Fundamentals of Active Inference (Chapter 7, Session 33) August 4, 2026

All right, greetings everyone. It's August 4th, 26 and we're in our third of four discussions on chapter 7. So, we're right in the heart of it in the fold of the book as it were where we have taken the jump from perceptual generalized filtering into active generalized filtering. We're in this 678 island where we're within the continuous time formulation which is pleasant because it has a lot of intuitive though also very sophisticated analytical features like extreme differentiability and a lot of other aspects that make the smoothness of perception, cognition, action, external state unfoldings things very clear and linked their formalism. And when we get to the next island of chapter 9 and 10 and we're more in the discrete time setting, a little bit more like chess planning, then we will pick up on some pieces that were not present in six, seven, and eights like related to explicit policy rollouts and other ideas that have to do more with explicit counterfactuals as a way of dealing with future states. and uncertainty rather than tailaylor series like expansions which have the advantage of this sort of single generalized coordinate vector representing indefinite future time scales but have the limitation that they don't explicitly admit counterfactuals and it becomes a second order question how confident should you be in that tailaylor series expansion through time whether you're only interested in the perceptual component or in the active component as well. Um, we can go to >> Hello. >> Yes, Babin, please. >> So, yeah. So, by the way, I mean, you know, you show the picture. So, it has like a dependencies and dependencies. So, it's like uh what it means like like you know cuz I'm from I mean I I used to develop Android app. So, you know, I know this one from Android perspective. It's like library that we can use. But here you know what this means like you know you can see the numbers and everything. >> The dependencies are which chapters each chapter builds upon. Beavenon please ask specific questions. I'm I'm happy to hear your input but please ask specific questions. >> Okay. Okay. Okay. So it's okay or it means it's like it's to build I didn't understand very clearly. >> This figure is from the beginning of the book. Do you have the book first off? Just to check. >> No no no no. I'm in. >> Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Please >> work with us here in this open science setting. If you don't have the book, work with the materials that we're putting out, but respect that you don't have the book or you don't have the full context. >> Yeah, of course. That's correct. >> Yeah, absolutely. So, the dependencies say which chapter each chapter is building upon. We're in seven and it builds upon all the previous six chapters. >> So, >> okay. Same like the Android. Okay, I got it. Now I can understand. Okay, >> cool. If you have a specific question, go ahead and unmute. Otherwise, I will take the honor to >> Thank you. Yeah. Okay. Is there any part of seven that anybody wants to go to like an example or a figure or a specific question? Cool. As as usual, Frraasier has added awesome notes to um the top of the page for chapter 7. So, if you're looking for these kinds of synopsis of different variables that are used and um different summaries, takeaways, this is great. Um there's so much gets brought up in this chapter and again it's just fascinating to see one way that action gets brought into the picture and how it connects to variational free energy and there's these deferred topics of adjacencies that aren't exactly addressed in this chapter. Um and especially in chapter 8 and 9 or 9 and 10 we see that more familiar policy roll outbased approach to doing active inference. Um so is there any question or topic that somebody wants to go to or just like a page or a line that they read that and thought it was kind of curious or they felt like there was something there that they got from it or that they had uncertainty about? Yeah, Andrew. Yeah, sorry. I'm I'm searching for the page right now, but there's um particular section in chapter 7 where it starts going into action and how we're sort of treating action and how it gets related to sensory observations since action is not actually part of the agent's generative model. um it's it's just like an interesting but also like important um kind of distinction how we're treating action. But then we also notice that it's following certain kinds of similar uh kind of gradient related updates with Uler's method uh as we're doing within states too. So, it's kind of like we get we get this like very similar uh almost pseudoisomorphic uh similarity in how we're treating action versus hidden state inference, but then they're still quite distinct, right? Yes. So let let's um pick up on page 161 and let's really get to the heart of how actions being brought in and what the formalisms for actions are. So first we have on the sidebar on the right side we have figure 7.2. So this is the basian network graphically that we're going to be discussing. X is the latent external state. Y is the observation. That's why it's shaded in gray. And we have some unfolding of the latent state. The temperature in the room is changing. And then we have observations that are being emitted from that latent state. And again, that was like the chapter 1 through five setting. That was the chapter six setting of how do you go in the generative direction from latent states to observations? How do you go in the recognition direction from observations to latent states? And what we're adding here is V, which is the force that is going to be acting upon these other variables. And again, we're not explicitly considering counterfactual actions. So, you're not going to see something that's like, if this happens, then the agent believes this will happen. If that happens, the agent believes that will happen. We're just going to be looking at this sort of like instantaneous kind of like two dance partners together without planning where it's just this unfolding action sequence. And these next equations are going to show how that action sequence unfolds. Okay. So on page 160 and 161 we have the fancy E and the fancy M's which are defining the um generative model of the agent and the generative process or the environment which is the fancy E. um that takes us up to through section 7.1. So that's just a a short section that's getting us to this kind of extended chapter 6 setting where we have the same generalized filter filtering on a perceptual setting plus this new variable V. So all we've done in section 7.1 is introduce V and the possibility for there to basically be this causal action of the observer. This is like chapter six perceptual observer. Chapter 7, it's an active observer. That's the active inference component. So that how is the action going to happen and how are we going to use the gradients that were established in the perceptual model to guide the unfolding of action without the explicit consideration of alternatives. I keep emphasizing that because when we think about reinforcement type discreet time policies like chess, go robotics, those are very often based upon explicit rollouts of alternative considered policies and the evaluation of different policies in some way either in a reinforcement or reward learning setting based upon like basically a score. How good is this policy? Single number or in the discrete time chapters we'll get to in D and 10 making a ranking or a waiting of policies based upon their contributions to utility or epistemic oriented goals. But here we're just looking at this sort of unfolding. It's almost like there's a um pinball in the pinball machine and it's just going to be engaging with its environment without consideration of alternatives. But you can get some surprisingly complex behavior through that type of setting. Okay, so we get to and raise your hand or or write in the chat if you have any um questions. Okay, so we get to 7.2. So first we see equation 7.5. So a is action triangle over the equals [snorts] means defined as and then here we have the change in free energy change in vf over the change in action. So we're just looking at this local gradients like we have the thermometer. we have for an incremental interval of increasing temperature or decreasing temperature. Which one of those will contribute more um directionally to taking the variational free energy in a gradient descending direction. Um and side note 9 at the top of 162 says a learning rate specific to action kappa sub a has been added to the equation. um which basically allows you to um control the inertia of action in this setting ranging from super high learning rates kind of like the infinite learning rate asmtote is just a memoryless process that instantaneously updates all the way to the direction of flow. So that'd be like if the pinball had no inertia mass, just instantly go the direction of the gradient entirely. Or a very slow learning rate would be equivalent to like a heavier pinball where the rate of change of action is relatively slowed down relative to the gradient. Okay. between 7.5 and 7.6 we get the usage of a um time step time interval incrementbased method for determining what is the action going to be at the next calculated time step. So, we're going to say action at the next time step is action at this time step plus a small interval of time and the rate over that small interval of time. It's literally like delta t, delta x, and delta t. And this is kind of an approximative way where especially if we're dealing with a function that doesn't have an analytical derivative, we can use different delta t intervals and just estimate basically rise over run and then add that rise over run to the current value of the action. So in this setting, action isn't like a event that occurs. It's something that's continuously unfolding through time like the um extent to which you're pushing on the gas pedal in the car. There's some unfolding through time. Even if it's at zero and then it's non zero and then it goes down and it's back to zero, there's still at every single time point, there's some value of that. Or like a joint in your body, every single time step, there's some value for that joint's action or a muscle, there's some amount of tension in the muscle, even if it's zero. And that's the characteristic of the continuous time formulation is that every single moment in this setting action is unfolding. Whether it's the thermometer um or whether it's the air conditioner's choice how actively to be firing or whether it's the um car engine's amount of RPM or the joint angle action is just something that's always occurring. Equation 7.6 Six is giving us this heristic approach where we don't need a d differentiable anything. We just take some tiny increment delta t look at rise over run and adjust our action that way. But as following we run into a problem. Variational free energy is not a function of a since action is not part of the generative model. So variational free energy is based upon the relationship between the observations and the inferences about latent states. So it would not be possible to differentiate VF in the manner specified by 7.5 by itself. Um but there's a workaround. We can optimize actions through sensory observations. The reason why this works is our actions have sensory consequences. So let's go back to that pinball bouncing around. Let's just say that its observation is like the G forces that are acting on it. And then it is able to make uh unfolding of action without considering alternatives just based upon its instantaneous experience of the gene forces. Um that intuition gets extended greatly and once we get to discrete time where we get like explicit cognitive counterfactuals we will get another layer of depth and richness. But it's really important to remember in this initial kind of kernel setting, we're dealing without the explicit consideration of alternatives, just a gradient. Like there's a set point for the room's temperature, there's a set point for the propriception of a muscle joint, there's a set point for the G forces experienced by some sort of pinball. And then it's just a question of more or less within this um value of that actions can take. And from 76 well really from um 75 to 77 we get the application of the chain rule. And you can sort of see how that works where the delta y of a so this is change in sensory observation as a function of action. So this is like will you experience more geforce if you will you experience more tension in your bicep if you lengthen your joint or if you close the joint. And then it's kind of like funny. It's just like if it were um regular fractions. We have delta f over delta n. That's what we want to calculate. We want to know how will the free energy gradients change as a function of change in action. And then by doing we can't calculate that directly. That was the previous paragraph. And so we multiply it by basically deltay of a over delta y of a which is just whatever that value is. It's one. It's like one over one. And then split that fraction back up so that we have these two values that can be discussed separately. So on the left side of the right hand side we have how do we expect perception to change as a function of action as a function of action. So this is like dy of x over delta x. On the right side, we have this more decision-making objective where we have the value, the free energy, which is what we're going to use as our sort of top level gradient to actually pursue because the units of the change in observation or the change in action are denominated in those currencies. whereas it's going to be the change in free energy that is going to be our guide in a generic way. Um so let's focus on this. So so this this left term again is saying how will observations change as a function of changes in action that relates to kind of like domain specific or observation action specific features of the world. Whereas how we're going to make choices about action is going to relate to how they their consequences are on the variational free energy. That's what it means for variational free energy to be used as a heristic or loss function or kind of unified objective for perception, cognition, and action. And this is just zooming in on what it means for it to do that in action in this continuous unfolding way. So we get to 7.8. So we have left hand side is that term that we're interested in. we have successfully um removed the the naked delta A which would have been nice to know. I mean if we could just simply know how will free energy change with respect to changes in action. Great. But here we're closing the loop and integrating the sensory consequences of action which keeps the variational free energy being about what it's about which is grounded in the sensory observation and we get to the derivative of VF with respect to sensory data which is to say this left hand side and we get this lambda I uh epsilon or e y and that I don't think those were defined earlier in the chapter but let's check if we have it in this All right. Yeah. Um, I don't see those specifically defined. Wait, let me I don't see them as specifically defined. So, I'm not going to um give them a name, but we can see that 7.9 has a really similar format to 7.5. We have this K A um something over delta A. So in 7.9 we have taken out from the top the delta f and replaced it with this delta y of a which is like the left hand side of 7.7 and then we have used 7.8 8 and plugged it into 7.7 to get this lambda y for the derivative of the vf which is what we we're symbolizing with the right side of 77. Um yeah thank you. Thank you Andrew. And that's described it's described in um 7.12 which is interestingly a equation but it's it's written as an equation. Well it has a letter in it. Yes but but it's but the variables are used before they're described. But we have 7.12 describing 7.9 which is change in action equals the negative learning rate multiplied by the forward model which is what are the change in perceptions as a function of action with respect to changes in action. That's that bold forward term. lambda y is the sensory precision that corresponds to the mapping between observations and latent states and the sensory prediction error. E pretty much means error. Um so we get to 7.10 10 a and b where we have change in beliefs about latent states mu mu dot subx and change in action. So it's like these are the two things. This is the change your mind and change the world type thing happening. This is the continuous updating of beliefs in a generalized filtering way just like chapter 6 plus this gradient that has to do with equation 7.5 which has to do with again this learning rate scaled gradient on free energy. So this is how free energy optimization can include perception and action within a unified objective here within this single observation channel, single latent belief channel, single action channel setting. But everything else with more channels and more other thought experiments is predicated on exactly these equations, which is why they're in a box. Any thought or comment or a question on this? Yeah, Andrew. Yeah, I guess I just add briefly um that we we see that through going this this like somewhat indirect route where we're relating action to sensory observations. Why? Um we're kind of meeting that definition of action that we're given in the chapter where the idea is you know we have action and perception. So perception is to sort of change your mind uh and action is to change the world. So you can see that if we you know if we produce an action that is then sort of evoked or you know elicited into the environment that changes the environment state and then that releases you know potentially new observations for us to receive. So it is we're quite literally you know undertaking action to change the observations we receive. And so Daniel gave nice experience uh um examples of that earlier. Another one that we're given in the book is, you know, whenever we take action to uh look a different way, right, in our environment that we're in that that tends to produce, you know, different observations. If I'm looking directly at my computer screen, I'm seeing my screen. Uh if I look, you know, at the sky, I'm seeing very different things. Um, so it's kind of like, you know, we we take action to sort of change the the sensory responses that we ultimately end up perceiving. And then finally, because we're still in this realm of using variational free energy, albeit in a slightly indirect route, um, we're doing this in reference to things like, well, what prediction errors are currently going on? What did I expect versus what really happened? So hidden states are still kind of part of the equation, though much more deeply baked into the variational free energy that we're already well aware of. So while there's some conceptual new things going on here whenever we introduce action, we're still in the same world of looking at variational free energy, we're still looking at the same kinds of prediction errors that we saw in chapter 5, like sensory prediction error and hidden state uh prediction error and and all the rest. Um, yeah, just wanted to add that. Great. Page 163, we basically get this written procedural list that should remind us of basically any loop from programming. But here applied to that single channel, single latent belief, single action setting. Initialize in step one. Generative process initialize in step two. The generative model just when it gets to your program, you can look at it. That's where like the variables are put into memory and they're checked for their shape. And then you get the for loop in step three where you step forward the generative process. The generative model computes a gradient, updates its prediction errors and then applies action and emits it which gets um into effect at the next time step. So >> any comment or or anything anybody wants to add here? Just getting through 72. Okay. So, I'll I'll I'll copy your question to seven. Um like just so it can be addressed is like yes active inference is applied often to physical navigation as well as to mental states. That's called cognitive modeling. And cognitive modeling can deal with a variety of inputs ranging from exterception, which are the external oriented senses, to interosception, which are like the internal body signals, on through introspection, which is like interosception for the mind. And that can take one into ideas like metacognition and awareness and things like that. Are there different design criteria to consider [laughter] in the sense that it's an active inference model? Thankfully, no. There's so much in similar to how the variables are going to be tracked and accounted for and optimized. But in terms of the variables being different, it's like yet as different as can be. It's kind of like we have two different um dishes we could make, but they're going to go into the same container. So they're similar in that we're going to put them in the same container, maybe run them in the same software, but they're as different as they can entirely be because you're studying possibly a different input, cognitive process, and action output. >> Okay. So, we get two >> and andrew's explored these these kinds of modeling >> topics um a lot. >> So, we get to 7. Yeah, Andrew. >> Yeah, sorry. I uh yeah this could definitely become a whole thing. So I'll try to be uh brief but I I really like that question that was asked and uh like in the cases of computational psychiatry and the other fields that Daniel mentioned um you know there are a lot of people who do things like try to simulate uh phenomenology uh or just more general like a effective or emotional inference without it necessarily having to be related to to pathology. um though it often can be um yeah like you can there's a bit of a distinction because of the wording of the the question I just wanted to clarify like um the idea of creating a model like an agent model that we have that actually treats a real person like an empirical existing human being um you know as sort of the environment where you have a model that captures or attempts to make different kinds of estimations off of this real empirical ical data that you're receiving from a person. Um, which is very interesting uh to do and people make like classification models for doing that or they do other sorts of uh like predictive modeling to try and capture the behavior of someone. That's that's a bit different from directly simulating a person where you're making a model that uh you know you think based on your best design and hypothesis about how a person might work or how they how a person functions within a particular context like oh I'm simulating a person in a particular environmental task or otherwise um that would be a bit different where you'd be you know giving it different kinds of uh proclivities and and hidden state representations and actions and all the rest that actually directly are supposed to simulate a person. In both cases, you can work with empirical data. Um, so you can actually like collect data from a person and for a model that is supposed to make estimations off of that. In the former case, much simpler. You already have the person's data. You just run it through and see what your model comes up with to gain some kind of insight into that actual person's behavior. Um the second case where you're trying to make a model that simulates a person, what you can do is you would also collect data from a real person and then cross reference that with the kinds of behavior and hidden state estimations that your model makes. And that's called model fitting. in which case you would just try and like incrementally perhaps or iteratively try to um get your your homemade model uh to to much more closely match the kind of parameters that we find in uh uh you know as as sort of retrieved or analyzed from the data of a real person. So yeah, it's a it's a really big good topic for any of those who are interested in like psychiatry, mental health, uh just general human psychology that want to do anything computational, very interesting things to think about. There's plenty of literature on this and active inference, but also just elsewhere more broadly. >> Yeah. in Sebastian's comment about like the endocrine system and hormones like that is the beauty of a unified approach is you don't need a special case model for every single hormone system or you can do direct model comparison with or without the effect of different hormones or you could treat a variable as an unobserved until you observe it. So you could develop like a whole model that includes EEG data, MEG data, you know, real- time glucose, real-time dopamine, all these different kinds of things that you don't have access to with the legibility to say, I don't have access to these. Here's my uncertainty interval on them. And then even if you got patchy data, like just one observation per day, you could quantify how much information is this providing. And then maybe in a different setting or in the future it's possible to get biometrics that are more high frequency or high accuracy. Babin's comment here. Um, once we detect the triggers of predictive error, then we can find the context to reduce the predictive error through dialogue and observation which could be for linguistic settings or it could be kind of interpreted as like a agent in the niche in dialogue more generically. Um, it can reduce the predictive error but I'm not so sure. >> Yeah, that's Yeah, >> you Yeah, go ahead. >> Okay. I mean, I think you can continue listen so we can talk about at the end. I mean, I think that's what Andrew said. I think so. So, yeah. I mean, you know, I think it won't be that much related to this entirely. So, yeah. So, we can discuss about this at the end. Yeah. >> Yeah. And and and like it's kind of like when they say there's the rub, which is like how does the model know which direction of action is going to be associated with which gradient change? That's not addressed here. That's the question of learning which is how does that forward model come to be learned? How does the thermostat come to understand that turning on the heater raises the temperature? what would happen if it was a setting where turning on, you know, it's just button A and B, how does it come to learn what the consequences of action are? So, like this reminds me of one little riddle, which is um there's an attic with three light bulbs in it and you're downstairs and there's only two light switches. What can you do with those two light switches to determine which one connects to which of the three lights and which one of the three lights is unconnected? So first off that that needs a life I mean that needs like uh you know a trained environment to know it actually or else it just like I mean it's an impossible thing you know even a human being also can't know it for instance if I arrive to office I I mean there is no way that I can understand which which which which I mean which kind of switch on the certain you know >> okay here's the the solution to the riddle um you turn on one of them and you leave it on for a long time. Then you turn it off and you turn on the second switch. And then you go to the attic and one of the lights is going to be on, >> one of them is going to be warm but off and then the third one is going to be off and cold. >> You can go into these second order questions like what if it's a fluorescent light? It doesn't heat up or how do we know, you know? Yes, you could always like break out of the box of a model and continue to explore these metacognitive questions. Okay, so the thermometer, you know, we understand the relationship between the heater and the thermometer, but how do we come to know about temperature? So like that is kind of like popping up a level which again brings us back to the importance and the relevance of having a generic modeling framework where we can actually model those kinds of ideas as structure learning and model structural understandings of the world within the same generative model family and type. that's going to eventually pay out at the very end in something that's like a body sensor. and and as the the the convo in the chat there often this is appealed to with like an evolutionary account or an ontogenetic a developmental account of like well that's just the kind of thing that comes to be largely structured to understand this kind of a causal relationship or even the causal relationship is built in and a lot of ways to go with that but um yes this is like describing the canvas of what the possibilities are. And then in the next chapter, we'll deal with learning and some other topics about how over multiple time scales [snorts] do these models come to be fit and used. But for here, it's just taking the model is given and it's looking at the fast time scale of inference. And so we get to figure 7.3 which is the one that um we all have written on our hand [laughter and gasps] right um where we have those two terms from equation 7.10 in the internal and active states. So both of them are minimization processes with respect to different variables. Internal states are obviously updating beliefs about latent states and active states are updating beliefs about action. But they're both a function of the same variational free energy of observations and beliefs about latent states mu [snorts] of x sensory states. We have observations being generated by the fancy E by the environment as a function of the true latent states Xstar. The true autonomous action V star. The true parameters that generate observations from latent states theta star subg plus the true variability noise of observations omega star subw. And then over here in that true external states we have their unfolding through time as a function of the latent states the action forcing from the agent and the parameters that define all those relationships plus a true noise term on the latent states. So this is like where we get this classic diagram and we can confirm that there are no causal connections coming from the outside world to the internal world except through observations. Why? And there are no causal effects of internal states on external states except through action selection. A that is the marov blanket property which is that two states are separated by a marov blanket if they are independent conditioned upon the blanket states. That doesn't mean causally disengaged. It's totally the opposite of that. It's conditional upon their engagement. There's no further engagement. So that's the sort of roller coaster. Oh, narco blankets are everything. Wait, it's conditional upon what it's conditional upon. By definition, they explain nothing. And then you make your way up the plateau of productivity where you realize that these are designable, they're legible, they're interpretable, they have positive computational properties, they're composable, you can apply category theory to it, you can develop software interfaces or interpret software interfaces as this type of a cycle. So that's sort of what's on the other side of that whiplash of it's just Markoff blankets. Wait a minute, they don't do anything. and then you meet what you can do with them once you play around with it more. So that's 7.3 classic figure and you'll see a lot of cousins of this figure. Sometimes the exact same equations perhaps with a little bit of different notation. Other times pretty different equations but still reflecting this partition or structured way to like disarticulate the imshment of the agent in the niche or two agents in conversation. Um 7.4 section 7.4 Four goes into three examples. Se example 7.1 is a basium thermostat. Example 7.2 is the countering of an exogenous force. So that could be like a sort of homeostatic setting. Or you could think of it as like someone's doing a prone hold and then someone else puts another weight on them and they have to counter this new force to come to a new isometric point and then the weight is removed and then they have to rebalance to there and experiment or example 73 which is countering a dynamic exogenous force which is basically just 7.2 two where there's changing forcing function. Section 7.5 goes into multivariate active inference in generalized coordinates which is like whoa because it's kind of like expanded in all the dimensions once you're like wait we have the generalized coordinates with all the higher derivatives and we could have any number of perceptual cognitive active and external channels. That's why it's so important to understand at that univariate level what is really happening because it's set up in a way so that the composability is all really powerful and that's what makes it flexible. Um and then 7.6 is the conclusions. uh Sanjie summarizes what happens in the chapter on page 176 and then 177 basically mentions eye movement and several other settings where this has been applied and connects it to alternative and classical methods in motor control and reflex arcs. And continuous time, this is something that that Thomas Parr and Fristen and others have highlighted is continuous time because of its lack of an explicit counterfactual tends to be utilized in mechanical and embodied type modeling because you're not dealing with this if this then that explicitly. Again, that's going to come up in the discrete time settings. So, um in the um let me just go to in the uh paper on active inference does not contradict folk psychology which I'll put in our notes. votes for seven [sighs and gasps] which is explored in these three live streams. That's the type of model that they study is explicitly this situation where um the body has this continuous time setting and then the mind as it were has this discrete time setting which allows like the explicit consideration of what would happen if I move my hand here versus here and then depending Depending on which one of those two choices are selected from the mind, there's a descending motor control that then implements that just like it is in this chapter with like the static or the dynamic exogenous force. Um, if there's a a comment or a thought that anybody wants to add, please go for it. Seven's a lot though. It's simultaneously hits very hard but also very soft. And it's like a lot but it's not that many pages. It's just very dense. Andrew, go ahead. >> Oh, yeah. Thanks. Uh, yeah. Chapter 7 is just like like if we take a a little bit of a closer look, it's just building on all the same things we've already seen. We just saw Uler's method in chapter 6. We saw sensory prediction errors in chapter five. We saw action perception loops like minus the action like way back in technically chapter 2, chapter 3. So we're like building on all these things from before. But then that said, action itself is a is a whole thing that absolutely warrants having its own chapter. Uh and then recognizing how we treat that distinctly from from other variables. Um and then yeah, I wanted to briefly hopefully comment uh on one question that came up in the chat that's uh I can't say anything too definitive about it because it's very broad, but this potential issue of dealing with like null values. So anyone who's like worked in like I I personally worked like professionally as a data analyst and scientist. So uh you know machine learning that that whole world um where we see values that could be encoded as like null values or missing values. And one of the appeals to something like active inference is that you might be able to have a model that on the one hand, you know, does sort of emulate a lot of these cognitive principles that that active inference purports to to simulate um and describe and explain. Um but then at the same time like being able to employ it in some kind of concrete situation like technical situation u industry or or otherwise. Um, we do have these things like null values or missing values from our data at times whenever there was a a hiccup in a in a sensor or some kind of error that happened during a pipeline that something just wasn't evoked that should have been. Um, and it's it's a it's it's a big question and uh but in reality when we think about it, if we're repeatedly collecting information through our eyes or elsewhere, it's like we aren't technically null values. It's like if you're looking at a bunch of pixels and we just consider black to be nothing and then every object that could take view uh you know anything that could be an object within our vision isn't black. It's going to be like combinations of other colors. Um that doesn't mean that the black is necessarily null, right? It just means you're probably paying a little less attention to it. Maybe you're weighing it differently whenever you you know um undertake different kinds of inferences is using that as your observations. But uh yeah, it's just to say that a distinction here is as opposed to some static data that has some missing values um you know a real organism is going to be continuously processing everything. And so they're they're going to be continuously reducing free energy and they're going to be taking in what information that comes in. So we we in one sense we don't technically have nullles. There might be others who have different perspectives on how that works but um that's how I tend to view it just like an actual uh like physical manifestation of these ideas and principles. >> I assume that you are answering my question. Am I right? >> Yeah. >> Okay. Okay. The thing is that you know uh this is actually solved problem in programming. Actually it's a billion dollar problem actually. It caused a lot of app crashes before this null pointer exception. So thing is that you know as we told you you are right mapping the null null variable is a hard thing but if you map the event into multiple parts it's extremely easy to take the null value. So let's take like you can check like English grammar. Okay. So thing is that since we have already mapped every parts of sentences it's extremely easy that if I talk like oh I football you will instantly know that okay he's bin is missing a web. It's very easy cuz you have you know that what sentence what the sentence will have actually. So if there is something missing you'll instantly find it. So like that if you pause the event into multiple variables like for instance like the definite things like we discussed on the previous like >> yeah if you have a very yes if you have a very structured understanding of the language yes you can interpolate and you might be very accurate or you might be totally off base but yes absolutely. >> Yeah. So think is that so we can take the event and we can take event is like a reality we can take it and we can pause it to like uh observe what okay it means what we are seeing and what we are hearing then interpretive okay I mean interpretive is not I mean it won't be a thing for robot but it's for human being so what they're internally feeling about the meaning you know so we can take these things then intended what it means like we can take like what we intended to say you know you know like for instance I just saying I want to drink water or something like that you So it's what to say we can separate these things then we can we will get the definite value of like what is the reality you can separate the part of reality here the tricky part is that context is the always the only thing that is dynamic actually so that's most uh trickiest part actually >> yes so far >> learn absolutely like the idea >> things are very bad >> no just the idea thing is >> we communicate through our intention We're not just doing descriptive statistics on the words. >> Exactly. So thing is that see you can have like these very contextual situations where like I mean if you are doing some kind of data processing for example like it might be very distinct to the particular problem you're working on like if I were um I I work in education for example so I'll frequently look at like students assessment scores. If a student is missing an assessment score, that means something very different from how I might handle a missing value in a different situation. In this situation, I probably shouldn't be including that student at all because I'm probably doing some kind of analysis to look at how, I don't know, like assessment scores correlate with, you know, future academic success or some some other thing where we're looking much more at just like kind of plain simple statistical principles of like, well, to find how two things correlate, you can't use missing values from either one of them, right? It's just you can't get any kind of correlation out of that. So, in other cases, you can do different kinds of value imputations, right? So if a value is missing, maybe you know something about the missingness in the data, right? Like if a value is missing, there's some pattern there where you impute a particular value. Or maybe you're trying to do like a big data kind of prediction problem where it where you found that it's savvy enough to just impute the value to be the mean of the data such that you want it to just kind of if a value is missing but we need to keep moving, just kind of treat it as the general average score. in that way it doesn't dramatically bias anything in one way or another. So yeah, it's uh it's interesting you phrasing it as a billion-dollar problem, but yeah, it's it's more about just yeah, what field are you in? What what domain are you in? So I think with my comment, my answer was mostly like given the textbook group and kind of the principles that we're talking about here, I think that in actual practice, we're usually not treating things as sort of null values. That's one of the the tricky things, right? We're trying to do something that aligns with computational anything. Computational neuroscience comput >> any any last comment otherwise we'll end it here. >> Okay. The thing is that so you know for instance let me I mean >> just the last comment just the last comment >> or lost the last lost comment. So mostly we can use effect labeling and cognitive reappraiser to find the context actually we can map it into a communicative syntax. That's exactly that's exactly what we do. >> Yeah. Okay. Yeah. That's a that's a most I mean efficient way as as per my opinion. I mean like as for the I mean as for the things I tested I recently model agent. Okay. Agent AI to map like multiple agents to do that. >> So it was successful in that area. So that's my that's what I was working on currently you know how to reduce errors. It it's exciting. Intense inference, belief, desires, intentions, and how those models adapt to the synthetic intelligence setting is huge. >> So, I'll stop the recording here. Yeah, of course.