 Hello and welcome. This is Active Inference guest stream, 68.1. It's January 23, 2024. And we're live with Darius Parvisi Wayne. It's not a podcast. I am confused, but it will be fun. Yes. Thank you, Darius, for coming on. Feel free to introduce yourself and the work. Looking forward to the presentation and discussion. Thank you, Daniel. It's a bit odd of an experience to be on this side of the camera and also for it to be live because I'm used to having the benefits of post-talk editing, although I don't do very much of that. So I'm Darius Parvisi Wayne. I finished my Masters at UCL last September. I'm currently working at Royal Holloway in a cognitive neuroscience lab. In my own private work, I work in Active Inference. You may have seen me on the Active Inference Insights podcast that I do with the Institute. And I've recently written this paper called Distrusting the Policy, How Inference Over Action Shapes Our Experience of Temporality in Flow States and Life More Broadly. This is kind of a follow-up from a paper that was submitted earlier last year with Lars Sambersmith, Riddy Pitlea, Jakub Lewinowski, Miles Toft and Carl Friston. And this is a follow-up piece to that. Before I start, I'm aware that it's a punchy title and I just want to clarify that I am not proposing to explain temporality or the experience of time entirely because it's an extremely rich area of study in which physics and philosophy and cognitive science converge and I do not have the time, the effort or the energy to really tackle all of those problems. So it's going to be rather specific, but I'll make clear where I'm purposefully missing stuff out. Okay, excellent. So there's kind of a central question before we even get going, which is what is time? And I want to clarify that I'm not talking really about so-called objective time. Now, I know that's itself a tendentious and tricky claim to make because of relativity, but we can have a kind of simple definition given to us by some wonderful researchers, Bogotay and Jabara, who have also worked on active inference accounts of subjective temporality. They define objective time as the temporally sequential development of worldly processes bound relatively to space, and so that will incorporate space time as a concept in physics. So it's spatial in the sense that it follows spatial laws and it's indivisible. And then it's directional. Now, this is a little bit more controversial, but this ties time to thermodynamic entropy, which stays the same and goes up in closed systems, second law of thermodynamics, and this is the so-called thermal time hypothesis. Carly Rebelli, Lee Smollin, those are the sort of researchers to go to. As I said at the beginning, there are obviously diverse accounts of time, especially in philosophy of time and the physics of time. So three main contenders are presentism, which states that only the present exists. The growing block universe, which states that the past and the present exist and the future is yet to exist, and the eternalism, which says that basically they're all happening now. So I believe that there are dinosaurs is still true. I don't really want to delve into those. If people do, I would start with the work of someone like John McTaggart, who was very instrumental in developing the philosophy of time. And of course, there's a lot of work being done by Carly Rebelli, among others, on the physics of time. But what I'm more interested in today is subjective time. Time is experienced. Time is a phenomenon in the sense of being phenomenological. Now, it seems to me when I started writing this paper that everyone had made the assumption that it flows, that there's what's called a passage phenomenology. And my account is predicated on the idea that time flows. And I said there is a question mark after that because it turns out that there's a fair dose of skepticism about this notion that time flows even within phenomenal states. So this could be considered skepticism over the passage phenomenology rather than objective passage. And these are people like Calender, Deng and Miller. So for example, Miller thinks that actually it's our linguistic structures which give us that impression. Deng thinks it is just successive temporal events without this notion of flow or passage. This paper and this talk, I guess, can act as a repost to that. Again, I'm not going to delve too much into that argument. I would just like to take it, at least for this talk, that it needs to be treated as some sort of axiom that it's flowing. And then the other thing is that it's thick. And that's a slightly strange adjective for me to use here, especially for people who aren't familiar with the cognitive science of time. What does it mean for temporality to be thick? Well, that leads us very nicely onto this man. This is Edmund Hussle, who was the godfather, the progenitor of phenomenology. He taught Heidegger, and so the lineage really starts with him. And I also have to mention William James because he's very important in this. And William James actually, he came up with the notion of what's called a specious present and what we could call a thick present. And this is a notion that, well, when we divorce subjective time and subjective time, we can cast objective time as following this kind of linear sequential order. So going back to the Bogelta and Jabara definition, we have A, then B, then C. Now Hussle, in his examinations of subjective experience and phenomenality, realized or recognized or postulated that the present moment is not just this this divisible unit of time that can be separated from the past and the present. Rather, what makes up the present moment is itself a constitution of what he called the primal impression, which we can argue is this, or we're going to talk about that in a sec, this very narrow slice of now-ness. But also what has just passed and what is about to become. So for example, this is a diagram, it's an adaptation from one of Hussle's diagrams because Hussle's also a mathematician. But as we can see here, let's take this time point B. So this is objective time point B. What Hussle is saying here is that B is also constituted by what's called a retention. Sorry. Yes, a retention of what's just happened. Sorry, we're looking at this line here. So retention of what just happens. A, so a retention of what's happened A, but see how here it's in brackets. It's a part of B and a sort of what equals a retention. What we could say an active influence is a prediction of what's about to happen. And so this would be C. Now C in terms of objective time hasn't actually happened yet. But the sort of expectations which are to be fulfilled or not fulfilled constitute the experience of B. And so do the residual memory in some sense of A. Now Hussle makes it very clear that this account of temporality is very different from explicit recollection or expectation or imagining. Rather, this structure is the very structure that defines time consciousness. And so we can kind of see in some sense how flow already comes out of this insofar as pretensions are fulfilled or unfulfilled and therefore become retentions which are, which received back in time but also have this predictive element i.e. they are pretensing what's going to happen. And I think to get a little bit more clarity on this picture it's worth just discussing this notion of a primal impression. I think the primal impression is from my reading and I highly recommend this Gallagher and Zaharvi paper because I think it's incredibly illuminating. The primal impression, most people take as real and it's constituting what Gallagher and Zaharvi call the component of consciousness that is narrowly directed towards the now phase of the object. Okay, so it's a little bit vague what that means. And actually I think it's more parsimonious and slightly more elegant to say that it's an abstraction. And so this is actually kind of in line with Derrida's reading of Hirstle where he distinguished the theoretical construct of the primal impression which is useful for us in modelling and in an instrumental way from the actual thing that appears in the phenomenological presence. I don't know what that is always horrible to say. And so my argument here in some way is that the nowness, what now is is in a sense already anticipated. It's already anticipated because it's being pretended and that pretension is in a sense a retention already because it's projecting into the next step and it's already predictive of what's going to come at t plus one. So we can treat the present moment itself as the result of prediction and then the fact, prediction about the next step and then the fact that that prediction is going to be fulfilled or unfulfilled. Now, this, before I get on to the flow of time, I recognise that this is complicated and it requires some non-linear ways of thinking. I highly recommend Bogota and Jibar's 2023 paper where they actually give a very nice base graph and active inference account of this exact way of combining objective time and subjective time in a Hustlian way and I think that should clarify exactly what I mean here. Again, I'm not actually too interested in deriving a perfect structure in an active inference sense, a perfect model of Hustl's time. I think that's been done very well by Bogota and Jibar and also people like Malal, Barasin and Herty. Well, I'm more interested in the flow of time per se. What's the difference there? Well, this is the structure of the present as is, as time consciousness as is. I'm more interested in the fact that time seems to flow, that the present seems to flow into the past and the future seems to flow into the present and that itself is a bit more of a diachronic process. So that takes me to the accounts of time's flow. Now, whereas the structure of time consciousness, as I said, has been explored quite comprehensively within the active inference community and beyond, the notion of time's flow perhaps because some are sceptical that time does flow has itself maybe been touched upon a little bit less. The most famous or the most... Yeah, I guess the most concrete proposal comes from Yaakov Howey and his colleagues, and they argue that the flow of time arises as the system gives up one hypothesis and settles on a new one due to its propensity for distrusting the present. So this is a rather famous 2015 paper called Distrusting the Present. Here is that every hypothesis, every state hypothesis I have about my observation is going to be really good for that observation, but immediately it's going to become less and less good because the world changes. And so because I have a higher order expectation that the world is going to change, it's actually adaptive. It's in my best interest as a free energy minimizing system to start to distrust this hypothesis because it's not going to be capable of explaining the sensations, the motivations at t plus one. So this is put very nicely by them. As the hierarchical system settles on this hypothesis, which for them is best at minimizing prediction error, this is continuous state space stuff, so that's totally fine, it needs to begin looking for a new and better hypothesis. And again, that's contingent on this higher order expectation that the world will change. Crucially, and I will explain this in more depth, that I think Jacob is proposing here is perceptual and attentional. And we would cast that as the A matrix within POMDP schemes in active inference. And that really casts or encodes the probability distribution of the likelihood. So that is, what is the likelihood of this state given my observations? What is the likelihood of this observation given my state, the posterior is x given y. And the idea here is that the observation at time t is yielding precise likelihoods that are going to govern a hypothesis. But because the potential spot is always moving around and the world is changing, that observation itself is going to lose precision. And so the new observations which yield different hypothesis are going to gain precision, and therefore the system needs to update. But I think most crucially is this notion that it's perceptual and it's about state or as he calls it, a hypothesis updating, a state hypothesis updating. And crucially for us, he also, Jacob and his colleagues mentioned also the speed of the flow of time. So we're now adding another dimension to this phenomenology. And I won't read out the whole thing because that's a bit dull. But basically he's saying that the faster or the more frequently one updates one's hypothesis about the world, the faster time seems to flow. So for example, he gives the example of a war zone. Well, in a war zone, and I think this is totally reasonable, the perceptual hypothesizing has to keep updating because it's volatile. So one moment I might be perceiving x and the next moment I might be perceiving y, but that's the frequency of that is happening a lot faster. In a different scenario such as waiting in a dull airport, the idea would be that I don't have to, I don't expect the world to be that volatile because maybe I'm just staring at a wall and so I don't have to do this perceptual updating or this hypothesis updating as Jacob describes it. Now, what's the problem? This is nine years old and I haven't really seen anyone say that there's a problem. But I thought there might be a problem and my problem in the sense is the fact that this account is fundamentally perceptual and so there are three situations I want to bring up here and then I'll talk about how they pertain to the question at hand. One is what I call static predictability. So this would be something like staring at a wall for a prolonged period of time, so that's our middle picture here. If agree that time would pass rather slowly there. Now, the second one is so-called dynamic predictability. Now, these are my terms, I'm very happy to be able to critique them, but I'm calling this dynamic predictability. Now, this would be the traps on the right. These are those people who may have played the video game level and enjoyed it the first time, but for some reason some torturous devil is making them play that video game level over and over and over again. Okay, so that's dynamic predictability. I'm not going to talk about how it pertains right now to the flow of time. And then the other one is flow, flow states. And again, I would point people in direction of the paper that I wrote with my great colleagues last year on flow states, but this is slightly different in the sense that the volatility of the world is inferred. There's a belief to be a sufficient level of volatility such that I have these narrow action policies, these temporally short action policies and this adaptive, these fast flows of motor responses to the world. So that's flow states. And you'll get that canonically in context such as surfing, rock climbing and so on. So this is the work of Success Mahali and others, John Devaki, and hopefully, we've contributed something as well. Let's just very quickly touch on the speed of time in flow. This is a nice study by Rutrex and colleagues in 2021. And they found that the more flow participants experienced the faster time pass subjectively. However, flow did not correlate with duration estimates. Now that might seem like kind of the same thing. So how are they dissociable? Well, I think the easiest way to think about this is what questions would you ask to discover these facts? You will ask for the first one, did time, how quickly did time pass subjectively for you? And then the second question would be, how long do you think that took? And interestingly, in flow, we get this sense that time is speeding up, but we don't actually, we actually have quite accurate estimations of duration. So for me, this shows that these phenomena are dissociable and motivates my focus on the subjective speed of time in flow because there's something going on. This is not particularly incontrovertible, I don't think, because Success Mahali, it goes right back to qualitative work. The notion of time dilation has been present in the literature for a while, but this is actually one of the first concrete pieces of experimental work that has validated that and showed that in flow, time seems to pass faster. Now, I want to go back to this woman in static predictability staring at the wall. There are some nice studies here which basically show that time does seem to slow down in these contexts, so fine. We both agree on that. What's the problem? Well, remember the model of distrusting the present is fundamentally perceptual, and so you can see here Jacob and his colleagues' account of distrusting the present in terms of perceptual attention and this notion of attentional spotlighting. Again, it'll be much easier for you to read it than for me to explain it, but I think the critical point in here is, to take it back to here, these video gaming boys, they're going to be in context where they're getting still volatile perceptual information in the sense that I might be playing a video game where things are changing the whole time, but I've still played it 100 times, so I'm bored, and I would postulate that I think time is going to move pretty slowly there, and in flow states, we've seen that time seems to move rather quickly, but according to this purely perceptual account, there should be no strong difference between the sense of time in flow and these dynamically predictable situations because in both of them, you could find a situation where there's the same or even more state updating to do in dynamically predictable situations. I think an even stronger example of this I actually put in the paper, so there is a paper on this where my thoughts are going to be a little bit more lucid because obviously this is not scripted, and the example I give there is of the poor person who has to stare at a screen with a dot constantly changing color. There's a lot of perceptual state updating there, and let's say it's randomly changing color, so we can't have an action policy that can pick out the patterns and say you can have these extended action policies that say, okay, it's going to go red, green, yellow, red, green, yellow, so it's constantly changing color. Now, I was hard pressed, there was no empirical work that contrasted flow in dynamically predictable situations, and so this is empirical work that I would like to do, and I'll get onto that, but I would still propose just heuristically and intuitively that time would seem to flow rather slowly. I think another issue of this is how if we go back to his original example of the war zone and the dull airport, even in the context of static predictability, there's always low level state updating that's going on. The world is never static, so even if it's the fact that the paint on the wall is very slowly cracking or there's a fly that passes in front of you, the actual, the world itself never changes and how he doesn't really give an account of exactly why or how that can be incorporated into his model. So this made me think, well, I don't think the perceptual hypothesis is necessarily the most elegant or fits the phenomenology perfectly, and I think it's incredible. I love Jacob. I think it's an incredible piece of work, but I just think there's a little bit more that could be added. So this is where I came in with this notion of distrusting the policy. Okay, so this is a Bayes graph. This is actually from Lars Sandberg Smith's 2021 work. I won't go too into it, although it's actually not too difficult, but basically what is shown here is the fact that in active inference and the references are there, we talk about action policy selection, so policies are the sequences of action. We talk about these in terms of tree searches. So I might have multiple potential policies that I could pursue, and I basically calculate expected free energy, which we can have as epistemic gain and pragmatic gain. If you want, I can talk a little bit more about that, but I don't think we need to get into the theory too much. I talk about, we have these tree searches where the agent evaluates the possible consequences of these different policies in terms of expected free energy. Now, they're not just doing this on a whim. As you can see here at the top of the graph that Lars gives us, Lars and his colleagues, because it was a collaborative paper, the expected free energy is constituted in part by or decided upon in part calculated by the so-called C matrix. The C matrix are the preferences about the sensory outcomes that you have. So if I have very highly precise sensory expectations about how my action is going, the sensory outcomes of my action, that's going to drive, and it's in favour with my generative model, that's going to drive that pragmatic action. Indeed, the pragmatic aspect of the expected free energy formula is the probability of why the observation given C. There is also this E matrix, which is habits. We can take as habits prior to certain policies, and we'll see actually that comes in a little bit more in flow. I think the main idea here is, well, there's a couple of things. One is that an action policy need not just yield, need not just be isomorphic or map onto just one observation. As you can see here, we can have action policies that stretch deeper into the future as long as we presume that the future is predictable, such that we have an N plus one of the so-called B matrix. The B matrix are state transitions. I might have an action policy where I go, I'm going to be in this state, then this state, then this state, and obviously this is informed in real time by the observations I make through the A matrix, through the likelihood matrix. The point here is that the tree search can, like a tree, be extended or narrow. That's one thing. The other thing to notice here is that every time I make an action policy, or every time I decide upon an action policy, these things like the E matrix and the C matrix are implicated. We can blindly select an action. This goes right back to the work of Chris Frith and his colleagues at UCL in the late 90s on actually how action happens and forward modeling. The idea is that beliefs about the consequence of our action are fundamental to even doing action in the first place. That's the idea here in terms of the actual theoretical side of it. What does that mean in terms of the flow of time? Coming back to the flow of time. Recall that in flow, I mentioned it rather briefly, but the idea is that the world is volatile. Although we have a good grasp of it, although we are occupying or embodying what Kivistina and others and Brunerberg and Wrightville are called a optimal grip, we still need to be aware that the world can change very quickly. I might slightly have my bow angle wrong if I'm playing the violin or the wave might not turn exactly how I expect it to. What that means is that deep temporal planning with these deep trees is precluded in flow. What that means is if we take five minutes of flow and we can assume that flow does last that long and there are some nice empirical studies on that, per objective time unit, I'm going to have to be doing more bouts of planning than I would, let's say, in so-called static predictable context and deeply predictable and dynamically predictable contexts. In that case, what we end up reworking is that the flow of time is not about distressing the present hypothesis about the sensory data per se, but within dynamically predictable contexts and in flow states that might be the same, but rather the iterative speed of belief updating about the current action policy. A question that came to my mind, okay, that sounds okay, but we've been talking about perception the whole time and arguably, if we're going to stick to the retention primary impression and retention scheme, we're going to have to talk about perception and indeed you can see Gallagher and Sahavi's sort of pointer that actually hustles analysis is very much characterizing perception and obviously the perception of time itself is a perception. We're talking about the phenomenology of time, the passage time. That's a perceptual act. So how are we going to ground planning and policy selection? How are we going to fold that into this picture? Well, it comes right back to what I said before. And it's the fact that policy selection includes the deployment or the precision of certain beliefs in a generative model including predictions over preferred sensations which is encoded in the C matrix and state transitions which are encoded in the B matrix. So if you look on the right side of the page here this is actually a base graph from my paper, something I did with Lars and you'll see here that if it's rather complicated actually I'll go through the text first and then we'll come back to the base graph. Again, the idea is actually at least in flow, so maybe it's actually worth looking at it. What flow really entails is the contextual queue for mental action. What you see here is it's in the mental state inference box. It's U with the 2 at the top and the 1. That's an action. That's triggered by E2 at the top with the superscript. And that's a habit of driving action but not a physical action per se. What that mental action is doing is driving the A matrix. And so again what that means really for this example or for the broader paper is that mental action is not limited just to the selection of policies but also these precision way to beliefs about sensory outcomes. So what does that mean? If I do iterative policy updating if I do more iterative policy updating per objective unit of time then what I'm also doing is I'm doing more updating about my preferred sensory expectations i.e. my C matrix is getting updated more frequently. That's a critical point here. So for example in flow what you're getting is you're getting this constant you can see here that the E matrix and the C matrix fold into the action policy because this planning would be taken as to be rather shallow you're going to have this process iterate more frequently over a given period of time. Now just to clarify for people who might not have any idea what's going on in this base graph as I said the idea here is that the mental action is deploying it's about deploying precision over these precision beliefs so you have the likelihood mapping, the precision of the transition mapping and the precision of preferences note that basically you can't, the idea here is that you can't necessarily perceive precision so the way that that's getting fed back up to the higher level is through the observations made O subscript to superscript so that's actually stated in the italics but I think, yeah, the point here I'm very happy to go into this, this is sort of deep parametric modeling which has been going on with Lars's work and Kasper's work Kasper Hess and I think it's wonderful and I'd love to talk more about it. In terms of this though I think it's a very simple point that the C matrices whether it's C with the two as a superscript or C with one as the superscript which is really going to be the kind of critical C in flow is the one with the one superscript because that's the one which basically drives the sorry, no, I'm talking about the precise beliefs of the tensor expressions, that's the one that's going to drive the action policy the you so to speak that's just going to get updated more frequently in flow states because of the shallow action policies that flow states and gender so therefore what's the kind of ultimate conclusion here, well my conclusion was that underline the speed of temporal temporal flows since beliefs encoded in the C matrix now this is kind of a critical point are fundamentally predictive i.e. they pretend and they are also retained over a period of time why is that? It's because my sensory expectations are going to govern what I perceive as we know from active inference because the posterior really is just this negotiation of a prior expectation and the likelihood distribution and yet they are also informed by just past experiences i.e. they involve a form of retention because as you saw with the flow that C matrix the sensory expectation about the consequences of my action are fundamentally downstream on this contextual queue so I think there's a nice way here in which we can wrap sensory expectations in coding the C matrix in the Houselian phenomenology again the really critical point here is that this can be independent of the pure perceptual inferences and pure perceptual inferences here are combining about prior beliefs about states before any observations the so-called D matrix and a precision weighted likelihood distribution in the A matrix so that's happening at the very top here of the contextual state inference in these dynamically predictable situations and in flow states we might have the same number of pure perceptual inferences but I would posit that in flow time seems to move faster in predictable than predictably dynamic situations and the reason I think that is is because we should be talking about perceptual inference updating about sensory expectations with respect to the C matrix and not perceptual inference about the combining of D matrix and the A matrix and again this is because the frequency with which we update those beliefs encoded in the C matrix is contingent on the frequency of policy planning so just to conclude in terms of going forward this is again so just to make it very clear it's a pre-print right now this paper I'm not claiming this is the comprehensive account of temporality or the speed of the flow of time but it's a suggestion so I'd love to get feedback from people in terms of going forward I think this would be an interesting thing to test whether we could set up a situation where we have one group while you have three groups you have a control group who aren't doing much a second group who are playing a video game for the hundredth time it's predictably dynamically predictable and then flow states which can also be done through something like a video game and what I would expect to see just at a very basic level is that the flow we could replicate that finding of time moving faster subjectively and confirm what I think is a reasonable intuition that in these dynamically predictable states time seems to move slowly or at least slower relative to flow states okay that's me Daniel feel free to fire away excellent alright three times it flew by yeah well well it's very another part of this I missed file or outlets part of the paper the paper is relatively large is that attention itself is an action and this is often forgotten by people but attention is an action it's a mental action and that itself will have its associated C matrix so even paying attention itself would constitute a mental action one note on that and then live chat I'll read policy selection under expected free energy is shaped by pragmatic value aligning future observations with your preferences and epistemic value information gain so that's very interesting that attention as internal action or covert action could be driven also by pragmatic attentive considerations here's the kind of regime of attention I expect to be in and also there could be epistemic or exploratory regimes of attention that maybe reflect us getting distracted just to gain information even when it seems to disalign us from what we expect ourselves to be doing yeah there are several things that come to mind that one is that we do have these nice meta attentional models given by people like Lars 21 paper which includes this parametric modeling so there is this what am I paying attention to how well am I paying attention and you can have these deep generative hierarchies where you have attention over attention this is a slightly separate venture but I've been thinking about the pragmatic value of attention and I think in some sense there's a really nice paper by Mitzer and colleagues 2019 where they talk about context dependent value driving attention so if I'm looking let's say for I mean I was talking about yesterday with Julian Kibbastien on the podcast if I'm in a crowd and I'm looking for my friend and he's ginger I'm going to up up the precision weighting of like a template of gingerness which means that anything that's kind of ginger whether it's a cat or my friend is going to go up and everything else is going to go relatively down and this comes back to original theories of attention bias competition theory there's no one Duncan but the idea is that even my very attentional schema there is being governed by a higher order pragmatic the pragmatic gain of finding my friend and that can be there could be a lower order element of that which is that there is value in paying attention to these things because it subserves that higher order inference cool all right I'll go to some questions in the chat but anyways welcome to have more okay Steven select asks my question is would you agree that self reflective self awareness is necessary for time perception therefore time requires the self okay so this is this is an interesting question I think in part it can be answered somewhat by the flow paper so in the flow paper it really does focus on self modeling and the idea here is that deep temporal planning in general involves the levels of the generative hierarchy which are propositional and conceptual and those are the very levels are associated with a rich epistemic agent model so this self conception of an agent that can use its knowledge from the past to inform future outcomes and its future action policies now in flow because you have these shallow you have these shallow decision trees you actually attenuate that part of the hierarchy now that means that when you're in flow people talk about the kind of self action merger we're not saying that the self is completely dissolved in flow states because for example the minimal phenomenal self researchers such as Anil Seth or Felix Blankenberg or Jakob Limonowski have grounded that in inter-receptive now I think it's an open question there are some people who think that all experience involves a fundamental sense of mindness so someone like Julian Kibbesteen has written a paper on this there are others who will say that inter-receptive inference is really at the heart of the minimal phenomenal self and if that dissolves then the self model itself might dissolve now whether you get time out of that is interesting question I would point people in the direction of a conversation I had with Carl where basically we had this there's this idea that these different temporal hierarchies are predicated on the actual dimensions of the Markov blanket so how nested the Markov blanket of an organism is because if you're very far away from your actuators you have to start making inferences about what your action is versus what the world's action is so I think there's a nice link there maybe not just physical size but dimensions in terms of embed in Markov blankets and time I don't feel like I'm necessarily qualified to say how these two are necessarily fundamentally connected because I think there's still an open philosophical and physics related question about the reality of time itself awesome cool a little more about trust and distrust it's very spicy to call the paper distrusting the policy because it feels like the interior dynamics of our thought should be like what we can trust so it's kind of a little bit of a stick in the wheel but that's kind of fun so where is their trust and where is their distrust in what in that kind of model I mean this is rift off of Yakov's distrusting the president so I want to clarify that as a master's student you've got to try and make some waves so it was just trying to get people's heads poking up but there is also the idea of distrust and I think you can cast that as in the form of precision waiting to go back to what Yakov was talking about in terms of distrust as understood by philosophers we're talking about subpersonal Bayesian beliefs which include precision waiting and it's just that your precision waiting over your state hypothesis is going to be down weighted over time because there's a higher order belief that the world is volatile and I think that high order belief is really useful to think about when we're talking about what trust is in this respect if you thought that the world wasn't going to be volatile such that for example you're staring at a blank wall then you could very happily trust that your hypothesis about that blank wall is going to be fine in one second two second three second four seconds the idea is that the trust comes from this kind of perhaps habitual inference that in this context I am permitted longer spans of deeper trees in this context however those deeper trees are precluded now in some cases that might be full active inference so there might be a higher order in the deep parametric sense a higher order action policy to deploy shallow action policies or temporally broad action policies in the state of flow we argued that this was habitual but I think the idea here is that for trust to trust or not trust in an action policy is not that we're not saying don't trust action where action is the fundamental currency in active inference we're saying you might want to agents have to be aware of the degree to which the action is going to yield them the outcomes they want not just here and now but also for expected free energy also for the subsequent steps awesome yes that clears it up it's not like policy selection isn't trustworthy it's that there's a variable speed at which policies like go stale that totally makes sense one really interesting and open topic is which phenomena either objective behavioral computational type of phenomena or experiential phenomena do we associate just coarsely more with the A matrix and load more on to the sense making component or more on to the B matrix more into the world transition model or the consequences of action in the world so like were there any phenomena that you saw different authors approach it like sometimes discussing that phenomena in the context of A matrix and sense making other times more like in this action oriented B matrix way honestly I haven't I think because deep parametric I might be missing something but I think because deep parametric modelling is a relatively new aspect of active inference a lot of the literature on this stuff especially in accounting for phenomenology adopt quite a strict or pure predictive coding language so if you read this distressing the present this Jacob and colleagues paper from 2015 there's no mention of the difference between the the D matrix and the B matrix the inferred state prior to observation and the state transitions based on observations which are contingent on action policies rather it's just that the prior is informed by observations and they talk about precision of observations where I would talk about precision of likelihood distributions so I would also they don't touch upon this but I think it's definitely worthwhile me pointing people in the direction of this Bogota and Jabara paper it can be a touch confusing because I think one difficulty that I was having is it's hard to hold both objective time and subjective time in your head at the same time I'm okay with one and then then go to the other but combining the two is a kind of confusing feat and they've done a great job so I would say just answering your question in terms of where did I see people providing different emphasis I would say it's difficult because I haven't seen that many deep parametric models of of time say and so that's one paper that I would point people to but I think it's definitely a definitely really interesting area of work about where we really put all the weight whether it's perceptual or action I just need to think of perceptual pure perceptual was going to do it for the flow of time but at least not positive by Jacob and his team cool well a few things strike me from this kind of presentation first how open and how many ways there are to address time phenomena in active it feels more like a toolbox or a mechanic shop for temporal or dynamical systems rather than giving us a singular answer like right off the bat we have discrete and continuous time now people have made arguments for continuous time also people have made arguments for discrete time and suffice to say that we are using a formalism that straddles both or can flip between both so there's many ways to describe it and another example of like how temporal phenomena can be brought into play differently is let's just say we were looking at a one thousand step time series we can model that as a thousand steps of one or steps of ten or we could do a nested model so that even though it's a one time step model maybe the slowest level is the whole thousand time step so it kind of course grains the long time into kind of compressing it into a perceptual moment even if it abstracted perception like is it possible to perceive the two thousands or is it possible to a decade I thought was a really important part of this was the fact that we need some kind of objective order on which we can like rest our rest our arguments so in the paper and in the talk I really emphasize talking per objective time unit and I know again that's controversial right but like we just take as a kind of miracle of our theory that we have seconds and minutes and hours if we can rest what we're talking about in terms of flows on that kind of standard that we can then apply to different theories then things become at least explicable because without that you're absolutely right it just depends on the level of the hierarchy or how deep you've gone into these time steps but if you can frame them within something that we can all agree on then it becomes relative to that and I think it becomes the science becomes a lot simpler and also while certainly is nice and simplifying to take the chronology or the timeline as an external given there's all this agentically constructed and then we might be looking for something like a gamma oscillation like we're able to distinguish intervals of time seemingly on the order of hundreds of milliseconds but not fractions of milliseconds and there seem to be some physical reasons like why we couldn't differentiate two things that happen within one millisecond for example so it's kind of like that gives us a bottom up granularity or at least atomicity for the time process that a given agent undergoes and for a tree that might be slower like its duty cycle might be slower and yet we could still use the same scale independent formalism to describe it. I wanted to make very clear in this that I am not capable of touching upon every element because this is one of the richest areas of philosophy and physics and I'm not a physicist and suppose I'm not particularly well versed necessarily in the philosophy of time what I wanted to do in this was stay really faithful to the phenomenology so this is informed really primarily by Emman Hirstle but I think you're absolutely right I think an ultimate theory of time if there ever could be such a thing he's going to have to integrate the bridge between the so called objective and the so called subjective and I think that's up for grabs you're absolutely right you can have a model where it's all emergent from the actual agentic being so it makes me think of sort of a Donald Hoffman Donald Hoffman would just say this is a screen this is an interface that's really useful in biological agents but the underlying reality has no time and of course you have physicists this notion of block time or eternalism that everything is now the past, present and future is all real again this is why I think it's important at least for me and other researchers when tackling these huge questions is demarcating exactly what you're going to be talking about and I don't have the skill or the acumen or the time if you pardon the pun the Einstein and the Carla Revelli story chats I would have written a shorter letter if I had different time quite still can't think at all I'll ask a question for live chat Soryus asks can the agent perceive time as the amount of parallel actions taken between events of attention I'll give a first thought on that yeah with what Chris Field shared in the physics information processing we talked a lot about the measurement and the preparation cycle and so in active inference at each time step sometimes we're just talking about like a vertical slice and there's an observation coming in and a hidden state inference it's all described kind of synchronously whereas in the quantum framework there's a bit more of a two stroke phase to the measurements which is that it's prepared and then measured and so if the external oscillator let's just say again give a seconds for now if it's going at one second and an agent we're preparing and measuring 10 times a second well then 9 of those 10 would kind of come back empty like the 9 sequential measurements and propose or preparations and measurements that didn't have any change would be kind of like in a timeless or same time moment they all would have like pinged the clock and gotten the same state back and then conversely an agent could use the rates and the type of change or interaction with the environment as a proxy of change and the example I think of there is like honey bees using the optic flow as part of their distance calculations and people have shown like if you have the honey bee move through a tunnel with very rapid black and white alternation it overestimates the distance it flows and then if you have it in large contiguous color blocks it underestimates the distance it flows so it's using the rates of optic flow change and the time differencing which is also the origins of predictive processing using those kinds of external oscillations or changes as part of its embodied time perception right the way I would approach this is exactly what you said which is when we when we build a base graph something that sort of became very clear to me is that we are taking time we are slicing time which is a detriment to talk about diachronic time whilst slicing my model and I think this really does just exhibit sometimes the discord that we have between our modeling and the actual reality so it's an inherent weakness or an inherent limitation of our instruments what I will say is I think it's interesting to talk about parallel processing rather than serial processing here because as I keep trying to make clear this generative hierarchy is rather deep and so I think for example I mentioned when one stairs a wall there is always some state updating the game is never paused and so I think there is actually an interesting follow up question here which is if we stack all of these regimes of attention up and we have rapid flows in all of them does that lead to an even more pronounced sense that time is flowing faster in the way that maybe now we can have this nice hierarchy of well I could predict a stable a stably predictive situations dynamically predictive situations of flow where the difference really is the actual number of attentional regimes in which you are seeing updates because I don't, there is bear in mind that to update the C matrix you have to know where you are in the first place to be able to do the state updating it's just that for me that was insufficient to explain the phenomena at hand but if you can now start taking that kind of base graph that I showed you know if we can take this base graph and in principle apply it up and down the generative hierarchy well if these narrow flows are happening all the way up and down there is an interesting thing there that this is really the top end of passage phenomenology and if these are really protracted at every level right down to the very lowest Markov blanket of our system and right up to the move rather slowly so I think there are some interesting ideas there and I think it feeds into this notion that we occupy or embody embedded sort of blankets that go beyond us and in many ways go kind of beneath us it depends on where you cast your identity but yeah I think the parallel thing here is really interesting but for instrumental scientific reasons it was just a lot easier for me to like and philosophical reasons just to take one slice at one level of the hierarchy although this does have some parametrization okay so looking at the figure your paper yeah so this one by the way this one's from the first flow paper okay I think this one yeah so this one's from the first flow paper I'm not sure I need to check I'm not sure it makes an appearance in the second one but it's that's where it comes from okay so for any parameter that you know or love on this diagram what would you say is like a prompt or a cue that might help us understand or come even to an intuitive felt sense pick anyone because I think that's like super interesting and it's all awesome to look at this as like the compilation instructions for a base graph which is in one sense what it is technically however also as we're talking about our own minds and experiences here it's like super fun and definitely of interest to tap into whichever letter and then let's see if we can go from there so it's sort of just forewarning this was like constructed with Lars really led on this project I should give him a lot of credit and so I also don't want to butcher his work because I know he's really into this deep parametric modeling that he learned from Carl but it's my paper so I have a good idea of what's going on here I think I could do an okay job or actually explaining what's going on more generally which is the idea is that let's just start right at the top at the top is not particularly it shouldn't be too confusing which is that basically you're just inferring the state that you're in right so what do you mean by state so the hidden state so I'm inferring so this given this observation of whatever has hit your receptors I believe with some high precision that I am in you know I'm about to start a violin concerto okay let's go with the violin violin concerto situation so I'm about to start playing my violin now what does that mean well that feeds into this E matrix this E with the two as the superscript and so what we're saying here is that that very state inference acts as a contextual queue or the deployment of this action you with the superscript 2 and subscript 1 so really this is all basically driven by this E matrix E with the two superscript now what does that look like well what's really important here is this S with the superscript 2 and subscript 2 because what we're saying here is we're posing a mental action and the mental action is to increase precision over the likelihood mapping the likelihood mapping is here the transition mapping and the precision of preferences so what we're saying here is that this action is basically in some sense this is it's kind of preferred state now the only issue as I said is we can't observe precision what we can observe are the outcomes of our action so what we're now saying here is that basically this observation here through this likelihood matrix will feed back up the hierarchy and inform the mental state inference here oh you've yes you've succeeded in increasing precision and the reason for that is as you know if we get increased precision over the C matrix and the B matrix we are driven by pragmatic action at least according to the expected free energy formula so that's kind of what's going on here um note 2 is there anything else that's kind of relevant um yeah I mean so you've got to bear in mind that like so again this is a mental action but these are the these are the physical actions so these are the actual playing of the violin right and they're going to contribute to this observation but this observation is really in terms of mental state inference during the job of saying you've done a good job at increasing precision over your sensory expectations the likelihood matrix and transition math things let me um try to try to echo this violin setting so at the purple top layer there's two contexts there's the presentation in the auditorium that's the real thing and then there's the practice at home so there's two contextual states and 90% of the time you're at home so that's your prior but then sensory evidence suggests that you're in the auditorium so that's the context the blue layer is the action policies like the angle and the motor activity of the arm and the hand of the violinist this is this is yes this is the physical actions yeah and the orange is the regime of attention of the musician now right what they expect in terms of this recital from sheet music yeah is that pragmatic value is playing like the most amazing and transcendent version of the sheet music as written so of course the map is not the territory however pragmatic value is defined as hitting those notes at the right order in the right time so it's not like an epistemic exploratory place so now how do we think about how those motor movements are skillfully done which can include not just deafness but also planning like a wind player has to plan their breathing and there's other things and planning the motions all kinds of things and so your kind of contention with the flow state is that in a skilled sesame holly setting where there's precision like yeah I am good at skiing I am good at violin personally I'm just saying when confirmatory evidence is coming in then that enables a shorter time horizon of planning so a vastly reduced cognitive overhead but relying on tacit and implicit heuristics that enable um professional outcomes to arise in that flow state even if somebody looking from the outside oh wow they must have really planned that however it wasn't really planned and then we kind of have the orange level the regime of attention meeting in the middle and mediating this like cultural contextual understanding top down with the body bottom up sensory component yeah so that's basically it there's plenty of things to point out there and one is I had this idea that these observations that you're making in real time are actually basically loop round into this contextual state inference so at time n plus 1 there's actually going to feed into our conversation about time what I've actually just observed so the very sound that has come out of my violin itself is feeding the contextual state inference which itself feeds another habitual action to cast precision over sensory expectations about the next thing that I'm going to play and then that cycles so we're basically saying that although the actual action policy the actual selection of the actual policy here at C1 is pure active inference the mental action that drives high precision over it is habitual and that itself is driven in this recursive way so that's one thing yeah the other thing is about this kind of shallow temporal planning that's exactly right, that's exactly what we proposed the reason for this and this was actually raised by Yackel Howey so I'm very grateful for his observation here and this is actually mentioned in his 2022 self-evidencing paper is that the like it's not the case that when I play my violin skillfully I can say okay I know it's not the case that I believe that everything I do is going to be highly predictable for 3 or 4 minutes because that would afford deep temporal planning rather as I said there are these volatilities in the environment that I always need to be adjusting myself to I need to be conforming myself to I need to be able to exploit those affordances so flow states are not just one extended action policy which is super rigid rather it's kind of because each state inference itself acts as a contextual cue over certain sensory expectations we have this kind of fractal nature where you might have like the first state well it might be fractal but it's like a tree you have the first state inference and then you have these bifurcations such that sure everything is constrained within the piece that you're actually playing but it's not going to be the same piece every single time now what we do say is is that if I have an observation for example if my string suddenly breaks what happens with this Bayes graph is it basically collapses and the reason for that is that that lower level perceptual state inference which feeds into the context that I'm in is no longer a sufficient context to trigger the habitual mental action so it's not just the case that you can be in flow in any context that's really important there's a set of contexts which are going to license this mental action and that's very nice to explain within active inference as just caching action policies as habits so the more you do them the more they just get cached and the sort of more precise that is as an e-matrix so that also allows us to talk about cognitive effort in line with Thomas Parr's 2023 paper about sort of it being diverge effort being divergence from priors over policies but yeah I mean the fundamental point here is that flow is a weird one because everything is like very predictable within its within its action policy regime but that's it the skill really is being able to have like a diverse or fractal structure where you can go to different nodes and they all link together through habits basically awesome okay few fun questions in the chat about precision goggles on I hope you're confident yeah okay Dave writes quote we can't observe our precision he says this assertion seems to need clarification or qualification eventually our precisions have profound effects on our actions and anxieties they must need attention somehow and in fact affective active inference have posited that there's something like the observation of variance parameter and that that's associated with anxiety and valence by HESP at all so which variances do we imagine are introspective and why so that's a good question at least in the way that this I don't think to be honest it makes a dramatic difference whether you have another level within this base graph which has the observation of precision our idea here is that just because things are happening very quickly it might just be more parsimonious to invoke that the evidence for that precision weighting comes from the observations so that's that kind of oh with a 2 as a subscript and 1 as a superscript but no I mean I'm not gonna die on this hill because also I don't think it makes a massive difference I might I would be personally inclined if he was here to pass this one off to Lars because that was the way he explained it to me and I'm not trying to exculpate myself but recall that this paper is a 6 or 5 author collaboration so it's not all me but that's the way that I would say I would say that there's no harm in saying that you can observe your preferences or your precision weighting but in some sense we're treating this extraceptively and so yeah this seems to work in a way that doesn't need to invoke a another layer to the hierarchy great that's not a hill to die on that's our research field yeah exactly but seriously it's a great question by Dave and those are the questions that we could imagine real studies to ask or address yeah so it's those are the great questions yeah I mean the other thing I would say here is it's something that's become quite obvious to me again I'm not the expert on this you might want to ask Thomas Paar but something that's become very obvious to me is that there's a difference between precision and precision weighting and I think there's often times in the community a conflation of those three and so when we're saying okay we need to pay attention to these certain things well attempt in the A matrix which is a certain type of precision but we have other precision in mental inference including these C matrices, B matrices, E matrices and D matrices and they do not all map onto attention so this idea that precision weighting is attention which is kind of an okay heuristic it's not really what's going on yes the terms not to make too general a statement here but the terms that refer more generally to statistic like state, belief, variance attention, precision, confidence usually without more qualification they're too vague just saying something is a prior are we talking about D in the purple D2, D1, are we talking about B playing the role as the prior for a future time step so like they are coarse-graining they're fingers pointing at moons and then there's the movement of the solar system and there's a lot of pointing and gesturing and also thankfully rapid accelerations of our understanding and our tool environment to help like move this model for example into an executable format or to use verbal descriptions of it and there's some really exciting and fun things like even just pasting this image or uploading the paper into a language model can give some fun and different insights now it may take an expert to then evaluate whether what it says is accurate or not however that's a super promising avenue for like developing context and metaphors or examples for looking at this yeah I mean I need to add that like I have I really do not, I'm not treating this as like a way to have a go at philosophers because my background is in philosophy and cognitive science so this emphasis on the computational model in many ways actually owes a lot to Carl because when I presented this flow paper for the first time in his group he was like well it sounds sort of philosophically narrationally good but kind of the emphasis now is on like can you put this into MATLAB can you build something from it that said I mean as I said this is informed, this paper at least was informed fundamentally by lived experience and phenomenology so I guess what I'm trying to become a part of is the so-called generative passage which is finding a way to get from phenomenology to Bayesian mechanics and hopefully this is one way but I don't I don't propose that I'm reducing the rich phenomenology of lived experience to these Bayesian mechanics because I think that would be slightly heuristic because no one knows yeah it's not even just approaching it from both sides I don't think there's only two sides to this there is though this just like you said grounding the felt experience and the phenomenology the ontology and the practices of philosophy and also formalism and wanting to touch grass but touch MATLAB as it were and then that kind of strange looping when our felt experience has a reference of a cognitive map and model and that kind of special there's a tri-partite structure here which is one the Bayesian mechanics to the model and three the phenomenon the target phenomenon and I think again this has actually really become apparent to me after speaking to Julian yesterday Julian Kibberstein has to be careful to not fall into the tendency to conflate them or reduce one to the other as you say the map is not the territory is a really crucial component of all of this I can provide a generative model I can speak about C matrices I really want to clarify I do not mean that in a reductive sense that the flow of time is the updating of the C matrix what the actual binding the relation is that I don't know whether it's emergent strongly emergent weakly emergent epiphenomenal is ontologically reductive is a good question but it's an open question I don't think we should just take for granted that they are one and the same thing that we're making an identity claim here of course many interpretations and we welcome them I see this as like a boundary artifact or a design artifact that brings us to the point of juxtaposition of all of those schools of thought that you just mentioned and more and then it's like through the bottleneck together maybe the distribution of beliefs on the other side is the same maybe it's different but we came together we synchronized our generative models and then we get to continue the discussion like from a point of contact or alignment otherwise different assertions might be made that don't even have the same underlying axiomatic set or referent or inclusion criterion at which point you might as well have one person saying well I designed a pickup truck and someone else is arguing about whether my red is your red someone else talking about ecology it's like blending up pages of a library it's very important to keep a philosophical eye on the theories that are emerging and to be honest Daniel I mean it's credit to you I think as a younger person entering into active inference you guys have all done a very good job at not making us subscribe to one philosophical position one ontological epistemological position I mean people will have their own I for one am personally happy to leave it open I mean I say that before doing my PhD I'm sure after my PhD I'll be an acolyte of some way of thinking but I think this is the exciting thing about active inference and hopefully that enthusiasm comes across in the podcast which is it kind of works whether you're a reductionist an epiphenomenalist it's just maths and statistics and it's informative for philosophical conversations it's informative to the maths and statistics and framing it but it doesn't constrain you like maybe having a super radical and activist cognition account might constrain you or being a pure computationist might constrain you so credit to everyone credit to Carl, credit to you thank you it makes me wonder there's so many ways to go from there so much of this is active inference itself if anything could even be itself and how much is merely optimal grasp of formalism and just seeing that whatever you write on the floor at the table or the page any equation category theory, dynamical systems statistics like anything is the diving board that you have to jump into the pool with for phenomenology and so what's the point of comparing diving boards only if it's about to dive you'll notice I didn't even mention active inference I mean like one because we're here it would be a bit silly to waste time explaining what the free energy principle is I mean from my perspective the way that I talk about it is I'll take the free energy principle as just unfalsifiably true as a principle of physics it's like it doesn't really get more fundamental than the free energy principle where it really gets interesting is the fact that if I am to be considered the same agent or same organism from time one to time two then I need to be abiding by the free energy principle now the real question is how am I doing that and predictive processing on POMDP schemes give a really nice way in but I kind of just take it that like I'm doing some form of active inference all the time and like from that I let it take me where it takes me whether I stick purely to a philosophical account or whether I stick to a more computational account like none of my commitments there are negating my ultimate commitment to my the ever present reduction of variational free energy so it's a nice building block but as you say I think the lovely thing about it is it can lead you in different directions sure it might be it's just an open season people have their commitments so if you're a predicted coding guy you have your commitments if you're a variational message passing guy you have your commitments for me I'm not I really don't have any skin in the game at that level because I've been in this game for a year so I'm happy to see where things go and just for me it's really important that it maps on in some way to the phenomenology if you're talking about something that clearly or analysis of my own experience is not the case well then then that's problematic interesting again I love this topic and there's many things we can continue on and it makes me think about how coming from an animal behavior and genetics backgrounds where first person experience was either implicitly or explicitly all of the phenomenology feels special to bring in kind of like the leavening of the bread a little bit because the animal behavior perspective is like everything but this even the orange layer here would be kind of like a proposal for a cognitive analysis but only to get a better handle on the blue never because there was something intrinsically like meaningful or experiential it would be just too speculative and a field so it's just interesting to come from the orange layer and descend into the blue and get into the nitty gritty with the motor reflexes and the isocades and the way the philosophers get into these embodied phenomena and then also from the embodied and the behavioral scientist side to come up into be like yeah actually we can't sidestep some of these philosophical questions it does matter how these different systems perceive and engage with time yeah I mean I think it's important to bear in mind the behaviorist for example they never actually excluded in terms of ontology in their case representational states they just said that they're uninteresting to science because they're unobservable so like an animal someone studying answers you do or bees they're behaviorists but that actually that doesn't commit them to saying anything about the representational capacities of a bee or an ant but yes I mean these philosophical questions are interesting I'm personally happy to start at the insights that people like Molo Ponte or Hustle or someone like Julian have derived from these very precise phenomenology is not just like a random science it's not woohoo it's codified and categorical and I'm happy to start with that and then when I find something like that clearly violates what at least seems to be pretty obvious in my experience now to go to your experience would require some philosophically slightly troublesome inferences because you know you couldn't be a philosophical zombie and I'm not a philosophical zombies but at least from my experience you know people get very angry about illusionist about consciousness not realizing that they're not actually illusionist about consciousness they're illusionist about phenomenal states or phenomenal qualities but for example you were an illusionist about consciousness wholesale well I just can't start there with you because you're saying something that is just so clearly wrong so again I'm not going for Daniel Dennett or Frank just because I know they actually don't say that but for me it's a decent starting point to say well I have very intimate knowledge of that which is happening subjectively and let me see how close I can get there with computational schema whilst not losing the phenomenology I think that's the important thing and then I think where you end up emphasizing is just like what your interests are right like I'm fundamentally a philosophy guy I'm not a computationist by training as you can see probably from my explanation of this graph so I've been really lucky to work with people like Lars and Carl who have kind of helped me and I think it's really really useful but again I'm not going to lose sight of lived experience and I don't think anyone, personally I don't think anyone should but it's difficult as you say when you're studying animal behavior because well that seems like a stronger inference to make because maybe they don't have public language to say what it's like to be an ant or a beast to paraphrase Thomas Nagel well very fun what are your next steps or directions so yes this is up on the preprint as I said I want feedback I want people to get angry and attack me with the spirit of scientific conciliance so I very much encourage feedback I think I'm hoping to get some more formal reviews from some people and then we'll try and get this out there somewhere I don't know quite where I'm working on some other things which I'm sure I can speak about I'm working on a paper about self-modeling in Tourette's with Libby Severs which is super exciting that's coming together I'm working at the very early stages of working on something about extended mind with Mao and others so that should be really fun and the podcast we've been putting in loads of work on the podcast so we have Chris Friff which is super exciting just a sort of pioneer in psychology Julian Kivistine as I mentioned Alex Orobia in a wonderful conversation about biometrics and cybernetics and I learnt loads and Lance DeCosta and that was the one where I was telling you before we went on live there was a lot of maths learning so to be honest Daniel I'm just learning and I'm applying for PhDs and hopefully that will kick off either later this year or the new year but I'm not going anywhere I'll put it that way awesome very fun very epistemic very chill a lot of good comments in the live chat you can review thank you everyone in the live chat thank you Darius see y'all thank you