{
  "video_id": "PNYWi996Beg",
  "channel_slug": "machinelearningstreettalk",
  "channel_handle": "machinelearningstreettalk",
  "title": "Your Brain Is a Prediction Machine, Not a Processor — Karl Friston",
  "duration_seconds": 4900.0,
  "url": "https://www.youtube.com/watch?v=PNYWi996Beg",
  "upload_date": "",
  "transcript": "Keith has been talking about this moment\nwhere we had a glass of cherry with\nProfessor Friston for about the last\nfour years and our dream has been\nrealized today.\nAbsolutely. So, pleasure to meet you.\nIt's absolutely been a pleasure.\nCheers.\nYou mentioned consciousness which you\nshouldn't really do with me but\nif we stay here for long enough for more\nthan 5 minutes uh we will ultimately\nbecome completely entangled. We will not\nbecome one. We'll still have a We'll\nstill have Well, actually Chris Fields\nthinks you would become one. But uh is\nevolution an intelligent process? It's\ncertainly a free energy minimizing\nprocess. You know, it is just basing\nmodel selection.\nWell, it's like is is the planet\nintelligent.\nI don't see the weather planning. I\ndon't see evolution planning. It doesn't\nthink about its future.\nLife is the intensive property of matter\nand intelligence is the extensive\nproperty. not an ensemble of universes.\nUm, well, of course, for the purpose of\nthis argument, uh, there's only one\nstate of the univer at any one time. I'm\nnot sure this is correct, and I don't\nwant to disappoint you. So,\ndon't worry. It's well beyond my\nlifetime when that will happen. If\nyou're going to make a difference, it's\nreally understanding how to use all the\nmarvelous engineering that we've\nwitnessed in terms of generative AI and\nlarge language models and the like um in\nthe service of understanding the\nprinciples that underwrite natural\nintelligence and deploying that\nunderstanding either to make life better\nfun\num or to underwrite sustainability in a\nslightly more sort of um uh um political\nway or to deploy it. you which is why I\ngot into this game in the context of um\nmental well-being and um um\ncomputational psychiatry to quote\nFineman that which I cannot create I do\nnot understand that means we have to be\nable to create things that suffer\num which means that we need to\nunderstand the principles of natural\nintelligence\nI don't know if you saw I interviewed um\nYoshio Benjio\nall right\nand um I used the term epistemic\nforaging a few times and he said I I\nlove that term I love that term And\nwhere did it come from? Professor\nFriston.\nAbsolutely. Epistemic foraging. It's\nwhat we do.\nIt's what we do.\nYeah. It's epistemic foraging. It's\nepistmic foraging.\nThere are certain causal forces in our\nuniverse that are at a distance. Those\nusually um mediated by electromagnetic\nradiation or magnetic fields in this\nparticular instance. And that really\nsort of complicates the way in which we\nconstrue all the cause effect structures\nthat we have to model in our brain.\nBecause if I was a simple little virus\nor a little bacterium, I wouldn't have\nto worry about electric fields and\nseeing things or hearing anything, my\nworld would just be that world that I\ncould touch literally. It can just be my\nnext door neighbors which leads to a\nvery particular kind of Markoff blanket\nstructure an ecosystem of Markoff\nblankets um that um precludes\noh certainly uh does not license um some\nof the deep structures that we were\ntalking about earlier on in terms of\nstrange things. Um so strangeness of a\nbeautiful sort,\nbeautiful loops uh to use um um one of\nmy colleagues um notion rests upon\naction at a distance of a very\nnon-spooky sort that is nicely\nexemplified by this compass. So, at the\nmoment, I don't know of any good maths\nthat allows you to work out what is the\nbest thing to do to disambiguate between\nthis structure and and that structure.\nYeah.\nSo, if you can do that, that that'll be\ngood.\nStructure learning.\nStructure learning\nto structure learning.\nStructure learning.\nSomebody out there help us out.\nYes.\nHuman data is shaping the direction of\nfrontier AI. Yet there's little\nvisibility about how teams are actually\nusing it. Our sponsor, Prolific, are\nputting together their first report on\nhuman data in AI, and they need\nvolunteers. It just takes a few minutes\nto fill out, and you'll also get early\naccess to their findings, so you can see\nhow you compare.\nThis podcast is supported by Google. Hey\nfolks, Taylor here, creator of Gemini\nCLI. We designed Gemini CLI to be your\ncollaborative coding partner at the\ncommand line. Think about those tedious\ntasks like fixing a tricky bug, adding\ndocumentation, or even writing tests. We\nbuilt Gemini CLI to handle all that and\nmore. It iterates with you to tackle\nyour most challenging problems. Check\nout Gemini CLI on GitHub to get started.\n[Music]\nMLST is supported by Cyber Fund.\nProfessor Friston, it is absolutely\namazing to have you back on MLST. And as\nyou know, um Keith and I are incredibly\nfond of you. You've been a big hero of\nours for many years. I think this is the\nfourth time\nthat you've been on show.\nSo, welcome, Professor Fen.\nI've lost count. Uh, but it's lovely to\nsee you both in person and\ncongratulations on your recent marriage.\nOh,\nthank you very much. Thank you very\nmuch.\nYeah. So, you know, we were we were just\nthinking about this. We we've done these\ninterviews. Um, the free energy\nprinciple originated around about 2005,\nright? And I just kind of wanted to get\na bit of a retrospective from you on\nlet's say like what have been the\nsuccesses so far of the free energy\nprinciple? How's it been going? What\nwhat could have gone better over the\nlast you know 20 odd years? You know\nreally just what's your what's your take\non the progress I guess of the free\nenergy principle? It's a difficult\nquestion to answer um in the sense that\num what are you if you ask me if I'd\nmade a meal and you've asked me how's\nthat meal going and did you really enjoy\nit when you're doing it you're sort of\nin in the middle of it and it's very\ndifficult to assess so it's been a\njourney um I um I do catch myself from\ntime to time thinking this is the the\nright way to think about things whenever\nyou come across, you know, a phenomena\nor a problem or an application\num that seems to just slip in quite\nneatly to the overall theoretical\nframework. Every time that happens, I\nget a little buzz of dopamine and that\nyou know that consolidates and yeah, I'm\non the right track. I haven't wasted so\nmuch time. There have been times in my\nlife when when I've realized I'm on the\nwrong track. So, I'm used to that. But\nfor the particular application of uh the\nfree energy principle, I have yet to\nhave that. Oh, well that doesn't work.\nThen think about it. You know, do do\nsomething else. Um what could have gone\nbetter? Um\nI'm not\nI'm about this, but um I am sometimes\nwell I am when I read the free energy\nprinciple is notoriously difficult to\nunderstand. Um\nnow half of me thinks good because it's\nalways important to have a slight degree\nof magic and mysticism to engage people.\nYou know if people don't think there's a\nchallenge that they have accomplished\nsomething by understanding it then\nthey're not going to be motivated to\ninquire and to think about it or indeed\ndebate debate it. On the other hand, um\nthe free energy principle is not meant\nto be complicated or difficult to\nunderstand. It's actually, you know,\nalmost toologically simple. Um so that\nbeing able to communicate the free\nenergy principle in a way that people\nfind a useful and obvious uh sort of\ntool or method to apply um could have I\nthink gone better. Uh, and that may be\nbecause I'm not very good at keeping\nthings simple or intuitive.\nWe can attest to that. I think,\nyou know, one thing that's interesting\nis I I feel the same way about\nprobability theory.\nLike, you know, conditional probability\ntheory is very easy to write down.\nThere's really that, you know, you end\nup with the two fundamental rules, the\nsum rule and the product rule. But the\nconsequences of it are very profound and\nprobability is notoriously\nmisunderstood, misapplied. there's many\nerrors of of reasoning, you know, that\nhappen when people try to reason through\nstatistical correlations or whatever.\nSo, I think there are things like that\nin life that are very simple and yet\nvery hard to to kind of understand the\nfull extent of their impact, right?\nYes. I mean, it's interesting you picked\nup on conditional probabilities. I mean,\nthat is the heart of the free energy\nprinciple at many many different levels.\nYeah. So the the classical formulation\nthe free energy principle just starts\noff with a partition of different states\nof being and that partition suddenly\nmeans you've got the probability\ndistribution or density over one set of\nstates of being that now can be\nconditioned upon another. So the your\nthe whole free energy principle is just\nbasically a principle of least action\npertaining to density dynamics. The the\nthe dynamics or the evolution of not\ndensities but conditional densities.\nThat's just it. This is before\nthermodynamics. It's before quantum\nmechanics. It's just about conditional\nprobability distributions. So it's\ninteresting you picked up on that as\nsomething that is so simple yet so\nabsolutely powerful.\nSo um Professor Fston, we were reading\nyour paper earlier. It's from 2023 um\npath integrals particular kinds and\nstrange things and it introduced a\ncategorization of particles. So for\nexample um inert to active and ordinary\nto strange. The strange um\ncategorization was particularly\ninteresting because you were um giving\nan account of how these particular\nparticles could um give rise to\nphenomenal you know conscious states or\neven agentic states. And we we felt that\nthis was almost an account of\npanagentialism. So you know historically\nwe've done interviews about the free\nenergy principle and we've spoken about\nself-organization and emergence and um\nthe agency and the phenomenal states are\nsomething which emerge when you have\nthis um sort of temporal and\ncounterfactual depth. And maybe we're\nmisunderstanding but it but it feels\nlike this is an account of low-level\nagentic potential and phenomenal\npotential. So to set the scene just to\ncome back to that partition that we were\njust talking about there are many ways\nof um arranging the the conditional\nindependences you know to um sort of\ndisconnect various partitions you know\nexternal states internal states my\nsensory states my active states um and\nthe combinations of disallowed\ninfluences give rise to an ontology of\ndifferent kinds of things and you could\ncall them natural kinds. I've been told\nby Maxwell I shouldn't use that but I\nlike natural kinds the things in nature\nthat just prescribe themselves just by\nbeing special instances of a lack of\ncausal influence. Um so if we start off\nwith uh you know the simplest case where\num there are no internal states and\nthere are no active states and what are\nwe talking about? We're talking about\nsome kind of causal black hole,\nsomething that's, you know,\nquintessentially inert and you can never\nsee it because you can only see the\nactive states that act back on the\nenvironment in which it is embedded. But\nthen we get to more interesting things\num that have a full complement of\ninternal states and external states and\na birectional coupling between the two\nwhich of course is mediated by the\nsensory states and the active states.\nAnd if you disallow action at a\ndistance, by which I mean I have to be\nin some metric sense next to you in\norder for me my active states to become\nyour sensory states and your sensory\nstates to be my um active states. Then\nyou have a very simple kind of markoff\nblanket which would be fine for\ndescribing active matter for example or\nany medium where I have to touch, to\nfeel, to influence or to sense. Um and\nthis kind of particle interestingly has\nits or this kind of system partition has\nits active states on the inside which is\nyou know I had to think about you know\nyou know and it can be no other way if\nyou you know just try to arrange all the\ndependencies.\nSo what would that look like? It would\nlook like something like a cell that had\non the outside its sensory states that\nwas supported by a layer of active\nstates that surrounded the internal\nstate. So you're complying with the laws\nof the Markoff blanket of of thingness\nthat you you you need to have that\nconditional independence to separate the\nthing from everything else or the self\nfrom from the non-self. and everything's\nfine because the active states are\nhiding behind the sensory states. So the\nexternal states can't influence the\nactive states. So that's tick one.\nThat's you know what we need. But we\nalso have to ensure that the um internal\nstates symmetrically cannot act\ninfluence the sensory states and that's\nfine because they're hiding behind the\nactive states. So you've got you know a\nnice um mathematical image of you know a\na cell if you like. But that doesn't\nwork for things like you and me. Things\nlike you and me have um um have a\nhierarchical structure. Um and what that\nbasically means is that the active\nstates which were hitherto in very\nsimple organisms um immediately\njuxtaposed to the internal states now\nbecome sequestered. So now effectively\nthe active states are no longer seen by\nthe internal states and then something\nquite remarkable happens mathematically.\nUh you it's really simple and it's just\nanother aspect of conditional\nprobability distributions\nbecause you can't see your active\nstates. So you can now only see your\nsensory states. It looks as if from the\npoint of view of the internal states\nthat the active states are now become\ncauses of the sensory states. So now\nyour world is caused not just by the\nexternal world by the environmental\nstates by my heat bath by my ex you know\nmy uh external miller but also my own\nactions so now I'm inferring the causes\nof my sensorium where I'm actually c a\ncause so I'm inferring myself so there's\nthis beautiful recursion which is\nlicenses this sort of strange loop um\nanalogy with with a bit of poetic\nlicense that\nanother way of putting that is that you\nknow for these simple structures for\nthese natural kinds that would be say\nsingle-sellled organisms\nthe internal states can be read as\nmodeling the causes of their sensations\nthat just are the external states. So\nthey have direct access to the active\nstates. So there's no conditional\nindependence that licenses the notion of\ndescription as of in terms of inference\nor sense making. Um whereas now the\nbasin mechanics that attends the the\nactivi the internal activity internal\nmachinations now covers both my action\nand the things that I'm acting upon. And\nof course this is a nice metaphor for\nplanning as imperance that I'm now\nthinking about and trying to infer what\nam I actually doing and that to my mind\num\nproduces a very unique and special kind\nof thing um you know things like you and\nme basically which are pretty unique uh\nwhen we look at all the different kinds\nof things that could be around. Well,\nit's it's it's another example of how as\nsoon as you had this recursion, right,\nas soon as you had these these strange\nloops, especially for systems that that\nare computational, you know, and I I\nforget who it was that said life is the\ncomputational phase of matter, but I\nmean, when you're at this level of\ncomplexity where you're having some kind\nof information processing and you\nintroduce this strange loop, right, that\nunlocks a completely different like\ncategory of behavior, you know, tier of\ncomputation. Right.\nAbsolutely.\nIt was David Krakow. He said life is the\nintensive property of matter and\nintelligence is the extensive property.\nNo, this is a different one though.\nSomeone it may have been Wolf Ram who\nsaid life is the computational. I I\ndon't remember. I'd have to look it up.\nUm but\nbut yeah, so it unlocks a completely\ndifferent category of behavior.\nYeah. Well, I mean, you alluded to that\nearlier on in terms of phenomenology,\nand you mentioned consciousness, which\nyou shouldn't really do with me, but\nwe'll read that as an elementary kind of\nsentience. Um, so I I think that's\nreally important. You know, that point\nthat you do unlock or you do manifest or\nnow at least you have a mathematical\ncalculus that allows you to talk about\nthings that model themselves. And as\nsoon as you're modeling yourself, you\nbecome an agent or you have an authentic\nkind of agency because now in order to\nmodel myself or to model the me as a\ncause of my environment, I have to infer\nwhat I am in do what I'm doing in\nparticular the consequences of what I'm\ndoing. And because the consequences have\nnot yet appeared, this has a a\nquintessentially future pointing aspect.\nSo, we're now moving from a thermostat\nor a virus or a single-sellled organism\nto things that actually have a future on\nthe inside, their private future that\njust exists for them. And of course, if\nyou can get f, you know, suitably far\ninto the future where you have a\ndivergence of particularly paths into\nthe future just as a nod to the\npathogical formulation, then you have to\nselect one. So now you've you've really\nunlocked again a calculus of selection\nof your paths into the future which I\nthink you know is not a definition of\nagency and certainly not consciousness\nbut certainly would be um I would\nimagine arguably necessary to support\nsomething that had true agency and\npossibly even uh consciousness you know\nof an elemental kind. Could we press on\nthe consciousness a bit because the\naccount that you just gave is a is a\nbeautiful account of agency and it's\nit's very plausible. So as you increase\nthis reflexive layering this recursion\nyou get these multiple paths into the\nfuture and you become the cause of your\nown action. So you become more of an\nagent. So it's almost um an account of\nstrengths of agency with with more and\nmore layers. But in the um the abstract\nof this paper, we were quite struck with\nthe language that convol consciousness\nwith agency and intuitively I feel that\nthat um phenomenal experience is is\nquite orthogonal to to agency. So why\nwere they kind of convolved together?\nUm I'm desperately trying to remember\nhow I referred to consciousness.\nYou got to be very careful when using\nthat word depending on who you're\ntalking to. Um\nthe C word\nthe C word. Oh naughty. Um so yeah and I\nthink you know people like Neil Seth\nmake a very similar point quite\nearnestly you know that intelligence and\nand consciousness are completely\northogonal that you're making agency and\nand consciousness are completely and I\nwould agree entirely. Um so um you know\nconsciousness um does is not implied by\nbeing agentic uh uh you know having\nhaving agency um\nyou would first of all well there are\nmany stories you can tell here um and\nyou know my mind goes to the people who\nmight be watching this to make sure I\ndon't offend anybody by not mentioning\nit. So, so yeah, the\nfrom from the point of view of three\nenergy principle, um the the way that\nyou'd look at consciousness is um\nin terms of a dual aspect monism which\num you get for free from the treatment\nof the density dynamics in exactly the\nway we were talking about before. But\nyou know so now my internal say brain\nstates um have a thermodynamics they\nhave that that can be written in terms\nof an information geometry that itself\nis just predicated on conditional\nprobability distributions. Um but they\nalso um have an information geometry\nthat inherits from the fact that my\ninternal brain states represent\nthe outside world including my own\nactions. So there's both a an\ninformation geometry and and by\nimplication of thermodynamics of my\nbrain activity\nwhich supervenes on exactly the same um\nsubstrate as does the information\ngeometry which is representational which\nis the inference the basic mechanics\nside of things. So that gives you\nlicense to talk about a sort of dual\naspect monism. But what is it what does\nit really mean to be conscious? Some\npeople might believe that it's just\nhaving a posterior belief, having a\nbasian belief which is literally just a\nconditional probability distribution\nthat is encoded or parameterized by some\nphysical state of being and in this\ninstance in the under the three energy\nprinciple it's the internal states of\nbeing. Um other people say well no just\nhaving you know just just being able to\nread a thermostat for example as having\nposter beliefs about the temperature of\nhis external world does not license you\nto ascribe consciousness to this thing.\nSo the next step would be oh it has to\nbe ignited in some way. It has to be\nrealized. It has to be emergent and um\nI'm sort of paraphrasing Yakobi here and\nto a certain extent Mark SS. So they\nemphasize now it is when these beliefs\nbecome sufficiently precise in a\ndynamical way that they are realized and\ninfluence other aspects of belief\nupdating in a hierarchal or distributed\nsystem. So Mark Zs would call this felt\nuncertainty. So he he would emphasize\nthat the the the feeling part of\nconsciousness\njust is or is mediated by the\nrepresentations not of the content but\nof the precision or uncertainty or the\nconfidence with which these beliefs are\ncurrently in operation. And you know he\nwill bring to the table all sorts of\nvery compelling neurobiological evidence\nas to why this is how you can switch off\nconsciousness literally by um I think\nhe's he says it's a you know a 3 mm cube\npart of the brain stem that is the cells\nof origin that these ascending systems\nthat encode and represent I repeat not\nthe content but the confidence or the\nuncertainty or the precision of these\nconditional probability distributions.\nUm Jacob um would would I think say well\nto be aware is to equip all those\nmessages that are providing evidence for\nyour current explanation with precision\nwhich has you know these physiological\nand you know bioelectric sort of\nmechanisms behind it. Um other people go\nfurther. People um in the world of um\nphenomenology of the kind that Thomas\nMetsinger pursues um would say that it\nwould be necessary to have control over\nthe prec precision. Um and furthermore\nto be aware that you are rendering that\nwhich was once transparent opaque you\nactually have to recognize that you're\nattending. So you again not only do you\nhave this sort of very simple recursion\nof modeling self-modeling um in terms of\nplanning as inference but now you've got\nthis hierarchal recursion where you're\nnow recognizing that you're attending to\nsomething and then to be if you're\nsomebody like LZ Sunvid Smith you'd say\nwell to actually be self-conscious I now\nhave to recognize that I was recognizing\nthat I was attending to something and\nyou get layer upon layer upon layer um\nand to join the dots with um Chris\nFields um who has brought to the table\nthe inner screen hypothesis. Have you\ncome across this yet?\nNo. No. Tell us about that.\nOh well, strictly speaking, you should\nget Chris to tell you about that, but\nI'll I'll give I'll give you a quick\npreamble. Um so um so Chris Fields is\nthe um the theoretician who has\nformulated the quantum quantum\ninformation theoretic version of the\nfree energy principle. So for him\nthe markoff blanket separates self from\nnon-self becomes a holographic screen\nupon which classical information is well\nto which it is written and from which it\nis read by some internal bulk and some\nexternal bulk. um and in the sense that\num strange loops and strange recursions\nare only allowable when you have this\nhierarchical structure on the inside.\nUm, you've now got lots of inner\nholographic screens, inner Markoff\nblankets. I mean, Markoff blankets, you\nknow, in the pragmatic sense just define\nthings like hierarchies, for example,\nand, you know, they define the\narchitecture of, you know, of any\nnetwork, factoraph or or or computer.\nAnd when you got lots of them, you've\neffectively got lots of inner screens.\nSo the idea, well my reading of that\nidea\nis that there is one irreducible inner\nscreen. So there's one in a screen that\nwithin it has no other screens.\nAnd this is an interesting screen\nMarkoff blanket from the the classical\nperspective because the only way that\nthe internal states of this irreducible\nMarkoff blanket can know themselves is\nby acting on the exterior which of\ncourse the exterior is the rest of your\nbrain. So you got this metacognitive\nself-recursive aspect to this inner\nscreen hypothesis.\num and that um\nthat has I think a lot of mileage that\nparticular notion has a lot of mileage\nin relation to other people's theories\nof consciousness. So we're talking about\nhigher order thought theory. So the very\nnotion of higher order thought and its\nimplicit appeal to metacognition and\nmeta meta cognition meta meta is again\nthis this notion of looking at oneself\nand inferring oneself but on the inside.\nSo looking at looking at the looking at\nthe looking um uh which is of course you\nknow a natural consequence of having\nthis this hierarchal structure um you\ncould also argue it's completely\ncompatible with um global neuronal\nworkspace theories that um you know\nthere is this sort of precision\ndependent ignition of certain sources or\num messages\nsufficient statistics of conditional\ndist uh probability distributions\nthat gain access to you to all of across\nall of these inner screens in a dynamic\nway because you're controlling the\naccess by attending by um encoding and\nrepresenting and optimizing the\nprecision or the uncertainty or the\nconfidence. So you've got this notion of\nignition and penetration through to the\nglobal workspace which you could read as\nthis sort of in irreducible inner screen\nthe core your the deepest part of your\nsense making your your brain. So I think\nthis this notion not only is it\nconsistent with the you know the\nconditional independences that are\nimplicit in the in the free energy\nprinciple as applied to strange things\nthat have this hierarchal and recursive\naspect. But it also um I think relates\nvery comfortably with ext theories all\nall coming from different perspectives\nseeing different parts of the elephant\nas it were. Um but it's you know it is I\nthink a comfortable um accommodation of\nthings that we you know presume um or\nthings that people have brought to the\ntable in order to explain consciousness\nyou know from from their direction of\ntheorizing. It seems that um if you if\nwe kind of split let's say the um the\ncamps or the the the the thought the\ngroups of thought about this there's um\nalmost all of what you just talked about\nnow are there is a minimal complexity\nrequired for there to be consciousness\nand then we can we can talk about what\nthe what that level is. Well, it\nrequires this or this or this. But then\nthere's also people who say no there's\nno there's no minimal complexity. It's\nlike electrons have some kind of small\namount of consciousness. And I fall much\nmore into the the category of people who\nsay there is a minimum like you have to\nreach a certain something. I don't\nreally know where that boundary is. You\nknow, maybe it's one of these that we've\nmentioned. Maybe we'll figure out a more\nelegant kind of categorization. But but\nit's just like you need a certain number\nof pixels in the game of life before you\ncan have a glider. You can't do it with\nthree. Like you need there's some\nminimum number. And so there's some\nminimum amount of brain material. You\nknow, three cubic millimeters is still a\nlot of neurons. I don't know how many\nare in there, but you know, hundreds,\nthousands or whatever it is. You need\nsome level of complexity before you have\nconsciousness or before you have, you\nknow, let's say um recursive agency or\nor something like or intelligence or\nwhatever. Would you agree with that?\nLike there is some some minimum. We\ndon't know where it is yet. It's hard to\nplace that, but there is a minimum or\nnot. No, I would agree entirely.\nOkay. So, so there there is a minimum\nand not only that, it has to have a\ncertain causal structure, right? And we\ncan we can kind of debate that and\nthat's really in line too with surl you\nknow in the Chinese room argument which\nis like look a dictionary is not doesn't\nhave understanding because it doesn't\nhave the right causal structure. You\nhave to have a certain causal structure\nor a certain minimum complexity and then\nyou reach this whatever it is whether\nit's consciousness we're talking about\nunderstanding agency all of these things\nright so I guess my question to you is\nwill we be able to build machines based\non our current computer architectures\nsome someday whether it's a hundred\nyears from now 200 years doesn't matter\nbut in principle can we build machines\nthat have understanding that have\nconsciousness that have all these\ncapabilities.\nYes, that's a that's a good question.\nUm, and the answer I think I think in\nprinciple yes. Um, can I just come back\nqualify that answer by reference to that\nwonderful example of you know um what I\nwould read as vagueness in a technical\nsense. So vagueness is you know how many\ngrains of sand constitute a pile?\nYeah. So it's not well defined. It is in\nphilosophy but not not not not in not in\nmathematics. I think that's absolutely\nthe right way to think about these your\nthe bright lines. And if I had to commit\nto um the dimension the number of grains\nof sand that you get before um you have\num consciousness or there is a pile. Um\nI would say it's something you actually\nreferred to earlier on. I think it's the\ndepth of your of your future.\nor your future in your head. So if\nyou're talking now about an algorithm or\nsome artificial um intelligence that is\nequipped with a generative or a world\nmodel of the consequences of its action,\nthere will be a time horizon\nassociated with that component of its\ngenerative model. And I think that it's\nthe depth of that time\nwhat what anal Seth would refer to um\nnot as counterfactual breadth which is\nthe number of divergent paths one could\ntake into the future the options that\nyou you select among but the the the\ncounterfactual depth\nthe temporal depth and um so that means\nthat you can you can say you can be pany\nyou uh you can you can you can say that\nyou know um a thermos at has um a notion\nof the future in the sense that it\noperates of you know sort of\npathological control and differential\nequations. As soon as you put a\ndifferential equation in play, you've\ngot a you know an instantaneous future\nbecause you've got some gradient with\nrespect to time.\nThat's not though the kind of depth that\nyou and I uh enjoy. So I would imagine\nit is really just the the depth you\nrather very very reflexive and I think\npeople like um Maxwell Ramanstead talk\nabout this as merely reflexive active\ninference um with you know very myopic\nvery shortterm self models for uh right\nthrough to fully well to be conscious in\nthe way that we've been talking about\nyou I think you need to have uh you know\na long depth so what does that mean for\num building um AGI or uh conscious\nartifacts or machine consciousness. It\nmeans first of all they have to be\nagentic because we've just said that\nhaving a you know a world model of the\nconsequences of your action um would be\nnecessary to be an agent. Um but more\nthan that you'd have to look quite a\nlong way into the future. So your your\nyour generative model your world model\nwould have to go quite a long way into\nthe future. Um, and I repeat, have that\nboth counterfactual breadth and\ncounterfactual depth at hand. Um, would\nthat be sufficient? Um, and I'm now\nremembering that I forgot to mention a\nNeil Seth in the list of people not to\nupset when reviewing theories of\nconsciousness. So this is an opportunity\njust to say that there are people out\nthere who would say well you know okay\nyou can write down the maths of all this\nand you can write down in a screen\nhypothesis or indeed simulate global\nneural workspace theories and and and um\ntry to produce um things that look as if\nthey have consciousness. Um that's not\ngoing to work unless you actually embody\nit. Unless you are in a Neil's words a\nbeast machine.\nUm\nI think that that sort of um coheres\nwith the argument for mortal\ncomputation.\nSo when you may ask the question can we\nbuild it on our um computer\narchitectures I would have to ask you do\nyou mean a vonuman architecture or do\nyou mean a uh memory uh processing in\nmemory or in memory processing\narchitecture which I would take\nsynonymous with the neuromorphic\narchitecture not spiking neural networks\nyou don't need those but you do need the\nprocessing in memory um to be you know\nto be mortal you need that substrate\ndependence 's um you know read you know\nin terms of Jeff Jeffrey Hinton's\ndefinition of mortal computation and you\nknow and Alex's uh subsequent\nelaborations of that. So if you've got I\nthink um I would subscribe to that. I\nthink you um largely to keep a Neil\nhappy. Um but I would sub subscribe I\ndon't think you can do this on a on a\nvon Newuman architecture cuz the the\nmarkoff blankets um of a von Newman\narchitecture where you're reading and\nwriting from memory um make it very\ndifficult for the memory to\nself-organize.\nI see.\nUm and people have written about this\nphilosophically. I think Vana Vice has\nhas has written a paper about this and I\nthink Anneil Seth um speaks to this\nargument in in a recent um behavioral\nand brain science paper. um you know it\nmay not be possible and certainly you\nknow from the point of view of\nefficiency\num it's highly unlikely that sort of um\nvonuman architectures are the kind of\nthings that would conform to let me\nreverse to be is to pursue a path of\nleast action in accordance with the free\nenergy principle to be conscious does\nnot excuse you from that so if uh\nIf you want a conscious machine, you\nhave to have a machine that pursues a\npath of least action.\nVia the Janinsky equality, that has to\nbe true both thermodynamically and\ninformationally in terms of the\nconditional probability distributions.\nIf you don't pursue that path of least\naction, you can't be and you can't be\nconscious for, you know, a non-trivial\namount of time. Um, and I can see you\nwant to ask a question, so I interject.\nWell, only to say um Maxwell pointed me\nto a nil's where he defends biological\nnaturalism and um of course even cells\nsaid that the biological substrate is is\nan existence proof. He's not saying it\nhas to be biology, but Keith and I\ncertainly agree that it there's\nsomething about the substrate which is\nwhich is very important.\nUm I wanted to talk a little bit about\nviruses. As you know, I'm a bit of an\nexternalist. I've never been able to\ncompletely pin you down, Professor\nPriston, because there have been so many\ninterpretations of the free energy\nprinciple that um that lean internalist\nand externalist and even um the hybrid\nversion which Maxwell also wrote a paper\nabout. But I'm fascinated in this idea\nof diverse intelligences and for\nexample, could a virus be intelligent?\nAnd um I I spoke with David Krakow and\nhe said intelligent things do inference\nand they have representations and\nthey're adaptable and we are a little\nbit you know we're a bit chauvinistic\nabout our brains aren't we because our\nbrains seem to have a privileged status\nbut what say you are viruses or actually\nwe we read your paper with um uh Calvo\npredicting green uh really radical plant\npredictive processing and in that paper\nyou gave a beautiful account of how even\nplants could be doing inferencing and\nand that to me seems it it seems\nincredible because I'm I'm amendable to\nthe idea but it it seems to me\nintuitively that plants are not as\nsophisticated as as we are. So it comes\nback to this line that Keith was talking\nabout before that if you go too far down\nthe stack, it's an account of we let's\ncall it pan intelligence where you have\na tiny little bit of intelligence even\nif you go all the way down the stack.\nHave you spoke to Mike Leaven?\nYes. Relatively recently about a year\nago.\nRight. I I mean you know it would be\nnice to um\nrevisit him. I mean that that that that\nis exactly his big question at the\nmoment. Um and and of course he um\nhas as an intellectual accomplice Chris\nFields with him as well. So that you\nknow the two of them are really you know\npursuing this notion of basil cognition\nthat there is intelligence everywhere.\nUm, and it's just a question of, you\nknow, how we conceive of it and how we\ntest for it. And indeed, his argument,\nMike's argument is that, you know, you\nhave to design the right experiments.\nIt's an empirical question. Is this\nvirus intelligent or not? Well, you have\nto, you know, design the right\nexperiments to disclose or evidence its\nintelligent behavior. And Mike would\nclaim that, you know, by many metrics,\num, many rule stick u, yard sticks that\nwe use to measure adaptive intelligent\nbehavior. Yeah. Slime molds and viruses\nand xenobots are incredibly intelligent.\nUm, and you know, I think he's fighting\nagainst the chauvinism that you\nmentioned that intelligent things are\njust properties of things like you and\nme, creatures like you and me and our\nbrains. um and that that kind of\ncognitive capacity and competence can be\nfound everywhere. Um so I you and I'm\nvery sympathetic to that, but it does\ntread on the toes of the vague argument.\nUm you know um I certainly don't think\nthat um viruses um have the same um\nexpressive kind of agency that we were\ntalking about. And indeed, you know, the\ncounterargument to the vague um notion\nof intelligence andor consciousness is\nof course um hitting you in the face\nwhen you when you read that strange\nthings paper. You know, I'm talking\nabout categorically different natural\nkinds that do and do not have, for\nexample, a causal power or influence of\nex active states on internal states.\nyou're either one of these or you're one\nof these.\nI'm not so sympathetic to the view that,\nyou know, a virus is intelligent, right?\nBecause I think this falls into the\ncategory of things like where you're in\na biology class and you learn how to\ndefine life and then somebody's like,\n\"What about fire?\" Then like it seems to\nmeet all these criteria here, you know,\nit it grows, it expands resource, you\nknow. So, for example, when a virus\nquote mutates, the virus doesn't mutate\nitself. it is mutated by a gamma ray\nhitting it or a you know transcription\nerror or whatever and and oil you know\nvinegar and baking soda undergoing a\nreaction is not intelligent right so\nthere's I think there's some level of\ncomplexity and you know whether it's\nrecursion or these other types of causal\nstructures that have to be there before\nI'm even interested in talking about\nintelligence or entertaining\nintelligence you know other things are\njust kind of dynamics that unfold in in\ncertain ways chemical reaction\nyou know that that sort of thing\nand I think that that that point um is a\nnice opportunity just to introduce the\nnotion of scale freess um or scale\ninvariance\nbecause I you know I can see you could\nalso take this towards collective\nintelligence and federated learning and\nfederated inference um\nso it's not it's not the single cell\nit's a single cell with its neighbors\nand its neighbors of neighbors and\nneighbors and neighbors um and So what\none's looking for um is some\nconservation of intelligent dynamics\nthat is preserved over different scales.\nSo you know the single virus is probably\nnot interesting. It's probably the\ncolony which is uh which which is which\nis interesting. So I thought that was\nyou know a nice point to introduce the\nnotion of of that kind of scale\ninvariance. And of course as a\nmathematician you you would be um\nlooking at the reormalization group to\nsee how that unpacked mathematically\nand interestingly um it also speaks to\nthis notion of a mutation. Is evolution\nan intelligent process?\nYou know it's certainly adaptive.\nIt certainly um has uh the level of\ncomplexity that you would uh require um\nin order to uh pass those vague\nthresholds. But is evolution in and of\nitself an intelligent process? It's\ncertainly a free energy minimizing\nprocess. You know, it is just basic\nmodel selection. Natural selection just\nis maximi or selecting those things that\nhave the highest model evidence or\nmarginal likelihood of being that\nphenotype in this kind of environment.\nWell, it's like is is the planet\nintelligent? You know, it certainly\ncontains 8 billion or so of us. So, does\nthat count? I mean,\nyeah,\nwhy not? Well, I mean, that's where\nWell, I would say like uh and and we\ntalked about this way back when we\ntalked about um you know, is a flotilla\num an agent or is it the pilot who's you\nknow kind of kind of controlling the\nconvoy or whatever. I would say like if\nthere's a more minimal boundary that\ncontains the intelligent then the larger\none is not like I'm talking about so the\nplanet for example I would say since I\ncan draw a smaller boundary which is an\nindividual person the individual person\nis intelligent anything that contains\nthat person is just more stuff\nKeith what is your the issue with with\nviruses is it because the individual\nvirus is inert whereas a cell for\nexample you can partition ition down to\nthe individual cells and the cell still\nhas some degree of agentic property as\nas as professor Fston was describing is\nis it that you see them as inert and\nbeing carried by something else?\nWell, I think it's that too much of\ntheir their causal structure and\nmachinery is basically outsourced to\nother things. It's like they you know\nthey're they're kind of just these\nlittle particles that go around and if\nthey happen to stick on a cell they're\nlike a mouse trap that activates and\njust injects some DNA but you know the\ncell does all the rest of the work,\nright? It does the transcription. It has\nall that machinery by themselves. A\nvirus doesn't change over the course of\nits lifetime. It doesn't really other\nthan this simple mouse trap kind of\nactivation. It does nothing, you know.\nIt doesn't doesn't have any machinery to\nlearn and adapt by itself like over the\ncourse of the lifetime of a virus, you\nknow. So I might be more amendable to\nlike if somebody says well like an\necosystem of you know viruses or\nsomething has some degree of you know\nintelligence but it still doesn't have\nthe the type of processing and causal\nstructure and minimal complexity that at\nleast for me is the useful concept of of\nintelligence. So I think we're coming\nback now to um the\nthe number of inner screens and the\ncounterfactual depth and the definition\nof agency in the sense of strange things\nthat we're talking about for this\nparticular scale. So if we just read the\nscale as if you like the size of the\nmarkoff blanket. If within at this scale\nfor this kind of thing um there is a\nsufficient degree of complexity read\nspecifically though in this instance as\num the counterfactual depth and breadth\nof your world model about the\nconsequences of your action. that\nparticular part of of of your um\nimplicit generative model. Um and that\nwould happily accommodate the fact that\nwhen you get something, you know,\nsomething um of the size or the scale of\na virus, there just isn't the machinery\nor the space to entertain that in any\nnon-trivial um in a non-trivial way. Um\nuh one could also argue and I got a\nsense that you were arguing yourselves\nuh towards this that if you go too big\nyou also lose that. So you know this,\nyou know, coming back to this wonderful\nquestion, we have 8 billion intelligent\nreally intelligent um u sound like Trump\nthere, didn't I? Um\nentities constituting our biosphere, but\nis is the biosphere is you're from the\npoint of view say that the K hypothesis\nis the biosphere in and of itself\nintelligent and I would I would say no.\nNot because it's simply because at that\nscale\nthe um the elemental all the complexity\nof the constituent elements disappears\nat that scale.\nSo you know it'd be a little bit like\nsaying um well let's just take it to the\nlimit. Let's just take it to the astron\num astronomical uh limit um you know of\nthe motion of heavenly bodies.\nNow the motion of heaven he heavenly\nbodily is is completely described by the\nposition of the planets and the moon\nright and the position of the earth. So\nat that scale you've averaged away the\nfact that the earth contains a biosphere\nand the biosphere contains human beings\nand human beings contain cells and the\ncells may or may not contain viruses\ndepending upon who you've been uh uh\nexposed to. Um so there will be no\nintelligence at that level. So it's\nperfectly possible to have intelligence\nat a particular scale that disappears\nwhen you get too big and when you get\ntoo small. Um another way of looking at\nthat is um the um going right back to\nthings like uh brigine and disputative\nstructures. You know you're talking\nabout complexity. If we just think about\nyou know how have people tried to\nunderstand\ncomplex complex systems and\nself-organizing systems that are open.\nSo we're not talking about 20th century\nphysics and equilibrian physics. We're\ntalking about the physics of\nnon-equilibria\nand things that are open in open\nexchange with with each other. And of\ncourse you get to the notion of\ndissipative structures. What does that\ntell you mathematically? Well, what's\nnot dissipative?\nThat's a school boy question.\nCan you remember? You probably\nforgotten. So what is not dissipative is\nconservative.\nSo yeah. Uh so I'm sure I'm teasing it.\nUm so um\nanother gift of Helmholtz of course is\nthat any dynamics\ncan be um partitioned into a dissipative\npart and a conservative part. And the\ndissipative part is that which um rests\nupon very fast random complicated\nfluctuations\num of the kind you might find in say\nquantum mechanics or thermodynamics\num whereas the conservative part does\nnot rely upon that and just goes around\nin circles. Basically it's you know\nliterally it's called solenoidal flow um\nor conservative flow. So that basically\nmeans that the that the um the\ndissipative structures have to have an\nad mixture of this circular aspect. this\num conservative classical um life\ncycles, reproduction, oscillations,\nplus the the random dissipative part\nthat that um um you'll find in things\nlike thermodynamics and and and and\nquant quantum mechanics and that is\ndefinitional of dissipative structures.\nUm as you go too small then everything\nbecomes quantum and random. It all\nbecomes probabilistic. There is no\nconservative stuff other than a scroing\nof potential.\nUm at which point I think you you'd find\nit very difficult to find something uh\nthat was intelligent in the sense we're\ntalking about because we have to have\nthis recurrence this solenoidal aspect\nin order to revisit the states. So you\nknow a virus is a virus in order for\nthere to be a non-equilibrium steady\nstate distribution solution to that. Um\nso but as you get get bigger and bigger\nand bigger you get to to to viruses and\nyou're still not quite then still not\nquite sort of complex enough to be\nintelligent in the way that we mean and\nthen you get to our size and that's\nperfect because we're both have both\nthis dissipative we deal with a random\nworld with an an itinerant world and yet\nwe keep revisiting states of being. So\nwe have this sort of conservative um\nbiometic kind of self-organization.\nBut then you get bigger and bigger and\nbigger. You get to the level of the the\nbiosphere or you know the size size of\nthe uh uh the moon or the sun. Uh and of\ncourse by averaging all the random\nfluctuations go away. So you're just\nleft with the solenoidal part. you're\njust left with the solenoidal motion the\nNewtonian motion of of classical of of\nclassical mechanics\nhere because there's there's so one\nthing that really fascinates me about\nthis is however it is possible for us\nand we do construct largecale things\nthat are intelligent like a corporation\nyou know a corporation is seen as like a\nsuper intelligence but we're only able\nto do that by putting in structure right\nthat maintains this balance of of the\ndissipative and and conservative flow,\nright? Like we have to put in\norganizational structure so that it\ncontinue to function at these larger\nscales, right? Like isn't that pretty\npretty interesting? I mean\nit's it's again this yin and yang thing\nthat we run into all the time, right?\nwhere it's the balance between two two\nforces like either between say uh\ncomplete order and complete noise or\nbetween dissipation and conservation and\nyou have to be almost on the edge of\nchaos and it has to have a certain\ncausal structure in order for it to to\nbe intelligent.\nNo, no, I I agree entirely. and and and\num you know that that that sort of\nGoldilocks regime where you are on the\nedge of chaos I think is quite specific\nto a particular scale certainly you\ncould you could invoke a sort of strong\nanthropomorphic principle here and say\nthat yeah the kind of intelligence that\nwe will recognize has to be at our scale\nbut I think there's something more\nfundamental than that I I think the very\nexistence if you subscribe as an\nexternalist to quantum physics and uh\nNewtonian physics Ronian classical\nmechanics. Um I you know I I I I think\nthat there is a Goldilocks regime which\nmeans that we can only exist at this\nscale with with as you say this sort of\nyin-yang this ad mixture of dissipative\ndynamics and conservative dynamics and\nyou know the just to reinforce this\nbecause the conservative dynamics is\nabsolutely essential because it is that\nwhich causes this ponare recurrence. It\ndefines these strange attractors which\nis the other way of licensing the notion\nof strange things. You always come back\nto somewhere near where you started like\nred queen dynamics you know um in even\ntheoretical biology. So you have to have\nthe conservative the circular motion\njust to re you have a routine have bio\nrhythms replicate reproduce many many\ndifferent levels at at many different\nscales but it's not it it's always\nremarkable in the face of a dissipative\nitinerant world everything is changing\nall the time um and yet we somehow sort\nof resist that change by being at the\nedge of chaos\noh no you you look more anxious than I\ndo to say Carry on.\nPlease go on, Professor Preston.\nNo, no, no.\nWell, I mean, I'm just so fascinated by\nby your observation that with Gaia\ntheory, for example, when we zoom out,\nthe the the apparent phenomena seems\nless intelligent. And as you said, maybe\nintelligence is just what we recognize.\nAnd um when we spoke with Wolf from that\ntime, he was talking about how we're\ncomputationally bounded as observers.\nAnd we we do this thing called\nabstraction where we ignore details or\neven idealization where we deliberately\ndistort the truth. And um David Krakow\nsaid that intelligence is doing more\nwith less and emergence is more is\ndifferent. And\nI mean the earth is not doing it's not\ndoing less. We just don't see what it's\ndoing. But if it's if emergence is more\nis different then that it licenses a\nfundamental reorganization of the\nunderlying substrate which means it it\nisn't actually doing the thing at the\nlower level anymore. It it's it's a\nfundamental course graining and it's\nactually changing at the higher level.\nSo which is it? Well, um I didn't\nrealize there's a choice there because I\nI agree with everything you said,\nbut I like the notion of poor grading\nbecause of course that is exactly what\nyou get from the reormalization group\ntreatment of these things and you can\nsimulate this. You can simulate um free\nenergy minimizing processes that are\nrunning at different scales and then you\nhave to ask the deep questions of how\nyou couple between the scales and and\nthere you get to supervenience and\nemergence mathematically so defined. But\nit all boils down to exactly where you\nstarted and where you ended up which is\na coarse graining in the right kind of\nway. So you know to to actually um write\ndown a renormalization group you have to\nhave a an RG operator. What is an RG\noperator? Well it has two parts to it.\nIt has a sort of dimension reduction the\nR part if you like and a grouping\noperator. Um and they basically do the\nright kind of coarse graining. they\nreduce dimensionality and they group\ntogether in the right way um those\nstates of being um at at the highest\nscale and so on and infinitum um so I\nthink your the notion of course graining\nis absolutely essential here uh and on\nthat level um you could you could argue\nboth ways so I I'm getting a sense that\nyour question is is something is there a\ntrue emergence emergentism here um or\nnot uh uh again\nI don't have a philosophical training so\nI'm not sure I can really answer that\nbut certainly there the the intelligent\ndynamics read as um self-evidencing\nperspective on the on a free energy\nminimizing process are recapitulated in\na at a completely different scale in a\nin a different way at each scale through\nthe recursive application of RG\noperators Um so yes there is something\nbrand new going on and yet\nin the spirit of say Haken synergetics\nthese RG operators\nare just taking functions of stuff\nthat's happening at the finest scale. So\nyou're not inventing anything new it's\njust predicated or emerging or\nsupervening on finer scale stuff. But\nwhat emerges at the high scale has all\nthe attributes of self-evidencing and\nyou could argue um consciousness or\nintelligence at some level. But coming\nback to to to um the point about\nbuilding bigger organizations,\nif the argument that as you get bigger\nand bigger and bigger um means that you\nare effectively to exist, you have to be\neffectively conservative. So you know a\ncompany for example um what is its\nmetric of uh goodness is how long it\nsurvives you know can you get from\nseries 8 to series whatever um\nin some recognizable form so as you get\nbigger and bigger and bigger I think\nthere's less opportunity for this\ncomplex scale-free within sorry not\nhierarchal recursive structure within\nthe scale so I do I don't see the moon\nthinking of planning. I don't see the\nweather planning. I don't see evolution\nplanning. It doesn't think about its\nfuture. It's too big.\nAnd I would imagine that you could apply\nthe same arguments to globalization and\ninstitutions that get too big for their\nown good\nbecause they can't plan anymore. So,\nwe're coming back to the pilot who's\nreally the intelligent person, not the\nflotilla.\nRight.\nYeah. It's in in a way it's um I guess\nmaybe it's slightly depressing to me\nbecause it means we can't build an\nintelligent intergalactic you know uh\ncivilization. It's it's almost like\nbeyond a certain scale we just have to\nleave it up to kind of distributed\nemergent you know um processes that that\nmay or may not end up doing something\nintelligent as a whole. um you know\nbecause because at some point you've got\nthe speed of light and that's that's\nbasically going to be a barrier to\norganization right across across light\nyears right so so maybe there is an\nultimate goldilocks limit right to to\nintelligence can't it can't get too big\nthat's certainly what the you know the\narguments from the physics part you\nwould so there is a goldilocks uh a\ngoldilocks of scale or zone own in scale\nspace. Um I'm not sure this is correct\nand I don't want to disappoint you. So\ndon't worry it's well beyond my lifetime\nwhen that will happen. But\nthere's a certain beauty though in\nfederating and and distributing and uh\nyou know thinking in terms of ecosystems\nof things at a particular scale. I mean\nevolution is a beautiful thing. Yeah.\nUm but you know it doesn't have to be\nintelligent to be beautiful.\nCorrect. Correct. Well good point. So it\ncould be it could be beautiful but not\nintelligent. We could have a beautiful\ndistributed civilization. That's fine.\nI'm happy with that.\nOn the internalism and externalism\ndebate, um I'm a huge fan of Andy Clark\nas as you know and um he famously argued\nin his was it 1999 paper with with um um\nDavid Charmer's that the phone extends\nyour mind. And even when we were talking\nabout the plant example before when we\nwere talking about plants doing modeling\num there's a there's an inactive\ninterpretation there and Keith and I\nwere arguing about this. He doesn't um\nagree with me but um is the model the\nthe physical morphology of the plant?\nYou know it fits the environment like a\nkey in the lock or is the model the\nsoftware? Is it the DNA of of of the\nplant? And and indeed it's just\nfascinating, isn't it, that that you can\nthink of cognition not being entirely\ninside our heads, but it's just this\nunfurling causal birectional process of\nof of so many things around us. So how\ndo we actually draw boundaries around\nthings where we say okay well here's the\nthe principle modeling and cognition is\nis in here. This is where bonafide\nbeliefs are happening. This is where\nbonafide thinking is happening. Is it\neven possible to make that distinction?\nUm, I I think it is. Um,\nI'm answering it in an almost trivial\nway that if you want to talk about\nsomething, you have to be able to define\nits markoff blanket. So, yes, you have\nto be able to draw lines, bounds and\nlines around things to talk about them.\nI think that your question though speaks\nto um something that we were that we've\njust been um covering which is the um\nthe separation of scales and the the the\neverything everything literally um from\nthe point of view of the free energy\nprinciple that is equipped from the mark\nMarkoff blanket uh is contextualized by\na scale above and the same rules apply\nto the scale above as to the thing\nitself. So you have to have a context in\nwhich everything is operating. So the\nvirus may not be intelligent but it\ncertainly has to be living in a you know\nin a world which is conducive to its\nexistence. And that world and the states\nof of that particular world at the at\nthe scale above. for example, the host\ncell and the host organism um has to\ncomply and be intelligent in some sense\num in order for the virus to be there\neven if the virus itself is not\nintelligent. So applying that to the\nplant um but we should come back to the\nextended cognition and Clark but to the\nplant thing you made an interesting\npoint you know is it in the DNA\num or is it in the morphology\num and the photoactics\nuh uh everything else that that plants\num um possess and we infer goes on on\nthe inside. Um\nfirst of all I think you're absolutely\nright to say yes you it is in the\nmorphology and in a sense that's what I\nwas getting at when talking about the\nimportance of mortal computation for\nmachine consciousness that you have to\nto realize a um a basin mechanics or\nself-evidencing you physically have to\nparameterize your conditional\nprobability distributions your basing\nbeliefs about the environment with you\nknow um to which you are coupled. So you\nit is the substrate is the\nparameterization. So by definition it's\nsubstrate dependent computation which\nmeans the structural form the morphology\nis the structure of the generative model\nin the spirit of structure learning. So\nthat to be part in very much in the in\nthe spirit of the good regulator theorem\nyou know to be in my world I have to\nphysically encode and embody a model of\nthe structure of my world which means\nthat I have to have if you like if you\nthink your world is scale free that\nmeans that my brain must have some\nscale-free hierarchal aspect and indeed\nit does and you could actually write\nthat down and model it with a\nrenormalization group um and that's just\na reflection of the fact that I um\nimmersed in a world that has some scale\ninvariance in it. Um so that is\ncertainly true for the plant. Um and uh\njust one little aside here that paper um\num was written before David Atenburgh's\num\nseries on the life of plants where he\nwas able to speed up by a factor of 10\nor 100. Uh and of course if you look at\nplants doing stuff they're con specifics\nby um sending their their little roots\nand off in a particular direction or um\nyou know eating insects or whatever. If\nyou speed it up the these things are\nvery anim animalistic and they look you\nyou'd be hardressed to say that they\nweren't intelligent irrespective of\nwhether they're conscious or not but\nthey certainly start to look much more\nlike like you and me when when you speed\nthings up. So yeah, they have their\nmorphology matters matters more than\nthan one could possibly imagine. Um\nbecause\nthis is how um you become a good\nregulator. You model your environment.\nyou are you install that that cause\neffect structure into your computer\narchitecture which is again the argument\nagainst vonneumman architectures which\nis you know um which is why um one might\nI think um look to all the um processing\nin memory neuromorphic photonics um\npossibly quantum computation but\nthat's gone off the boil recently uh but\nthe um all that processing in memory\nstuff memory for example I think that's\nwhere where the answer the answer will\nwill be is it in the DNA so is it does\nthe DNA cause a structure so\nyour comment about the expression make\nthat point\nyeah so I think um I I think some of the\nhangup or crux may actually be like what\nwe mean by model and I'm sure this is a\nwhole philosophical debate that I'm\nprobably not qualified to you know uh\ndecide side. But I think for example um\nyou know if you're flying an airplane\nlike the airplane internally has a\ncomputational model of the airplane and\nyou're interacting with this computer\nand through controls that then are\nenacted through all kinds of gears and\ncables and levers and things like that.\nBut I think of the model as being the\nthing that's in the computer and then\nit's modeling the the airplane, right?\nAnd so in that sense um you know the the\nDNA of of a plant species must\nincorporate the code for the model right\nbecause because that's what unfurls is\nthe plant it's the same DNA in every\ncell that's enacting a program that's\nessentially happening through gene\nexpression you know and so a particular\ncell maybe some salt content is in there\nwhich causes it to release a hormone\nthat binds to all the other cells in the\nplant which activates certain genetic\npathways case. So I kind of think about\nin that same analogy, you know, the the\ncore of the model is is the code, you\nknow, that's running and operating and\ncontrolling these processes. And then I\nthink of the embodiment of the plant as\nthe stuff that the model is modeling,\nright? But I guess, you know, there's\nsome some vagueness there, too. So\nyou could um if you wanted to simulate\nthese kinds of things um you could read\nthe DNA as the code that specif the\nprize that specify the structure um and\nyou know I think that um if if if you\nthink of DNA\nas prescribing the structure of your\ngenerative model or your world model or\nyour factor\num that will be fit for purpose and is\nlearnable.\nyou know, so you know, you start off\nwith with say DNA. DNA does not tell you\num what particular kind of plant you're\ngoing to be where you it's not going to\ntell you how to engage in your photoax\nsystem and point towards the sun or\ncompete with other uh plants that are\ntrying to deny you sunlight. That is\nsomething that you have to update and\nlearn during your particular lifetime.\nBut you're equipped with the basic\nstructure the prior on the structure of\nyour generative model and of course that\num is most e most gracefully\naccommodated I think again with respect\nto the reormalization group which is\ntalking about two scales. So you got a\nslow scale where you got the slow you\nmentioned before viral mutation. So the\nthe viral DNA should it have RNA or DNA\num um is changing very very slowly on a\ntime scale that is greater than or\nequivalent to the uh the lifespan of any\ngiven virus. And that's if you like\nspecifying the initial conditions, the\nstructure for the specification of a\nparticular instance of a virus.\nMaybe just jump in for a minute because\nI think actually it's not even\nspecifying the structure. I think it and\nand this is what you pointed out before\nabout the difference between say active\ninference versus modeling is actually\nthe DNA DNA is specifying a policy.\nYes.\nIt's effectively just specifying a\npolicy that every single cell in the\norganism follows. And so this policy is\nnot instructions for for uh a structure,\nbut it's in just instructions on what to\ndo in response to a certain set of\nsensory states. Right? So it's almost\nit's the policy that the DNA specifies\nand this policy applies to every cell\nand then somehow miraculously it works\nout to create a structure that is fit\nfor purpose in its particular\nenvironment.\nYes. Yeah. And of course to work that um\nspecification\nhas to change at the same rate at which\nthe environment changes which is varies\nvery very slowly. I think that's\nabsolutely right. I mean just\npractically um that hits you in the face\nwhen it comes to um thinking about how\nyou commit to one policy or another\npolicy. So in my world that would be the\nexpected free energy. The um expected\nfree energy comes in two parts. It has a\nsort of expected information gain and\nthe expected cost or constraints or\nutility and that is at the point you\nneed your DNA. You need to tell you need\nto know what it is to be the kind of\nthing that I am. What do I not do and\nwhat do I do?\nWell, the fascinating things in\norganisms is this is different even on a\ncellbyell level, right? because\nmorphologically the same code evolves\ninto all the different organs and parts\nand and roots and you know bark and\nwhatever else which is pretty\nfascinating. Well, and yeah, this is\nexactly the kind of, you know,\nfascinating issue that preoccupies Mike\nLeaven. Uh, you know, and how how do you\nuh\nI spoke with our our mutual friend\nMaxwell Ramstead the other day and he\nactually gave a wonderful description of\nthe free energy principle because you\nknow what you want in a good pitch is is\nfor it to be replicable uh easily\nunderstandable and I I'll try and\nrecapitulate it to you to just to test\nhow replicable it was. um he's you know\num second law of thermodynamics closed\nsystems what if we have open systems\nwith boundaries and he was saying when\nthings can't merge together they instead\nshare information with each other so\nthey kind of there's thisformational\nsynchrony and that was his way of\ndescribing the free energy principle but\nfirst of all is is that a good\ndescription but number two the\nboundaries where do the boundaries come\nfrom because we we want to have\npractical implement presentations of of\nthe the free energy principle. And of\ncourse, we can we can um implement it\nwith computer games or um contrived\nenvironments where the boundaries are\nclear. But what if I want to build a\nrobot and I and I have a camera and I\nsee all of these pixels on the screen\nand I need to divvy that scene up into\ninto objects so I can start doing the\nfree energy principle. How do we do\nthat?\nRight. Five minutes, five questions.\nGood.\nIt's got to be practical. That's why\nwe're only giving five minutes,\nright? Uh first of all um the um notion\nof\nbeing separate but part of a universe\nthrough synchronization I think is\nabsolutely uh absolutely correct. Um and\ncould be unpacked in two ways. one, you\ncould say that, you know, to find a a\nfree energy minimizing solution to your\ndynamics, namely the path of least\naction, just is to um events a\ngeneralized synchrony between the inside\nand the outside. Um generalized\nsynchronis also known as synchronization\nof chaos. So it's basically that two\nsystems that are loosely coupled uh in\nthe sense of um dynamical systems um\nwill ultimately converge on a\nsynchronization manifold and they will\nshow chaotic dynamics on that manifold.\nThat manifold is a synchronization\nmanifold and it is being on that\nmanifold that is the minimization of\nfree energy. So you can talk about the\nfree energy principle without mentioning\nself-evidence, without mentioning bay,\nwithout mentioning predictive processing\nor even extended cognition. Um and you\ncan just talk about it as um a\nvariational um specifying a variational\nbound on the lronian that specifies\ngeneralized synchronization\nor synchronization of chaos. and you\nknow what you're talking about is is\nexactly this sort of um separate but the\nsame um I think quite quite nicely\nalthough I don't fully understand it um\num recapitulated in Chris Field's\nquantum treatment so what he would talk\nabout there is not the synchronization\nbetween the inside and the outside or me\nand you and everything else like me and\nyou um but entanglement\nso the principle of unitarity is just\nbasic well you can read the free energy\nprinciple as just the principle of\nunitarity which just means for if we\nstay here for long enough for more than\nfive minutes uh we will ultimately\nbecome completely entangled um which\nmeans classically we will um we will be\nengaged in a generalized synchrony there\nwill literally be a synchronization of\nour interament dynamics we will not\nbecome one we'll still have a we'll\nstill have well actually Chris feels\nthinks you would become one but uh we\nwill we would be indistinguishable\nbecause we're just playing out on\nexactly the same synchronization\nmanifold. So I think that's absolutely\nright. Um your um your last question and\nyou had about three in between. Um was\nhow would you how would you practically\nuse the free energy principle or the\nnotion of thingness inher that inherits\nfrom Markoff blankets practically if\nyou're building robots. I think you're\ntalking here about sort of physics\ndiscovery or Markoff discovery\nalgorithms which of course you know\nMaxwell and Jeff Becker you know sort of\nfuriously working away on to you know to\ntry and uh assuming you didn't know\nanything about that stuff I mean how\nwould you approach that problem I mean\nfrom from first principles if if you had\na robot and a camera how would you\npartition the scene into boundaries\nyou would be appealing to to um the\nnotion of markoff boundaries and uh you\nI mean you could turn it on head and\nlook at image segmentation as just you\nknow what has emerged in computer vision\num as a way of identifying things. So\nwhat you're doing is you're committing\nto um an implicit world model or\ngenerative model for your robot. You're\nassuming that it is going to be\noperating as a good regulator as a good\nmodel of its environment in a in an\nenvironment that is composed of things.\nYes. But I think um my intuition is\nwe've been talking a lot about\nunderstanding and creativity recently\nand I I intuitively feel that\nunderstanding is about knowing the\nhistory of something. It's about knowing\nhow you got there. It's not about\nknowing the state. So it's a little bit\ndifferent from segmenting an image. Um\nto understand the history and the\ndynamics of a system seems to be much\nmuch more powerful.\nAbsolutely. Yeah. Yeah. Yeah. No.\nAbsolutely. Which is why you know your\nmark of B blanket discovery is not image\nsegmentation because to do markoff\nblanket discovery you have to look at\nthe dynamics and the history.\nYes.\nYeah. Um I mean there's quite a\nfundamental point to be made here that\nyou know we're talking about we have\nbeen talking about from the beginning\nright through to the end about\nconditional probability distributions.\nProbability distributions over what?\nBecause there's only one state in the\nuniverse at any one time. There are not\nmany worlds there, not an ensemble of\nuniverses. Um, well, of course, for the\npurpose of this argument, uh, there's\nonly one state of the at any one time.\nSo, you can't have a probability\ndistribution unless you invoke some sort\nof, um, ensemble canonical ensemble\nassumption as they do in thermodynamics\nwhere there's some exchangeability you\ncan swap swap around universes.\nYou can have an epistemological\nprobability distribution, right? I would\nargue though that what you're implicitly\ndoing is actually doing that over over\ntime that there's a history a past and a\nfuture. Um that that epistemological\nnotion um really has to refer to what\nare what is the what what is the the\nbackground over which these multiple\nrealizations occur from which you can\nselect to build a probability\ndistribution. And of course from the\npoint of view of the of of the free\nengine principle on also building robots\num this has to be time. So by definition\nit has to be in the dynamics and the\nhistory and not just the short-term\ndynamics but this recurrence the\ncharacteristic states that I keep\nreturning to as a robot as a good robot.\nUm uh so yeah absolutely. So\nsegmentation algorithms are, you know, a\ngood start, but they're going to be\ncompletely useless when it comes to um\nautonomous vehicles, for example, unless\nyou've built in the fact that there is\nconservation of this particular markoff\nblanket over time, which of course they\ndo. um uh but it probably better start\nwith the generative model that this um\ntwo-dimensional RGB feed has been\ngenerated by Markoff blankets that by\ndefinition persist over time because\nthat is that time is a support of the\nconditional distributions that define\nthe existence of the markoff blanket. Uh\nand then it's a question of how many h\nwhat kind of prize do you put on these\nthings? Are they things are this stuff?\nYou know, um what's the markoff blanket\nof um water? uh or well perhaps up fog,\nyou know, if we're doing autonomous uh\nuh um and do they uh to what extent\nwould I now apply uh my prior to these\nthings and you can go right through to\nsort of object centric prior um you know\nof the kind you know sort of physics\nengine based stuff of the kind that say\nJosh Tenbal uh pursues assuming that you\ngot sort of Newtonian behavior or you\ncould be much more relaxed Ed um um you\nknow I would relax but not beyond the\nreormalization group.\nProfessor Fen, it's been an absolute\nhonor. Thank you so much for joining us\ntoday. It's been amazing.\nI've really enjoyed it. It's lovely to\nspeak to you both again.\nYeah. And it's it's really been an honor\nfor me because this is the first time\nI've met you in person. So uh and you\nknow I've been thinking about you for\nmany years. So it's been it's been\nreally a pres pleasure to to meet you\nand thank you for uh coming to the\nstudio today. Well, thank you for coming\nto England.\nAbsolutely.",
  "transcript_chars": 73387,
  "ingested_at": "2026-05-12T00:42:39.212729+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 27723,
    "like_count": 912,
    "channel_id": "UCMLtBahI5DMrt0NPvDSoIRQ",
    "categories": [
      "Science & Technology"
    ],
    "tags": []
  }
}