 Hello and welcome everyone. It's active inference guest stream number 48.1 on July 19th, 2023. We're here with Arthur Giuliani and Adam Saffron, also joined by Michael Edward Johnson in discussion. We're gonna have a presentation and then a conversation about deep canals, a deep learning approach to refining the canalization theory of psychopathology. So thank you all for joining and looking forward to this presentation and discussion. Thanks, Daniel. I'm very happy to be here with Adam and Mike to talk about this work and to discuss more broadly this general direction that's emerged over the past few years to try to utilize some of the concepts from the active inference world and computational modeling more generally to understand what's going on both in psychopathology and in psychedelic therapy. Yeah, so I'll give a brief presentation and we'll move into a kind of more general discussion of this work. It'll start with some kind of general considerations around psychedelic therapy, some of the neuroscience of psychedelics and then move kind of more specifically into the model. So there is this class of drugs, the serotonergic psychedelic drugs, which includes psilocybin, DMT, LSD, and others. And the past decade has seen a resurgence of interest in psychedelic therapy, psychedelic assisted psychotherapy where one or multiple of these drugs is used in conjunction with a kind of psychotherapeutic protocol. And there's been a lot of investment of money and time and clinical resources into utilizing these drugs for treatments, depression, anxiety disorders, substance use disorders, and others. And there seems to be quite a lot of positive signals suggesting that these kinds of therapeutic interventions can actually be quite effective. So there's a number of studies suggesting that psilocybin assisted psychotherapy can be helpful for depression, anxiety, substance use disorder, and others. And as this work advances, there are more and more larger clinical trials comparing it, for example, to the standard treatments with SSRIs with multiple studies suggesting that psilocybin assisted psychotherapy is at least as effective as the lexapro kind of standard care. All of this seems to suggest, and the timelines are pointing at this idea that psilocybin treatment for major depressive disorder may be only a couple of years away from becoming approved as an actual treatment in the United States, which means that it's quite important for us to understand what's going on mechanistically in the brain so that these kinds of treatments can be best optimized, best determined, how to be used, for which kinds of mental issues people might be having and under what kinds of treatment protocols we might expect the most success. I think this will also ultimately hopefully lead to the development of even better drugs, even better treatment protocols, and much more personalized care. So some of the basics of how these drugs work on the brain, for those who are unfamiliar, so these are the serotonergic psychedelics. They are serotonergic because they act as agonists for a number of serotonin receptors in the brain and the central nervous system or broadly. It's thought that most specifically, these drugs like psilocybin work through agonizing the serotonin 5-HG2A receptor, and this has been demonstrated multiple times by using blockers, antagonists of the 5-HG2A receptor, which seem to largely remove the effect of psilocybin or other psychedelic drugs. And what this 2-A receptor does when it's agonized is it excites the neuro and making it more likely to fire, and this is kind of essential to some of the downstream effects. So given that it seems that this 2-A receptor, that psychedelics are mainly targeting the question, then becomes, well, where in the brain are these 2-A receptors most widely expressed? And research has focused on, I would say, three kind of key areas over the past decade or so. The first being the thalamus, which is a region involved in the gating of sensory information, into the cortex. There's a lot of expression in the cortex itself, expression, for example, in the visual system, which is perhaps not surprising given the prevalence of visual hallucinations and psychedelic use, but therapeutically, a lot of focus has been on the prefrontal cortex, where there is also a dense expression of these receptors. And then also more recent work has looked at the clostrum, which is a small brain region, responsible for coordinating the transition dynamics of cortical activation, and this too is likely of essential importance to the psychedelic effect. So kind of moving up one level of abstraction to ephemeris studies. There are a few kind of key takeaways, which have been identified from a few different ephemeris studies looking at psychedelic use. Kind of the main theme has largely been this idea of increased entropy or increased complexity of the cortical activation in the brain. So there are various measures of this. On the right here, there's a figure of one of the common measures of complexity and looking at various different kinds of states that an individual might be in and how that relates to complexity. So you have individuals in vegetative states or anesthetized states having relatively low brain activity, complexity, you have individuals in awake states having higher complexity, and then individuals who have ingested either psilocybin or LSD showing the highest brain complexity. Now, along with this, rather than just this kind of single unified score, there's work looking particularly at the changes in functional connectivity between different brain regions or different brain networks. And the general trend seems to be that you have decreased within network functional connectivity an increase between network functional connectivity, which this can be interpreted as different brain networks which typically don't communicate as much with each other more likely to communicate with each other under the effects of psychedelics. And the internal coherence of networks is reduced under psychedelics. People have maybe heard of this default mode network, which is responsible for a lot of self-referential processing, which is responsible, which is typically activated, for example, when individuals are off task, mind wandering, not actively engaged. The default mode network has been kind of like consistently implicated in being having less of this kind of like within network connectivity, less of this functional integration. And this has been hypothesized to be one of the kind of key mechanisms of the therapeutic effect of psychedelics. In addition to looking at kind of like brain network activity on a high level, you can then also zoom in much more deeply and look at what's actually going on at the neuron by neuron level or the neuron connectivity level. Here we see kind of very certain very consistent story around this psychopathogenic effect or this neuroplastogenic effect, where there are a number of different markers of increased plasticity in the brain. So one that's very common is this increased brain derived neurotropic factor, which we see hours after psychedelic use for a variety of different psychedelic drugs. And this is then accompanied a few hours later by visible increases in the dendritic spine growth in cortical of pyramidal neurons. So this is both the number of dendritic spines and the complexity of the spines. You can see a very clear example of this here on the right, where you have on the top the vehicle and then the bottom the GMT treated pyramidal neuron and the difference is very, very stark. You don't even have to do any sort of actual computational analysis for it to be clear that there's quite significant dendritic growth from a single dosing. And the changes to functional connectivity, which I mentioned in the previous slide, tend to persist though in a less extreme form than what we see during the acute effects. So the question then becomes given all of this, how can we attempt to tell a consistent computational story of what's actually going on during both the acute effects, the post-acute effects, and in terms of the actual downstream therapeutic outcomes? So what's kind of been proposed as one of the leading models of this is the relaxed leaf under psychedelics model, the rebus model of Carhart-Parris and Pristin. And so this builds on the kind of like active inference, hierarchical predictive processing framework where we can understand the brain as the series of predictors stacked kind of like hierarchically along brain networks and predictions are being made by higher level networks. That are attempting to predict the errors of lower level networks. And the process of basically doing this then produces this series of belief landscapes at different levels of spatio temporal abstraction. And these belief landscapes of course can encode beliefs with various levels of precision and it's the modulation of this, the precision of these beliefs which is believed to take place under psychedelic use. So here's kind of like a very high level version of this rebus model and how belief relaxation might actually happen. So given that you have this series of cortical networks which are representing beliefs at different spatio temporal scales, we can kind of think of the thalamus as being kind of like the entry point where sensory information is being gated into this hierarchical system and the prefrontal cortex is being kind of like one of the highest level networks, right? Which is then responsible for both representing the most abstract kinds of beliefs as well as being kind of the primary inhibiting force for all of the downstream beliefs in the network in the system. Right, so what is this heavy organism of the 5-HG2A receptor supposed to be doing in this case? Well, it's going to be decreasing the prefrontal cortex's ability to make coherent predictions and coherent representations of belief which will then disrupt its ability to inhibit lower levels of the cortical hierarchy. At the same time, by disrupting this activation of the thalamic or this disruption of the thalamic gating, right? It's going to result in additional sensory information entering the cortex that wouldn't underwise into the cortex. You have these kind of disruptions at these two levels which then result in beliefs being spontaneously represented in ways that they wouldn't normally be represented. And the Rebus model at least is suggesting that this mainly is going to manifest itself as a kind of relaxation of these beliefs. So here is a kind of like very simple diagram of what this might look like. So you have some landscape of beliefs here and you can kind of interpret the position along the X and Y or I guess the X and Z belief in A as being corresponding to some particular belief that might be represented by the network at this layer or at this level of representation and then the Y axis, the height as being the actual precision of that belief, right? So here we have this kind of very deep part on the right part of this network around this landscape and this would be a very strongly represented belief, let's say. What the acute effects of psychedelics are then understood to be doing right is to be flattening this landscape to be less strongly representing these strongly represented beliefs to be making it easier for the reversal of this landscape. So to move from one attractor state to another attractor state which is meant to be understood to be kind of therapeutic in various ways. And then the post-acute effect being some kind of return to normalcy but one that is not totally complete. One that it's still incorporated some of these newly relaxed beliefs, these newly explored beliefs. Building on this Rebus model, Gareth Harrison colleagues last year proposed this canal model of psychopathology which basically kind of takes the intuition that through relaxation, if the Rebus model is correct, if the relaxation of beliefs as a therapeutic outcome, if you kind of take that way of thinking to its logical conclusion, you might then propose something like this which would say, if relaxing overly strong beliefs is therapeutic then having overly strong beliefs in the first place is pathological in some kind of way. And they refer to this development to these overly strongly represented beliefs as catalyzed beliefs or and this problem generally is catalyzation of this process and developing them as catalyzation. And this is the canal model of psychopathology. So they kind of build on this idea, this set of previous work in the literature suggesting that there may be some primary factor of psychopathology and they suggest that this primary factor is catalyzation. And in this work, they discuss a number of different psychopathologies which seem to quite clearly fit this catalyzation type characterization. Though it's very relatively straightforward, I think to understand things like addiction, depression, anxiety disorders, obsessive compulsive disorders, all of these involved in one way or another. Having some strongly held beliefs, typically negative beliefs, in the case of like depression, anxiety, obsessive compulsive body image disorder or in the case of substance use disorder, some very strong beliefs about what kind of behaviors should be taken at any given time in any given context. And then of course, the effect of psychedelics is to undo catalyzation and the better able any sort of therapeutic intervention is to undo catalyzation, more likely for a positive therapeutic outcome. So I think this work has a lot of strengths and there's a kind of elegance to it. In our Pinal's work that we had been, we've worked on, we kind of identify a couple of limitations of this canal model and attempt to extend and refine the model in ways that I think ultimately make it more powerful while kind of keeping the essential elegant insight that catalyzation is kind of like intimately tied to like mythology. So the first kind of limitation I think that we identify is that in the canal model, it's very heavily implied that all forms of catalyzation are ethical in one way or another. And as I mentioned just now, this is clearly the case for a large class of very common psychopathologies. It's when we start to look more towards edge cases, I think that these exceptions starts to problematize this understanding, right? So if we consider psychopathologies such as bipolar, borderline, schizophrenia, depersonalization disorder, it becomes much more difficult to kind of easily classify these as simply problems of overly strong priors. Many cases here we have problems of priors that are not so strong or priors that seem to drift over time in ways that are problematic for both the individual and the people that the individual are close to. And if we think about psychedelics then as being this kind of panacea to undue catalyzation and undue psychopathology, we also run into this other issue, which is that there are certain kinds of psychopathologies which are typically issues for use with psychedelics, right? So schizophrenia being a very clear one where there's evidence of psychedelic use being a trigger for the onset of psychosis and individuals who are predisposed to schizophrenia. There's also evidence of the development of depersonalization disorders with strong psychedelic use. And this is not what we would expect if psychedelics were simply universally undoing or getting rid of psychopathologies. There's also this idea which I think Adam's gonna develop quite a bit later that the stability of beliefs is typically correlated with positive mental health outcomes not negative ones. And what is a stable belief if not one that is highly canalized in some kind of meaningful sense within the system? And a second limitation is that if you think about an actual dynamical system, one that is parametrized in some kind of meaningful way like a brain, which is parametrized by synaptic connections and then instantiates neural activity patterns over time. And we think about what kinds of optimization landscapes are actually being induced by a psychosystem. We find that there's not just one kind of canalization there's actually two kinds of canalization. And in our deep canals work we tend to kind of take inspiration from deep learning literature to tease apart these two kinds of canalization and understand them a bit more clearly. And they are just to kind of give a preview of these two different things overfitting and plasticity loss. So overfitting is the problem you have your canalized particular kind of way and whatever sort of policy you've learned whatever sort of set beliefs you have these don't generalize to some other context. You're in contact say you may be very well adapted you move to context B and whatever you've learned doesn't apply there. Then plasticity loss is a little bit different. Plasticity loss has to do with one's not one's ability to generalize for other ones ability to adapt or to learn. So imagine you're in context A you may or may not be well adapted to context A if you move to context B the question becomes are you able to adapt to context B or not? Not are you immediately already ready to deal with context B but given time in context B can you learn and change? And we use a very simple recurrent neural network model to kind of make sense of this. And I think this is important because there can be different kinds of canalization on these different kinds of optimization landscapes and these have unique implications for the treatment of psychomethology. And I think even more importantly psychedelics work on these different landscapes in different ways. And so understanding how they work is essential to developing better drugs better treatment protocols and more personalized treatment. Okay, so what are these two optimization landscapes? How can we make sense of them? This is kind of like the big overview figure of the deep canals work. So we kind of start with this intuition of our current neural network, right? And in our current neural network there are these kind of two there are these two sets of parameters that you have at any given time. You have the actual weights of the network and then you have the hidden activation of the network, right? And so the weights of the network are theta and the hidden activation is H and there's some input into the network at any given time step. And then the output of the network is a function of both the input and the hidden state and the hidden state is updated at every time the hidden activation pattern. Right, and this gives us essentially these two different landscapes. One, what we refer to as the inference landscape or the type A landscape and this is equivalent to the neural activation pattern of a brain at any time, right? So what gets read off from NFMRI and what gets read off from EEG is going to correspond to this activity landscape. But kind of like underlying or supporting this activity landscape is this other more fundamental landscape which is the actual optimization landscape that is induced by the synaptic weights themselves. And this is what is the result of right click synaptic connectivity in the brain between all the sets of neurons in the brain. This is what is changing when we have various forms of like heavy in plasticity, heavy in learning. And this is what kind of like instantiates the possible sets of neural activity patterns, right? So this is kind of like a way we can start to understand this which is when we think about belief landscapes, right? As I mentioned before, here's a very simple diagram where we have a two dimensional belief landscape and let's say there's someone who's depressed or potentially depressed and they might have a negative self image. I'm a terrible person. I can't do anything right. Something like this, this would be encoded as a particular position in this belief landscape. When the neural activity pattern looks like this, this belief is what gets expressed, right? And there's another set of neural activity patterns which would correspond to a positive self image and positive self belief. I'm great, I'm wonderful. Or even just I'm okay. I'm totally acceptable as a person. And this would be a different set of patterns. So based on the synaptic the actual connectivity structure of the brain, right? These two patterns may be more or less likely. So here, these are being represented by these two different sets of neural activation landscapes which are the result of two different positions in this synaptic weight landscape space. And in this top space, the one that's in the middle of here that then is represented here at the top of part C of the figure, we see that there's very high precision assigned to or very high likelihood assigned to the negative self image and very low likelihood here assigned to the positive self image. But in some other part of the synaptic weight landscape a very different kind of neural activity landscape might be instantiated. And this one kind of has the inverse properties where negative self image is unlikely and positive self image is much more likely. So we can kind of then understand positions in the neural activity landscape is kind of corresponding to specific phenomenological experiential instantiations of certain kinds of beliefs, right? So like the actual experience of holding a negative self image or holding a positive self image. And we can understand positions in the synaptic weight landscape as being something more general. So in the case of depression, there might be a specific set of synaptic weight connectivity which would correspond to what you might call a depressed individual, someone who might qualify as having major depressive disorder and another set of synaptic weights connectivity pattern that wouldn't qualify as that, right? Okay, so before I get into this psychedelic, how psychedelics relate to this, I think Adam is going to talk a little bit about the work he's done extending the repus model. Hello. So yeah, I've been thinking about this a while and then along the way, I met Arthur who also has been thinking about this a while but in different ways. And so like Arthur's background is in basically consciousness science and a very peculiar field of computational consciousness science in terms of actual models of what it would be to have a conscious system, even potentially an artificially conscious system. And so I'm coming from more free energy principle and active inference in my background and Arthur's coming from more mainstream machine learning. And so we're coming together to explore this beast from different perspectives and been finding a lot of, I think, very interesting overlaps. And so in terms of albis, the basic idea is to take rebus and expand on it. I love rebus, it's a great framework. And I've been trying to make it, I guess, more flexible in terms of being able to handle like a wider range of assumptions. So not just talking about relaxed beliefs, but also potentially strengthened beliefs and in different combinations. This was always part of the rebus framework and that you can have all sorts of both direct and indirect effects when you're dealing with a large predictive hierarchy or any complex system. You've changed something in one way but then you'll get an effect elsewhere. So you have this hierarchy of predictions. For instance, you can potentially relax beliefs at higher levels and that can indirectly cause strengthened beliefs at lower levels. So you can both get relaxed and strengthen beliefs and psychedelics in all sorts of different combinations. And so with albis, the idea is, so I've been calling as opposed to a belief relaxation when you get this sort of, basically, bottom right here, you might see it. So this red would be basically something like a synchronized complex of neural activity and you've often involving the thalamus, which would be thought to encode your predictions or your top-down stream of priors, which then are suppressing the ascending stream of prediction errors. And so under rebus, you're seeing here relaxed predictions, a smaller complex of synchronized activity and the gates are down, just like Arthur described, more things are coming in. Rebus will sometimes be described in terms of, so when we're talking about rebus, this idea of relaxing beliefs, we're thinking at the functional or computational level. Like David Maher and analyzing the mind, you have these three, I suppose, supervenient levels that they're compatible and synergistic, so computational, functional, like abstractly, what is the system doing? Something like the pseudocode of the system and then implementational, mechanistically how it's realized and then in the middle some sort of algorithmic, intermediate abstraction layer, how are these computations realized via these mechanisms? Some bridging principle description. And so when it comes to rebus, you can adopt it in different ways and relate to it differently. So you can relate to it at the functional level or you can relate to it at the algorithmic or implementational levels in terms of where you're taking the account. So for instance, you might think something like rebus effects is a very valuable model of something like what they call quantum change was psychedelics, but you don't necessarily believe in predictive processing, but you still might be, you could say like a rebus type theorist or you might be thinking in terms of like rebus, like in a more narrow sense of thinking of predictive processing information flows, but then when you get to like the overall person and integrated organism in the agent, you might have different interpretations there. So with the work with Albus basically, the idea would be to take predictive processing and models of psychedelic action and suggest that sometimes not only do we have these relaxation belief landscapes, but we might have the opposite of strengthening. And so specifically if you're activating these five HD2A neurons, which would be at the deeper levels of the cortical sheet and also the deeper levels of the cortical hierarchy would be more inwards and higher up and these would be the units that would loop with the thalamus and thereby be able to form large synchronous integrated complexes, but that when you excite them too much they don't wait for each other to sync up. They fire whenever they can out of sync and this relaxes the belief landscape. However, it's not clear this should happen the whole way through for any amount of increasing the gain on these neurons. So in theory, just increasing the gain a little bit unless that might help them to sync up that might strengthen the belief whether it's a situationally specific suggestion of you in this context or maybe something old coming to the surface and if you will manifesting. So Albus basically in a nutshell would be saying to take rebus and keep going with it and look for different combinations of both strengthened both directly and indirectly strengthened and relaxed beliefs. That's basically it. And so I'm not gonna go into too many more details. Arthur and I just submitted this for publication of Wish us luck. Hopefully the reviewers will be useful. Maybe even kind, but perfectly useful. And yeah, I'll kick it back to Arthur but before I do like this work I think it's immensely important that we get these details right. These are extremely powerful medicines that could revolutionize things or we could have a backlash if we do it wrong if we're careless, if we don't see the risks but also if we're overly beholden to like a single way of viewing things prematurely we might also miss opportunities to help potentially enormous opportunities. For instance, like you might not think of helping people like let's say micro dosing for cognitive decline. You might not have thought of that. And that might be a really big deal that you do think of that. So I guess before going back to Arthur one more thing would be just in terms of at least for me like experientially like the way I sometimes think of it is like with relation to wellbeing and suffering both things matter. Like the strength of your positive beliefs and the strength of your negative beliefs and sometimes the negative beliefs have positive aspects and vice versa. So it's both like you might relax your beliefs and like to use like a Harry Potter metaphor like be somewhat like out of the so much out of the grip of the dementors that usually that under some circumstances might be pursuing you that you have room to be different, to become different to think different things under this different state. But additionally and not mutually exclusively not necessarily incompatibly maybe synergistically the strength of what beliefs you're able to have in that moment and maybe draw upon from the past could be deeply important of your values, the kinds of beliefs that you call upon in times of need and they give you hope. Maybe those are just originally suggested to you right there in the session but maybe there's something else, another belief and that strengthening it could be a good thing. And even without a clinical context like micro dosers for instance it's not clear that exactly they're trying to relax their beliefs like the range of use cases the way we wanna relate them and how we wanna be in the world it seems complicated to me. So I will get back to Arthur. Great, yeah, thanks so much Adam. Okay, so with that context of all this I wanna talk a little bit about how we might understand at a very high level what psychedelics are doing to these two belief landscapes. So before I showed you very quickly I guess I can go back to it. In the rebus model there's only one type of optimization landscape and so there's only one type of canalization and what is supposed to be happening is a relaxation of beliefs which is strong during the acute effect and persists in the post-acute effect. So with Elvis though we might be able to think about all of this a little bit differently. We can in during the acute effects of psychedelics interpret what's happening to the type A landscape as being less of a kind of universal relaxation and more of a kind of destabilization where under certain circumstances beliefs are being strengthened other beliefs are being relaxed. This is changing quite rapidly over the course of time. So a certain belief might be transiently strengthened a few minutes later it may no longer be supported or supportable and this actually functionally enables a kind of exploration of all the of a wide space of belief instantiations during the acute effects of the drug. As far as the interface landscape is concerned post-acute effects as Adam was saying it could be the fact that certain positive beliefs are explored and can be strengthened. Other negative beliefs might be weakened. This is kind of like the ideal psychotherapeutic outcome. It may be on average that something like a general relaxation of beliefs is what you might see but the way that you get there is likely going to be through a kind of much more complicated process. I think what's then interesting and related also to what Adam said about thinking much more broadly about how psychedelics might be used in terms of neurocognitive disorders in addition to kind of like traditional psychopathologies is to think about what's happening to the learning landscape that kind of like underlying synaptic weight landscape and what do we understand is happening here kind of quite clearly is there is neurogenesis, psychopathogenic effects. There's dendritic growth and what is this enabling? Well, this is enabling greater connectivity within and between networks and so kind of from a functional learning perspective, learning theory perspective what this means is that gradient information can flow more easily through these networks. And what this means is that it's easier to get from one part of the network to the other part of the network. There are fewer of these kind of optimization local minima. There are fewer of these kind of like tricky subtle points and what this means kind of like functionally is that this plasticity loss that I discussed earlier is going to be less of an issue. And here it seems to be kind of like regardless of the acute, the actual phenomenological effects. There are very strong downstream effects on this structural connectivity. And I think there's some interesting work suggesting that there may actually be independent mechanisms that are affecting these two landscapes. So there was recent work suggesting, for example, that psychedelics as well as ketamine is able to kind of like particularly bind to receptors in the brain that produce that are a result in the kind of release of the brain-derived neurotropic factor and then the downstream changes in synaptic weighting and that this may be potentially totally independent of the 5-HT2A agonism, right? And this has very interesting implications for future drug design and development. Okay, so I'm going to, before my next slide, turn it over to Adam once more to briefly talk a bit about personality theory and the relationship between psychedelics and personality. Hello again. So over the past few years, I've been doing some work with a personality researcher, Colin DeYoung, on intersections between the way agency is modeled in the Free Energy Principle and Active Inference Framework and the way that persons are considered in different situations in personality science and looking for points of intersection. And so personality is actually an extremely deep concept and if you think of it, it's, I guess we call it like the normal form of a system in terms of what is the way you can describe the most of the system with like the minimal message length, the account for the most of the least. And what we're trying to account for is the states that people tend to occupy. People aren't always happy, people aren't always sad, people aren't always friendly, they're not always angry. It's all sorts of things that are, there's a range that people aren't counting but they tend to occupy some states more than others. And you might call that their personality. And you might call it the personality, many systems would have their own personalities in terms of trying to describe something with their essence. And so they're attracting states. And so with Colin DeYoung's work, he interprets personality and these major domains of personality that people have discovered, a reliable factor structure, in other words, you can chunk the variance, the big five as a trait hierarchy, but where these different traits are interpreted as cybernetic control parameters, specifically cybernetics in terms of the study of goal seeking systems governed by various forms of feedback that different stages of the process of pursuing goals, evaluating outcomes, updating that, and understanding that basically personality as different modes of being in the world for cybernetic systems interpreting them as parameters for goal pursuit, because we are goal seeking systems trying to pursue the goals we value. And so I guess moving to this, and with relation to what we're talking about now, both in terms of personality formation, reformation and the structure of personality, the PF, so okay, to get to the PF factor, within personality theory, this big five traits, the set of big five traits, there's a richer or more detailed factor structure, a hierarchical structure for a person that's been discovered. So above the big five, neuroticism, agreeableness, conscientiousness, you'll have a meditrate of their shared various that's observed called stability, it's also been called alpha by I think digmen, then above the traits of extroversion and openness intellect, you have this meditrate of plasticity, and we're called beta, and the shared variants of these two. So what's interesting is that within the field of, within Hytop, which is this effort to try to characterize psychopathology using the same kind of data-driven methods and personality science, a principled account of the structure of psychopathology, the way that personality has a structure in trying to account for this, there seems to be a common factor across many forms of psychopathology that's referred to as the PF factor, and it tends to be the inverse of stability, it's high neuroticism, low agreeableness, and low conscientiousness, and this is actually what's referred to in the article as this PF factor, thinking of it as a lack of flexibility, you could make a case for it, but more straightforwardly, it seems to be a lack of stability, or something to be a lack of flexibility, you would maybe be thinking of something more like plasticity, like someone having low plasticity, but empirically that's not what's observed, that's a very important detail we should stop and take note of. So you might say that rather than emphasizing the first principle component of psychopathology, it's possible that this canal model is actually focusing on the second factor. It's not a contest, it's still enormously a big factor, like the work of Stephen Hayes and others showing that a lack of flexibility is trans-diagnostic, but just empirically speaking, the first principle component of psychopathology might be more closely related to a lack of stability rather than insufficient plasticity. Well, I guess one more thing to think of in terms of stability and plasticity, like as I said, it's not exactly a contest, they're both important, both need to be considered, and they're deeply interrelated, so Colin describes them as being in the state of something like dynamic tension. Like if you're insufficiently plastic and then flexible, you're not gonna stay stable for long in a changing world, but if you don't have a stable base to move from, it's gonna be hard to be plastic and reconfigure yourself in a sustainable way either. However, you can also see someone being overly rigid and excessive stability in a way that undermines their plasticity or overly flexible in a way that basically is destabilizing. So a dynamic tension, but involving synergies also. And I think in terms of the structure of personality and well-being, this is what we need to keep in mind that both are needed, stability and plasticity. And this is also just within, it's actually what this was inspired, I think of, and Colin's thinking to consider these meditrates this way from machine learning, I think actually reading Grossberg in stability plasticity trade-offs. And that's why he interpreted these meditrates in those terms. So I guess one more thing is, there is a Petersonian backstory here. They don't know if I wanna get into, but we can discuss it later if there's interest. So yeah, back to Arthur. That was great, thanks Adam. Yeah, very, very helpful contacts, which should hopefully make this slide here much more clear to everyone. So given these kind of like two meditrates, stability and plasticity, which as Adam beautifully described you're in this kind of like dynamic tension with one another, we can kind of go back to these two optimization landscapes and these machine learning concepts that I discussed earlier and Colin tried to understand how these things might all relate. And ultimately relate to psychopathology. So I had hinted before that they're were these kind of two types of canonization, we might say. One being overfitting and another being plasticity loss. And we can interpret these as kind of like being extremes, the kind of like canalized end of how these two optimization landscapes might manifest. So in the Taipei inference landscape, this looks like overfitting on one end, heavy canonization and underfitting on the other end. In the middle of the course, underfitting on the other end, in the machine learning context, right? So overfitting being having very strong rigid beliefs that are in a particular environmental context, adaptive, that are very unlikely to be adapted in any other context. Underfitting being having set of beliefs which are not adapted to the current context and also unlikely to be adapted to any other context. Awesome. As far as stability is concerned, I think we can understand stability. This Metatrate is kind of representing the balance of these two things, right? You have high stability when you are neither overfit to a particular context or underfit. So this means your well-fit to this context here and now and you're also likely to be well-fit on a future context. I think this is also in complexity science sometimes described as metastability, right? So you're at a stable point that will also likely be able to take you to other stable points in the future. And this kind of like ability to stay stable over time is then a ability Metatrate. We can then look at the type B, the learning landscape where the kind of extreme version of canalization is this plasticity loss where it's very difficult to adapt to new contexts. The inverse of this is then catastrophic forgetting where the problem is not one of adapting to new contexts. The problem is one of meaningfully retaining adaptive strategies and beliefs from previous contexts. And so obviously, nine of these are desirable. What is desirable is some kind of balance between these two, right? Or one is able to adapt just enough to new contexts without being without forgetting previously learned things. But here the relationship to the Metatrate of Plasticity is a bit different that stability, right? So as is being represented on this graph, the more plastic you are basically the less analyzed you are in the space, the more easy it is to move around. But actually extreme versions of this end up manifesting is catastrophic forgetting, at least in the machine learning context. And I think this in particular has relevance for psychopathyology and relevance for this idea maybe complicates this idea, right that psychedelic therapy is in and of itself likely to always lead to positive outcomes. Okay, and we can attempt to understand this by making an effort at trying to map or understand various forms of psychology in the context of these two different learning landscapes. And so before I get into this too much, I think it's important to say that this is very much a kind of like preliminary and provisional attempt at making sense of these things. I hope that this is kind of the start of a a much more fleshed out dialogue looking at how we can kind of understand psychopathyology with respect to measurement and concepts. But I think this at least gives like a good starting point for maybe trying to understand where and why psychedelic therapy is useful, isn't useful which is a bit more nuanced than this simple idea of like psychopathyology means canalization and psychedelic therapy means reducing canalization, right? Okay, so let's try to make sense of this a little bit. So for the kind of classics that psychopathologies for which there has been a lot of great research done with psychedelic therapy and a lot of clear benefit has been seen, I think we can kind of class these within this upper left-hand corner here where these are disorders of both overfitting which means in the moment over the strong beliefs get represented, you know, these are strong attractor points in the neural activation landscape and the plasticity loss in the type B landscape which means that not only is a certain kind of dynamics instantiated, right? Which make, for example, in the case of medri-dispressive disorder extremely negative self-evaluations very common but it's the case that even given the correct changes in environmental context, even given in the case of psychedelic assisted therapy a lot of the work with major depressive disorder is being done with treatment resistant depression, right? And what is treatment resistant depression but a kind of a set of beliefs a set of neural activity patterns which are resistant to adaptation to new environmental context being, you know, whatever sort of psychotherapy or previous drug was used. So this is a high plasticity loss hyperbidding, right? I think we can then, and these are, you know, kind of fortunately, you know, for psychedelic therapy and I think maybe for psychiatry in general these are all very, very common and very, very costly psychopathologies, right? So if it is the case that psychedelic therapy can treat these effectively, that's fantastic. But I think we can maybe look at it more complicated. You know, if we start to expand other psychopathologies the story becomes a little bit more complicated, right? So let's look, we can look at the bottom left-hand corner here where we have cases of in the inference landscape still overfitting. But in the learning landscape, something else happening form of catastrophic forgetting or form of kind of drift of the belief landscape that's being instantiated here, right? So I think two kind of representative pathologies here typically referred to as some form of bipolar disorder or some form of our borderline personality disorder, right? So in both of these cases, in the moment to moment neural activity as it unfolds there are very strong sets of beliefs being instantiated at any given time and these are very difficult to change, right? But the key differentiating point here being that over longer timescales over the timescales of days or weeks the kinds of beliefs that are instantiated change quite a lot, right? So someone who's bipolar, if they're in a manic episode or in a depressive episode the nature of the belief landscape being instantiated is very different. For borderline individuals from a daily or weekly basis the nature of one's beliefs about the people that are close to done can change very rapidly very extremely, very in certain cases. And so here the benefit of psychedelic therapy is a bit more mixed where there is likely the case that to the extent to which the kind of overt type A beliefs can be updated and changed to be more positive be hopefully more relaxed, that's positive but to the extent to which plasticity may be being increased when it's not actually needed when there's not actually like meaningful therapeutic benefits of that it's a bit questionable. I think there's clear borderline or rather bipolar for example is often individuals with bipolar are often excluded from psychedelic files under the assumption for example that psychedelics may trigger a manic episode or may trigger psychosis and I think that is a valid concern. So on kind of that topic we can then this bottom right hand corner would be cases of both under footing and catastrophic forgetting with I think schizophrenia being kind of like the most clear manifestation of this and here there's very little work currently looking at trying to treat schizophrenia with psychedelic therapy I think for pretty good reasons though I would say to the extent to which there are secondary manifestations secondary issues like depression and anxiety that people with schizophrenia might have these actually may be treatable with some form of psychedelic therapy. And then lastly this upper right hand corner here would correspond to under fitting in the Taipei landscape along with elasticity loss. So I think a good example of this for example is something like ADHD where there are not very particular strong beliefs that are being instantiated at any given time but it's difficult to change these out of beliefs nonetheless. I think here there's a case where it's not totally clear how the benefit will actually play out. In both of these cases I've heard interesting work around like micro dosing may be of interest. I think yeah more research has to be done to be seen. And then what would micro dosing be doing? Micro dosing would be like very slowly kind of like decreasing plasticity loss just a little bit potentially also decreasing under fitting as Adam had kind of suggested in the Albus model. And here we have just kind of like a brief overview of more or less everything I've described about how these two landscapes are different. I think most of these things I've said before pretty clearly, right? So the Taipei landscape it corresponds to the neural activity pattern. This is the hidden state of the RNN. Over-canalization is going to be stereotyped mental circuits very strongly held beliefs under-canalization is going to be these inconsistently deployed circuits. It looks like over-fitting in the canalized case under-fitting in the non-canalized case and then correspond to stability in the Metatrate. I hadn't discussed this before but I think there's a relationship between cognitive flexibility and this inference landscape. Psychedelics are mainly going to affect the work at the acute level care. And this is by directly exciting the neurons or inducing forms of how to be in plasticity and the outcome is going to be a mixture of relaxation and strengthening of the kind that Adam described in the Albus case. The Taipei landscape then is working on the synaptic weights and these are the actual network parameters or RNN model over-canalization is this inability to learn or adapt over time. Under-canalization is the inability to retain information, right? And those correspond to plasticity loss and catastrophic forgetting. This also is kind of the traditional plasticity Metatrate which I think corresponds to this longer term constructed psychological flexibility. The effects are going to mainly be mainly post-acute with psychedelics and these are the more downstream metaplastic effects, the cycloplastogenic effects. And I think on the whole the outcome, the effect here is a relaxation of canalization. And yeah, these are just the references. So those are the slides I had that's kind of the overview of the deep-canals model. I think now I'll turn it over to Mike who's gonna share some thoughts both about this work and some of his own work and kind of this approach to understanding psychedelics more broadly. Great, yeah, thank you for that overview. That was great. So I guess I'm just in general very optimistic about this corner of neuroscience that there are these very intuitive models that are very sort of clinically significant. And I also think there is a lot of responsibility to sort of get things right here as well as I don't know that. So yeah, I have just a few comments. And for the viewers, I'm not speaking for Arthur and Adam, I'm just grateful for the invitation to join the discussion. So I wanted to note quickly three research themes dealing with the implementation of this class of model. So just broadly kind of pointing at both deep-canals and androids and albis. So in 2019 I wrote neural annealing and that was sort of developing the canalization frame within the context of energy buildup in the eigen modes of the system. So you have the nice annealing metaphor, you have a temperature parameter and entropic disintegration. Robin Carter-Terra said a beautiful paper on that to interrupt the brain. And then you sort of have a post-heating result where the microstructure of your system is sort of a result of the sort of how hot and how long the heating was. So it's an alternative take on belief relaxation. And I do think that like there are some pluses in terms of sort of explicitly making the annealing metaphor. And I think that's like two things that I'd point out. One, top-down models or top-down predictive models as energy sinks. And eigen modes as the places where energy can build up for almost sort of hide from top-down models. So you get this phenomenon of semantically neutral energy or semantically illegible energy you can say, which is sort of kind of how perhaps psychedelics, music and rotation all sort of allow energy. The energy that they contribute to the system can evade the top-down models. And yeah, just to note, one item that I'm looking at now is what is deep canals has the inference landscape versus learning landscape distinction. And how would one sort of build that distinction within the neural annealing framework? And it's one metaphor that I'm considering as the self as sort of a magnet to sort of apply a constraint, a uniform focusing constraint on the system as it's cooling. The second item that I wanted to mention, so I have this piece, autism as a disorder of dimensionality. And this brings in the concept of network density that it may very well be that people on the spectrum have literally more neurons and or more synapses. And there's this great Michelangelo quote, I saw the angel in the stone and I set him free. And sort of the thicker networks you have, the more stone you have to work with, but you actually have to do the work. So you sort of have to do sort of, well, normally you sort of inherit these highly pre-trained optimized circuits, most like ASICs, applications specific integrated circuit. And then this is kind of more an untrained network and you have to train yourself. And within the deep canals framework, this would show up as both under fitting and higher capacity for plasticity. So I think that's gonna be an interesting thread to pull. Finally, so I just posted that last week, principles of vasocomputation and as an implementation level account of active inference. And this is a big rabbit hole. I don't think we have time to go down today, but I just comment that I really liked the way deep canals explicitly split out the learning landscape and the inference landscape. That I think that making this distinction is increasingly going to make sense back to Arthur. Yeah. Awesome. Well, thank you all for the presentations. Arthur and Adam, feel free to pick up where Michael left off or I can bring in some questions from the chat or from what I've written down. Just a quick comment. I do think that in the deep canals framework, it's going to be interesting to sort of look into metabolic demands and like sort of different like metabolic disorders are pretty hot in brain science right now. And I do think that it'll be interesting to look at sort of what the metabolic profile is in terms of each of the four items on quadrant. If one quadrant or another quadrant is particularly associated with more metabolic demands in the system. Arthur. Yeah. Yeah. No, that's fascinating. I was I'm actually, I'd love to kind of hear a little bit more about this comment you made about the cell as a magnet that's applying these constraints. Oh, yeah. So there's a the like psychedelics, the effects are not or sort of the you sort of get pulled back into position. That's a position after after some time. And, you know, I can't say that this is a full full theory at this point. But in the neuroscience notation, 2018, I talked about the self as kind of the the God of the gaps in terms of aligning different scales of like you have circuits which kind of pull in different ways. And you have the mesoscale systems which pull in different ways. And you have the overall brain wide systems that pull in different ways. And kind of the self is sort of the thing that tries to make everything work together. And an alternate frame would be sort of this self is just kind of this slow gravity toward this alignment. Yeah. I don't have anything particularly more to share. That's it. Yeah. It's interesting. Well, to combine the heating up and the magnetism, different metals and connections there to alchemy and everything, different metals when heated up and cooled down in specific ways, changed their magnetic profile. So, of course, this is all just concordances, but they're very interesting. And then also to the metabolic angle, that reminded me of the trilogy by Dale Pendel with pharmacnosis and the other two books where treated alongside traditional psychoactives, we could say like caffeine and nicotine and some of the other substances that were mentioned here. Also, there's tea, sugar, and these are treated as psychedelic or at least as in the same pantheon of neurocognitive modifying in the sense that the destinies of individuals and of empires ride on the sense making and decision making influence of sugar, for example. So, it definitely broadens our perspective to include the so-called classical molecules, like LSD, but also maybe we could think more broadly about antioxidants or about different lipids or about different metabolic connections. Michael Pollan's recent book, This Is Your Mind On Plants, I think kind of has that what you're saying in mind because it includes caffeine. Not like introduces a psychedelic, but just discussed among also psychedelics. I remember he focuses on mescaline, but it is a different way of being in the world, potentially radically different, like if caffeine is in your life or it is not. It's a different life world. And different extended niche. I'll go to a question in the live chat so anyone can give a thought. So, Matt asks, is there any way how to approach the individual reactions to psychedelics in terms of positive or negative results of treatment using the ideas of deep canals? So, how can we approach predictions and explanations at the individual scale with what you're sharing today? One thing for, this is kind of like a meta comment, but if we're actually going to do computational psychiatry, and we actually are going to, and precision medicine. So, the goal being to create models of sufficient detail that they could actually guide our practice and guide our practice in a way where we know for an individual what they might need in any given context. So, precision medicine with respect to like standard antidepressants, you know, people sometimes will go for a long time before they find ones that work and sometimes things are made worse and they might deal with antidepressant withdrawal as they're looking for the right one. And if they're in a state of acute distress, that could be a hard lot. And so, like if you can help to speed up that process of finding what works, the ability and what might hurt also to avoid that. I mean, the ability to help people is immense. And so, for that, it's like part of this, so the meta comment is like, I have to just say I'm not convinced that an active inference, we're doing the job. I think we're not acknowledging our uncertainty, and I think the models are under specified with respect to the complexity of the effects who would expect on different scales and in characterizing the systems, and we need to be more patient and modest. But what we don't know, before we've gone rushing in headlong into telling the rest of psychiatry how they should comport themselves and what models they should use based on psychedelic science and plausible models, hopeful models, but very early days. And with the reality being complex in the ways we're seeing here, like this was not an easy set of lectures or a set of talks. Like what we're going through, this is like, and yeah, if we're actually going to be able to handle these powerful medicines responsibly, we've got to go into the weeds. And so the one more meta comment was like, I think I saw like some say, oh, how can something like help everything? You're like, yeah, I don't know if this is right, but it's almost like a violation of like no free lunch theorems or something. It's like, how could something just be good across the board without, you know, something is off. You've calculated wrong somewhere. You had to pay the price either in like some trade off that wasn't acknowledged or, but that being said, there are some things which are generally good for a lot of things. Like being more flexible, like if you can get something like metastability, like, so there's almost like, I think I'm a church member of the, I'm a faithful member of the church of criticality. I think Mark Miller might have indoctrinated me with like optimal grips, but basically the ability to have like an optimal grip on existence in your soul, not too tight, not too loose and being able to move around and being able to move around and be like flexible and itinerant. It's like something kind of like mindfulness is just like Sriracha for the soul. It's good for everything almost. You just get to put on this, put on that, makes it better. I buy that there might be some things that are good across the board, but I wouldn't call the psychedelics necessarily putting you in this like state of optimal criticality or anything. It's like, in general, like the rhetoric I'm uncomfortable with of saying psychedelics do X. I would prefer us to say like, which psychedelics tend to do X, but just a little more qualifier, like not weasel words, but just like a little bit more like, this psychedelic might do X in this context with this dose and this setting, but that being said, there do also those seem to be across the board, broad sweeping, not super, but overarching statements that we can make that express a lot with a little and so like I don't want to be so much of like a splitter and like a prickly person finding all the details that like these powerful things we can basically give to people to guide like our research. I don't want to leave that on table either. So yeah, so I guess some, that's a bunch of metal of a commentary that did not address the thing. So. I think that's just so it's very helpful. I think to have all that context with respect to the question itself, I would maybe offer this, which might be helpful, which is, you know, I think you mentioned it, Adam, and it's very frequently mentioned this idea of sudden setting, right, as being like heavily predictive of the quality of the acute psychedelic experience, and often also to some extent the quality of the post-acute experience. And I think we can understand this in the deep canals model quite, quite straightforwardly. So I'll just quickly pull back up one of these slides here. Yeah, right. So what is sudden setting? So set is mindset, which would basically be some combination of the current, your current position in the synaptic weight landscape and the current state of your neural activation landscape. Right. So your place in both landscapes is your mindset and the setting is the, what is generating this input, right? What is generating the X into the network? And that's, you know, some complicated function of the environment, right? And then we can understand, right, the evolution of movement through the inference landscape during the acute effects of the drug as then kind of like married directly and straightforwardly being a function of these two things, right? Because both of these things are basically what's affecting how this landscape evolves over time and how you move through the landscape. And so, you know, with this in mind, I guess we can then start to characterize and think about like, okay, well, what kinds of, you know, we can maybe be a little bit more mechanistic about like how does input from the environment, the setting actually, given, you know, that we relatively fix the actual synaptic weight landscape, how is that changing the dynamics over time? Or the other way around, right? Like how are, is fixing the input, the environmental context relatively stable and looking at individuals who are at different fixed points at different points of time going to change the dynamics of the event, right? And I think given that this is a highly stochastic process, given that these dynamics at Albus are at play, right? Where like certain beliefs are being strengthened, other beliefs are being relaxed. I think at least having this hold on, okay, well, setting setting is really important and it's important because like these are explicitly conditioning the way in which both the landscape, the infant's landscape is shaped and how you move around it and how it's being, you know, dynamically perturbed. Yeah, hopefully that answers a little bit the question. Nice. I just want to jump in there just thinking about biomarkers. That, I mean, if we're looking for something that's of high clinical specificity, we want to sort of be able to measure what's going on at a very careful, in a very careful way. So like, first of all, what is temperature parameter? You know, we talk about, okay, psychedelics increase the temperature parameter, but like how high and do different substances and dosages sort of differ in interesting ways? Probably yes. And then also sort of measuring temperature parameter, measuring the temperature parameter in different parts of the body. So you might have a different temperature, like not literal temperature, but computational like entropic integration in different parts of your body and also the plasticity status in different networks throughout your body. So I guess I see a big frontier here as doing some careful thinking around biomarkers and proxies and inferring these sort of things that we are talking about as core, and I think they are core, and we should be able to find a good marker for them. Yeah, that's totally correct. And yeah, I just want to repeat it for my own sake, which is that these two things, which I think were maybe, one is I think already a core part of the deep penance work, which is that I was showing these single images of these toy optimization landscapes. But of course, the human being, the nervous system as multiple, many of these landscapes, hierarchically organized, and they influence each other in a complicated way. And the project for it is to understand where these landscapes are, what they look like, and how we can find biomarkers to them. And again, if these two kinds of characterizations are true, this overfitting, underfitting, plasticity loss, and a catastrophic beginning, then it's a very meaningful project to be able to quantify that objectively somehow and individual. And that's when you can start to get to this point that Adam was talking about precision medicine and understanding, okay, we'll actually turn that right. These networks are penalized in this kind of way. And because of that, we can say with much more certainty that this kind of psychedelic dimension is going to likely be helpful. So I'm wondering, I don't know if this potentially brings in some of the work with like, vascular and neurovascular and visceral integration that you're doing, but like, so there's like a sense in the free energy principle where the overall phenotype and the extended phenotype in terms of its constructing and coupling with and constructing its eco niche, there's its success in manifesting this phenotype. There is one free energy functional for this overall effort as a kind of prediction that's enacted through the dynamics. But you could also like break it down in terms of like subsystems with like local gradients and like local auto poetic processes that are in like, they're brought together into an overall system, hopefully with synergies, and they have to be an organism and an agent we call this like a thing to the extent it has thickness. It's coming together. But I guess, so you could, I guess think of maybe like a energy landscape, you know, something like, let's just say like right now in this moment, like my explicitly felt beliefs, but my body also though might have a different opinion, different parts of it in terms of like what it's implicitly optimizing its revealed preference from its inaction that may or may not correspond to how I feel about it. I might be different to the opinion. I'm not even be tuning in. I might be ignoring it because I don't want to. And so it's kind of like the score kept by the body might have its own, be describable as its own like information geometry, which is being tracked to degrees and is in different degrees of alignment with different ones and different levels of control hierarchy. So plausible, right? Absolutely. I think so. Yeah. And like I think of, I think of the body as sort of a parliament of different networks and you know, alignment is the scarce resource. Coordination is always the scarce resource and like how does the body keep alignment and when does it go wrong too? And I guess I see this as sort of, I mean, harmonically mediated. Each part of the body has some choice of what else in the body to listen to. And depending on the particular sort of location in the information topology, different parts of the body have different carrots and sticks by which to try to control other parts of the body. So if your stomach is very unhappy, it's going to start grumbling and then it's, it'll let the body know, okay, I need some food. Like I really seriously need some food. And then the rest of the body, they can sort of turn away from it and then it'll just get louder and then you have to turn it away from it more and it'll get louder and that's a, of course a negative dynamic. If it gets what it wants, maybe it sort of gives a very harmonious hum and sort of tries to give, give carrots the rest of the body. So yeah, I just, there's an ecosystem there. I'll go to another question in the live chat. So Burt asks, what do you think of the average of solving within a generative hierarchy and the impact of temporary openness moving it upwards? And then I asked to clarify, what do you mean the average of solving? And Burt wrote, where predictions meet incoming signals like disgust and rebus, where top-down predictions are softened in precision and thus solved for. I think Arthur will be able to get into the details of that probably better than I would. But the one thing I'm wondering there is sometimes like, again, I love the entropic brain, but sometimes I wonder in terms of like the connections, whether I don't, so how we're having more or less entropy and whether that matters the how. So for instance, like you can think of like a free energy landscape that's been flattened, or you can think of one where like you're better able and you have like more like energy for like this, you know, the manifold activities better of like climbing up or down the peaks that are there for the more enduring belief landscape. So it's like in the moment, it could be that my previous way of thinking is unavailable to me, and I no longer believe those things at any level, or I can support more novel inferences that I wouldn't otherwise. I can kind of move into an area of mind space that would otherwise be difficult, potentially because of greater energy into the system, greater throughput. Yeah, so as I think I understand the question, right? It's basically asking about trying to optimize for this free energy variable. And then the context of deep canals, right? Like this is the loss function that's driving the optimization of the learning landscape, the synaptic wave landscape, right? So like the entire connectivity structure of the nervous system is like trying to get to this point of like low free energy. And I think there's lots of, there's, you know, for any given individual in a given environment, assuming the environment is relatively fixed, right? There's like an optimal. There's a global optimal. And it's very likely that the individual is in some local optima that is maybe close to it, maybe far from it. But that's just inherently, you know, what we're, you know, within the context of active inputs. That's what we're assuming that the nervous system is doing, is doing this gradient system in this space. And openness then is going to correspond to this, plus this theme at a trade or be one of the components of it. And it is in the deep canals model, then going to correspond to an ability to move around this optimization landscape more easily, right? It's going to be a flattening of the landscape or more precisely, I guess, a smoothing of the landscape, such that it's less likely that the individual, the nervous system, won't get stuck in this local optima. So maybe it won't mean that the individual will arrive at the global optima. I think it's difficult to say what that even, you know, might be or however, how individuals might get to such a place at all. But I think it is likely that the individual will be more likely to find a more global optima and more easily move from one option to another. And I think that's kind of what this openness is corresponding to. Yeah. It's okay. Oh, I... So in Mural Nealing, I mentioned the idea of psychedelic extrapolated deletion, kind of a play on a coherent extrapolated deletion from the AI safety crowd. And it's this sort of sense that if temporarily you have more control over sort of low level shaping of your cognition, your emotion, and maybe sort of based on this attractor that we call the self, or I guess like based on what is a good question, then what are good signs that you should be trusted with that power or you should trust yourself with that power rather? And what are telltale signs that you shouldn't trust yourself with that power? And I don't have an answer, but I think it's an interesting question. Yes. I am normal and can be trusted with a key exception. Let me in. I have a short question. Where is action in this model? So I imagine many places are filled into that, but for what I was thinking with Albus is that actually the 5-H D2A system, I talk about two regimes potentially. One is more towards overt inaction and modifying the world and strengthen beliefs potentially of that variety and the other might be more of adapting to the world. So Robin has a really excellent paper called A Tale of Two Receptors where he talks about 5-HT1A receptors, the ones that antidepressants tend to act on because they tend to reduce this excitatory signaling, talking about them as a kind of passive coping. And so this would correspond, and there's some machine learning models of this by Diane and colleagues where they talk about a ponency between dopamine and serotonin and they're focusing on the 1A as a kind of patience-promoting parameter, keeping you from confidently charging into your policy. So that would tend to inhibit, I guess, patterns of either overt inaction or potentially even some kinds of mental simulation too. If it's like, you could maybe even think of, even if you're not overtly inactive, enacting something externally, there might be more or less a brave type styles of thoughts that you might be willing to entertain. Going places you might not otherwise be willing to go and being able to handle the swings and expected pre-energy that might happen as you try to pursue a given path of policies. So, but the idea would be that there might be something more like a micro-dosing mesodosing regime, which might be more of what initially sculpted the 5-HG2A system, which evolved with the gene duplication event around the time of the Cambrian explosion with the advent of jawed fishes. I'm personally interpreting this as having the significance of evolving in the context of predator prey arms races. I suspect that could be wrong, but that basically it's locally trying to strengthen patterns of policy pursuit to pursue particular goals in a given context. And this to come back to what was actually earlier that worked by Robin of like a kind of active coping. But it seems with Rebus, there's almost like, not that kind of story and that it's more you're adjusting through all and actually if under a very altered state you are overtly enacting, that's actually kind of a dangerous thing. I don't know how much to make of this that the 5-HG2A receptors are less expressed on like motor cortex. I don't think we know enough to make too much of it, but it's like some dreams you might not want to act out as they're happening, but they still might powerfully impact you. So I'm currently actually wondering whether this other Rebus-like regime might have been specifically sculpted in the context of bonding and actually opening the self to include others and that might have been some of the primary selective pressures. I don't know how to think of that in terms of an action, but one more detail that might kind of like, not sure what to make of it, that kind of goes through both of those, excuse the phone, would be anthropologically, I believe, like psychedelics are oftentimes used in ways that would be both like creating a kind of group solidarity and might be used in the context of pursuing goals together in the world. So for instance, like a hunting party, they might all do psychedelics together or something that looks a lot like breath work before going out, maybe going to war. So that's a very different type that kind of includes both a relaxation of the normal boundaries of the self and a kind of like gearing you up for pursuing goals in the world. But there's got to be another story of like, so Arthur's expertise is machine learning and so it's like, you want these systems to act intelligently and not act in other times and so basically now we're in the language of policies and when you deploy them and how and it's complicated. Yeah, I would say that action manifests itself at a few different places. It's true, I didn't mention it here. Yeah, I think it's very important obviously. I mean, I think what the kind of like most implicit place where it exists is that these belief landscapes are being instantiated ultimately for the sake of adaptive behavior, adaptive behavior, adaptive action in some sort of environmental context and that's the kind of like framework, deep learning framework that we're using or like the artificial agent framework. Right, the assumption is that this system is supporting some sort of agent that's embodied in some sort of environment which is changing over time such that it can either be adaptive or not adaptive. I think then the question of how action actually plays into the evolution of these belief landscapes during psychedelics or not, acting in the world is a very clear way to modify one's setting, right? So, action is a way of modifying the function that's providing the sensory input into the system which is going to change how the activity patterns are evolving over time, right, which is going to affect the acute quality of the trip, affect how beliefs are either being strengthened or not. And I think, you know, it's already clear that there's a kind of whole set of like folk psychology, right, that's been built around like what is or isn't a good way to cultivate a helpful setting and that of course all is a function of action in the world. I think then there's also this kind of more subtler form of action which is how one thinks, how one deploys attention, right, essentially during this experience and I think this deployment of attention has a huge effect on whether there end up being rebus or seabus like effects, right? So, whether one allows oneself to go down certain thought puppet holes, whether one allows oneself to get kind of stuck in certain kinds of thought patterns. I think this form of mental action, yeah, has a very big downstream effect. And I think a lot of, I'd love to see much more work on this. I think it's very important. I think it's kind of essential. And I think this is kind of one of the most interesting things about these psychedelic experiences is that in some sense, like the most extreme rebus model and the most extreme seabus model could be true for two different individuals, right? Someone could say, you know, I went to hell and I was there for eight hours straight or say, you know, I had a completely selfless experience. I let go of all beliefs and experienced nothing for eight hours straight and it was great, right? And like clearly those involved like two very different neural regimes. And I think the nature and obviously, given that they involve the same drug and they involve not such a different setting, it's obvious that they have something very important to do with how attention is deployed during this experience, yeah. Yeah, just, I do think that this frame of attention, I mean, attention is so important for everything. And I do think that there's something interesting kind of digging into the attention and albus framework that, yeah, I'm really curious to see how that'll develop. I would also say that in terms of active inference and action, there is a cost to holding an attention. And then you once, you know, you discharge your intention and you can release that cost. So I guess I would have to think about how to apply that to the albus framework. Fred Barrett, I don't know if he's published it yet, but this is kind of like a follow-up to a claustrially central model that Arthur alluded to, is emphasizing basically a lack of executive control and placing that as central to the psychedelic spirit, like it's not controlling your mental action in directed ways as much. And then trying to come from more of the types of cognition and phenomenology sort of psychedelics based on just releasing of the normal, more agentic modes of policy selection around attention as mental action, although not using active inference speak. So. Yes. Yeah, very interesting, Adam. Your introduction there of, of course, attention as covert action, which was echoed by Arthur and Michael, both of you said it really introduces a case where we can use active inference to find continuity between what's happening inside the agent and outside the agent, because on one hand we could kind of take a belief oriented frame and we could say, well, what's happening inside the agent are beliefs about things, beliefs about hidden states, beliefs about observations, beliefs about policy. So that's kind of a belief oriented view where everything is framed basically cognitively, or we could take a very action oriented view and then we could think about the interface in a more inactive way. And then also even the realm of where, where many active entities dare not tread in mental action with attention as covert action. And so then we have like a, a total active account that is mathematically isomorphic with the total belief based account. They're both just maps. They're not the territory. The territory is the person in the room. So all of this is just talking about the formalism and how we choose to interpret the formalism in terms of it being more inferential or it being more active. And so there's a lot of ways to go. I guess in our closing minutes, I would love to hear what are the next moves or what are some areas that you would hope that people look into. Michael, and then the authors. Okay. So I know is, I think my, my favorite mental move is to try to drop down on our level of abstraction. Okay. How, how is this implemented? What's, what's going on here? And then I do think that details of implementation are really generative to look into. And like the biomarkers and the proxy is for, you know, figuring out how to have parameterized. Okay. Exactly how hot is, are we talking about with this temperature parameter and, and so on in different parts of the body? I have kind of my, my niche that I'm, I'm working on the vasomuscular system, which is a rabbit hole. But yeah, I just think that like, I do think that this, like the deep canals framework specifically and the canalization paradigm in general, it's good enough that we should be able to dig into mechanisms. So I'm, I'm looking forward to that. Great. Adam or Arthur. So, um, or Arthur, you can go first. Okay. Yeah, okay. Sure. Yeah. I mean, first I want to echo this idea. I think mechanism is very important. This is one of, you know, the people we've talked to about, talked with about deep canals. This idea of like, well, this is great. We have these two axes. You know, it'd be wonderful if there are some like biological correlates you could look for. I think that's incredibly important. My background is also in machine learning, in deep learning. So I'm currently doing some work around like simulating some of these things. I think, especially for these albus dynamics, being able to kind of like study under what circumstances you get these different kinds of effects, even in kind of like, simple toy models is kind of important to see how they unfold. And then I'm also, you know, at the end, I think it's really important. It's very interesting to me is the role of the deployment of attention in this. And there's lines of work where people are trying, you know, various researchers are trying to connect, for example, meditation to psychedelic research and kind of certain people are trying to tell unified stories about what's going on in the brain and people meditate, what's going on in their psychedelics. Some of them using these active imprints models. But I think in all of these cases, the deployment of attention and how the deployment of attention changes over time is kind of a key aspect of this. And it's not as present in the deep canals model right now. But I think any really robust future kind of like model of these things is going to have to talk about attention quite a bit. I guess next steps would be seeing how the review process goes. But also, I guess, helping Arthur to establish a field of psychedelic machine learning, psychedelic inspired machine learning where we use, we try to make systems more flexible and capable better. They can handle edge cases better, imagine better in terms of real time policy selection, learn in ways that not too much, not too little. And then maybe, so that's useful, eventually earlier, like you have a driver with this vehicle, do I train slain or don't I? You send it back to the Tesla dojo. What do you want it to learn? What do you not want it to learn? But then down the line, well, before we even get there, as you're going into this, you could see the capacity for like a rich bidirectional flow between machine learning and psychedelic inspired machine learning. So it's like you get a bunch of machine learners together and a bunch of people who know a lot about the cognitive science and biology and maybe phenomenology of psychedelics. And in addition to designing better machine learning systems, they could probably tell you more about how psychedelics act. And so you can have this kind of virtuous cycle between the biology and the field. But you keep going with that. I guess before we get there, we'll keep going, but before we get there, very interested in the implications for best practices. So for instance, like there might be like when, so oftentimes you might want semantically neutral energy to kind of get them, you're just turning up the heat. You're just maybe at first breaking up, you know, I call them defenses, but the usual structure. But then the question is, at what point do you try to direct the canalization process? Do you wait till after? Or maybe when the, when the metal is hot, is that a really good time to do it? And so these are the kinds and maybe there's no one who will question this might depend on the substance, the set, the setting, the dose, the person. So getting clarity on that could be faithful for many people in terms of the outcomes they have. And I specifically what I'm gunning for is I'm on a mission to make compassion meditation, treatment as usual for all psychedelic psychotherapy and preferably all therapy. And so the question is, is like if you're in a state of enhanced imaginative capacity, maybe you're having more ability to tune into your entire receptive system, whether you've just increased the channels or maybe you're better capable of because maybe your inflammation went down a bit and you can tune in better because it doesn't feel, it feels better to do so. Some combination, that might be a really good opportunity to do certain meditations that might update you in particular ways. Like try to direct the re-annealing process, maybe not just like afterwards integration, but during the session. And so if we could get clarity on something like how to canalize and re-canalize and the like annealing schedules, it's going to be complicated. That's I think where we need to head. And then finally, for all of this, I'm very, that would be enough. But I'm re-centering the conversation on psychedelics on basically meaning and connection would be enough. But I'm wondering down the line, when we talk of alignment, we're talking about our alignment with our values and maybe the alignments of machines with our values. And so does this actually end up going to some of the conversations that are happening now about increasingly advanced AI and our relationships to it? Could there be a story of canalization and re-canalization? You want some things to hold fast and you want some things to be more fluid. This might be, it's always been a timely conversation. We've always been dealing with this with us, the alignment problem of us, which we often fail. But the alignment problem with these technologies were developing. Wow. Well, so many great topics. Thank you all. All I have in closing, turn on, tune in, act and first serve. Thank you. See you all next time.