Watch Episode Here
Listen to Episode Here
Show Notes
This is a ~48 minute discussion between Alexey Tolchinsky (https://scholar.google.com/citations?user=tiBKmrsAAAAJ&hl=en), Patricia Silveira (https://scholar.google.com/citations?user=VexSZ40AAAAJ&hl=en), and me on the topic of molecular markers of trauma and how to extend ideas about the psychology of trauma to minimal models of cognition such as Xenobots and Anthrobots.
CHAPTERS:
(00:00) Early adversity and EPGS
(07:02) Trauma biomarkers and networks
(15:29) Designing anthrobot validation experiments
(19:42) Xenobots, rodents, and trauma
(27:10) Optimal developmental stress window
(33:41) Too little stress question
(41:49) Uncertainty, self-models, and play
PRODUCED BY:
SOCIAL LINKS:
Podcast Website: https://thoughtforms-life.aipodcast.ing
YouTube: https://www.youtube.com/channel/UC3pVafx6EZqXVI2V_Efu2uw
Apple Podcasts: https://podcasts.apple.com/us/podcast/thoughtforms-life/id1805908099
Spotify: https://open.spotify.com/show/7JCmtoeH53neYyZeOZ6ym5
Twitter: https://x.com/drmichaellevin
Blog: https://thoughtforms.life
The Levin Lab: https://drmichaellevin.org
Transcript
This transcript is automatically generated; we strive for accuracy, but errors in wording or speaker identification may occur. Please verify key details when needed.
[00:00] Patricia Silveira: I'm an associate professor at the Department of Psychiatry here at McGill University in Montreal. I'm a pediatrician by training with a PhD in neuroscience. My lab works with the long-term effects of childhood adversity or early life stress: how that affects neurodevelopment, behavior, and risk for disease. We use different types of approaches to try to understand these effects. We use rodent models, large data sets of humans, genomic studies, and neuroimaging studies, trying to address the question of how early life stress modifies development and risk for disease in the long term. That's the broad question of the lab. I think that we got together because of your interest in the expression-based polygenic score, which was a tool that we created to address this question. It ended up being applicable to many different other questions, but it was originally created to test this and to bridge what we found in the animal models, where we can really explore the molecular mechanisms associated with early life stress — what happens acutely, what happens in the long term — and in children and humans, where we can't really measure things in the brain. We can't really study molecular mechanisms with the same level of detail, so we started using genomics to try to bridge and study individual variations in genes initially. But genes do not work alone, do not work in isolation. They work together in coordinated networks and help each other in biological processes. We started to understand that we needed to go beyond understanding the effect of a single gene or variation in a single molecule to understand variation in several molecules working together in networks. That was the basic idea: first try to identify a co-expression network that was involved in a biological process and then use this information to inform the development of a polygenic marker to understand individual differences.
[03:17] Michael Levin: How happy are you with this? Do you think that it's basically a good tool now, or is there a lot of room for development of new markers?
[03:29] Patricia Silveira: No, I think it's a nice tool. There's a lot of room for creation from here. We started by simply selecting genes that were co-expressed with a specific gene, like an anchor. One of our first papers was on how metabolic signals act in the brain and modify neurotransmission. I'm always very interested in how insulin, for example, produced by the pancreas and important for peripheral glucose metabolism, modifies neurotransmission in the brain. So we wanted to understand individual variation in the expression of the insulin receptor network. We identified all the genes that are co-expressed with the insulin receptor in a specific brain region, the mesocorticolimbic system, and identified this network and created a polygenic score that reflected variation in the expression of this network in this brain region. We did this in different brain regions, showing that the networks are brain region specific. We continue doing this anchoring and trying to find the co-expression network of a specific gene, but you can also find networks that are associated with an exposure. For example, we have models of early life adversity, and you can identify gene networks that are linked to the exposure by looking at genome-wide gene expression and doing analysis like WGCNA, where you identify co-expression modules associated with this early life adversity exposure. When you do that, you have a network that is nameless. It's just associated with your exposure. There's no anchor gene that you use to start your process with. This was a different way to identify co-expression networks that were relevant to any sort of condition. We are interested in early life adversity, but it could be in response to a drug or a certain phenotype that you identify in your model or in human tissue too. For example, we are still working on a project where another scientist identified genes in the blood that were associated with early life adversity or early life trauma. They did RNA sequencing in the blood, identifying networks that are linked to trauma. You can use this to inform the development of a polygenic score. What you're measuring in the end is really individual variations in this group of molecules or this network or these biological processes and how people vary on this and link this to disease or risk for disease.
[07:02] Alexey Tolchinsky: I was hoping to say a couple of words about why I thought that you two talking would be useful and then I'll be quiet and listen to both of you. I attended a presentation by Michael Meaney, your colleague, esteemed colleague in New York City at the Congress of Neuropsychoanalysis. He gave a talk. The whole Congress was about development, human development. He talked about the profound influence of early childhood adversity or enrichment on the development of a child. He presented this work that you're mentioning. We essentially have a biomarker, perhaps one of them, and he talked about the EPGS score with glucocorticoid receptors in the hippocampus. The very interesting result was, to be metaphoric, the sort of dandelion children and orchid children, where there were these very sensitive children with this particular genetic profile, who if they're exposed to childhood adversity, they would become susceptible, but if they're exposed to enrichment, they would become resilient. More interesting, Michael was able to articulate the entire causal chain of events of what happens in animal research when a rat is licking the pups, what's going on all the way through with demethylation and other mechanisms of why the CRF expression and the whole HPX is functioning, why this actually worked this way. So there's the entire story, which is coherent and makes sense. When I mentioned this to Michael, with your work with Xenobots and Anthrobots, Michael said we can test this. We can try to create enriched environments for Xenobots and Anthrobots or adverse environments. They're made from xenobots from frog skin cells and heart cells and anthrobots from cadavers, which is epithelial tissue from trachea, I think. I thought, is there any possibility to do the EPGS as just a data point? We don't know what's going to come up from it. If we were to look at the donor and look at the EPGS and then see what's going on with the anthrobot composed with a certain profile of EPGS, that was one of the ideas. The second idea: I think that you both share this -80 Celsius freezer, where you have a lot of neuronal tissues in McGill at Ladmer Center. I don't know if the neuron survives freezing and thawing, but the glial cells might. I was thinking, what if an anthrobot is composed, if it can be composed in a stable way from these glial cells, being informed by your work with EPGS may allow us to get additional data. Michael mentioned the interesting experiments with tadpoles where when exposed to GABA, you saw a bimodal distribution where these ones did this and those ones did that. I immediately thought of your work — what if these are susceptible and these are resilient? These were just some preliminary thoughts.
[10:22] Michael Levin: When you said they are specific to regions of the brain, does that mean that these are typically not expressed outside the brain at all? Are all of these things brain specific or can you find them in other tissues?
[10:44] Patricia Silveira: You can find it in other tissues. It's because we start with gene expression data, and gene expression is tissue specific. When you do that, you identify, for example, for the insulin receptor — we did this in the striatum and PFC, in the hippocampus — when you identify the co-expression network for the insulin receptor in these different regions, you find completely different networks because the function and the expression are dependent on the region. We have created EPRS starting from peripheral tissues too, from the liver, from the adipose tissue, from the pancreas more recently. If you enter this gene and do it in different regions, every time the network is very different; it's tissue specific.
[11:43] Michael Levin: When you've defined the network, have you or anybody else parameterized the network into a model with numbers on the degree to which the different genes turn each other on and off? Is there an actual GRN model for something like this?
[12:09] Patricia Silveira: Not like that. It's more simple. We just see the relationship between each of the genes, if they are co-expressed or co-repressed. We consider this in the polygenic score calculation. For example, if one gene inhibits the other gene, we switch the sign on the calculation so that at the end, when you have the sum of values for all the genes, the higher score will reflect the higher capacity to express that network and a lower score, lower capacity.
[12:48] Michael Levin: The reason I ask is that if there was such a model, it's a lot of work, but you can infer GRNs from data. So in theory, we could do this. We have a bunch of work over the last few years looking at how gene regulatory networks themselves store memories. And the latest, which isn't published yet, but there's a preprint looking at how certain stimuli can delete memories. They can make them forget specific memories. I think it would be very interesting to analyze that network in particular to see, maybe the memory is somewhere else, that's always possible, but just to ask that network alone, how much of the stressors are actually stored in the dynamic system's memory of that network alone? And when we infer stimuli, what we do is not look for gene therapy or things like that, but stimuli that can be delivered by drug, pulsed pharmaceutical interventions and so on, to wipe certain kinds of memories from that network and what the somatic and then the psychological effects on the body would be of doing that.
[14:04] Patricia Silveira: I think that applying more elaborate or newer mathematical models to infer the design of the network or the components of the network would be fantastic. We have so far been very concrete, using RNA sequencing to identify genes that are correlated. I agree, using this other type, there are so many ways to identify networks. For example, some students are now working not with gene co-expression but protein-protein interaction and creating a network from there. Other types of co-localization involve selecting only proteins that are in the same microspace. Most of the work we've done so far is really based on co-expression from RNA sequencing and the basic relationship that you find from these values from crude measures of gene expression.
[15:29] Michael Levin: Have you guys done any of this in vitro? I'm not sure that any kind of in vitro culture is low stress as far as the cells are concerned, but we can crank up the stress if we wanted to.
[15:54] Patricia Silveira: We're starting to do these. Exactly as you were saying to Alexis, to do this validation work when you have the genotype that you can calculate the score, but you can also measure the expression and see how much the score is linked to actual measure of gene expression. This work is at the very beginning. It would be fascinating to do this either using post-mortem tissue; that's what we're doing in collaboration here with the brain bank, but also using your models as well. It would be a nice validation for the method too, for sure.
[16:36] Michael Levin: I think it'd be what Alexi was saying. I think it's very interesting that when we make our anthrobots, occasionally they come from cadavers, but often they come from living patients who donate tracheal biopsies and cells. So many times we do have a living patient. I've not done human studies before, so I have no idea how difficult or easy this actually is, but presumably we could ask some of them for a psychological survey or how you determine it. I've wanted to do that, to do behavioral testing on the bots to see if they correlate at all with the psychological properties of the donor. So how much do the bots reflect? It would be interesting to look at certain kinds of addictions. It should not be too hard. But something subtle. How doable, and what might be the cost of such a project to take some reasonable number of patients, get some kind of profiling on them, get the samples, grow the bots, and do analysis?
[17:53] Patricia Silveira: I think for us to do this investigation, what we would need from the EPRS side would be the genotype of the participants and gene expression level, RNA sequencing, which I know you already do, transcriptomics.
[18:17] Michael Levin: That you could do on the cells. If you're going to do that anyway, you don't need to involve the patient at all because they're already donating the cells. You get the cells, you do your genotype, you do your transcriptomics, and then we make Watson profile them. That may not be so hard.
[18:42] Patricia Silveira: I read your papers and I think that even existing data could be used for an initial analysis. I don't know how much you can access in terms of genotype from the materials, but I'm sure you have.
[18:59] Michael Levin: It's not hard. Nowadays genotyping them is not hard. You're right, we've published the transcriptome. That's probably worth trying.
[19:20] Patricia Silveira: And yes, of course, starting from human cells is fascinating, but a lot of the work that we do starts from models as well. So even the bots that you have that are created from cells, like animal cells, also could work.
[19:42] Michael Levin: We don't have any mouse bots. We probably could make them. We don't have an easy source for the tracheal epithelial cells. I'm not sure what that would entail. But the other thing we do have is frog. This is frog embryo epithelial cells, which make centibots. There we have the adult animals. We have the cells. We could do that. We can also, as Alexi was saying, start looking at environmental features. We can definitely do enrichment. We can do some sort of adversity. We've recently published that they're sensitive to vibration, so acoustics kinds of things. We're playing with light. It's not clear to me yet how much they care about light, but acoustics, they definitely care about electrical signals in the water; they care about chemical signals. This is not published yet, but they can sense various things in the water. There's a wide range. Also, we can make mazes. We can 3D print mazes for them to play with. I wanted to do this. If you provide loose epithelial cells in the environment for the Xenobots, what they do is they run around and they collect them into piles and they polish those piles and those piles become the next generation of bots. So it's really weird. Basically we've made it impossible for the Xenobots to reproduce in the normal way because they don't have any of the frog organs, but they found a different way of making copies of themselves, which is just wild. What I'm interested in is comparing bots that had a chance to do that and bots that never had a chance to do that. What are the differences? Do they, is there some long-term consequence of "yes, I've done kinematic self-replication, I've made copies of myself, I am happier, fulfilled, more stressed"? I don't know what the answer would be, but do they have some marker of that? Do they know that this is what they've done? As a control, you could do passive beads or nothing and things like that.
[22:10] Patricia Silveira: Can you vary nutrition too?
[22:12] Michael Levin: We can. So normally what powers them is maternal yolk. The cells are from a fairly young frog embryo, which means they still have a lot of yolk and this is what they're powered by. They live not much longer than a week based on that. But you can put nutrients in the water that they will take up. Our record is 84 days to keep a bug alive. It's extremely interesting because it's started turning into something. I have no idea what it's turning into. It's very odd. It becomes this long transparent thing with a black spot on one end, which I think might have been a nice spot. I don't know exactly. The problem is we don't have a lot of these because it's a funny, practical reason. The nutrition that we use for them is so rich that they get immediately overrun with bacteria. Keeping them clean while continuing to feed them is very hard. We've tried all the obvious stuff. We've tried antibiotics and titrating them out. It's really hard. We're still fighting it because I would very much like to keep them going and find out what happens six months later. What do they turn into? Yes, we can feed them in the short term, but we're still fighting this bacterial issue.
[23:43] Patricia Silveira: Can you stress them by providing less nutrition, or do they die?
[23:48] Michael Levin: I think they're pretty stressed when they start to run out of nutrients. You can modulate that. There are a million ways to stress them. You can give them heat shock, put electric shock in the water, everything. It's interesting. Every life form hates electric shock. You can shock them. The vibrations, we're still trying to figure out how they feel about the vibrations. My guess is that we're going to find some kinds of frequency/amplitude combinations that are neutral in terms of this is signal, I can interpret this versus this is really annoying because it's shaking everything. We can find some kind of version that is annoying to them.
[24:40] Patricia Silveira: That's interesting. I ask about restricting nutrition because that's one of the models I work a lot with in rodents. We have worked on identifying networks that respond to this kind of stress during prenatal development. Prenatal food restriction to the dam has long-term consequences for the health of that animal; we see resemblances with low birth weight effects in humans. That's why the model is interesting for us. But having worked with the type of stresses that we know well in a model may help because we find some similarities in terms of networks.
[25:36] Alexey Tolchinsky: I may make a quick comment on that part, Patricia. Michael and I met Frank Putnam, who is a traumatologist and psychiatrist. He was one of the authors of a multigenerational study of trauma on very sad cases of women who were abused in early childhood, typically complex trauma. He saw very high obesity rates related to insulin. Of course we're talking about the mesocorticolimbic dopaminergic system, exactly what you study with a DET1 transporter, and also faster aging. We talked about that. Where this work gives hope to human clinical work is that if there are genetic studies and you see the EPGS score suggesting susceptibility for that child specifically, early environmental intervention would make much more difference. This is actionable: while the genetic profile cannot be changed, the environment can. We can eventually have either therapeutics or suggested interventions. This is where the work can go, and it has application to clinical work. In Frank's paper he mentioned the demethylation region that you and Michael mentioned—he mentions exactly that. Your work is known in the trauma community.
[27:10] Patricia Silveira: What's nice about this, Alexa, is that, coming back to the cost of the experiment, if you have the genotype and transcriptomics, you can actually explore whatever network, and it's not a one shot. You can calculate all kinds of polygenic scores and different types of networks that you identify that are relevant. So it's work that can be investigated in many ways and answer different hypotheses if you have the data.
[27:50] Michael Levin: That gives me an idea of what you were just saying about the transgenerational business. I wonder if we can see evidence of that in the kinematic self-replication of the bots. Would that work as a transgenerational thing? It seems fairly simple. All they do is push the cells together, but maybe there are different subtle ways of doing that. Bots that had been traumatized would do it differently, and the second and third generations would. I think it would be very interesting to look at potential transgenerational effects in kinematic self-replications.
[28:28] Alexey Tolchinsky: And I had a simpler thought. The two kinds of fitness you mentioned, they live on average about three weeks and the regeneration doesn't go forward other than three generations. So it's grandfather or grandmother, daughter and then granddaughter and that's it. So you could look into, again, the environmental applications early on. Let's say you just composed a Xenobot and there's an enriched environment, and see if they live a week longer or if maybe their regenerative fitness goes one more generation. But in addition to a very concrete measurement of susceptibility and resilience, where you apply a certain number of stresses and you see the reaction, where high reaction could mean susceptibility, low reaction could be resilience, you can also look into life fitness and regeneration fitness.
[29:19] Michael Levin: I have a weird philosophical question for both of you from your field. Obviously there is such a thing as terrible trauma. But on the other end, what is the best-case scenario? If you ask older people who have had kids, "How do I prevent screwing up my kids?" they'll say, "Well, you can't. There's just different ways to do it." But we're talking about the very minimal best possible scenario. So my question is, what do you see as if we took care of all the social stressors, the resource stress, and you take care of all that, you're in an ideal environment. Still, there are going to be some things. Kids are going to find out about certain things that are stressful. What is the actual least amount of stress? And do you think that there is a least amount that's too little? What does that look like? What does the left end of that spectrum look like?
[30:32] Patricia Silveira: Do you want to say something, Alexei?
[30:34] Alexey Tolchinsky: We talk about the window of tolerance. This is the Yerkes-Dodson law, a pretty old thing, which has been refined, but when stress is too much, we're moving toward the trauma category and that doesn't improve any resilience. When stress is too little, there's no motivation to grow and to apply effort, and so in the middle, when there's enough stress within the window of tolerance, then there's certain growth and development. There's a dose dependency there. If we talk about the ideal environment in humans, first of all, the programming starts during pregnancy where the mother's heart rate goes up, the fetus's heart rate goes up. When the mother is very stressed out during pregnancy, the child's stress regulation network starts getting programmed right there. Since birth, we have nutritional and biological factors, but we also have attachment style, which gets formed from about 6 to 18 months of age. These are interactions between the child and the primary caregiver. Beatrice's baby's work is relevant here: what exactly happens between the baby and the mother, the preverbal baby, and the whole goo goo gaga dance and how they interact has a profound influence on life going forward. The highest correlation we have with psychopathology is a disorganized attachment style. When a mother is very traumatized, it is objectively very hard for her to do that with a baby; not to blame the mother, but this is one of the mechanisms of transmission — not genetic. It's what happens in the interaction, what happens in pregnancy. That lays the foundation, and from there you can see what happens to the child in the environment. Then there's parenting, and we call it the zone of proximal development, where a parent and the child, when they do things together and the stress is gradually increased as the child matures, we have optimal development. When the child is thrown into the ocean to learn to swim, that's not optimal, or when there isn't stress whatsoever and the parent is helicoptering and doesn't allow any frustration, there's no development. It's a very delicate dance.
[32:56] Patricia Silveira: I agree. Everything is a balance. You should think, as you just said, if you think of the most precious thing, maternal care that Michael studied so much, it's so special, very important. But you can always go to overprotection and helicopter parents — it's not good. Same thing about nutrition. If you lack nutrition, awful for development, the individual can die from this. But overnutrition is also not good. The balance is the key and not easy to say what's optimal there.
[33:41] Michael Levin: Do you think that there are systems in the body that react to insufficient stress? Let's say in an extreme helicoptering thing, is there some system that reacts, paradoxically, as if another kind of stressor is present when there's not enough stress and we're not growing? Is there anything, or are there only sensors for the other side of the spectrum?
[34:11] Patricia Silveira: It's interesting you're saying this because we're writing about this just now, but I think that stress is inevitable during development. Even overprotective parents can't really avoid this. But there are very well-established control systems for stress responses. They are adaptive. They are there to help the individual survive, and they are necessary. For example, inflammation is one of the systems that are part of the stress response. And we need that. We need the system to be exposed and trained so that we can face infections and fight disease. So there's a level of stress that is inevitable, part of development, and adaptive when the system is ready to face it, to learn from this and to grow. I don't know biologically whether there is a biological basis for "lack of stress is bad." I can understand this from a psychiatric perspective. But biologically speaking, I think it's more about stress being part of our nature to explore and we need to go out there and survive. The stress system is ready to face these things. We'll learn from these experiences. And this is part of normal development.
[36:02] Alexey Tolchinsky: Maybe add a psychoanalytic perspective. Freud said ego grows in frustration. When there is no frustration, then there's underdevelopment. Winnicott added to that; he had this beautiful concept called good enough mothering, which is just gradually letting go because babies are completely dependent and adults are completely independent. Our job, as the baby matures and grows and develops capacity, is to give them space to grow into and to let go. So when we don't let go, when we helicopter, then they're underdeveloped, which perhaps in contemporary terms means lower allostatic complexity. So when they leave the world, the reality of the world hits them hard. It's a rude awakening because they don't have enough allostatic complexity to handle the unpredictability and the other things. If they're gradually exposed to increasing amounts of stress within the window of tolerance, then they become optimally prepared for the environment to go on their own and adapt.
[37:07] Michael Levin: I was thinking back to those rat paradise experiments. Somebody basically created a paradise for rats where there was as much space as you wanted, as much food as you wanted, as much nesting material as you wanted. There's absolutely nothing to worry about. He documented some very interesting genotypes of what happens in the long term. As I recall, you ended up with this interesting phenotype of male rats that would spend basically all day grooming themselves and making themselves beautiful. They lost all interest in mating and doing anything else. It was a new behavioral phenotype that came out where they had absolutely nothing else to worry about. This is what they focused on. I wonder, under those circumstances, is there any stress being activated or is there some kind of group phenomenon where the individual animals are totally happy because they have nothing to worry about, but as a group there's something detectable going on? This is what I think about for the future as robotics, regenerative medicine, and anti-aging come along: what does that look like and what does the transition look like? It seems to me that a mature civilization somewhere presumably isn't still worried about all the dumb stuff that we have to worry about. What does that look like psychologically? What is the level of stress when there really isn't any stress about food or resources or anything else? What, if anything, should there be stress about? We've never had to deal with that because we've had plenty of stress forced on us by scarcity, but eventually presumably we'll get past that. Then what? I think it's interesting to think about.
[39:30] Patricia Silveira: The immediate thing I think about when you say this is relationships. Even if you fix all the scarcity in terms of nutrition, environmental stimuli, exercise, metabolic health, we're still social individuals and relationships can be complicated. How can we fix them?
[40:13] Michael Levin: Have you seen some of those robots they're making? People are addressing this, people are on top of this problem, right? This is one of those things that I think we're also going to have to deal with. If you want a simpler relationship, I think you'll be able to get one. And if it's too simple, I think you'll be able to go in the apps and dial up, I need a little more, three arguments a week, no more, no less, and you'll be able to have it.
[40:43] Alexey Tolchinsky: I was also thinking, Michael, of this video you recorded with George Bongard recently about how all the robots and AI we have now are very oriented to certainty. They're optimized for certainty. In fact, we have minimization of surprise and all the other things under FEP and whatnot, while in biological systems, all of them, including single cells, deal with uncertainty, as you put it on an interview, on different spatial and temporal scales. When you have this red pop in the red paradise, it is very certain because the food is in the fridge, the roof is over your head, there's no need to fight, there's no threat. He's leaking and grooming. I'm a beautiful narcissistic rat. There are no other things to tackle. I think the amount of uncertainty is also playing into it. We have this term uncertainty tolerance, and if you think about AI and robots, because you looked at frustration tolerance in bubble sort, with robots and AI, if we expose them to gradually increasing amounts of uncertainty and see what they do, that's part of the question.
[41:49] Michael Levin: I think that's very interesting. I'm slowly working on a paper I call the fog of agency, or maybe the fog of life. It's related to this concept of the fog of war. When stuff really goes down, you lose certainty about what's going on. It's hard to make decisions. It's hard to have long-term planning. It's hard to know what everybody's doing and collate different kinds of information. All this starts to go to hell as soon as things really get tough. As you point out, living systems are in that state all the time. You never actually know what's going on. Maybe some of the systems know certain things in certain spaces, but for the vast majority of them, there are many more unknowns than knowns, and living things are dealing with that constantly. I think it is interesting what happens when you try to lift some of that and where are the optimal points. We've been looking at a theory of aging that you could summarize as the boredom theory of aging. What happens is there's a homeostatic process where you have to make the body. You've made the body. Now what? Without being given novel goals, you upkeep it for a while, but then eventually you start to decouple; all the subsystems start to decouple slowly because there isn't an overarching goal keeping them together. It would be interesting in those kinds—aging, senescence, cancer, those kinds of things—what really is the optimal. We know exercise: some level of tearing up your muscle tissue is good for aging. Can we find the optima for these things? For the different subsystems. For the metabolic system, obviously there is one; maybe for the immune system.
[43:50] Patricia Silveira: It's fascinating what you say about predictability, because in this child development world, it's one of the worst stressors for a newborn or for a young child: unpredictable caregivers. So it's interesting to think about that in the context of AI. It's nice.
[44:19] Michael Levin: There's another fundamental unpredictability too. Josh had some of the earliest work on this back in 2006. One of the things about living things coming into the world is that they don't know ahead of time where the border is between them and the outside world. So as a brand new being, I have these things. They don't seem to be under too good control. Maybe those are mine. Maybe they're not eventually. But at the same time, you make some noises, mom comes and gives you food; that's gotta be, I have good control over that. A couple months go by and this stuff gets better. The other thing gets a little less reliable. So that uncertainty of what do I have exactly? In 2006, Josh had this amazing paper with these robots that didn't know what their structure was. They had no pre-existing. They would just flop around and eventually build an internal model of themselves based on actions they took and what happened. They would infer, I must look like this because that's the model that fits what I've been doing and what's happened to me since then. So they would infer their own structure. What was super cool were two things. One is, they call it the evil starfish because it looked like a big starfish with these legs. If you take off one of the legs, it very rapidly adapts because it just does the same process again and it says, now I have fewer legs. So it's much more plastic. Chris Adami wrote this amazing commentary on it because what they found is that during periods where the robot wasn't doing anything, it was internally pruning the models and figuring out I definitely can't be a big ball. I must be something else. He wrote this whole thing about how they were dreaming, that that was their sleep phase where they would throw away the stuff that wasn't helpful to them and consolidate their picture of themselves in the world. It's really interesting. Almost 20 years ago now. Amazing.
[46:22] Alexey Tolchinsky: And one more thought related here is the learning part, and Australians joke that when a kangaroo wakes up, it has breakfast, goes to pee and then has a look around. And so it's the look-around business that is important. That's epistemic foraging. What's out there now that my tummy is full and everything's okay, what's out there? If we look at a baby, it doesn't learn much; the whole resources are mobilized to save life and protect oneself. You really need a safe, secure base and safe haven to explore. You need mommy's lap, and then the three-year-old crawls and looks what's around. So the enriched environment means the baby feels well enough, all the needs are met, and the baby is learning a lot, and that increases allostatic complexity. Conversely, when there's adversity, then you're not learning as much. I think learning is right there with this resilience.
[47:23] Michael Levin: I would love to know what the physiological and transcriptional versions of this are like. So the lab story tells us what the 3D world of motility looks like. But transcriptionally, I'm a cell. I feel pretty good. My needs are met. I'm going to explore transcriptional space a little bit. I'm going to crawl out. What happens if I crank up the BMP a little bit? No, I didn't like that. Or bioelectrically too. Everything is good. I've settled into a nice rhythm. Can I try some things and see how much play? And that's something I've been thinking about a lot: how to recognize creative play in these other spaces. We know what problem solving looks like. We know what hard work and response to stressors and repair look like. What does playing creativity look like in these other weird spaces? I think it would be really interesting to try to have some way of quantifying it or at least formalizing what we're looking for.
