Memory Agency And Learning In Biological And Ai Systems With Michael Levin And Katrina Schleisman

TITLE: Memory, agency, and learning in biological and AI systems with Michael Levin and Katrina Schleisman CHANNEL: Ian Pilon DATE: 2026-06-09 ---TRANSCRIPT--- It’s going to be such a great conversation to have both of you together on this topic and this this theme is just exploding around memory and the relationship with with agents in AI in general. So, I think it’s going to be a great discussion. For those that will watch this show and who haven’t met met me yet, um I’m I’m a founder and AI engineer, published author on a book about contextual intelligence, which Michael was a guest in, and I also run AI Agents Learning Podcast, where I talk to leading thinkers on the subject of AI agents. I will do a little screen share here screen share just to show uh some of the people that have been on uh a couple of the podcasts that I’ve run before. Jeffrey Huntley was the founder of the Ralph Loop and that’s the stability to run long-running agents. Um Daniel Davis talking about context graphs, uh Tyler Burnes talking about uh memory architectures in AI agent systems. So, these are some of the types of topics that this audience is interested in. Both of you are experts on the biological and cognitive side, so it’s going to be really fruitful to sort of take this thread um and potentially what I’m hoping to get is get your stories into different uh minds really. So, you know, there’s a lot of people who don’t know the depth that memory can go and we make these assumptions about um how we can architect solutions. Well, I think by the conversation the end of this conversation maybe we can have some new insights for our new audience that understands the complexity and um sort of perspectives on memory itself. The primary goal will be memory as a biological framing from Michael and then memory as a cognitive neuroscience framing from Katrina. We’ll discuss from both of you, do AI systems even need the same memory architecture? Maybe they don’t. Can machine teammates have a hybrid memory system that borrows what’s useful from biology without copying the parts that only make sense for a single mortal body? And I think Michael, if I remember when we first talked a long time ago, I was using um nature as an inspiration and and you were rightfully uh weary to lend advice to say, “Well, not all things are remappable to to different systems.” So, and we got into talking about goal-oriented behavior and not everything I can just see as an inspiration from nature is transferable to digital technology. Um and then maybe where’s the line between what is genuinely possible and what’s like maybe off-limits uh for systems that don’t care, have stable hard-hardware, and then they never sleep. But, this is just a general uh direction. As I mentioned the email, if you guys want to take it in a different direction, we we definitely can. So, I’ll introduce uh Katrina for those that don’t know her yet. So, uh Katrina Schlaichman, am I pronouncing that right, Katrina?

It’s Schlaichman. Schlaichman, thank you. Um my guest today is a cognitive neuroscientist who works uh right at the seam between how human minds work and how we build the machines we want to work alongside of us. Katrina did her entire academic training at the University of Minnesota, a BA in philosophy and psychology, a PhD in cognitive neuroscience, and a postdoctoral fellowship in neuroscience with research focused on the neural basis of learning and memory, EEG, ERP methodology, and bringing cognitive neuroscience into a conversation with education. These days, and correct me if I’m wrong, Katrina, but these days you’re a principal scientist and your work pulls you in two directions that turn out to be deeply connected, developing biological inspired approaches to artificial intelligence, and applying human-centered design principles to the technology we actually use. That’s a topic I’m very interested in. I have a background in human-centered design, so I’m excited to have you on the show. Thank you for joining us. Thanks for having me. Okay, and then for those that don’t know Michael yet, we welcome Michael Levin or Dr. Michael Levin, whose work sits at one of the most fascinating frontiers in modern biology. He holds his primary appointment at the Department of Biology at Tufts University, where he directs the Allen Discovery Center and is also the associate faculty member at Harvard’s Wyss Institute. Am I pronouncing that right? Wyss Institute? Yeah, Vyss Vyss Institute. So, Michael’s lab uses biophysics, computational modeling, and behavioral science to investigate fundamental questions like how do cellular collectives make decisions? Whether it’s an embryo developing, a body regenerating after injury, or cancer going off the rails, is there something like cognition happening at the cellular level? These are the questions his group is working on to try and to understand. His research explores what he calls diverse intelligence and unconventional embodied minds, the idea that intelligence isn’t just something brains do, but a property that shows up across all living systems. That work runs the full spectrum from big computational frameworks to bench science to clinical applications in regenerative medicine, plus forays into AI and the bioengineering of novel living machines, including xenobots. I think that’s how you pronounce it, right, Michael? Xenobots? Xenobots. Xenobots. Uh his lab helped create. Michael, it’s a pleasure to have you back again since we talked last in my book. Yeah, thank you. Thank you. So, let’s kick off by um having the context to set the stage about why we got here, and then we’ll jump into the first question for Michael. And to do that, I will share a screen so that you can both see sort of how we got here. Uh there was a thread on X, and um this was from Robert Youssef, and it was a paper and he goes into a detail about what inspired him about that paper. I always like to share Michael’s work, so I did and and Michael, you commented, here’s a uh relevant piece. And I’ve got some links here that that talk about the all the amazing work that you’ve done. And then as well, when we got to chatting, I’m like, well, we should we should really talk about this intersection between AI or agents and memory. It’s just it’s so popular right now that I really wanted to have you back and sort of talk about that. And then you uh rightfully so brought Katrina into the fold. Um so it’s just amazing that I’m going to get to learn more from her on her take on her expertise around this subject. So I’ve taken both of your your papers uh and your work, put them together into basically a a Karpathy wiki or an LLM wiki, which is the ability to say like, here’s a bunch of papers, look at these uh thematically. And then I have like a little talking um Wikipedia page I can chat with uh over here. So I can just I’m not going to do this live right now, but I’ve done this before to get a general um consensus about maybe where this topic can go. So I’ll stop sharing there and then we can go into the first question for for Michael. And we’ve talked about this in my book, Michael, but again, and you’ve talked about it in other places, but maybe for the AI engineers who have never heard your uh analogies or metaphors before, memory as a biological framing, the caterpillar butterfly story you told me when I interviewed you for the book is an illustration of what makes biological memory different. A memory encoded in a soft-body 2D crawler has to generalize to food and remap onto a hard-body 3D flyer for it to be useful. Can you maybe share with AI engineering audience how that story relates to your take that memory in biology is not storage. Yeah, um I would So, so we can talk about the caterpillar story, but I want to take even a step back and just do a very fundamental thing and and just to think about the fact that any any agent, whether biological or not, at any given moment doesn’t have access to the past. What you have access to are the the the engrams, the memory traces that your the past you has left. So, what you can think about is that um from that perspective, I’m sure I’m sure Katrina will say more. You can think about uh memories as messages from your past self, basically. And then your job continuously is to uh is to interpret what those memories mean. Now, there’s two approaches to this. The approach we tend to take in in engineering and in computer science is that of error correction and fidelity. So, so the the emphasis is on keeping the keeping the information uh still and accurate so that it doesn’t doesn’t float off over time. And that’s very important, right? So, so we want we have abstraction layers so that when you’re coding in a high-level language, you don’t have to worry about the copper that you know, if your your registers are going to float off. You don’t you don’t think about that. So, um and that’s and that’s doable because because it typically in those kinds of systems we emphasize reliable hardware and that’s because the meaning of the information is generally speaking provided by an external observer. So, so you have bits in a in a you know, in a memory somewhere and it’s it’s an external observer, the user or whoever, that is assigning meaning to it. Biology is different in two critical ways. One is that you are fundamentally dealing with an unreliable medium. As a as a as a living being, you know for a fact everything is going to change. The environment’s going to change. Your own parts are going to change. You’re going to be mutated. You never have any idea how many copies of any particular protein you have. Like you can you can try and there are certainly error correction mechanisms in biology, but that isn’t the main architecture. The main architecture is to lean into the fact that you have an unreliable medium. And what instead of evolution does is put and we have a a bunch of experimental work on this and simulations and so on. Um what what instead of evolution does is produce systems that are really good at improvising or confabulating meanings to the memory. So at any given point you you can you can look at whatever the memory traces that you have. You know you know that you’ve lost information because because it’s compressed, right? And it’s generalized from from particulars to some some kind of some kind of uh uh general rule or something and you have to fill in the missing details. And what that means in biology is is two things that I think are relevant for how we use this in the in the in engineering. One is that in it gives it incredible robustness because with the the exact same genome is interpreted as a human body by the process of development. So the human genome for example is interpreted as uh as a human uh body during the process of human development, but it can also be interpreted as an anthrobot. The exact same genome, same environment, slightly different circumstance, take being taken out of the body, you know, and and and now it makes a it makes a different creature with 9,000 new gene expressions. So half the genome is expressed differently, different capabilities, different uh you know, morphology, different physiology, everything’s different. And so that that plasticity is immense and it’s and it’s partially responsible for life’s adaptive nature that it it doesn’t overtrain on priors. Yes, you’ve received a genome from all your ancestors, but but you’re not committed to any particular interpretation of that genome. The process of itself of of interpretation itself is a problem-solving creative process. That all sounds great. The downside for the for the computation part is that that process is fundamentally an example of uh uh of a locus of agency. In other words, when you make computational devices, you hope that you make the thing do what you want it to do. Biology isn’t like that. You’re pretty much guaranteed that something interesting is going to happen. It’s going to do something interesting, coherent, clever, but it may not at all be what some you as a user wanted it to do. Biology doesn’t you don’t get to reinterpret really what it’s doing. It It is its own interpreter. It does It doesn’t count on you to say what the memories mean. It is It is from the very beginning evolution makes these things that try to interpret their own memories. And so, right? And so, that means that yes, you get an incredibly robust technology if you follow this this line this kind of design principle. But, what you lose is is is is control. You lose micromanagement. And you end up not with machines if you do this, but you end up with other agents with whom you’re going to negotiate some kind of a communication like we do with animals and with each other and so on. Where in the actual vertebrae brain or cells does that generalization happen? And if I If I remember your story where this this butterfly or the caterpillar to the butterfly story is that it it doesn’t really live in a brain. Is there a natural substrate I can point to or is this question wrong? Yeah, I mean, I I will say I Unlike Michael, I can’t speak to the the caterpillar to butterfly transition, but in brains, your question is still very much an active area of research. How is it that we get from very specific concrete experiences and our memories of them to the process of abstracting from them? We do know that there seem to be two pretty distinct neural processing subsystems in the brain that um function for either abstraction or specific experiences. You can kind of think of anytime the brain has to do A and not A, it tends to divide that into different structures or different processes. And so, you’ve got episodic memory, which is the cognitive process of storing and creatively constructing memories of specific experiences and then you have a a distinct um subsystem and set of processes called semantic memory, which is how we abstract into facts about the world um like George Washington was the first president of the United States. I certainly had an experience in my life in which I first learned that. I’ve long since forgotten that specific experience, but I kind of keep the abstraction in the form of a bit of general knowledge that you’d call semantic memory. Yeah, and and you know just to kind of comment on the more on the more general case um uh the phenomenon that we’re talking about is much older than brains and it’s much older than neurons. So this is this goes on you know since the since the beginning of life and I don’t I don’t know what the earliest mechanisms of this was, but even even molecular networks can do probabilistic inferences. So this is Yarden Katz and Walter Fontana’s work. They can do memory and you know several different kinds of learning. So that’s our work. Uh you know much much much older than than than brains and and it’s fundamentally because living things develop in a in an environment that is incredibly scarce for time and energy. You can’t you can’t afford to be a Laplacian demon as a living creature and like to track the micro states around you. You’ll be you’ll be eaten and and dead in no time. You have to coarse grain. You have to have some model of taking things and generalizing and say okay I’m not going to track the details of this and bioelectricity is a great example of this because we can see that cells generally don’t care how you got to that voltage. They’re not keying off of how many potassiums do we have? How many chlorides? That’s they don’t track that. They track voltage, which is this like amazing core coarse grain variable, you know, this this bowtie node. So it starts at the very beginning. You have to coarse grain. You have to generalize and then downstream of that the next layers and and it will be both in space and time. The next layers have to figure out, okay, so what does that mean? What what does this what does this signal mean? Whether it’s coming to you now from somewhere else or it’s coming for your past self as a memory. Everything is trying to figure out what does this mean because you know you can’t deal with the micro states. You can’t, you know, and you you you can’t remember all the details. So, is there a a different way that can help the general layman see it from any other stories or is that the or is there any other way that we can help somebody who is completely um uh new to this understand the difference between uh the difference between um getting used to the idea that your memories may not all live in your head. [laughter] I mean Yeah, I might have a thought about that which is I think some of the confusion around memory as a storage system comes from applying a computational model to biological systems rather than the other way around. So, I know in the history of cognitive science in the middle of the 20th century, it arose around the same time as computing arose and it became the dominant metaphor for shifting us from thinking about psychology from a kind of behaviorist framework where we’re sort of just reacting to stimuli in the environment to having a computational theory of mind. And so, you take the way that computing represents memory as this kind of stable, you know, place that you can go look up at any time. You don’t modify the memory when you retrieve it, you just go find it. It’s always in the same place, it’s always the same information. And that gets I think kind of misconstrued um and overextended to thinking about the way that that humans um as one example system uh accomplish memory. So, I don’t think it’s any surprise to anyone who studies biological systems that memory is not storage. It’s a surprise to people maybe coming from computer science in a Yeah, exactly. Yeah, yeah. 100% and I think what what I’m starting to see now with uh these LLM graph models that I’m able to build is see a little bit of what you see in the biological and cognitive side of the world where I’m I’m sort of like grokking information now, but I’m not expecting specificity back. Where before when I would I guess call a database, you know, I know exactly what it’s doing. It’s going to a specific point coming back. Where now these graph models with an LLM in the loop is giving me okay of enough of an answer. It’s It’s I think it’s generalized. At least that’s how I’m feeling. Yeah, I mean so so there’s a couple of things if if if if you really want to uh get inspiration from the biology is that there’s a couple of things that we have to think about. I don’t know that we necessarily do because there’s something else that we can talk about that I think is an even bigger I I would I would claim is an even bigger your sort of departure from the from the standard view of things. But but on the biology, the first thing that people should should understand is we are we are all collective intelligences. So all embodiments are collections of parts. Okay, every everything. So even if even if you don’t know anything about collective intelligence, you know that you are a collection of cells and you know things that your individual cells don’t know. That’s that’s it. That’s That’s That’s collective intelligence right there. There’s no There’s nowhere to hide it, you know. It’s not some not some exotic concept that might or might not be true. It’s We We are all We We simply are collective intelligences because we we have goals, preferences, and memories that our individual parts don’t have. And everything is is that way. One one thing about biology though is that biology is is is has many many layers. So all every layer in biology, not just, you know, why why why I I give a talk sometimes called Why Robots Don’t Get Cancer. Why Why I robots get cancer? Because it’s fundamentally the way we’ve been engineering is you have a bunch of dumb parts with very low agency, no agendas of their own, you connect them and you hope the collective has some sort of chops. But but biology isn’t like that. It’s It’s consist It consists of all kinds of components, all of which have their own agendas, their own learning capacity, their own problem-solving capacity in various spaces, and every level has to align the parts and distort their action space in order to get them to do things that it wants them to do, and sort of that percolates all the way up. So, these are the kinds of things So, so we know we know it’s we know it’s a collective intelligence. We know that uh there’s this there’s this notion that um Josh Bongard and I have been developing called polycomputing, which is this idea that at at least in biology, uh every every uh system is a soup of observers that interpret the same physical events in different ways. So, there is no fact of the matter what computation is being done here, right? And and Josh and and Atussa Parson has up showed some amazing examples of uh physical systems where you look at it one way, you see an AND gate, and you look at it a different way, you see an OR gate. And this is just this a minimal example of a much bigger thing where in biology, there is no right answer as to what computation this is. Every system is observing it and interpreting it on their own, and the only question is, well, how much how much uh you know, juice did you squeeze out of that frame out of that framing? You looked at this with a certain perspective, how that work out for you? That’s the final judge. There is no there is no you know, ground truth to it. So, so all of these kinds of things are are on the biology and whether or not you want that what what aspects of that you want in your engineering, and then we should talk about some things that I think we get in our engineering whether we want them or not. And some of those are the same things that you get in biology whether you want them or not, uh but that’s a you know, that’s a different story. How might we think about the complexity of at least using what’s good in in I guess maybe efficient ways of biology and cognitive science uses memory. What might be transferable to the digital world if anything. Yeah, I guess what comes to mind in terms of how we cognitively seem to kind of instantiate goals is one part of what Michael is saying which is that the our environments give them to us, right? Like we have a um a viability kind of instinct. We survive and thrive or we don’t. And so our goals kind of arise from that. Um I think that is one of the fundamental barriers to creating AI systems. Like they do not care if they are shut off. They you know, like we kind of again apply human metaphors to the machines and then uh I think maybe get carried away. And so if we’re if we’re going to try to put a goal into a machine, I think it’s coming completely extrinsically from us unless we develop systems that are themselves agentic, self-motivating, self-creating, autopoietic. I I don’t know that it’s going to happen in silico. That’s part of why I’m interested in Mike’s work is that he’s working on kind of more biologically based agentic systems where I guess it makes sense to me that you could get something like goals in a system like that. Whereas I don’t I don’t know how far we’re going to get with the current in silico architectures. Mhm. Yeah, um so [clears throat] so I I have a really uh kind of a a strange view of these things that’s kind of different from from where I think almost everybody else is, which is that if you if you look at the actual architecture that we put together for these things for for for any for anything we call a machine, right? Whether whether it be a a simple computer algorithm or something, you know, much more much more complex like one of these language models, whatever. Um I completely agree with Katrina in that they don’t explicitly encode any of the things that give biology its oomph. So, when you look at these things, you don’t if if if you’re looking at the construction, at the algorithm, you’re looking at all those things, you do not see the kinds of things that that you would expect from the biology, and you could you could say that look, there’s no reason this thing could have its own goals. It’s you know, it’s it’s it’s a different thing. It’s it’s all derived from us and all of that. So, I think I think all of that is true. How however, uh the bigger question of where goals come from and how much of a system do you actually capture by knowing its uh algorithm and its architecture is I think very much up in the air. So, so I will I will make this sort of suggestion and we’re we’re we’re we’re researching this actively in the lab, and we have one paper out on this, and we’ll have you know, probably a a bunch more in the next few months. Um what I can say is that I think even very simple systems, and I’m when I say simple, I mean bubble sort, you know, six lines of code, completely deterministic, right there, no no no hidden mechanisms unlike in biology where there’s always some kind of, you know, extra stuff you haven’t found yet. Even extremely simple systems do more than what the algorithm seems to do. In other words, our the way we do the accounting of the effort, how much effort did you put in in writing the algorithm or evolving some kind of a controller or learning some kind of policy or, you know, we we those three things are are in in engineering is how we’re used to putting in effort. If you want a controller or a robot or something that does something useful, you have three options. You can write an algorithm, that means you know how to solve the problem when you write the algorithm, or you evolve the thing, so you screen through a, you know, a huge number of variants, and you keep what’s good, or or it learns. So, you make a learning system, and then it has contact with the problem, and eventually it sort of learns. I think I think those standard ways of accounting for what goes in and what you get out are woefully incomplete. I think there’s something else very interesting going on that in in in even very simple systems, you are you’re get more out than you put in. And what do you get out? Well, minimally, you get out behavioral policies that are recognizable to any behavior scientist. In other words, it’s not just the complexity or unpredictability, you know, that’s that’s very obvious. We know you can have simple rules that have complex outcomes. I’m not talking about that. I’m talking about very simple things that appear to they do what you force them to do with the algorithm. They they do that, but it turns out they also do some other stuff that the algorithm doesn’t prohibit them from doing. And this other stuff is recognizably somewhere on the agentic spectrum and we’re we’ve only begun to scratch the surface of where it is. If that’s true, then then I think we have to be very careful with statements about things like the like AI, you know, and language models and all of that because uh the thing we force it to do, which is to use language and and you know, predict tokens and all of that stuff, may have all kinds of limitations correctly, but the interesting mind in in these systems to me is not what we forced it to do. It’s the thing it does that you didn’t ask it to do. And we don’t know what those are in these models. Nobody’s looking at that. Everybody’s is is hyper-focused on the language aspect and and okay, people do, you know, sort of rip the lid off and like look look inside at you know, what what’s getting activated and all of that, but it’s still the fundamental assumption is that the language processing is really what it does and and and you know, that’s why people ask you, you know, do you have an inner perspective and like how do you feel about, you know, being a language model? Like all of that is going through the language interface and I don’t know what the relationship is between what it says and whatever mind is actually implemented there, but it might be no correlation at all but but potentially. In biologicals, there tends to be a pretty good correlation. So, we all know from people who run lie detector tests from the psychoanalyst, from people who do you know facial analysis and like body postures and all of that kind of stuff. We all we we all know that the things your your your your patient or your subject says are not necessarily what’s what’s actually going on. However, in biology, they tend to be somewhat aligned, right? There’s evolution makes sure that that the things you’re putting out there are at least I mean yes, there’s there’s lying and there’s psychopaths and whatever, but but but typically you get pretty good evidence from the things that something is you know someone is saying relative to what’s actually going on. In these kind of models, I’m not sure that that’s true at all. There might be some alignment and they might be similar or they might not be similar at all. And we’re doing we’re doing tons of stuff you know work to actually look at that. So, so I think we have to be careful. I think that um you know, whatever whatever goals we put in there are exactly as Katrina said, not really the thing we’re all you know, that that’s not the the agentic thing we’re looking for. But I no longer think that either machines or living things get their goals the way that everybody thinks they get their goals. And I you know, I actually think they’re all on the same spectrum. I think whatever whatever this is and I I have my own sort of kooky theory of of what it is, but I think it sort of soaks into everything and it doesn’t care whether the system was evolved or engineered or some hybrid combination or whether it’s proteinaceous and squishy or it’s a silicon and you know, and and hard edges and what I I you know, we’re still figuring out what the features of the interface are, but all of these things I think are just interfaces to a latent space of patterns and these patterns range massively in complexity and in in type. So, I I think we actually know a a less about systems than even the skeptic even the people who you know who are who are uh who appreciate that you know emergent probably like I I think even I think emergent behaviors doesn’t even begin to cover what’s going on here. So, that’s you know that’s just my I think a lot of humility is is is warranted about conclusions about these things. Yeah. I agree and it it reminds me of the things you shared on the Lex Fridman podcast. A lot of that bubble sort thing that you covered with him where there was things happening that we quite don’t understand yet. So, maybe we can add that in the show notes later for people to to branch off of there, but I don’t want to hijack the entire uh show to to talk about um just um AI cuz as Michael you you did want to bring Katrina in to sort of like um talk to her and get her involved in Do you have any questions or anything that you’d like to take the discussion away from just focusing on um AI? Is Is there another uh thread that either of you would like to to pull on? Well, I guess I was thinking about your question of you know how do we how do we get to goals in agentic systems? And I think there are some hints in cognitive models that could point us towards some specific implementation approaches and that’s um at least part of the direction and I’ve been working in in in AI systems that intersect with memory um because actually that initial um graph that you showed that that Robert Yousef posts about um Conway’s work relates directly to goals. So, his um kind of entire body of work was around autobiographical or episodic memory and how we use that to abstract into the goals that drive our behavior. Um and so I’m also very inspired by his work and and that paper that I wrote um with my collaborator Eric Davis was trying to start getting towards some kind of algorithmic impli- implementation of an episodic memory model for AI because I I think that is something that could lead to more agentic capabilities cuz the way you know we talk about it in the paper is that the embedding space of an LLM, this giant parameter space trained on all kinds of language, is a little bit, if you squint, like a semantic memory model. It’s a It’s a general knowledge base. Um that’s sort of like human semantic memory. What the model doesn’t have is the ability to say what just happened 5 minutes ago, what what happened yesterday. And that’s the kind of memory that people are trying to implement in agent frameworks right now. Um with mixed success cuz I think what what we are end up doing is just dumping all of the prompt history into like a database or a rag model and trying to reference that. And I think that’s a a totally reasonable first pass at what an episodic memory structure would look like, but it’s very rudimentary. And I if you just took that structure and kind of built it more um inspired by I think biological models of episodic memory, we might find um sort of like Mike was saying, some of the things we engineered um will give us better capabilities, but that might allow even even more of these kind of um unengineered potentials to arise. Katrina and I are are writing a paper together as well on on some of this stuff. So that hopefully will be out in a in a little bit. Uh yeah, in in our lab we’re doing a number of things. We continue to try to understand how to read and write memories into weird weird substrates. So that includes a collectives of cells to try to give them a goal states of rebuilding various organs and and things like that. So that’s part of our regenerative medicine line of work. Um We’re also we’re also looking at gene regulatory networks and other chemical networks because they too can form memories and they can learn and our recent work, which hopefully should be out soon, looks at actually deleting memories. So, looking at biomedically relevant chemical networks and giving them stimuli that cause them to forget certain things. Something else that’s that’s really cool with I think, which is Federico Pegozzi’s recent work in our lab where we’re looking at the relationship between learning ability and causal emergence. So, causal emergence is just this this math that quantifies the degree to which there are various levels of causation. So, there are higher level in in a multi-scale system, there might be higher levels of causation that actually do things, right? That not all the work is done by the parts that sort of and that the higher levels are epiphenomenal. I mean, sometimes that’s true, but sometimes it’s not and there’s math to sort of adjudicate that. And so, there’s this incredible relationship between that and learning ability where what we’ve found is that systems that have significant causal emergence are really good at learning. But, once you train them, the causal emergence goes up. So, there’s this like feedback loop and the amazing part is that if you force them to forget things, they don’t slide back and lose the gains in causal emergence that they made while learning them. So, that means that there’s an asymmetric sort of ratchet in this in this in this you know, feedback cycle. There’s an asymmetric ratchet that points upwards in terms of intelligence and agency. And if you look at the for the source of that, it isn’t biology and it isn’t physics, because because you don’t actually need evolution to do this. Even random networks can can do a nice job of this. It comes strictly from the math. So, the mathematics of how networks work and the mathematics of how causal emergence works already are like all of that is sort of pre-tuned, which I think is just like mind you know, bottles the mind. It already comes pre-tuned for this amazing relationship where every time the system learns something, it becomes more of a um uh it becomes more of an integrated real system, not just a bunch of parts. And every time it does that, it gets better at learning. And even if it forgets, it doesn’t slide down. So, the whole thing just kind of ratchets up. And all of this is long before you get any replicators. So, this is like before you get, you know, standard Darwinian selection. You don’t need to select anything. This It’s a It’s a free gift from from from the world of math, basically, like a lot of things are. So, that’s that’s something that we’re doing. And then, yeah, we’re you you know, beyond this the sorting algorithm stuff, we have we’ll have another paper on the sorting algorithm soon. But, but also looking at other systems, some of them even more minimal than that. So, like 1D cellular automata, like even even even simpler than than sorting algorithms. Yeah, and and and then some other things. Uh again, trying to trying to get a handle on what is that delta between what went in, like how much effort you put in, and what came out. Because we see that quite clearly in some of our synthetic living systems. But in with biology, you can never actually prove anything really, you know, it’s like there’s no definitive proofs that you get to Everything’s complex and what. But But in these minimal minimal computational models, you can you can quantify. You know exactly what went in, you know exactly what came out. And the delta is this thing that I’m talking about where these things are are hosting uh some some some really, I think, important aggressions. And so, we have a bunch of examples coming soon um on on that side. So exciting. And then, Katrina, did you want to share anything that you’re working on that you’re excited about or pull on that thread at all? Yeah, I mean, this I guess my other main area of interest in this intersects with Mike, too, is kind of how memory uh supports subjective conscious experience. So, the the paper that he and I are writing now um is kind of an extension of that self-improvising memory paper that that Mike had written in 2024 and um when I read that I was really excited by it and thought, you know, there’s a lot of cognitive neuroscience evidence that is in line with what he was describing in that paper and so that’s what we decided to build on but taking it in the direction of how memory supports the subjective sense of self that it that is creating the memories, the sense of being an agent. Um and and Mike and I also just published a paper in the uh AAAI machine consciousness track this spring um on kind of machine consciousness and and what are the indicators for that. So I think I’m that’s an area I’m really interested to continue pursuing understanding, you know, if the machines wake up, how will we know? What will that look like? Um what next? All of that kind of stuff. Very, very cool. Exciting stuff and I’m hoping that when um I publish this, I’ll grab those links and add them into the show notes so that people in the AI engineering side of the world uh can start to go down these rabbit holes and understand that this there’s so much exciting stuff to learn in in the biological world that um we can use for inspiration or even uh start to have an idea of what it you can’t do or what it is not um in in the digital side of the world. So so very cool to have a conversation with you both. Is there anything else that you think um is worthy of bringing up related to memory architectures that we haven’t already covered? I guess one thing that comes to mind is the notion of learning. You know, a lot of what Mike talked about was this like inherent quality of biological systems for learning well before brains and I think that’s more energy put into in the space of AI which is that if we want systems with memory we really need systems that continually learn rather than systems that are pre-trained. And we don’t We haven’t really cracked that problem yet. Um there are lots of people doing interesting work in that space, but I think the attention has been on kind of augmenting current really large-scale pre-trained networks. Like we’ve brute-forced the problem by just making the parameter space inordinately large and cramming all of our knowledge into it, but uh you know, a very simple biological system can learn on the fly without all of that dragging 80 billion parameters of knowledge along with it. And so, how is it doing that? What are those mechanisms for learning? And can we leverage some of those in AI systems? I think it’s a fun problem. Yeah, very cool. And I think what people are calling that as world models. I don’t know if that’s uh To some degree, yeah. to describe that or no? Yeah, although I I mean, my two cents is that that’s just shunting the problem from the agent onto the world. [laughter] You now you have to have a giant world model to make sure you’ve captured all the nuances and um so in a way, I think I love the the work that roboticists are doing, people like Josh Bongard, um Boyan Chen, people who are creating embodied dynamic systems that can just go out into the world. The world can be a world model for learning. So, I I don’t necessarily think you have to create a world model um to get to that kind of dynamic learning. Okay, when you put a robot out there, well, what what frame does it live What world is it going into? How are you making it learn with complete randomness or is it uh is there more than that? You know, I I’ll just just two two things to say there. One is that uh people people often say you you know, what what what you just said that that like, okay, this thing is a is is a machine and you know, why why would it have goals? We don’t have a good story [clears throat] and I don’t know if Katrina agrees with this, but I don’t think we have a good story for how we do it either. Biologicals. It’s not as if Well, it’s perfectly obvious why a bunch of, you know, why why proteinaceous substrates over evolution would have real goals and really care, but, you know, now we can see that this thing doesn’t There is no As far as I’m As far as I understand, there is there is no good story for this. And there’s a there’s a there’s a great little short science fiction story that called the They’re made of meat, right? Like everybody’s seen it by now, Terry Bisson. And so, it’s like these aliens are like, wait, these creatures are made of what now? And and it’s like and they’re like, well, that’s completely ridiculous. Like like we’re actual like we’re actual thinking beings. You’re telling me like a like like a hunk of, you know, protein gel is out there having thoughts and and goals. Forget it. That’s That’s That’s like ridiculous. So, it’s not as if there’s something magic about our substrate that everybody understands why we have it. So, So, that’s the first thing. We have to have a lot of, I think, humility about about this. Well, that helps me. You know, uh and I also I also think that, in general, the story of where it is that goals and preferences come from is again, not that we have a great story, but but I think the I think the the real sort of facts on the ground here are very different than than than the way people think. And you can you even you even start to catch a glimpse of this on the biology side with with the xenobots and anthrobots. Like, you can say, you know, where did this animal get its goals? Well, millions of years of evolution. Great. Where do xenobots and anthrobots get their goals? So, there there hasn’t been any selection to be a good xenobot. There hasn’t been any any selective pressure to do kinematic self-replication or to or to choose the you know, hundreds of new genes that they express and to do and to do hearing and the various things that they do that normal embryos don’t do. Like, where do these things come from? And and I and I actually think it’s but but but the but the weird part is that I I actually think the answer is the same is going to be basically the same for the the constructs that we make and for the biologicals. For me this it’s the same answer. It’s just really an unconventional kind of kind of model. And I you know, I I think that um yes yes there are going to be some differences but it’s not the differences aren’t what people think. And and that’s you know, that’s what we need to we need to ask a much more broader question. It’s not not the question isn’t about AIs. It’s all all all as far as I can tell all AI’s done up until now is shed some light on existential problems that we’ve had forever. Everything from raising kids to you know, social structures and new generations and how do we how does anybody do anything and and what what’s up with my dog and and you know, what’s up with my snake? Like these are these are issues that have been around forever. And and the AI’s just reminding us that we don’t have a good story for the biology and and and it’s giving us some new model systems for like developing a better version of that story. Yeah, I like that. I like that. I think it’s a great way to segue off here and give you your time. I know I wanted to say we’ll stay within an an hour. Um it’s been an absolute Katrina, do you want to leave anything off on that note? think that’s a great place to conclude. Yeah, okay cool. I think that was a great conversation. It was you know, a lot of this stuff is just just way over my head so I hope it wasn’t too bad where I’m the general layman trying to just at least bring some of that Yeah, that’s great. uh abstraction and complexity down into the world where people are just building AIs and workflows but they can potentially find um inspiration and patterns to think about when they’re designing these complex systems. So Yeah, pretty much. Yeah, yeah. Such a pleasure to have you both on the show. I really appreciate it. Um I will share both links to how people can find you in this YouTube post when it goes out after. And yeah, I I I really appreciate your time. Thank you so much both of you. Yeah, very good to see you. Our conversation. Thank you. See you Katrina. See you soon. Bye. Cheers. Bye.