics is a physics but it's a it's a physics for the the kind of dually constrains uh the system itself the physical system and the beliefs that the system encodes about the things to which it's coupled right and that I think is really the key uh for the the kind of underwrites this whole construction uh this is also what makes it unique among other uh you know candidates and the cognitive sciences and the biosciences uh it's that it is connecting the thermodynamic entropy of the states the system is made of to the information entropy of the beliefs of the probabilistic beliefs that those States encode so the free energy principle is all about this hinge between the two and the equations of motion that you write down using the free energy principle are constrained in both of these spaces and that I think is absolutely the key to understanding what's going on here like the the the system trajectories that you get uh when you're writing uh when you're writing down an fep theoretic uh model uh of the system that you're considering it is constrained both by the the physics and thermodynamics of the system and by the physics of the representations if you will uh the the system is entertaining about the rest of the universe in some sense um and I think it's it's not without precedence in the sense that for example on statistical mechanics you know back back when we're driving Boltzmann's distribution and there were questions about are these particles do they behave as if they're identical or if they have the same attributes are they are they are they the same particle and the same thing with um poly exclusion you know it applies to some particles and not others so there those are examples of where a statistical um calculation a statistics fiscal property does map to the actual underlying physics you know that's happening as well this is almost a big step up from that it's like you know it's a generalization to a much richer set of um I don't want to say conservation laws but a much richer set of these uh flow laws you know absolutely and Dynamics yeah that's exactly the right way to think about it um in fact it what's transpired over the last few years is that really you can think of the free energy Principle as a kind of generalization of the second law of Thermodynamics to open systems um so you know the kind of universality uh that the second law uh has with respect to closed systems uh the free energy principle has with respect to systems that are far from equilibrium and as Carl has you know pointed out the Markov blanket is precisely the apparatus that allows you to move from the equilibrium to the non-equilibrium regime in the sense that you are now specifying the interface through which the system is coupling to its environment or the particle so the word system is a bit ambiguous usually um you know because we could mean you know the whole in agent environment system or we could mean uh the specific kind of compartments that we're considering as agents uh yeah so the uh you can think of the the free energy Principle as applying to systems that are not at equilibrium and as giving you the Dynamics of like particles within the overall system uh if there exist particles in the system if there are these things that you can poke and re-identify uh over time right then the free energy principle will basically tell you how they behave on average yeah and it's it's interesting because almost the Markov blanket really provides the concept of stability because when I think about equilibrium right when when um you know the thermodynamic kind of equilibrium what comes to mind at least visually to me is things that are kind of unchanging and just kind of quiescent and they're just sitting there like a goo you know that's kind of all thermally stabilized and never never moving around but that's the point the key point there is stability like it's a it's a stable sort of unchanging system and what the Markov blanket gives you is a way of defining a property that's stable about the system and yet the system as a whole may be far from stable it makes me influx and moving around all the time and jiggling it's the Eternal Flame where the boundary is you know kind of changing all the time but there's this property that you can map from from time to time that that's still there and so it's the stable the state blanket even though it may have a different form it's still got a boundary between some stuff that's inside and outside that's identifiable you know from frame to frame if you will yeah and I I I I go even further and say that it's the self-identical pattern that you see at all scales of self-organization uh so it's got this kind of fractal quality to it where uh you know the it's blankets of blankets of blankets of blankets uh so I mean you can think of the brain for example so the brain has this nice Markov blanketed structure uh at several different scales so you can start with neurons and neurons have their own uh Markov blankets that yeah the actual membrane yeah precisely and it's probably the most obvious but you can go downwards in scales and upwards and scales and what you recover uh are uh you know similarly Markov blanketed structures uh you know from uh the uh the from the voltage-gated channels uh on a cell membrane to dendrites uh to the you know arborescences that form up to the the the uh the neurons Soma and then you know uh you could go the other direction too right from neurons to canonical little micro circuits to more specialized brain regions to brain networks and so on all of these things are things and they can be reliably re-identified with their own kind of uh you know uh properties and features and connectivity to other things and they all have this same pattern that repeats and so it really is yeah so and and I think anybody can see that pattern you know I mean look around we got we have trees and they have boundaries and you know yeah even higher in scale you have planets and galaxies and but at the same time if you if you dig down into it and you try to really precisely Define it then it it slips away it's like if I zoom into the surface of my skin you know fine enough well it's no longer a surface it's this ragged thing with cell membranes and then the cell membranes or molecules with tons of space you know in between them and heck atoms you know are mostly empty right like there's this this weird you know vagueness and difficulty in defining you know boundaries so they're not sharp like how's the concept of of a Markov blanket how does it evolve to that kind of fuzziness or this great set of questions your questions are still on point key exactly this is uh so um I would say the the first thing to say is that so this is sort of Mind physics or brain physics if you want to think about it like that and um one Hallmark of explanation in physics is simplification or uh you know you might say over simplification um so I I'm sure um a biologist reading a paper on the free energy principle might look at this and go but you know this is way too oversimplified whereas the biological detail um I think physicists also do this to physical phenomena right like that's that's sort of the joke a physicist would respond yeah but we do that to physics as well right like it's uh so yeah there is a deliberate idealization going on uh if you define this Markov blanket strictly as the the set of degrees of freedom that render some in inside independent of the outside right so if conditioned on the blanket States you can speak of statistical Independence between the inside and the outside well uh that that is too strict to describe most physical systems right so I assume you myself and our listeners we all had a bit of coffee this morning we've all used the washroom uh so there's clearly a kind of permeable exactly there's clearly a kind of permeability at play um and so uh yeah we we know that there is material turnover in most of the kinds of systems that we find interesting especially those that self-organize far from equilibrium so the Markov blanket is necessarily an idealization um having said that um there are good reasons to think and we have some results coming out uh later this year uh a lot of them are due to our senior mathematician Dalton shaktivadivel who is an absolute Dynamo um so we have seen I've seen some of those papers yeah it's very impressive stuff he's been working on um weak or fuzzy blankets uh and precisely this context so the the idea is uh can we get really rigorous about the mathematics of approximate Markov blankets or fuzzy Marco blankets um and the idea in a lot of his work is to construct this this quantity called the blanket index to gloss over some of the technical details just for in the interest of uh our uh our audience uh basically uh they're in if you consider a given dynamical system there exists a Markov blanket in that system if a specific inner product is equal to zero uh so in particular this is the the Hessian of the and the uh the solenoidal flow the product of those two things being zero well it basically is a way of quantifying the the force or the curvature uh that a system is subject to uh and uh yeah if if the entries in this inner product are zero then there is a strict Markov blanket but there's a way of constructing uh an index or a measure such that you can accumulate the non-zero entries um and basically quantify how far from perfect blanketedness a system finds itself um right and so uh yeah this blanket index has a number of interesting properties um one of which is that it it it tends to zero as systems increase in size hmm so under what kind of assumption so um it's very very very general um so uh locality assumption or yeah you get the locality stuff from from the background right so you get the rest of mechanics going right so okay okay so then you already have like you know relativity in the background and you have statistical mechanics classical mechanics and all that stuff so yeah uh you do get uh this kind of nice uh locality so that's interesting so in in a universe like ours that has the the basic physics that the Universe like ours has as the scale of a system gets larger and larger you you generate Markov blankets you're balances with the probability one yeah absolutely fascinating uh and you know most of the systems that we consider in physics are large in the appropriate sense right uh so think about how many molecules are in in a drop of water right it's 10 to the 26 right some some something times Avogadro's Number yeah exactly itself is 10 to the 23rd yeah yeah so 10 to the 23 sorry yeah so that's just for a single drop of water now as you consider the brain the brain has like something on the order of a 100 150 billion neurons uh Each of which make thousands of connections if each of those connections can encode a parameter then you're talking about like a very large system right we're way way way Beyond like you know 20 50 1 000 different states that that are coupled together we're talking about like billions and trillions of different states so there's reason to think that just due to the physics of the situation uh most relevant things that we might want to consider will have Markov blankets um and uh I mean it it so Dalton is going to be releasing a few papers having to do with large fluctuation theorems okay and so let's let's pause for a minute and appreciate this because this to me is a fascinating a fascinating result so okay so we start from the point of a Markov blanket is kind of this intuitive concept right like it's you know a boundary and and that sort of thing but there's no reason to believe that they're inevitable and I'm finding it fascinating that there's that there's this work right that says that as a scale Inc so we have a measure of blanketness it's kind of between zero and one zero has a a blanket one doesn't Okay and yet as the scale the system gets larger and larger blanketness approaches you know zero you get blankets um no matter what and in a sense there's a sense in which that's recapitulating what we see if we just look around like everybody out there listening look around yourself and you're going to see blankets all over the place you're going to see things and those things have have boundaries right but it's it's remarkable right that that there's a mathematical proof that that's inevitable in this sense isn't it well I I think it's remarkable in part because we have approached uh the question of self-organization and emergence from a false starting point uh so I I've been going around saying recently Aristotle was wrong well the hole is much less than the sum of its parts it turns out um so uh yeah there there are a bunch of things to unpack from that well the first is that um what makes you the kind of thing that you are is the sparsity of your coupling to the rest of the world right if you think of a gas right everything is coupled to everything else then it's just this fuzz and it's all one system and there's there's no you can't really identify particles within the system uh particles or things are defined by their sparse connections to everything else so I am in some sense what I am not or I can be defined in terms of what I'm not connected to as opposed to what I am connected to I mean if you were to create like a giant adjacency Matrix for the entire universe most of it would be empty the whole the whole is less than some of the parts which is there's more if you if you get rid of the parts you have you have less right and but but there's more um the uh so think of an engine right like an engine functions as an organized Hole uh because you're constraining its parts to behave in very specific ways so like you know if you think of an engine more specifically like a petrol engine uh well the the mechanical effect of the engine you get by you know moving these pistons in a specific you know Direction up and down and the best way to wreck your engine is to introduce new degrees of freedom into it right I would not want to introduce new degrees of freedom into the Pistons it's a that's a great way to just tear your engine apart and I would uh submit to you that this is uh you know an accurate way of thinking about all self-organization we exist as holes because our parts are constrained to behave in very specific ways so it's not merely that I am what I am because I I am not what I am not just a nice tautology uh it's the the what makes me what I am is the way that I remove degrees of freedom from my parts such that they conspired to create you know to generate me as an overall pattern uh so I mean it's it's counter-intuitive from the perspective that we inherited from you know a traditional Aristotelian metaphysics but um I mean uh yeah it's so exciting I mean it really is I don't know I don't know why I get excited but to me it's it's really excited and this I mean so this is work that's coming out of verses correct that's right yeah that's right and so we're coming out of yeah I would say uh coming out of uh Carl friston's group more generally and let's not forget you know Carl uh proposed the free energy principle in the middle of the 2000 knots um and so yeah uh but definitely this is uh the the PowerHouse behind uh versus AI Technologies uh so yeah let's talk about that for a minute because um yeah we generally don't talk about you know companies on the show but this to me this is an exception because it's fascinating what you guys are doing thank you and um so versus versus technology is trying to operationalize this understanding really and this is technology and I guess and I I'm super excited by what we've been talking about conceptually but tell me why should anybody care like what does this what does this fascinating you know view of philosophy and Mathematics and the relationship to really I think it's really about understanding complex systems in a new way and new mechanics of complex systems what does that mean for us what does it mean technologically well thank you for your excellent question um so I I think one of the things that we want to do at versus is to uh apply active inference to artificial intelligence so active inference is basically the kind of machine learning that follows from adopting the free energy Principle as your kind of core method our contention at verses is that active inference will be to the 2020s what reinforcement learning was the 2010s effectively so it's going to be we think the way of doing machine learning uh the ethical scalable if most efficient way of doing machine learning um and there are a lot of different aspects to that um so uh one of the main differences between artificial intelligence built on the principles of active inference and more traditional approaches is that we start uh from an explicit generative model so-called um so we talked about the Markov blanket the generative model is another core piece of the free energy principle uh puzzle or uh constellation if you will and the the generative model is basically um a statistical description of the dependency relations within the system that you're considering um so when we're talking about Markov link it's actually what we're saying is the generative model of the system contains a Markov blanket right so all of these dependency relations that we're writing down uh like once you do write them down you get this nice sparseness structure where some parts of the system do not affect other parts of the system um and so yeah the this this generative model um is is really key to the doing of the free energy principle um and so uh yeah what we do in active inference is write down generative models explicitly label generative models that then allow us to uh you know perform inference uh they allow us to do that because the uh variational free energies that we were discussing earlier um basically the the uh the gradients of the free energy that you're following they come from this generative model basically um right so uh yeah the the model itself is this explicitly labeled structure and this is where you get like a huge explainability uh Advantage we we actually have a paper that uh came out uh on designing artificial intelligence explainable artificial intelligence using active inference and uh yeah what what you get just immediately from flipping to an active inference framework is a way to write down uh yeah generative model such that it is explicitly labeled and thereby auditable uh by human users and stakeholders so you don't have this unlabeled you know 10 trillion parameter net as it were right what you have instead is uh yeah a model that explicitly represents all of the different factors in the situation that you want to control so let me look into this let me dig into this a little bit um so and I because I'm curious there's a connection here to some some work I used to do a long time ago so writing down generative models my experience has been at least is uh it's it's actually relatively easy to do that so for example a long time ago I was contracted to do some work on mad cow disease so we could try and figure out what interventions to do to reduce right the spread of mad cow disease and it was pretty pretty easy to learn about the uh the food system the supply chain the food system to model you know the processing of um of cattle and that sort of thing and write down a large simulator so this would be a generative model it can sit there and can generate trajectories through this space right like we knew we could do that we couldn't tell you anything about the large-scale kind of thermodynamics of the system or anything right we could write down this generative model though and then using um you know various sampling and techniques like that we could then compute statistics from it try out interventions see which of those had had kind of a beneficial effect but that was all ad hoc you know architecture that we we designed and produced by ourselves if I'm understanding you correctly what you're doing is producing a technology that one formalizes that much more and applies the free energy principle I think to help guide like the sampling and the optimization and you know really just the effective use of these kind of generative you know simulations and models is that close to what you're talking about that's right um so I mean you can why use active inference it is demonstrably the most efficient machine learning technique so it's sort of like uh you know a car no cycle analysis but for for an engine but for AI uh in particular uh with the free energy principle uh allows us to formalize uh is the thermodynamics of information writing onto the boundary um so uh in in some of the newer work on the quantum information theoretic formulation of the free energy principle uh which we don't necessarily have to get into in detail but uh there are these kind of new scale free extensions to the free energy principle that have been developed uh that appeal to the tools that have been developed on in quantum mechanics right so the theory of very small fast things but Quantum information Theory so the kind of information theory that gets augmented to handle things like probability amplitudes which are the uh the roots of probability densities and so you can get your wave equations moving in place and all that so the uh that formulation of the free energy principle allows us to formulate uh the computations carried out by a system in terms of like a per bid read and write cost so okay there's a there's a there's a sense in which like you're you're bringing it down to the like to the the the bear kind of you know machine elements of the of your computations and you're you're writing things down in a way that is demonstrably the most efficient way of doing it so if you if you set up you know some simulation system using active inference you are uh and this kind of brings the conversation full circle in some sense you are generating a model that is as predictively accurate as possible but also that expends as little energy as possible due to this you know controlling for the uh complexity of the model so there yeah we we have this um this preprint uh up uh that we'll be revising soon um called the map territory fallacy fallacy which is precisely about the kind of a canonical nature of uh fep theoretic modeling uh yeah uh you know the one of the reasons why the fep is optimal um is that it's it's another way of writing down James's maximum entropy principle so for our uh audience uh the maximum entropy principle uh you can think about it from from the point of view of statistics and also from the point of view of statistical mechanics from the point of view of statistical mechanics the maximum entropy principle is the principle According to which things dissipate um so uh from from the point of view of Statistics it's the principle According to which give me a data set and a set of models from which I might have sampled that data set the it says the model with the highest entropy is most likely to be the uh the real model from which you sampled so right that's the way of kind of saying like what is a maximum entropy probability density it's a flat density right and so basically a maximum entropy probability density encodes no information because all of the outcomes are equally probable and in some sense what Occam's razor would tell you is that like you would want your model to be as flat as possible Right you want to build in as spread out as possible exactly you want to build in as few uh assumptions as possible into your model this brings it back to the whole keeping your options open thing that you were saying uh you were discussing earlier you know if you're thinking of a probability density over different courses of action uh unless you're really sure that you want to do this you probably want to keep things as non-committal as possible and keep your options open so the just to back up then the free energy principle is a is a way of writing down uh the principle of Maximum entropy they are effectively the same thing you can move from the one to the other and we know that the principle of Maximum entropy is is is the principle of parsimonious explanation in some sense right so if the fep and maximum entropy are the same principle then all of the epistemic virtues that accrue to maximum entropy also carry over to the fep and therefore uh yeah a a free energy principle theoretic model of the belief updating of a particle can be shown to be the optimal Dynamic systems model for the whole system that you're considering like there is a kind of canonicity uh what we're calling Jane's optimality that yeah the basically the fep it allows you to write down the best model that you could for your the system that you're considering given your current state of knowledge in that system it's just the optimal way of writing that down full stop uh so that's why we you know care about active insurance yeah it might be useful to give you know some examples of um of you know maximum entropy distributions to understand so for example if you have a data set um and let's just suppose it's continuous you know data and it's and it's positive only so I know that it's continuous data it's positive only and I know that it has a particular mean you know then the the distribution that has the maximum entropy under those examples is an exponential distribution right like it's it's sort of spread out as much as possible and yet it has a particular particular mean and and and for example if we go to the case of it can be a real number anywhere between minus infinity and infinity and it has a mean but it also has a finite variance then you wind up with the gaussian you know distribution as being you know the maximum so it's really it's it's a distribution that captures what you know about the system I.E the constraints right and yet is as spread out as it's possible to be while satisfying those constraints right yeah absolutely and and you know I I hope our audience is able to kind of see the pattern that's starting to form here you know uh like all of these connections are non-accidental right so the the FED is all about balancing your predictive accuracy and your complexity so all these things are kind of connected at a deep level uh and yeah I mean maybe the audience also can see why we're so excited about this like this is highly non-trivial and I don't know if this is crazy or not but it actually even seems to have implications for uh fairness of of models because so for example um suppose I'm trying to train a model that does anything of of human interest you know diagnose uh or prescribe Medical Treatments or you know give out give out loans or that sort of thing and we need to train it on some type of data whether it's a generative model that we calculate or data that we actually observe well we want we want the system we want the machine to learn only what's relevant for that particular task and like nothing else you know we don't want to to learn kind of extraneous things so for example if it's deciding to hand out medical diagnoses we want it to be based solely upon the uh medical attributes that are in in the you know the data set and not some spurious you know correlation to like the geography or you know where where you came from or or um the letters in your name you know of your file or anything like that and so in a sense um what maximum entropy helps you to do is force out that stuff because that stuff if it's not useful for the actual prediction it'll get you know ironed out because it's smoothed out by the demand of Maximum entropy that's right and and that brings us back to sparseness right how so tell me Well in the sense that um you know this this even has to do with like situation or task definition uh like situations are sparse uh they're not all not everything is connected to everything not everything is relevant to everything else uh so there's even like a kind of it's really a kind of uh meta methodology because you you can even define specific situations in terms of their sparseness and then you know yeah so um so it sounds great but I'm always the Eternal skeptic because I know a lot of this type of computation you know the the generative starting with a generative model doing inference on it just computationally is so difficult like what what is the magic that verses is found and and how are you so sure that uh how are you so confident it's going to work great question um so I mean in particular one of the big questions that uh plagues generative model based methods is where do your priors come from right so what is the structure of your model why what are the parameters that you're using what's the relevant State space all of that uh is often I mean often you have to hand design this stuff and it's very labor intensive um so at versus we are a uh a contextual Computing company so we draw inspiration from the architecture of the brain uh and uh basically we're proposing a kind of General uh standards-based solution to the problem of where do your models where do your priors come from um so just to uh you know do a crash course in Neuroscience um so we love Neuroscience here so uh please do I'm sure well also a source of inspiration I'm sure everyone is familiar with the idea that the brain has a layered or hierarchical structure um so it's not the case that the brain is just a soup of connections where everything connects to everything else um no to the contrary there is there are very regular structured patterns of connectivity uh in the brain um okay again this takes us back to the theme of sparseness right uh like the to say that the brain has a hierarchical organization is to say that it advances a specific very special kind of sparseness where connections uh are directed in specific ways where you're connected to layers immediately above and below you but nothing beyond that um so we've spoken in the past to Jeff Hawkins like is this related at all to the concepts of cortical columns like the way in which they're connected like it almost has these you know components that are reused and put absolutely in terms of different layers uh yeah uh the and not all the layers speak to each other um from a an active inference effect is this layered or uh level uh level involving structure has a specific purpose um so what each level is doing is providing priors uh or expectations to the level below and receiving context from the level above and what each layer is doing in turn is shuffling prediction errors to the to the level above and receiving prediction errors from the the level uh below um so what what you have is basically uh a set of layers that contextualize each other and the way that each layer contextualizes the next one is essentially coarse graining right so uh there's some fast uh fast small scale activity uh the lower levels are tracking closer to the sensory end and with each successive layer is doing is finding basically the the uh set of hidden States or latent States uh that explain variance in the data that it's receiving and so on in this kind of hierarchical fashion um so the the brain is not one monolithic system rather each layer of the brain is specialized in encoding specific features of the situation pitched at a specific scale and it functions by kind of providing context so we like to say that the brain is an organ of context effectively where uh really what it embodies is successive layers of context that are each coarse graining each other in this kind of fashion so I know I know our neural network fans out there in the audience are gonna they're gonna be hearing that that's just what a neural network is like it has kind of layers you know they're connected what's what's different about just any other you know kind of neural network architecture if you will so what we're proposing is an infrastructure project in some sense uh we have uh we're working with the IEEE um in the U.S and so when I say we um versus the has a sister uh organization a non-profit called the spatial web Foundation which was uh to whom we gifted uh the the first massive chunk of research that came out of our group um and what we are developing with the spatial web Foundation is a stat is a set of public standards uh that people can use to build out uh basically shared knowledge graphs so we're building an ecosystem where folks can basically uh think of this as sort of like wikidata or Wikipedia But for kind of shared contextual compute context uh so you know obviously uh we are building this out so we're going to be the first to put things onto this network but what we're trying to build is basically a spatial web or a hyper spatial web that kind of in some sense reflects the structure the kind of graph structure um of the the various kinds of situations that humans uh deal with it so it's so it's really an application of of interesting so it's an it's an application of the free energy principle at multiple scales absolutely not okay okay so my mind was going in the wrong direction here which is I was going to like the scale of of uh you know the individual thing and and and how it does it's it's modeling its compute but this is this is is more than that this is well it is that you're absolutely it's more than that it's it's that it's and it's actually like a multi-scale I guess arbitrarily nested really uh framework for communication across absolutely it's it's more than a mesh it's um what's the right word for it's a hypergraph I guess it's it's a hyperspace we call this the you know the the name of the protocol that we've developed uh for modeling is called the hyperspatial modeling language hsml it it's it's meant as a an homage and a nod to uh HTML and uh what we hope it doesn't use the same syntax please no no well um well what we've also built is a transaction protocol and a querying language that live within the hyperspace so so-called uh hyperspace transaction protocol hstp and uh the hyper space hyperspace querying language hsql um and so yeah what we are basically in the business of doing is on the one hand building out these graphical models of knowledge basically of the knowledge that is involved in specific tasks domains and situations uh and then we've developed methods to take these uh knowledge graphs and to flip them into graphical models of inference uh so that's really that's that's kind of where the the secret sauce and the magic happens is that this is a two-step process and the overall versus technology stack combines the active inference based AI with explicit generative models on the one hand and this kind of nested hyperspatial representation of the problems uh to be solved on the other uh and the the Technologies uh kind of are uh married uh through the their kind of Reliance on uh on graphical techniques basically so it's it's knowledge graphs meets graphical inference in a nutshell um yeah and um we uh we really are committed to developing these in terms of uh you know open publicly available standards uh you know there's a lot of as I'm sure you're acutely aware there's a lot of um you know hype and doomerism going on right now uh with regards to AI there's a lot of I think maybe uh over inflammatory uh uh you know dumerism and over utopian hype going on um and uh in terms of you know the different options that we have to uh develop these Technologies in a responsible way that there seems to be one call for you know government oversight which is interesting but comes with its lot of limitations for example governments are limited to their jurisdictions uh and so you know you can't uh you can't coordinate an International Community of research and development in r d merely through uh National regulation uh so that's limited on on the other hand you have you know uh markets and companies that want to uh you know solve these issues in-house uh they are maybe faster and more flexible but there is this necessity of you know how do you how do you constrain the activities of corporations in such a way that we develop these Technologies uh responsibly ethically transparently in a manner that's Audible and that's uh you know acutely aware of and sensitive to the potential harms that might be caused by these Technologies so what we are uh proposing is a kind of third path a middle way not to say that we shouldn't pursue you know a private development of these Technologies and regulation I think this is all uh you know a great idea uh there's some interesting uh legislation coming out of the EU uh the AI act that everyone is talking about that I think are interesting paths forward but it to really consolidate the International Community around these Technologies we have proposed a standards-based approach um and the the IEEE group uh where the standards uh will be housed uh is an open Group folks from anywhere can join uh we have some pretty high profile corporate Partners but the idea is to build these Technologies in a manner uh where we avoid silos basically and where we can kind of coordinate the entire world's techno technological and intellectual prowess towards solving these uh issues so I think the um and and yeah there has been definitely um uh the threats of AI the the risk of AI have been uh quite quite heavily quite heavily discussed um as I think with good reasons yeah well and you know I mean recently we had a show about this and I was uh Hillary you know quite a bit in the comments because my my role as uh Devil's Advocate but I mean I I mean for me personally I see the damage of of AI happening right now I mean you don't really have when you have kind of um let's say social media algorithms that have been highly engineered and optimized and no small part by by Machine learning to suck up everybody's attention you know it's even before well before we get to the the possible point of of super intelligence you know in in a general sense um uh it's already damaging right and so and and I personally think the path forward are openness and transparency and making sure that that the good guys um uh that it's easy for them to do the right thing and so I think I think like approaches like what you're advocating for the right way to go I mean for sure going back in a bottle it isn't and um I think you know you you said all the right words transparency I think is is more than just um transparency in the decision making of stakeholders and parties involved in the research and development of AI Technologies in our case it really means the transparency in addition it means the transparency of the models that we're using right uh so you know right one approach one approach to training up AI systems might be uh you know give AI uh access to extremely curated data sets so that it learns only the right things you know for in in extremely like controlled settings um there's reason to think that that won't generalize easily another thing that you can do is to equip your system with the capacity to form inferences about its its own inferences uh and to evaluate itself with respect back to you know things that we value so you could it's just design an AI that had an explicit notion of like discriminatory bias and then train it to identify discriminatory bias in data sets for example and you know uh you can use active inference Technologies to allow the system to access and report on its own inferences and I mean it's it's even more General than that I think I'm understanding correctly because for example you have these nodes right in this in this hyperspace you have the you know you have you have all these nodes in there and and you can learn for example that say a particular node uh is is racist you know it like it'll it'll give you great answers to a particular question so long as it doesn't think you have certain you know demographic characteristics or whatever and then it gives you like you know bad answers well then you can the network can learn how to mitigate against that it can learn how to you know compensate for the biases like inherent in in nodes so not just in particular actual algorithms yeah exactly and you know the algorithms that we're using are based on explicitly labeled generative models so these are systems that can be audited by Third parties right right right because everything is labeled explicitly right so like you you can really calculate the incidence of this or that node on this or that part of the inference and it gives you a kind of tractable uh interpretable uh you know auditable method of constructing systems uh such that you understand what went into the decision-making process I said earlier that my suspicion our wager is that active inference will be to the 2020 is what uh reinforcement learning was to the 2010s my my gut tells me that if legislation the legislation that they are uh you know writing up in the U.S and in the EU goes through uh it may be that active inference will be the only set of AI technologies that we're allowed to use uh in the sense that um so you know neural Nets as they're used right now are black boxes they're not explicitly labeled and they are built to be black boxes like the the they are not built they're not designed to be interpretable the um the kind of uh you know these kind of uh privacy security uh issues issues around confabulation issues around you know the the uninterpretability of these models these are not like bugs in some sense they are features of the approach that we're using to design the systems uh you're not using an explicitly labeled model there is no way to render this tractable post-hoc uh whereas if you start from an an approach that you know it's it's explainable it does what it says on the tin then you get around these these uh these issues uh through your choice of architecture in effect um so you know I I think there's you know tremendous ethical import to the manner in which we're designing these systems we care a lot about ethics and verses as we discussed my PhD even though most of my Publications are in computational uh neuroscience and kind of theoretical biophysics you might say like my PhD is in philosophy uh we have a lot of properly trained ethicists really at the core of this team and we take these things really seriously uh adverses uh so there's no accident here uh and I really think that active inference uh plus the standards-based approach uh is how you're going to get something like responsible scalable ethical AI okay so and so far I'm on board with everything everything I've heard but I have a question for you though which is if it's all open and and ethical and you're trying to do the right thing how exactly is that profitable like what you know what's going to keep the the lights on it uh versus well so we have our own in-house implementation of hsml um so you know we uh we are providing the standards so that anyone can build a version of hsml We're providing the kind of core infrastructure uh for uh you know domain registry and this kind of stuff uh but we also have you know a very Advanced highly engineered and developed versions of these things that can actually do things um so uh yeah uh yeah yeah I mean uh you can think of uh you know red hat and the Linux Foundation as a kind of similar aspirational model right so uh you know Red Hat do generate a profit from a commercial point of view that Linux is still open source the Linux Foundation is the open source custodian of the Linux operating system uh and yet they are able to operate our contention is the our stock is organized in a similar manner where the spatial web Foundation is the custodian of these Open Standards that we are uh you know Distributing hoping uh everyone will widely adopt them and we are the more kind of hard-nosed um kind of Architects who know how to build things using these tools who for having built them know how they they work uh and you know we are probably at this point I mean almost certainly the world's Premier active inference research and development group so uh there's a whole Powerhouse of you know there's a lot of great academic papers that that uh that come out there I mean you know I know Tim and I have enjoyed looking at uh you know quite a few of them um well thank you yeah you mentioned uh Dalton you know earlier I mean you know there there's an example of some some very refined and and quite deep philosophically and mathematically you know papers right well we have basically hoovered up the kind of core uh luminaries of the active inference tradition as it stands currently I mean Carl Tristan himself is our chief scientist and is joining us uh in an increasing capacity over the next few years uh yeah uh so when you combine that with you know I I I'm fairly uh well known in in the field uh yeah we have really scooped up like uh you know the Lance DeCosta Conor Heinz Brennan Klein Mao alberison there's a uh it's it's it's pretty um sometimes it's a little bit uh I guess the Lunchables must be must be interesting though right yeah absolutely uh yeah and I mean um this was this was my dream uh you know uh back back in Academia to uh have a centralized research group with with all of this Talent able to like work together and yeah uh we're definitely doing some interesting stuff and you know at versus we are committed to uh continuously also contributing to uh the public domain and open scientific uh publication uh as you said you know we're we're a fairly productive research group uh we published a few dozen papers last year for example uh you know so we uh I think there's a way of striking the balance between uh contributing to an open community of development having an open core and this kind of thing on the one hand um and also being able to continue existing as a profit generating entity on the other uh but really the I think the key strategy is this open core right so like have the standards open yeah make sure that everyone can contribute to it like uh you know there is a kind of selfish Dimension to it as well because then we are able to harness the entire power of the into the intellectual International Community uh you know to build this uh stuff out I think that yeah there there are there are certainly some advantages uh we also uh for instance uh maintain and contribute to the pi NDP package uh which is a python package for uh partially observable Markov decision processes that power a lot of the active inference technology uh so we use that as our core um and it's uh it's on an open uh license so anyone can just you know download these packages go to the GitHub uh and use it so we're trying to build these Technologies such that you know everyone can start to use them uh but we definitely have uh you know some key uh differentiators and I think a pretty uh pretty unshakable Market Advantage uh well so so speaking of Market Advantage let me um ask you about this question which is uh as as you know we've had we've had uh Professor first one on Michelle a couple of times we talked about the free energy principle a few times and and there seems to be um a lot of uh what's the right way to say this you know misunderstanding or even negative press you know if you will not you know around the free energy principle right like like just kind of push back against it is either something of of trivial trivial interest or you know a tautology that's that's of no value and we talked about some of this too in our in our intros so like what if any what do you think the biggest misconceptions are about the free energy principle and or active inference that that really acts you know potentially just intellectual barriers to the adoption of the technology and this is your paper I believe is you know the Matt fallacy fallacy right which is this this enduring kind of um difficulty in understanding that A system can can follow these Dynamics okay it can it can it can behave as if it can behave as if it has beliefs right about the world and a model about the world and it can behave in those ways and it's okay to point that out like it's okay to say yes this thing is behaving as if it has beliefs I'm not literally saying that it's like a conscious mind you know that has has beliefs what we're saying is that if it continues to exist it must have have these behaviors so that it doesn't dissipate into equilibrium right yeah I think that that's really maybe the the most important confusion that people have is they think of the free energy Principle as some part of like philosophy or metaphysics but it's not metaphysics it's just physics it's physics physics uh it's mathematical physics in some sense uh so you know uh this isn't really a statement about the way that how how things really are in some kind of deep kind of philosophical sense it's about how we can come to know them given the kind of the kinds of modeling tools that we can deploy uh you know so it there there is this kind of deflationary aspect to the free energy principle uh like it it is a way of writing down canonical models for the Dynamics of systems that we find interesting given our state of knowledge about it uh it's it's not it's not necessarily going to tell you about the ultimate nature of mind or something like that unless you take a super deflationary approach and you think that physics in at the end of the day will be able to tell us everything we need to know about the mind uh so yeah what makes this do you think that um yeah the the the the the physicist in me wants to say yes the philosopher in me wants to say there are still a few issues that we need to sort out uh like uh you know where does Consciousness come from but we're working on it again using the free energy principle uh I think one of the things that makes this difficult is the the free energy principle uh is um ontological so it's about things but it's not metaphysical in the sense that it's not about like these fundamental philosophical principles that tell you about thingness it is a it is a theory of every thing without being a Theory of Everything do you see what I mean um right well I think you hit upon this earlier which is um I don't know if it's the only example I but as far as I know it's the only one I know of an example that directly ties physics to to inference or to you know belief updating like this this is the first example that that I know of so like you just said you know it is it is a physics principle and it just so happens to correspond to Bayesian updating that's right absolutely yeah yeah or approximate Daisy and variational stuff uh which is to say basically the same thing um I'm I'm being a bit cautious here because I don't I for those of you who uh in our audience follow me on Twitter you'll know that I have uh to put it diplomatically some reservations about some of the recent literature that's been published on the free energy principle I think a lot of the issues with the literature is sociological um it's you know and uh it's difficult to talk about this without seeming like that you know a bit like deprecating or negative but like a lot of this work was uh written especially the critical work was written by early Courier researchers uh who did not necessarily have the formal familiarity with the free energy principle that might have been required so I mean look for example um I I heard a lot you know Circa 2017 2018 2019 you know uh people say well you know the free energy principle can't be true because uh some systems maximize uh their entropy right they move towards more in Tropic States uh now from the from the perspective of our conversation that might seem nonsensical because we've just spent like you know about an hour and a half talking about how the free energy principle is a way to write down maximum entropy uh but um yeah the the the free energy principle says something very specific right it says that if I maximize the entropy of my beliefs then I can keep the thermodynamic entropy of my physical States at Bay but these kind of sophisticated kind of hinge statements are not necessarily fully appreciated so it can lead some people to right right just say false things about the free energy principle um well you just you just made another statement that I don't know that it's hinge but it requires paying careful attention you said something to the effect of you know the free energy principle is a theory of all things but it's not a theory of everything that's right and I think and the way I interpreted you there was to say that like the free energy principle applies to all things but it doesn't necessarily tell you everything about all things that's right right is that is that what you meant yeah and um you know like uh I think that there there are some states and Affairs uh uh that are just not that are not directly free energy principle adjacent sure so I still don't know why you know the so-called hard problem of Consciousness right why does red feel like red why does a middle C sound like a middle C but it also seems to act as a kind of lightning rod that attracts you know multi-disciplinary criticism let's let's say it that that way and in fact um you know so we we over the course of kind of studying up on the free energy principle you know we've we've read critiques right in the of it you know so for example um those by Beale and others and I'm wondering like what you think about the criticisms of it you know do you define validity and then do you find do you find the quality of the criticisms to be good has has the fep just moved on from it like what's what's kind of the state of the art if you will of of criticism of the of the fep well the first thing to say is that we appreciate uh the critical engagement uh that the free energy principle has received uh and you know like any good scientific framework uh it it stands to benefit from serious uh adversarial engagement um uh I think the quality of criticisms varies quite widely in the literature so to take the example of uh you know Beale and colleagues uh the P I think the paper you're referring to is a technical critique of some parts of the free energy principle yeah uh yeah it was a very I think important paper when it came out it pointed out some of the inconsistencies uh in the way that the free energy principle had been formalized Circa 2012-2013 or so um and I mean since then uh the mathematics has been corrected um and I think we've moved uh beyond the criticism I I have this uh directly from Martin Beale himself he says you know my paper the the the lesson to draw from this paper is that you should incite uh friston's 2013 paper Life as we know it to make claims about the free energy principle which is fair um on the flip side I would say that we have moved uh beyond the formulation as it was as it stood in 2013. one thing to keep in mind is that uh so the fep is sort of like brain physics or mind physics depending on how you want to think about it and uh physics and Mathematics have a strained relationship that I think is important to think about so um the history of uh you know developments in physics often uh goes as follows a physicist um you know borrows some tools from mathematicians more or less uh you know rigorously uh applies the tool to explain a bunch of interesting phenomena uh but then that leaves mathematicians wanting um so for example you know the Dirac Delta function uh was introduced in the context of quantum mechanics uh and the the Delta function is this weird probability function uh that concentrates all of the probability mass under one outcome uh so you got like basically a probability of one for one outcome and then zero everywhere else when Dirac introduced uh this measure into the literature statisticians were not pleased it just didn't seem to them to be a well-behaved object and it took a couple of Decades of work in mathematics and statistics to make sense of and kind of tame uh you know the direct measure um and uh yeah you often get this you can think of a lot of res the history of recent theoretical physics as this kind of back and forth between like sloppy uh mathematical physics that gets then tightened by uh some rigorous mathematical work and I I think you know we're in a kind of similar back and forth here um where you know the fep was effectively developed as a kind of brain physics or math physics and uh what we are witnessing now is an attempt to uh red arrive all of the core theorems in terms of more well-established mathematics and it effectively recapturing all of the core intuitions uh but within a kind of mathematical receptacle that uh you know passes the uh mathematical smell test as it were um what's interesting too about uh the the direct Delta function and correct me if I'm wrong here but I think um I don't know if it was the first but it definitely helped to push along what later became known as generalized functions right so so at least spurred some significant you know work and research in mathematics right oh absolutely and the fep is similar in that you know we're now working out uh you know some some cool stuff uh having to do with you know generalized coordinates um and uh things like maximum caliber which are which is like maximum entropy but over paths and all of this investigation is in effect opened up by the free energy principle uh and by the can of worms so I've never heard I've never heard of Maximum caliber but let me see if I can guess what that is so if you have paths going through you know some State space I guess it's like the the you know the uh the cross section the cross-sectional area the paths that they Traverse or what it's it's more like um well it's similar to that that you're what you're basically doing is uh you're considering um the entropy not of individual states but uh the entire paths throughout the system so it's it's actually the entropy of the whole path okay and so yeah uh it's an extension of uh you know James's principle of Maximum entropy uh but in such a way that uh we can talk about like the counterfactual histories of the system uh that is like all of the different paths that it can take through its state space in terms of their entropy and then the principle of Maximum caliber uh is that the the real path is the one that maximizes entropy so it's not just about finding yourself in a in a low entropy configuration it's about finding yourself along a path that has the lowest entropy um yeah and and this uh turns out to be important um because the the free energy principle uh evinces all of these interesting dualities uh to the space that James is describing so in the literature there are roughly speaking two main families of application of the free energy principle uh the so-called density Dynamics formulation uh and the uh path based or path integral formulation in the density Dynamics formulation what you're considering is uh States and how surprising those states are per se so uh when you're trying to talk about that what you do is you appeal to this construct called a generative model and the generative model is basically a joint probability density and what it describes is the relations of dependence between the variables in the flow or dynamics of a system and so in the density Dynamics formulation uh what what the surprises about out like I was saying is how implausible is some configuration of states of a system so this is different from the uh from the path-based formulation in the path-based formulation you're considering the trajectories of system over time and given the kind of thing that you are for example uh then you're going to have an inertial path through your system just given the kind of thing that you are you know I'm the kind of thing that wakes up at 6am and has some coffee and then gets progressively more tired and then goes to bed at like nine or ten and then you know me so well so if you're that kind of thing then there is a characteristic or inertial path that you'll take and in the path integral formulation the surprisal scores the deviation from the inertial path so uh yeah these are slightly different objects and you can think about uh dualizing these to an entropy context where you know in in the density Dynamics formulation you would be talking about the entropy of states or configuration of states or indeed of the beliefs encoded by those States or the entropy of entire paths so this caliber notion um but yeah all these are kind of uh joined up and as I'm sure we'll discuss a bit later uh yeah yeah the uh yeah the the free energy principle turns out to be a way of writing down the principle of uh of Maximum entropy uh yeah so all these things are really uh like deeply connected I would say now uh you know previously I kind of prodded you about whether or not the physics uh could explain you know Consciousness and you said the physicist in you was really really uh you know leaning towards hoping it hoping it could so I want to take this chance while we have you to ask you about one of our favorite perennial topics on the show here which is emergence you know weak emergence um strong emergence now um you know there's there's different ways of looking at it different definitions of it I think um what I want to ask you though is that there are clearly certain behaviors right that are best modeled and talked about and described mathematically at these higher levels of um of abstraction like thermodynamics okay I mean you know just there are these bizarre kind of properties of you know entropy and temperature and free energy and not somewhat related to the free energy we've been talking about that you can that you can develop these equations on right the first law second law third law of thermodynamics and they apply to these kind of bulk systems now like uh in principle in principle and that's really the the key word there which is in principle if you could write down all the you know details excruciating details of every particle and solve you know wave equations and kind of massive massive Dimensions you may be able to uh formulate those laws right and predict the emergence of those laws but the fact is nobody ever has like they they don't do that we can't do that and and there's even the possibility that in mathematics like mathematically these effective theories right where you try to do that you try to write down all the particles calculate integrals and averages they can sometimes end up with these singularities where you can't cross you know that divide so I guess my question to you is is this which is that and if you couldn't cross a divide that would be a strongly emergent phenomenon versus a weakly emergent one where you could mathematically cross right so I'm just curious if you think um first of all if you think there's such a thing as strongly emergent phenomenon or if we're always be able to cross that mathematical divide if you will and then secondly even if we could is there really any point or is it better just to develop you know descriptions at different levels and and just be happy with that at the end of the day that's a really great insightful question um I would be inclined to say that the free energy principle gives you a kind of heuristic or a map to tame weak immersions so maybe we could start there for a second so the free energy principle applies in a kind of uh multi-scale manner to uh things composed of things composed of things and the kind of key Insight um that you get from the study of Ensemble Dynamics in the free energy principle is that things can engage in emergent Behavior provided that they have a shared generative model so if if you and I kind of encode the same dependency structures then we are going to react in characteristically coherent ways to whatever we're experiencing and therefore we will end up coordinated even if we're not directly communicating so there is a story that you can tell and I know you've had Mike Levin on the call here on the show here a few times uh yeah he's he's absolutely amazing so this kind of like um yeah Ensemble behavior that Mike is interested in explaining has to do with the emergence of a shared generative model um so this again is the same core story you know you have things interacting over time and the the fep says that if there are boundaries in the system between things then they will track each other across the boundaries right so this means that over time things end up sharing a generative model and can then begin to display uh coordinated shared patterns of Behavior Uh so I think that's really key um the you know and we we discussed um Aristotle a bit earlier this is again removing degrees of freedom from the parts right if you and I are aligned on what this or that means it means that we're aligned on with this or that doesn't mean so we are you know coordinating by uh becoming models of each other becoming good predictors of each other kind of sparsifying the set of things that we expect each other to do and over time we can arrive that some form of like you know coordinated behavior and I think the the contention is that this explains you know biophysical emergence at every scale where we observe it um so there there's an argument to be made that really this is really a theory like a a formal theory of uh well I mean the the fep itself is not a theory as we discussed it's a principle but it's a formal approach to that begins to give you a grip on how to model nested systems of systems of systems of systems uh the cool thing is that the whole stack operates on the same quantity basically so you have one objective function it's the same pattern at every scale it's the Markov blanket that repeats and it's this free energy minimization behavior um you know and uh so that's I think powerful you know this segues one of the applications of the free energy principle uh adverses is to design systems that are able to perform inference at different scales using the same uh objective function so if the variational free energy and if you introduce this time scale separation into the mix then you know at the bottom of the stack you have state inference right so uh I have hypotheses about what might be causing my data I am generating uh data through my actions and I I end up selecting the hypothesis that Accords the best uh with the data that I'm generating but you can do your parameter learning using exactly the same architecture just on a slower time scale right so you accumulate counts of State estimation and then over time you're able to uh you know estimate what the best configuration and value of your parameters are for your model and then similarly you could do the exact same exercise at the level of the entire model's structure and then you know that's where you get into uh the natural selection story that the free energy principle brings to the table I embody a model in my existence I am kind of generating evidence for that model good models persist and leave copies of themselves bad models are destroyed and dissipate uh so you can really tell a kind of I think powerful multi-scale story uh so yeah I would say like there's not much that remains in the weak emergence stuff that you can't you know address an attractable manner using this kind of approach uh your other question about strong emergence is a bit more vexed um you know thinking about this I think the only thing that might really be strongly emergent is our conscious experience um and there are reasons to think that that actually isn't strongly emergent but just weekly emergent um in particular are you familiar with Andy Clark's notion of basing qualia um not not quite what is it it's very cool cars he calls this The Meta hard problem oh okay yeah and what what Andy is basically saying is that well uh what we experience is qualitative Sensations are just further inferences namely inferences that I am feeling this right right well that's that's very similar you know I I really liked um Professor friston's take on this and I'm I'm blanking on on the uh on the article where he described this but but he said that you know when you when you're doing this free energy principle and you're doing this this inference once you're once your inference has reached a sufficient temporal depth and counter factual depth if you will then that's when Consciousness can emerge because it can start to model itself and its own internal States and you know this sort of thing I think it makes a lot of sense um absolutely yeah yeah well you know you're you're preaching to the choir here obviously I find that kind of an explanation extremely compelling um yeah and so you know we also have some interesting in my view uh work on Consciousness that we have uh you know in a free energy principle adjacent uh kind of area I've been putting out for you know nearly two decades but more recently we've been looking at the question whether you can uh derive a theory of Consciousness directly from the free energy principle um and so we believe that we can and the idea there is that Consciousness corresponds to something like an inner Markov blanket so an inner screen um so to back up just a little bit um oh oh you said inner screen now yeah I think of the homunculus well so uh that's why I want to qualify this carefully um you can think of any Markov blanket as a screen of sorts so it is um a screen and you mean a screen in like a mathematical sense yeah exactly like in the sense of the uh holographic principle uh in physics right so the holographic principle is a principle uh that originated in black hole thermodynamics and it relates to uh our ability to encode information uh within the system and what it basically says is that you can from the perspective of an outside Observer a given system can only contain as much observable information as you can fit onto its boundary um the reason being that if that if that bulk collapsed into a black hole uh and there were more information than that uh then you classical information would be destroyed in the process yeah exactly and you would violate basically the principle of unitarity in quantum mechanics so uh I you know I mentioned this Quantum information theoretic formulation of the free energy principle that I said I wouldn't get into but it's probably worth getting into for like two minutes um so according to this formulation you can think of the Markov blanket as a kind of holographic screen that separates the inside of the system from its environment and it's a slightly more kind of computational take on what we've been talking about so uh you know the active and sensory states of the Marco blanket would correspond to uh reading and writing onto the screen uh or a measurement and preparation operations in the kind of quantum mechanics uh perspective okay so the interesting thing about the Marco blanket in this formulation is that by definition um because it is constructed or contains all of the degrees of freedom that a couple uh you know systems across the boundary all of the classical information that you need to describe the coupling lives on the boundary in some so uh Chris Fields about three years ago along with Mike Levin and Jim glazebrook uh proposed that well maybe uh the core architectural feature of systems that have Consciousness uh is an internal Markov blanket right the idea being the the classical information that I use and bring to bear in parsing my you know perceptual streams and deciding how to act has to live somewhere and so the idea is it probably lives in this kind of internal uh screen or Markov blanket interestingly uh there's a way of just taking the stuff that we were saying about you know levels in the uh brain and just directly Translating that into the Markov blanket talk between any two levels of brain architecture there is a Markov blanket by definition whereby these messages are being passed so it you can basically according to the model that we're proposing so you have an external Mark all blanket right that separates the uh the internal the inside of the organism from its outside and then if you look at the these internal States they have this structure of a nested hologram in some sense right where like you have a series of screens that are successively coarse graining each other and that are therefore kind of resonating with each other and so there's a whole story that you can tell about yeah the the emergence of something like uh uh yeah uh Consciousness uh just directly from by appealing to the free energy principle the kind of key to this architecture is that at the very top of this nested hierarchy there is a write-only layer that doesn't get further contextualized from by anything else right and so there's a sense in which that layer is constrained to write down and can only uh kind of infer itself into existence vicariously by acting on the other layers of the network um and so the idea then is that this is where the uh kind of um well this is how you resolve the homunculus Paradox first of all is that there is a layer that is not observing itself it just exists to observe right layers below um yeah and so yeah we have a new paper that was just published it's in a series of two or three papers that were articulating um so I say this because like I think if if anything is a candidate for strong emergence the only really serious one is consciousness and the the absolutely utopian over uh confident part of me thinks that like by uh combining you know some of the philosophical work that say Andy Clark has done around Bayesian qualia with these computational architectures for conscious systems that we're developing based on the free energy principle somewhere in the vicinity of this if you if you you know if you look at it for long enough there'll be something that'll emerge to resolve this issue so well well I guess uh David Chalmers will at least be happy that um you've identified Consciousness as the only if there is something strongly emerging it's only Consciousness I can't think of anything else frankly um and I mean Consciousness May ultimately be inexplicable uh but it also may not and so you know I I'm uh an optimist until I'm I'm proven wrong uh and we're going to keep working on this uh I think you know what we're proposing is just the the very kind of basics of a sketch of what a mechanism for the generation of Consciousness might look like we're very very very very far from like a comprehensive question but uh yeah so um oh yeah go ahead well as I say we weak emergence definitely a cool idea I think the fep gives you some tools to handle that strong emergence I I'm not sure uh like if if anything is it's definitely Consciousness uh but maybe not okay well fair enough and you know I was going to say what's funny about the free energy principles it's this onion that just keeps on giving you can just keep pulling back more and more layers and I think we probably have uh I don't know maybe centuries of mathematics you know uh inspired by it most likely well yeah I I mean I I think I think it is one of the singular most important discoveries like uh of the last maybe 200 years like I think you know the you you've had long discussions with Carl so uh yeah yeah I mean I I I think this is like you know the this is physics of mind you know we are we are I think this is the same you know to bring it back to mechanics um so Galileo uh basically destroyed uh you know it's the Galilean moment right but like Galileo destroys thousands of years of philosophy inherited from the Greeks right the Greeks thought that there were basically two kinds of stuff there was the sub lunar stuff and the super lunar stuff and then the sub lunar stuff was governed according to these four elements where you know like uh uh things uh the the uh that were uh Earth uh fell down and things that were fire Rose and stuff like that right uh and the super lunar uh was all about like Eternal perfect circular this cyclical circular motions of like planets and so on right and you know you can if you were around 400 BC and you you looked up into the sky it really would look as if there were two different kinds of things two utterly distinct you know the the stuff around me you know rocks fall to the ground and uh smoke Rises and and water is cold and so on uh and the stuff out there it just moves in perfect Eternal Spirals and so on um what what Newton and Galileo but I mean maybe in particular uh Newton end up doing is to say well no actually uh it's all one kind of stuff uh like it's it's all you know subject to the same fundamental laws it's it's classical mechanics and the Newtonian kind of Galilean Newtonian moment is sort of like the split where like you know the old way of thinking yeah uh started to show uh signs of being incomplete and a new alternative arose so it kind of unifies all of reality under the auspices of classical mechanics and I think the same kind of thing is at play here where you know we used to think that there was physics and then there's biology or maybe you know there's information Theory yeah exactly or maybe there's like physics and biology and then there's the mind but what the fep is ultimately telling us and I I you you've heard this you know from from Mike and Carl before I'm sure is that there really isn't a distinction to make here like it's it's all just you know physics in some sense and biology is just slow physics and psychology is just even slower physics and culture is just even slower physics and so yeah I think we have the same kind of like radical Quantum Leap in the way that we uh you know are able to think about the world that opens up with the free energy principle especially when you combine this to like the philosophy of sparseness and emptiness that I was referring to a bit earlier like I think what emerges out of this is like a real powerful constellation to to think about the way that intelligence is expressed in physical systems and you know that's also why versus uh has adopted the approach like here is finally we have the physics of intelligence so let's let's build our AI systems on these bases I mean it hits exactly the phrase that you said that it's so impactful and meaningful which is the whole is not greater than the sum of the parts exactly it's radically less than the sun in the park this is beautiful well listen Maxwell it has been absolutely a pleasure to have you on so thank you so much for uh taking the time to join us and um and talk and you know I hope you come back on again seems like there's so much to discuss I'd love to and you know I I am a long-standing fan of your podcast So This was oh thank you so much no really uh genuinely this was uh a pleasure and also an honor uh so thank you very much for having me uh this discussion was extremely exciting and fun and yeah let's do this again sometime I I I'd love to come back absolutely and best of luck at verses I'm making supportive of what you're doing there very glad to hear it

Channel: Machine Learning Street Talk

Views: 9,361

Keywords:

Id: 8qb28P7ksyE

Length: 125min 51sec (7551 seconds)

Published: Sat Jul 15 2023