Combining Biology and AI w/ Michael Levin and Dr. Leo Pio Lopez

Hey, Upstream listeners. Today we're releasing a conversation between Nathan Labenz, Michael Levin, and Dr. Leo Pio Lopez, which originally aired on Turpentine's show, The Cognitive Revolution. The conversation touches on technical and big-picture questions of the relationship between biology and AI. Without further ado, I'll hand it off to Nathan to give you a little more context about the episode. Please enjoy.

Hello, and welcome back to The Cognitive Revolution. Today, I'm pleased to share a conversation with Professor of Biology Michael Levin and Staff Scientist Dr. Leo Pio-Lopez of Tufts University. Professor Levin's previous appearance on the show has become our most popular guest episode ever, and for good reason. His perspective on the ongoing convergence of biology, computer science, and philosophy are fascinating. I invited Michael and Leo again now because they recently published a groundbreaking paper that uses advanced embedding techniques to combine multiple biological datasets spanning the modalities of genes, drugs, and diseases into a single unified network model of disease. I've been watching out for new machine learning approaches that can help pry open the black box of biological interactions and causation, and I was impressed that this approach has already generated meaningful new insights. Including most notably a predicted link between the neurotransmitter GABA and the cancer melanoma, which was subsequently validated by laboratory experiment. In this episode, we first discussed the technical details of this work, including the application of random walk with restart algorithms to the challenge of learning associations across a multimodal, multi-layer network, as well as the potential for this approach to uncover additional new therapeutic targets in the future. From there, we zoom out to consider broader topics in AI for biology, including the current limitations in biological data collection and standardization, the shortage of data related to the aspects of health and development that matter to us most, and how emerging technologies like robot scientists and cloud labs might accelerate progress. We also touch on the concept of multiscale intelligence in biological systems, a recurring theme in Professor Levin's work. Including the remarkable observations that even simple gene regulatory networks are capable of some forms of learning, that biological systems can often survive and thrive despite major defects in their own hardware, and that humans have historically domesticated and effectively trained wild animals despite knowing essentially nothing about their biology. We briefly explore how such results and other biological models might inspire more robust and adaptable AI architectures. Toward the end, we get into more philosophical territory as well, including the future of human enhancement and the ethical implications of outsourcing aspects of cognition to AI, the blurring of categories like living things and machines, often thought to be mutually exclusive, and the potential for digital life, the challenge of envisioning a positive future, and specifically what humanity might look like as a mature species, the potential incoherence of the concept of high-level AI alignment, considering the lack of alignment amongst humans, The possibility of transcending individual identity, but also the extreme risks associated with overly powerful collective intelligences. And finally, the existential importance of developing a new science of where agency and goals come from and how we can effectively manage them. Overall, it's a super thought-provoking episode which serves to remind us of how quickly the future is coming at us and how much work we still need to do to be ready for it. As always, if you're finding value in the show, we'd appreciate it if you take a moment to share it with friends. And we always appreciate your reviews on Apple Podcasts or Spotify. We welcome your feedback via our website, cognitiverevolution.ai, and you can always DM me on your favorite social network. Now I give you Professor Michael Levin and Dr. Leo Pio Lopez on the intersection of AI and biology and the future of intelligence.

Professor Michael Levin, professor of biology at Tufts University, and Leo Pio-Lopez, staff scientist at Tufts University. Welcome to The Cognitive Revolution.

Hey, Nathan, good to see you again.

It is great to have you back. So, so much continues to happen. You guys have put out a lot of great work recently, and I'm excited to pick up where we left off last time. If folks haven't heard our last episode, which, by the way, is our most viewed episode ever on YouTube of any guest that we've ever had, then definitely go back and check that out. I think a lot of really interesting tie-ins were made in that conversation between all the fascinating work that you're doing in biology and all of the stuff that people are just starting to wrap their arms around and minds around on the AI side. So I kind of want to do more of that, but I want to start also with a paper that you guys put out on July 4th of this year, which I thought was really interesting and to me seems like a major sign of things to come. It's called Universal Multi-Layer Network Embedding: Reveals a Causal Link Between GABA Neurotransmitter and Cancer. And what stood out to me a lot about this paper was that it is integrating multiple different data types and multiple different scale of systems, genes, drugs, and diseases into one dataset, one network, one sort of unified understanding, and driving new insights from that. The techniques are a little bit unfamiliar, and I do think it's worth spending a little bit of time, or maybe unfamiliar to people that have only been in AI in the transformer era, but I think it is worth taking a little time to kind of unpack how did you get the idea to do this? How did you have the sense that these datasets could be combined, and how did you actually go about it? Leo, if you want to start, that's great.

Yes. Yes.

So, well, the idea came from also my previous postdoc because I worked on this kind of data, multilayer networks. Because what we have in the omics era, if we have so many kinds of data and we don't know how to integrate them in the same representation, and networks are a very good way to do it. So until now, we had big limitations on the kind of multi-layer network we could have, only, for example, 2 multiplex networks. Multiplexing software, just a different kind of network. So I realized that, I think it was in the middle of last year, that I could finally have this kind of universal multi-layer network because we had new extension of different techniques to do it. So I did it and I explained what I did to Michael and it happened this way.

Let's go a little deeper into that. Can you break down the different datasets that you started with and how you bridged these different modalities together?

So yes, with this now we can have as many kind of nodes that we want. So for medical applications or to understand and more biology, we use what we had. So genes, drugs, and diseases. And in networks, you have interactions that are— this is how you link different nodes. For example, with genes, we have so many things. We can have protein-protein interactions network, we can have co-expression network, and so many different things. So we used, well, the most interesting one, the one that we thought were the case. For drugs, I use DrugBank, for example, which is a very big dataset, and then I add as many things I could find for diseases. That is actually pretty difficult. We don't have so many interesting diseases networks. And the beauty of network embedding is at the end, you have the possibility to apply machine learning on this kind of data that could not be integrated in the same space, let's say, before.

So I think Figure 1 in the paper, and we'll put a link to the paper in the show notes, is very helpful for developing a little intuition for this. As I'll just kind of describe it. There's the 3 different kinds of data, each in kind of their own space, so to speak. You've got the drug data, which consists of clinical drug interactions and also drugs that are often combined together, I guess. And then with the disease side, you have a sort of network of shared symptoms across different diseases, and then also comorbidity or the co-occurrence of different diseases.

Yeah.

And then in the gene side, it's a little bit less clear exactly what that is saying, but there's a pathways and a PPI pathways, I guess, being like what interacts with, you know, what sort of promotes or inhibits what. And then there's these links between these different spaces as well. So this would be the indication of what drug is used to treat what disease and what drug is used to treat what symptom and what gene seems to maybe cause, or at least correlate with different diseases. Was this a matter of All these datasets existed and you had to like do a big sort of reconciliation of all the different datasets to make sure that you were actually making all these linkages? Because all these different datasets, like they don't connect automatically, right? I guess a big part of the work was identifying like which element corresponds to which element in other things.

Yeah, actually I have to say first that this figure is a simplification of all the data that we have. We have more in the paper. And well, all these databases are already existing. My work here was just, yes, to translate them in the same language, in a sense, because we have so many. For example, we have different kinds of naming of the genes. So I just had to use the appropriate language, and then it's easy to build this kind of network. I think here the important idea was to take the most interesting databases that we have in the omics, all the omics databases that we can have right now.

Okay.

So now let's talk about the actual learning technique. You've got all these datasets, you've assembled them all, you've figured out how to identify the connections between them and sort of unified them into one giant multimodal, multiscale dataset. And now the challenge is, or the goal is to say, okay, if we can take all these links that exist, maybe we can learn those and maybe we can use those to learn a model which can predict new linkages, right? So I think the techniques of multi-exverse which is new in this paper, and then the random walk with restart, which is more established, but I do think not super familiar to a lot of folks who've been following AI for only the last few years. I think those techniques are probably not super familiar. So take us through those techniques. Just give us a little sense of the setup, the motivation for the technique, how it ultimately works.

Well, yes, the main motivation, as I said, was to apply machine learning on networks. But with machine learning, we can't apply it directly on networks. We need vectors. So network embeddings has this ultimate aim to open this big machine learning toolbox for different kinds of analysis. But to do that, we need to translate the nodes in the networks into vectors. So network embedding, vector embeddings are the vectors. And to do that, we need what we call a similarity measure in the network. Because the network embedding will be to have embeddings in this vector space. That can relate to the similarity measure between nodes in the network. So to have this similarity measure, we use random walk with restart. Random walk with restart is a method that is, I think, very well known in the field of network theory, network analysis, but not really in machine learning. And the general idea is to use like an imaginary particle that from a seed node will explore this big network a lot of time until convergence. And this will allow to retrieve like a probability distribution of how much you can, you will go through this node during your random walk on the network. And this is the similarity measure that we use and that will allow us after the learning phase to minimize the distances between the embeddings according to the similarity measure that we have in the network that we just computed.

Okay, cool. And then what's the next step with the Multi-Xverse? Take us, take us the next step through the process.

So when we have all this, we can do so many things in terms of analysis. So that was the question. So after I did that, I explained to Michael, well, I have this system now. I can do a lot of things. I could, for example, do link predictions between new drugs and diseases. We can find new targets. We can understand more diseases. So many different questions can be asked on this kind of system. And we did clustering, we did link prediction, and we found this link between GABA and cancer. But if I have to talk about a new, let's say, next step after this method, I think it will be to link it to some sort of LLMs or transformers because they use also embeddings. And we can use these embeddings that I have with this system as an input now for new kind of applications.

So just to make sure I understand correctly, the process starts with this random walk with restart. And that is essentially a way of figuring out through many different starting points and many different starting with random traversals through the network, what should be close to what. So you're essentially minimising a sort of distance travelled through the network by pushing the positions of the embeddings that are often traversed when 2 nodes are connected, they're traversed often, and so they need to become closer. Is that kind of the intuition for how the embeddings get developed through that process?

Yes, it's like you have this similarity measure between nodes. So between all the nodes in the network, we have a similarity measure. At the beginning of the optimization process, the learning process, we are going to initialize, copy randomly the vectors. So the distances between the vectors at the beginning won't reflect the similarity measure. And this is what we will minimize during the optimization process. Gotcha.

And then you hold out, if I remember correctly, the network is trained on like 70% of the data, and then the learned network is then used to make predictions, which you validate on the other 30% of the data.

Yes, this is just to validate the quality of our embeddings. We want to be sure that, well, we did something that is really coherent. So this is the validation set. So then you, you did that, you just keep the 30% of the network to be sure that everything is good.

And then after having validated, then you can say, okay, now let's go take a look at the predictions for some new links. When you're doing these new link predictions, are you just making like a brute force process through possible links and just scoring them all and just then looking at ones that have a high score, or is there any sort of heuristic for the search that guides the process toward particular links to look at?

No, no, I will feed machine learning method with all the links that we know, and then I will test on new links that we don't know. And I will see if I have a high probability of new links, yes. And so you can do that between cancer and new drugs, for example. There are different ways to do that. There is also, well, the clustering is a bit different, but the clustering has some, most of the people use it for visualization. But in the case of network biology, we also like to do that to understand the notion of biological modules. So biological functional modules that very rarely have a biological function. And sometimes when you look at clusters also, you can find a new drug or a new target associated to diseases. So some diseases. This is how you can do that.

Cool. We'll continue our interview in a moment after a word from our sponsors. You know, like many of you, I spend a good chunk of my day managing emails and phone calls. And despite being careful about where I share my contact info, somehow the robocalls keep multiplying, and my inbox is flooded with spam. It turns out there's a reason for this: data brokers. These companies don't just collect basic contact info; they're gathering everything from your Social Security number and financial records to your shopping habits and employment history. Some health insurance companies are now working with data brokers to track your online activity, which could actually affect your insurance rates. Another huge privacy risk are people search sites. These websites publish detailed profiles about you, including your home address, family members' information, and even sensitive details like religious beliefs or political affiliations. This got me thinking about how exposed we all are. Whether you're sending a wire transfer, want to protect against identity theft, or just trying to keep your family's personal details private, That's why I started using Incogni. Here's what makes it different. They automatically handle 5 different types of data brokers: marketing, recruitment, financial information, risk mitigation brokers, and those people search sites. Instead of you having to track down all these companies yourself— and trust me, there are hundreds— Incogni contacts them on your behalf and demands they remove your information. And here's the key: they keep monitoring and removing your data automatically, because these brokers often try to collect it again. Incogni protects your data across all your devices, whether you're working from home, a Plus, you can add up to 4 additional family members under their Family and Friends plan. And they're so confident you'll see the difference in your day-to-day, they offer a 30-day money-back guarantee. Take your personal data back with Incogni. Use code Upstream at the link below and get 60% off an annual plan. That's incogni.com/upstream. Hey listeners, it's Erik. If you're looking for more expert-level conversations, but specifically about top negotiation strategies, Then I have the perfect podcast recommendation for you. DealCraft: Insights from Great Negotiators, hosted by Jim Sebanias, Harvard Business School professor and renowned negotiation expert. DealCraft features interviews with the world's greatest dealmakers and diplomats, from former Secretary of State Henry Kissinger to Blackstone's Steve Schwarzman, on their most challenging negotiations, from which Jim draws lessons for listeners. Not to mention that The Guardian recently called DealCraft a best podcast pick. New episodes release weekly, and I'd recommend starting with their first episode. Which reveals how legendary entertainment attorney John Branca was able to successfully negotiate for Michael Jackson's ownership of his master recordings and the Beatles catalog. Almost unprecedented feats at the time, so don't miss out. Follow DealCraft: Insights from Great Negotiators on Apple Podcasts, Spotify, or wherever you get your podcasts.

So do you want to talk for a minute about the specific finding that made the headline, the, the title of the paper, and kind of what makes that notable. Maybe give us a little sense too of like, okay, so the network spits out that, hey, there appears to be a high chance of a link between these 2 things. For the non-biologically informed, there's a little bit of a question of like, what is this GABA thing? I wouldn't have too much to say about that. So I'd love to hear a little bit of like, okay, you saw this, what did that mean to you at first? Then what did you go do to validate it? What did we ultimately learn from this? And then we'll zoom out and talk kind of, you know, how this may be a preview of things to come.

I can start with that and maybe Michael then can add. So my first problem was to develop the method technically. I wanted everything to work and to be validated. And then I talked to Michael. I said, okay, I have this. What do we do with it? Because we can ask so many different questions about not only about cancer, but also rare diseases, ion channels. And I remembered, so he started to talk about the links between serotonin, I think, and cancer. if I remember correctly, and that we already have published work on this. And then he told me, yes, we could look also to everything about GABA and cancer. And this is how it started.

Yeah. I mean, so to take a step back a little bit, we are interested in cancer obviously because of the biomedical importance of the disease, but also because it teaches us about multicellularity and the failures of the collective intelligence of cells. You know, the reason that we have anything but cancer is that, meaning why cells normally in the body operate together to form specific organs, specific anatomical structures? You know, why do they work hard to do those things that individual cells and unicellular organisms don't do? There is a set of communication policies among cells, which I call a kind of cognitive glue. There are many different kinds, but what it enables is for all of these cells to join together into networks, and these networks store very large pattern memories of what it is that they're supposed to be doing. So, all of the cells then are aligned in that, okay, we are building this particular structure and there's information in this network about what it is that we're building. It's a liver or a kidney, a heart, whatever it is. And so, the breakdown of those mechanisms results in cancer. The cancer phenotype is that individual cells disconnect from that network and then they roll back to this kind of unicellular lifestyle where the size of the goals that they can pursue instead of big organ and tissue-level goals, they become little tiny single-cell goals like metabolism, eat as much as you can, go wherever life is good, that kind of thing. So, we study cancer as a breakdown of this collective intelligence almost like a dissociative disorder of the morphogenetic intelligence. And, that means that we have to really understand how do the cells communicate and how do they get aligned into these networks. And, of course, bioelectrics is a major part of this. But, one of the things that is downstream of the electrical signaling is neurotransmitter signaling. Same as in the brain, this happens in the rest of your body too. So, we studied for years, we've studied the way that voltage changes cause the small molecule neurotransmitter serotonin to move from cell to cell and have various effects. And, it's kind of a downstream currency of the bioelectrical information processing that occurs in the network. And so, if you mess with the serotonergic signaling, you can get certain kinds of cancer phenotypes and so on. So, what was interesting about this prediction is that GABA-A, which is one receptor for a different neurotransmitter called GABA, GABA-A is a chloride channel. And, a chloride channel is a little protein that sits in the surface of cells and it regulates how much chloride ion goes back and forth. And thus, it also regulates the electrical state of the cell and the resting potential of the cell membrane. So, that suggested that we could use GABA signaling to perturb the electrical communication among cells by changing their potential. And, we've had work before from a number of people in the lab using chloride channels to manipulate these kinds of properties. And, it turns out that when you use muscimol, which is a drug that is an agonist of these chloride channels, then something very interesting happens. You get basically a kind of melanoma where all of the normal pigment cells in the body, they're called melanocytes, these little black pigment cells, they basically go crazy, they transform, they overproliferate, they dig into other tissues, they dig into the blood vessels, they become incredibly invasive. They change shape. It's really weird. They almost start to look like neurons. They make these very long skinny projections and they just start moving around and invading into everything. And, it's basically like a metastatic melanoma phenotype. A couple of interesting things to note about that. One is that there is no primary tumor. So, the one way that people often think about cancer is that there is a single cell that got a bunch of mutations that took it over the edge to transform and then it releases clonal progeny that have those same mutations and they kind of take over. And so, it has to be— there's an origin in one location, and then from there, it spreads. That's not how this works. There is no primary tumor. What happens is that pretty much every melanocyte in the body converts. They all go at once, and you have this phenotype. And then, the other thing, of course, in this case is that there is no genetic damage. There's nothing genetically wrong with any of these cells. It's a purely physiological change. What you've altered is the ability of the cells to work electrically with other cells. So, it's a very interesting example of how the dysregulation of cancer can be initiated without any kind of genetic damage, so no carcinogens. no oncogenes, just this dysregulation of development, of these developmental mechanisms. And it can be triggered by this kind of GABA mechanism that was predicted by Clio's network.

Okay. So there's at least a couple of different directions that I want to go there to unpack that a little further. Let's come back to the multiscale intelligence in just a minute, but just to stay on this network and sort of the use of machine learning to understand these biological systems for a little bit longer. As I've been kind of watching the space and I've been increasingly fascinated with AI for science generally, AI for biology in particular. I'm coming at it as somebody who took like 2 biology classes in undergrad a few years ago now, all that information long since outdated, I'm afraid. It seems like the grand challenge of biology, to take the most zoomed out lens, as I have kind of come to understand it, is that we just don't know how things work. And so it becomes very difficult if you don't know how things work, to identify what an effective intervention is going to be. It seems like at this point we've made a lot of progress where if we do, through often painstaking techniques, figure out exactly where we want to intervene in the overall biological network, then we're getting pretty good at designing proteins or whatever to go in there and hit that target. But we still have just a vast black box problem of what affects what, what interrelates with what. Obviously, all these things are kind of messy and sort of overlapping in all sorts of ways. And so I've been noticing that there seems to be the beginning of a trend toward using machine learning systems to model the black box. And I don't know that there's like a sharp distinction, but it seems like some of the early systems, like your early AlphaFolds, and maybe even still through to AlphaFold 3, arguably, are like highly engineered systems where there is a very kind of detailed theory of what is going on that is then being sort of encoded into the network structure and ultimately being modelled. And then there's some other things, like I look at the Evo model where it's trained purely on base pairs and it seems to learn these higher order concepts purely from the DNA sequences. And that seems to be a bit of a different thing where it's like the higher order concepts seem to be more emergent as opposed to really engineered into the structure of the network. Because, you know, this might call this like how sort of opinionated is the inductive prior of the architecture of the network. It seems like this work is like more on that latter side, right? If I understand correctly, it's like this is just sort of taking a ton of data that existed but was maybe in disparate places, figuring out how to represent it in one network form, and then kind of learning on that in a way that seems rather unopinionated, right? It doesn't seem like you've encoded too much into the structure of the architecture. It's really just, we have all these links, let's see what we can learn from them. And that seems like a big deal inasmuch as now we seem to be entering into a sort of bootstrap type dynamic, perhaps, where the knowledge we have today, we can encode that, we can make new predictions from it. All of a sudden, if these new predictions start to be validated at a reasonable rate, you can imagine we could really start to feed that back into the network. And this starts to feel like a beginning of a way in to really open up the black box and sort of solve the grand challenge. I realise I'm speaking as a rank amateur to masters here. So critique that story. Like, would you describe the black box challenge differently? Would you say I'm missing some inductive prior in the way that these networks are structured? And ultimately, like, does it seem to you that we are heading into this sort of bootstrapping into more and more knowledge through this loop of machine learning validation and back into the network?

Yeah, maybe Leo, if you want to say something about your approach and then I'll talk about the big picture here.

Yeah, well, our approach is more data-driven. Yes. So it's like we do a lot of big data mining of— we try to aggregate all the mixed data that we have and try to understand a bit what's happening. But in systems biology, you also have model-based approach. So we try to give explanations more than just predictions. I mean, both are absolutely needed, but only data, I think we will miss a lot. In addition, when we do this kind of approach, even mine, if I— so for example, we can find only what is already into the data. But for example, there is no bioelectricity data in it that I think is super important. So there are also limitations. I think the future of this kind of approaches will be more and more important, mainly multimodal AI, but we will have to think really deeply about what kind of data we want to—

Yeah. I want to take a step back and kind of give a slightly different perspective on, I think, where this is all going in the future. One of the things that we and others study is the collective intelligence of the body, taking seriously the idea that what you get in biology as you go up scales is not simply emergent complexity and unpredictability. It's very easy to get complexity and unpredictability out of simple rules. But, I think there's actually something much more interesting going on, which is that you don't just get emergent complexity, you get emergent cognition, meaning that at all levels is already starting from the molecular pathways. We have a couple papers on this. You start to get systems that have problem-solving capacity. They have certain kinds of learning behaviors. For example, pathways and gene regulatory networks by themselves, never mind the cell or the organelles or any of that stuff, just the pathways alone have several different kinds of learning capacity. They can do habituation, sensitization, associative conditioning, meaning Pavlovian learning, and so on. So what you have from the very beginning is a kind of intelligence, problem-solving baked into the different layers of the body, solving problems in physiological space, transcriptional space, anatomical space, and so on. So not just complex behavior, but actual goal-directed Problem solving. So if that's the case, then something very interesting happens. Let's imagine a graph. So on the x-axis is complexity, and on the y-axis is how hard is it to deal with it. So at first, what happens is the more complex something is, the harder it is to— for you to deal with it, right? It's exactly the problem that you laid out. When things get complicated, it's very hard to micromanage the situation because there will be unforeseen consequences and so on. But then something interesting happens. There's a cliff that I think you drop off of, and the cliff is When the system becomes significantly intelligent such that you can communicate with it, you can use the tools for dealing with cognitive systems, the tools of cybernetics, control theory, behavioural science, and so on, then things become much easier. And the dumb example that I use for this is people have been training dogs and horses for thousands of years knowing zero neuroscience. Isn't that remarkable? Why is it that we can get these animals to do all these amazing things? Without having to understand what are all the neurons that you would have to tickle to get them to do a particular kind of action. You don't need to know any of that because all of that is encapsulated by the system that offers to you this amazing interface. It offers you this learning interface. You can train it, right, by rewards and punishments and with more complex animals, other ways and so on. So what this suggests to me is that the AI tools that we are really going to build are not just systems for inferring better ways to micromanage a complicated system. We will get that. Too, I guess. But, the bigger power here is to use these kinds of AI systems as translators, communications devices. I think what we're actually going to use AI for is not better ways to tickle specific pathways because hitting specific pathways has massive limitations. We're not dealing with passive matter here. We're dealing with an agential material that has agendas and various homeodynamic loops. It's going to fight you on the various things that you try to do to it. Some of them will be recognized as hacking attempts by parasites and other biologicals that are trying to control the cells, so they will resist. That is, you'll get side effects and all kinds of loss of efficacy, drug habituation, and things like that. The goal is to try to reset the set point of the system in a way to get the buy-in of the cells that this health state is the new set point, and we're going to help you achieve that set point, right? We're not going to try to clamp the various microstates and try to fight symptoms. We're going to try to get the whole system to do what it does best, which is to adjust towards a complex outcome given simple signals. We don't want to try and micromanage everything. We want to give very simple signals. So I think that the power of AI is to build a theory of mind of the system. We're not trying to make a black box that tells us how to control it bottom up. We're trying to get the system to learn what is the kind of proto-cognitive system that we're dealing with. And it's a little bit like, you know, there are people who use AI to communicate with whales. And things like, you know, people are going to be using it to process radio telescope data to look for extraterrestrial life and things like this. The idea is there is a communication event going on. You want to communicate with a fairly alien intelligence. It's not alien because we're made of the same stuff, but it kind of is because we're only used to communicating with like one kind of creature pretty much, you know, big brains and sort of like us and then that kind of thing. This is different. But the good news is that AI, as exactly as you said, if we let go of the priors a little bit. If we don't force it to think the way we think, right? I think that's the power of AI is that precisely that we should not try to make it exactly neuromorphic or human-like or anything like that. We should let it do what it does best, which is to find unbiased patterns and become a translator tool so that we can communicate with these complex systems in whatever way they're competent to receive. So that's how I see this challenge is we have to move beyond, you know, we start out with these systems that look like they're discovering facts about molecules, but I think ultimately what they're really going to do is discover facts about goals, about competencies, about measurements, about memories, preferences, these kinds of things that are present in tissues. That's really what we're trying to manipulate in biomedicine.

We'll continue our interview in a moment after a word from our sponsors. Hey everyone, Erik here. You know, at Turpentine, we're always looking for ways to streamline our work. That's why we use Notion for all our internal documents. It's been a game-changer for us, and I want to tell you why. Notion is this amazing all-in-one workspace that combines your notes, docs, and projects. But what's really cool is their new Notion AI. It's like having multiple AI tools built right in. You can search, generate, analyze, and chat, all without leaving Notion. What sets Notion AI apart is how it personalizes responses based on your content. It's not just another generic chatbot; it understands the context of your work. Plus, it can search across thousands of docs in seconds, even connecting to tools like Slack and Google Docs. And get this: Notion is used by over half of the Fortune 500 companies. Teams using it send less emails, cancel more meetings, and save time searching for their work. It's designed to keep everyone on the same page while protecting your privacy. Wanna give it a try? Head over to notion.com/upstreampod. That's all lowercase. notion.com/upstreampod. You can try Notion for free, and when you use our link, you're supporting our show. Give the powerful, easy-to-use Notion AI a shot today at notion.com/upstreampod. Hey everyone, Erik here. In this environment, founders need to become profitable faster and do more with smaller teams, especially when it comes to engineering. That's why Sean Lanahan started Squad, a specialized global talent firm for top engineers that will seamlessly integrate with your org. Squad offers rigorously vetted top 1% talent that will actually work hard for you every day. Their engineers work in your time zone, follow your processes, and use your tools. Squad has frontend engineers excelling in TypeScript, React, and Next.js ready to onboard to your team today. For backend, Squad engineers are experts at Node.js, Python, Java, and a range of other languages and frameworks. While it may cost more than the freelancer on Upwork billing you for 40 hours but working only 2, Squad offers premium quality at a fraction of the typical cost without the headache of assessing for skills and culture fit. Squad takes care of sourcing, legal compliance, and local HR for global talent. Increase your velocity without amping up burn. Visit choosquad.com and mention Turpentine to skip the waitlist.

So has this work gotten over the top of that hill and over to the other side of the cliff, or are we still— because when I read this as a sort of naive person that knows a little bit about biology, I think, okay, here's this link. It's been discovered. This is amazing. We're going to discover all these links. This is going to give us all these targets. Then we'll develop these drugs to go hit these targets. And then, to use your term, we'll kind of micromanage our way out of problems and back to health. You're interpreting this in a different way, but I'm not sure if that interpretation is falling out of what the network is telling us, or if that is all based on all the other vast knowledge that you're bringing to understanding it. I guess it's the latter though. Yeah.

Yeah. I mean, everything that I just said right now, you can't pull all of that out of this one paper. Like, this is my view of what the future is going to be. And I don't think we're anywhere near the top of that cliff. I think We're just starting the climb upwards. And I think, you know, in fact, this is a very, I mean, to be clear for everybody listening to this, that what I just laid out is not the mainstream vision. This is a controversial kind of take that I think a lot of people wouldn't agree with. You know, most people see the biological layers here as a kind of more or less mechanical machine. And it's not considered standard approach to start asking, what do your cells, or never mind cells, what do your pathways think about? health and disease and your prior history and all of that kind of stuff. So yeah, the thing that I'm describing is a long-range vision.

So what is the limiting factor? I mean, Leo mentioned a minute ago that there isn't any bioelectrical data in this dataset. I've been hearing more and more recently that AI for biology is data limited, and I've sort of been like, that seems counterintuitive because I look at just the scale of Llama 3, for example, and it was trained on 15 trillion tokens, which is Like, yes, a lot of tokens, but I would think we have in aggregate kind of sitting around in various systems, we've got way more biological data than that. So it's not just a matter, of course, of raw data, it's the different modalities and the kinds. And maybe there are just missing modalities entirely, like the bioelectric data. Do you think we are in fact data limited, or is it that we don't have the right kinds of data? Or more generally, what is the current limiting factor? That would prevent us from like scaling this thing up and going over to the other side of the complexity cliff?

Leo, you want to, you want to start?

Yeah. The answer would be yes and no. I think we have so many biological data right now, but we don't have maybe enough data on bioelectricity or the, but the main problem is maybe not about the data also. It's about how we transform this data into information and knowledge. So what is the theory behind it? And I think you, Michael, just presented a new one that is super interesting. So in omics data, we have so many. This is why we do this kind of integration of all these databases in the same network representation. But also we have clear lack on things that we thought are super important. But this is still very important information. And in the kind of AIs that Michael just described, we could use, for example, the embeddings that we have right now as an input, and they will use it after how to communicate with cells, for example. So yes and no, I would say.

Yeah, I think it's very much data limited because of the kind of data. So some years ago, a former postdoc in the lab, Daniel Lobo, and I started this idea of bioinformatics of shape. So one thing that you really want, I mean, let's think about the kind of data that we have, right? What you basically have is micro-level data. So we have tons of omics in terms of protein, gene, RNA, DNA, all those kinds Never mind the issue that we're lacking all the physiomics. So, the bioelectrical data is completely lacking. So, that's a whole other thing that we can talk about. But, what we actually care about is not any of those things. What we really care about is the anatomical and functional outcomes. So, we want to know, okay, how many fingers? Why this number? Why not that number? Why are there 2 bones in your arm? Despite all the work on limb genetics, still nobody has any idea what determines the number of bones and all of that. So one thing that we're missing, which would be incredibly important for the AI, is here's the way we stimulated the system, the change we made, and then here's the outcome, right? And we started as basically we scoured and then hand-coded all the data for planarian regeneration, crustacean regeneration, insects, and amphibians. So that means we had to come up with an ontology. We had to come up with a graphical way to describe what happened. You know, here's how your bones got rearranged. Here's the length of the different parts, right? So, the topology and the geometry of all the outcomes. So, you have to come up with a language for that. So, we did that and be able to encode that in the database. And so, now, in that database, you have— and so, from each study that we saw, there's functional experiments that say, this is what we did and this was the change of shape. This is what we did. This is the change of shape. Now, this is great, and you can deploy AI on this to say, okay, so given that functional relationship, what can we say about the control structure? You gave it a stimulus and this is how it changed the shape. What's the language? What's the morphogenetic code that leads from this to this? If we crack that, then, you know, birth defects, regenerative medicine, aging, degenerative disease, cancer, all of these things would be solved, right? Fantastic. So, but now we worked really hard and a whole bunch of undergrads helped and we populated this database. It took us about 4 years to build all these things. The amount of data in there is in the hundreds, hundreds. That's not even remotely close to what AI needs nowadays. And so that's the problem, right? Is that not just getting the data out of the journals and into this database, which is already hard. And I'm not sure, I mean, back then certainly AI wasn't up to it. I'm not sure if it is now. And there is no standardization. In fact, the journals, if you copy your methods section from one paper to another, they call it self-plagiarism. And these publishing houses try to boost the value, you know, the amount of novel content they have. So they won't even let you really reuse the methods section verbatim. You have to keep changing it. But that all works against this idea that what we want is to standardize it. We want these papers to be readable by AI. We want it to be very clear on what you did and how this differs from this other thing you did. And having it be in different words because you're trying to supposedly not plagiarize yourself, what a weird concept. All it is is confusing for any kind of standardized method. So that doesn't exist. But even worse than the data standardization is the fact that getting the data is incredibly expensive. It's hard. I mean, this is what we do as developmental and regenerative biologists all day. Is we make perturbations to test some hypothesis, and then we get some kind of outcome and we measure the heck out of it. But that, that's really hard and expensive and time-consuming. I don't know how to get that functional data in a way that gives you the amount of material that you need for current AI approaches, right? So, you know, hundreds of thousands minimally, and then millions typically is my understanding of what you need for these kinds of things until we really get the robot scientist platforms. Off the ground. So this is something else that a number of labs are doing. Something that we are doing is to try to automate some of that process and to really enable this kind of feedback loop where there's laboratory automation that does experiments and records these outcomes. But really critically, not just we keep sequencing the RNA, the DNA. I mean, that's great, and that gives you information about the low-level components of your material, but to really get functional data on the things we care about— the shape, the structure, the function. The ability to problem solve in anatomical space and physiological space, all of those things. That I think is going to be a really important part of really getting AI unleashed on all of these things.

So I wonder if we, I've had this kind of pet idea for a while that this may shed some light on. And so I'll just pitch it to you and then you can tell me what's missing from it. I've had this sense that when we look at the giant scale models, right, they clearly are developing some sort of proto world model that's still debated, but I think it's like Increasingly hard to argue that there's not some sort of world model in there. It is a flawed world model. That doesn't mean it's not a world model at all. Anyway, bracketing that debate, I feel like one of the, you know, I just look around at the world, right? There's so much confusion even among people as to like, what makes us feel good one day to the next. It's very striking to me when I just say like, how will I feel tomorrow? And then how will I feel next Thursday? Like, I really have no idea how I'm going to feel next Thursday. It's kind of crazy that it's like, it's not that far in the future. And it could be anything, right? It seems so inaccessible even from my own individual subjective experience. And so people are constantly taking supplements and doing fad diets and who knows what, right? All these kind of different things which largely seem to be born of that confusion. And then I think, man, there's just all this data from the omics, as you said, to stuff that we can increasingly collect through wearables, like to take a picture of your food, but also monitor your heart rate on an ongoing basis and track your exercises and so on and so forth. And so I have this sense that we're getting probably pretty close to where we might be able to create a large-scale model of human wellness or human well-being that could just be like sequence data in the same way that the language models are in sequence data, predict one token at a time. I think of the human wellness thing as predicting the next wellness check-in. Like, you've got all these different data points. You ate this for breakfast and you slept for this long on your Eight Sleep mattress, and your Strava run was this today, and maybe you're on this medication or not or whatever? And then what's the next wellness check-in? And you train the model to just predict how well you are at the next time. Those could be objective maybe, and subjective measures. It seems to me like that could work. I mean, it'd be a massive coordination problem, massive data collection problem, so on and so forth. I wonder if you think that could work or if you think there are— one counterpoint to that might be like, well, you've got really low-level stuff and you've got really high-level stuff, but you're missing these sort of intermediate layers, and this goes obviously to your multiscale intelligence worldview in general, maybe that gap between what I can measure at the low level and the high level is just too big for a model to bridge that. Or maybe these sort of middle scale things could sort of be learned as emergent properties or emergent middle representations that arise because they're useful for predicting those next wellness checks. Do you think that that paradigm has promise? And if yes, why aren't we doing it? And if no, what do you think is missing in the middle there?

Well, a couple of things. First, just a kind of a funny observation. I think that because— so there's a prediction paradox here, which might actually be real in the sense that to the extent that you deliver to some people an application that can predict how they're going to be feeling, that will immediately make them depressed because the whole idea that they are determined to, in a way that they don't know, is depressing to a lot of people. And then the question is, well, did your system predict that they were going to be depressed or not? Right? It's, I mean, it's the old prediction paradox. But so we have to think about the second-order effects of people knowing that this is happening. But overall, I absolutely think that something like that will have predictive power, no doubt. And I think it can have uses, but I think it's going to have the same failure modes that we now see with language models. In the ways that they occasionally fail, which means that they put on a pretty good show and they do a nice job for a lot of things. But the one problem is that we don't know what we're actually optimizing for. So for example, if what you're optimizing for is feeling good, I would think the model will very quickly hone in on the fact that you should be on a lot of heroin and that'll take care of things in the short term, right? So the question is, What is it that you're asking this model to do? You know, is long-term health and success that it's got to make you kind of miserable for a couple hours every day with high-stress exercise because long-term that's going to be good for you? Or is it going to try to say that, well, here's as many donuts as you want because I know you're going to feel good about that? So I don't think we have a clear specification of what success looks like in these cases. You know, if you live to be 90 and you're lean and mean and boy, you never had a donut, is that the life you wanted? Like, I have no idea what the answers are to these questions, but what the AI will be good at is optimizing whatever you want it to optimize. But let's be really careful and humble about the fact that we're not actually sure what we want to optimize here. And I think it's very easy to, you know, they say whatever gets measured gets managed. And I think that's absolutely true. If we set it a fitness function, it will— Absolutely shape us to that. But do we know what that's supposed to be?

Yeah, we definitely don't. I've been really intrigued by the notion of multi-part reward design as another interesting area of machine learning recently, because it does seem like there's a paper on this that's like, scalar reward is not enough. And the basic idea is there's no single metric that you can really optimise for in probably any sufficiently powerful system, but certainly one that we would focus on the problem of our own will. being, there's not like a single measure that we could use. So at a minimum, we probably need like multi-timescale different—

Yeah.

And that's okay.

Right.

You can have, you know, vector or whatever quantities that you're trying to optimize. And Mark Solms talks a lot about this in terms of the drives, basic biological drives, where you can't just boil it down. You can't rank order them because they're orthogonal. They're different dimensions and they all need to get met. The problem comes when they conflict, right? Is like, it's all well and good to try to optimize multiple variables when they're more or less independent. But sometimes they conflict, and in the real world, they're definitely going to conflict. And then the question is, what's the optimum prioritization? Yeah, and I think there's a lot to still be done to kind of develop what it is that we want to have happen here. And not to mention, this always seems to butt up against issues of agency and autonomy and will, because, for example, already you hear people like some of these new drugs coming online that prevent obesity, and eventually you'll have something that helps you build muscles and all that kind of stuff. Already you hear people saying, Well, that's not fair. That's cheating. You know, you should get it the old-fashioned way. You should be in the gym. Like, well, how many hours should you be in the gym? Is there some God-given number of times that's, you know, amount of time per day that's the correct, and anything more than that is unnecessary, and anything less than that is cheating? Well, what's the correct amount of time? And so if you had a system that managed you towards a specific outcome, I'm sure you would get people that are resentful of it because they're saying, well, now I don't need my willpower anymore. I was really good. I was the one who avoided all the donuts on my own, and now everybody's managed it. But that's not they've given up their humanity because they're not needing to control it anymore. So we're going to run into all those kinds of issues, right? So how much help are we willing to accept from these prosthetics? Because that's very quickly going to merge into other areas of life. So once you have your, I don't know, glasses-mounted AI assistant that's seeing what you're doing and kind of helping you to a helpful. Whatever, somebody's going to make a plugin that's a relationship consultant that will say, no, don't say that, you idiot, say this other thing and it'll be much better. And then every facet of your life is going to, you know, objectively lots of it will get improved, no doubt, right? No doubt some stuff will get improved. But overall, what do we think about those kinds of prosthetics, you know, that are taking over not just the boring stuff like digestion and whatever, but some of the things that people think about as core aspects of themselves, like willpower, relationships, whatever.

It reminds me of that Google commercial during the Olympics where the dad is using Gemini to help his daughter write a letter to her hero. And it's like, this isn't maybe where we want to start. Nick Bostrom also has some really interesting work on this recently. What does a solved world look like? And how do we think about the value of people doing things for each other if you could just have an AI do it or whatever? But there might still be some meaning that we don't want to give up that comes from the actual origin story of what has happened as opposed to purely that it happened?

Yeah. I mean, I think we really need to be careful because it's very popular and it's an easy tendency to say, look, this is where we are now as humans and let's not give that away. But, but let's be clear. This is what, what we have now is in no way some magical optimum. You know, I had a friend who was a forest ranger and a very kind of outdoorsy guy. And he was astounded to know that I don't know how to hunt down my own deer and kill it and eat it. And he would ask me like, what would you do if there aren't any supermarkets? I'd be like, I have no idea, you know? And he would say, well, how are you a grown man and you can't go out and like catch your own meat and whatever? So this idea of giving up what used to be basic core competencies, right, in favor of division of labor and of outsourcing some of the things that we used to have to do on our own. All the prosthetics that we use now, canes, glasses, a toothbrush, education, workouts, all these things that we know how to do, we've already done that. We've already outsourced an incredible amount of stuff. And so, yes, potentially you could outsource things you shouldn't and don't want to, but let's not pretend that we are in some pure state now and only beginning to start to think about it.

Let's flip things around now. I have been very fascinated by this idea of multi-scale intelligence. And when I look at the current prevailing paradigm in machine learning, much of it obviously is driven by scaling up relatively simple kind of single-scale architectures, right? Like, the transformer is obviously just the same layer stacked time and time again, and it doesn't have this sort of multiscale complexity that we see in biology. Do you have any intuitions for what people might ought to be looking at to try to bring in concepts of multi-scale intelligence to machine learning? Like, what would a transformer look like if it had access to something lower level? I feel like there's a seed there in terms of what you said about even just simple dynamical systems having learning capability, but I'm not quite able to crystallize that. So maybe you can help.

The multi-scale competency architectures, we are starting to see it, for example, with neural cellular automatas that have been developed very recently. And we have also hybrid Well, they call it hyper-NCAs. So now we have several layers of this. And I think we are starting to be really close to what you call truly biologically inspired architectures here, because they reflect much more what is AI as collective AIs, as we have neurocellular automata, or just that we have neural networks inside each cell, and cells are communicating between each other in a grid, for example. And yeah, so just to say that I think it's really beginning.

Well, 2 things. So we are working on some of this in our lab. And this is what I call sort of truly bio-inspired AI because neuromorphic architectures and all that, I actually don't think that's where the secret sauce is at all. I think there are some really fundamental things about the way life is organized that have to be the basis of the kind of AI that we're talking about here. And we can talk about a few of those things. I do want to point out that some of those things I actually am not going to talk about anywhere. I started writing a paper on this at one point about Like the specific things that I think differentiate true morally important agents from the kind of systems we have now and how you could achieve that. I actually think we have a pretty reasonable guess now as to how you would do it, but I don't think we should, and I don't want to be responsible for it. Somebody else will figure it out, no doubt, but I don't really think there's a way of escaping it. But that's why I don't want it to be me. So some of these things I'm not going to talk about, but one thing that I can talk about. Is this— the issue of one specific difference is that, you know, biology is always dealing with an unreliable medium. Fundamentally, everything in biology is full of noise and you can't really count on it. So never mind the environment's going to change, like if you're an evolutionary lineage, right? And of course the environment's going to change, but your own parts are going to change. You're going to mutate. There's going to be all kinds of things that are going to happen to you. You So what biology does, I think, is not overtrain on its priors on the past. And I think evolution makes not fixed solutions to specific environments, which is what we're taught, you know, in terms of selection and so on. But I think what it makes is problem-solving agents at all scales, specifically because from the very beginning, it's dealing with an unreliable computing medium. And I think that's the basis of the intelligence ratchet that we see in life is that that you start right away from having to— the only thing that survives are systems that can reinterpret whatever is going on now in whatever is the best way, not how it was meant to be or some locked-in idea of how to interpret your genetic memories, your physiological memories, but a creative, dynamic reinterpretation, which leads to problem solving, and then that scales up. So I think if we're really going to have AIs that are multi-scale biological kinds of I think we have 2 pieces to this. One is the multi-scale architecture. So I do a talk sometimes called Why Robots Don't Get Cancer. Like, it's not an issue that anybody's had before, but in a proper bio-inspired system, you will absolutely have subsystems that decide to go on their own. They disconnect from the rest of the network. Like right now, you make an AI or a robot or something, and you hope the higher level is intelligent, but you know, all the parts underneath are kind of dumb. So it's a very sharp kind of discontinuity. In biology, it's not like that. Every layer has some competencies, and each layer tries to bend or behavior shape or coerce or hack the layer below it, and sometimes the layer above it, to get it to do what it wants to do. So that's the kind of thing we would have to have in these architectures, and I think we would have to give up what we work at very hard in computer science, which is isolating the layers so that each layer can count on the bottom layer being correct. You know, when you're writing high-level code, you never worry about that. Well, these variables sometimes degrade and the data kind of goes away. Like, yes, there's unreliable computing where we typically don't code that way. You know, you try very hard so that each level doesn't have to worry about the metal, and then you don't have to worry about your machine code. You assume that what you put in your variable means exactly the same thing that it will mean later on, and it hasn't floated off and whatever. I think that that's what keeps these architectures from going to the kinds of things we expect them to see in biology. I actually think it's remarkable So what we've learned in the last few years with the AI that we do have, I think is remarkable and showing us that if you're very clever, how much you can get with this kind of fixed architecture. But I think it's absolutely going to plateau if we don't take it in this other direction.

You see some kind of previews of that already, right? With like dropout and certain like data augmentation type techniques where a bunch of different types of noise are added to training data. But between those sort of just randomly deleting neurons or randomly deleting weights and kind of systematically but randomly messing with the data, that seems to be like a step in that direction. Those things do. I think they're used generally because they help with robustness, right? They help with generalisation. What's kind of next there? Is there a— and maybe this gets into the sort of detailed stuff that you don't want to talk about, but if it does, then I guess the next sort of zoom out would be like, I think all of this is very fraught, right? Last time you spoke about being worried about creating trillions of descendants, not being able to vouch for their well-being, and maybe you shouldn't do that. I think the other angle that I'm seeing more and more right now is that this sort of pure, distinct intelligence that kind of comes out of the lab and is like a lightning bolt into a generally sort of somewhat equilibrium-finding ecosystem seems like a destabilising force, I'm wanting to find ways to make it a little more organic, to start thinking about, instead of how can we find the perfect recipe for alignment of some superintelligence, is there a way that this could be more integrated holistically into the world so that it's not like an all or nothing, like the thing is aligned or it's not aligned, but rather it's in a more sort of buffered type of dynamic with the rest of the world? I also see some really interesting things. I just recorded an episode with a couple of guys from AE Studio where they're looking at neglected alignment techniques, and they've recently put out a paper, one on self-modelling and one on what they call self-other distinction minimisation. Both of these are biologically inspired. Self-modelling is, obviously there's theories of consciousness that come from that or that are predicated on that. And they basically just have a very simple thing where they give the model 2 jobs. Instead of just having to do the task, it has to do the task but also predict an internal state from an earlier layer. And fascinatingly, what they find in that paper is that the network gets simpler. There's like 2 ways that you can make the self-modelling succeed. One is you can get good at learning how to self-model. The other is you can become easier to model. And they do see a reduction in this complexity metric when they add the task of self-modeling. The idea there is that like maybe it'll be easier for other, you know, this sort of could be a pro-social mechanism.

Yep. That's super interesting. And if you wouldn't mind actually dropping me a link to that, I'd love to see it. I mean, this gets into something that we were mentioning earlier, which is that one of the reasons that animals offer this amazing interface is because Not only does it make it the learning interface where you don't have to know what all the neurons are doing and so on, is not only is it because it makes it easier to relate to others and, right, and have a theory of mind about others and to communicate, but also it makes it easier to control itself. So a biological system that wants to control itself and its parts, it also behooves it to have parts that are modular and encapsulated and have competencies in a way that you don't have to micromanage them. You can do some delegation and things like that. The trade-off you get there though in biology is that if you are too understandable and too easy to control with signals, then you're too hackable. And, there will be parasites and cheaters and exploiters and competitors and whoever that are going to hack you. And, we see this all the time.

There are—

I don't remember if I spoke about this last time, but there are these amazing structures, these galls that are these red spiky just incredible-looking things that are formed out of leaf cells on an oak tree. Yeah. Because the wasp is this like non-human bioengineer that can prompt the leaf cells to do a completely different thing that they apparently know how to do. Who would have known if we hadn't seen it? But it's completely taking them over because the cells themselves are so competent at other things. So I think there's going to be some kind of drive, at least in nature, to optimize this and not to be too understandable and too controllable and whatever. But in our technology, maybe it doesn't have to be that way. When maybe we can put pressure on the kinds of things that we develop to make sure that, that there are better ways for us to interact with them. And then just the final thing to say about alignment is that I don't think there's any hope of a single alignment because we're not aligned among ourselves. Not only are we not aligned culturally, right? If you ask different people on the face of the planet what they see as an ideal future, you're going to get 100 or more different answers, right? So we're not aligned within ourselves. We're not even aligned really within ourselves as a single human. With split-brain patients, you find out very quickly that there are actually all kinds of multiple opinions, sometimes radically different opinions on important things. You know, do you like your spouse? Do you believe in God? You know, that there are modules within one person that don't agree on the answers to these things. But typically they somehow cobble together a fairly consistent, or maybe not in some cases, outward functional persona that kind of hangs together as a coherent, consistent thing. But it's a pretty thin covering, as we know from All kinds of situations in life. So being aligned is a whole other question that's not as simple as it sounds.

So this and another episode that I just recorded sort of pushes me toward this idea that it seems like we might kind of need to, and I'm not sure I believe this, but just to present the case, it seems like we might need to go to a sort of computational or digital life and try to sort of merge the like normal environment that we have developed in with the digital in some way to try to create an overall dynamic that can hopefully be in some balance or some stability. It seems like digital life is not super far off. The episode that I just recorded with 2 guys from the Google Paradigms of Intelligence team showed that in a remarkably simple— and you've done some similar things with your sorting algorithms. You have these little cells that perform remarkably sophisticated behaviours. They had a similar kind of thing where just Very minimal computing environment, random strings, random interactions between strings. Before you know it, self-replicators are arising, and then you've got all these dynamics. The complexity metric spikes, and then there's sort of takeover, and then certain things go extinct, but then there's like parasitism and symbiosis, and like all these sort of familiar paradigms are emerging in. And these are just 64-byte strings that are randomly interacting. There's not even an optimization function.

Yep.

So it seems like that could happen very easily. Part of me is very scared of digital life because I don't know what that would entail. It seems like it could be invasive, it could be hard for us to deal with. But then I'm like, what is the alternative? Is there any alternative to a sort of ecology of AI long term? I don't know, you react to that however you want, but I struggle to see something that doesn't have sort of ecological qualities being long-term sustainable for us anyway.

Yeah. I mean, about the digital life, I think one important thing to keep in mind is that, well, to me, making these things, it's not a goal in and of itself. I mean, we make synthetic life forms and then other people make digital, you know, digital agents and so on. I think the reason we do it is to better understand the agents that are all around us right now. So I think we have an incredible amount of mind blindness at the moment. Because we were evolved with firmware for detecting intelligence of a very particular kind, you know, medium-sized objects moving at medium speeds in 3-dimensional space. That's what we're okay at kind of noticing. But I think that what all of these efforts in diverse intelligence and artificial life and AI, what they're all helping us to do is they're serving as prosthetics. They're helping us to recognize and get a little broader in our ability to recognize, communicate with, and ethically relate to other kinds of minds that are all around us right now. And, and we're, we're just not good at recognizing them or, or at figuring out how to relate to them. And so I think that's gonna be really important because they're already here in surprising places. And they're crumbling. They were never great, but at least in the olden days, they kind of did well enough, but they gotta go. And they have to be replaced with principled, empirically useful models of what it is that we mean by the space of possible beings and by the different distinctions that we try to draw between them.

So only a few minutes left, just being mindful of time, I guess. Do you have a sense for where that ultimately takes us? I mean, I'm very compelled by your idea that we shouldn't create trillions of descendants that we can't vouch for, but I also feel like that's in tension with the idea that there are multiple companies out there right now that seem to be trying to create a sort of monolithic superintelligence. Ultimately seems to be their goal. And And then they have the question of, like, well, how do we make that safe for ourselves? And nobody has a good answer. And so I feel like we have very tough choices on either side. I also take— I'm open-minded enough to imagine that I could be pleased with long-term descendants that are not just like me. It's not all about just cloning myself into the future. But I do feel like I want some version of our substance, our values, whatever, to go on, or I would definitely feel like something is lost if it's just a sort of mindless, super powerful optimizer terraforming planets or whatever. So what do we do with this? I mean, it seems like we're in a very tough spot.

Yeah, well, there's a simple, easy question. Okay, a couple of things. First of all, I think everybody should take the time to paint for themselves the picture of the future that they want. Maybe there are some people that think that 1,000 years from now, we're still walking around with lower back pain and astigmatism And the limited lifespan of whatever kind of body you happen to have been born into and whatever your IQ limitations are. I mean, that seems outrageous to me, but maybe some people want that. It seems to me that once we've given up the idea that this particular form with all of its limitations, cognitive limitations, physical limitations, and so on, if you give up the idea that that was planned out for you by a benevolent, all-knowing intelligence that, you know, in some way set this as the optimum and this is how it's got to stay and don't mess it up. If once we give that up, and I think we have to give that up, then you're free to imagine, well, what does a mature species look like? What are your rights as a being being born into this world? Do you get dumped into a random body with whatever the cosmic rays did to its DNA and physiology during development, and that's how you have to live your life with those limitations? Or is part of your rights or some kind of freedom of embodiment where you can live in this world in whatever way you want? And more generally, this notion of You know, like, how can we not be replaced? I mean, each one of us is guaranteed to be replaced. Our children are not going to be exactly like us. They're going to do things that blow many of our values out of the water. We are all going to be replaced. There's no question about it. I mean, we have been for millions of years. The question is, what do you want to be replaced by? That's the only question. And, you know, you said mindless terraforming. Agreed. That doesn't sound too exciting. But I think we should all formulate what should the universe be filled Should it be filled with humanoid-looking things with our DNA? That doesn't seem particularly optimal to me. But, but, you know, magnificent agents with freedom and creativity and an exploratory instinct and cooperative, compassionate values and whatever, like, great. And if they look nothing like us, also great. I'm not particularly tied to this form in the slightest. I don't see why anybody should be. So yeah, I think it's paint negative futures of things we don't want. That's easy. And that's what everybody's doing. I want to see everyone specify, what do you want? What do you think things should look like in the future for us? And more generally in the universe, what should the universe be filled with? And if you can make that statement, then we're closer to being able to work towards it.

We haven't even got to the paper that you have on agential memories, but maybe one way to kind of tie that in with this line of discussion is I'm just reminded of the Derek Parfit idea that identity is not what matters. We obviously, as individuals, tend to be very wrapped up in our identity and our individual experiences. I wonder if there is a way to start to prompt you for a vision of like what a good future might be like. Is there something where we can sort of dissolve the boundaries of individual identity and maybe sort of share more of this like cognitive glue that is Memory, that's an opportunity to give a little bit of a tease of like that whole notion. But then I wonder if you could zoom out from there into something that you would personally find inspiring.

Okay.

In the last couple minutes, I'm not going to try to paint an entire picture of the future, but I do want to say that you have to be really careful with the notion of gluing everybody together. And what I mean by that is a lot of people who hear my story of cancer as a breakdown of gap junctional communication of the physiological connection between cells. One of the things that allows them to cooperate and be aligned is that there is an anonymization of memories where cells connected by gap junctions share physiological traces of past events to the point where it becomes hard to have individual memories. You have kind of a collective memory and that gives you more of a collective intelligence. And then people say, oh, well, there you go. We should all, you know, sort of gap junction ourselves together and then life will be great. Well, so I just want to point out that I think that's very dangerous. I think that's been tried in human history. a few times, and in modern history, it goes the same way always. It's pretty routinely terrible. And part of that issue is that the different layers have different goals in different spaces. And one example that I sometimes use is if you go rock climbing and you meet a bunch of social goals, a bunch of personal development goals, you're happy, whatever, all the skin cells you left on that rock face didn't get a choice about whether they were going to have a good day or not. It's, you know, you're not worried about them. And so that reminds us that simply forming together into a collective intelligence has no guarantee that that intelligence is going to have the welfare of the pieces high on its priority list. In fact, we don't know where the goals of collective intelligence come from. We must develop— I think it's an existential task for humanity and flourishing of life on the planet, is for us to develop a mature science of figuring out where do novel collective systems goals come from, how do we manage them, how do we predict them, and so on. So I guess the only thing I'll say right now is that I think essential for the future is a development of the emerging field of diverse intelligence, of understanding where goals come from and how we can predict them and how we can guide them towards a positive—

That might be a great note to end on. I know you want to get out a couple minutes early. So much more to your work. Anything else that we didn't touch on or any other closing Any closing thoughts before we break?

Yeah, no, just, you know, we can put a couple links in the show notes where people can follow up. There's a ton of work from us and others on this, so we can put some links on.

Cool. Well, I hope we can do this again in the not too distant future because your rate of output continues to be incredible and there's always something fascinating in every paper I find to discover. So thanks for all the great work. And Professor Michael Levin from Tufts University, Leo Pio-Lopez, staff scientist at Tufts University. Thank you both. For being part of the Cognitive Revolution.

Thanks very much. Thanks for having us. And yeah, Leo's got some great papers coming in the next few months, so stay tuned for that.

We will keep our eyes out.

See you then. Thank you.

Awesome.

Upstream with Erik Torenberg is a show from Turpentine, the podcast network behind Moment of Zen and Cognitive Revolution. If you liked the episode, please leave a review in the Apple Store. Hey everyone, Erik here. At Turpentine, we're building the first media outlet for tech people by tech people. We're the network behind the show you're listening to right now. We have a slate of hit shows across a range of topics and industries, from our AI and investing cluster of podcasts to shows that drive the conversation in tech with the most interesting people in the industry. Interesting thinkers, founders, investors, and influencers like Econ 102 with Noah Smith. We're launching new shows every week and we're looking for industry-leading sponsors. If you think that might be you and your company, email me at erik@turpentine.co. That's E-R-I-K at turpentine.co, and let's partner together.