The AI Fundamentalists
A podcast about the fundamentals of safe and resilient modeling systems behind the AI that impacts our lives and our businesses.
The AI Fundamentalists
Complex systems: What data science can learn from astrophysics with Rachel Losacco
Our special guest, astrophysicist Rachel Losacco, explains the intricacies of galaxies, modeling, and the computational methods that unveil their mysteries. She shares stories about how advanced computational resources enable scientists to decode galaxy interactions over millions of years with true-to-life accuracy. Sid and Andrew discuss transferable practices for building resilient modeling systems.
- Prologue: Why it's important to bring stats back [00:00:03]
- Announcement from the American Statistical Association (ASA): Data Science Statement Updated to Include “ and AI”
- Today's guest: Rachel Losacco [00:02:10]
- Rachel is an astrophysicist who’s worked with major galaxy formation simulations for many years. She hails from Leiden (Lie-den) University and the University of Florida. As a Senior Data Scientist, she works on modeling road safety.
- Defining complex systems through astrophysics [00:02:52]
- Discussion about origins and adoption of complex systems
- Difficulties with complex systems: Nonlinearity, chaos and randomness, collective dynamics and hierarchy, and emergence.
- Complexities of nonlinear systems [00:08:20]
- Linear models (Least Squares, GLMs, SVMs) can be incredibly powerful but they cannot model all possible functions (e.g. a decision boundary of concentric circles)
- Non-linearity and how it exists in the natural world
- Chaos and randomness [00:11:30]
- Enter references to Jurassic Park and The Butterfly Effect
- “In universe simulations, a change to a single parameter can govern if entire galaxy clusters will ever form” - Rachel
- Collective dynamics and hierarchy [00:15:45]
- Interactions between agents don’t occur globally and often is mediated through effects that only happen on specific sub-scales
- Adaptation: components of systems breaking out of linear relationships between inputs and outputs to better serve the function of the greater system
- Emergence and complexity [00:23:36]
- New properties arise from the system that cannot be explained by the base rules governing the system
- Examples in astrophysics [00:24:34]
- These difficulties are parts of solving previously impossible problems
- Consider this lecture from IIT Delhi on Complex Systems to get a sense of what is required to study and formalize a complex system and its collective dynamics (https://www.youtube.com/watch?v=yJ39ppgJlf0)
- Consciousness and reasoning from a new point of view [00:31:45]
- Non-linearity, hierarchy, feedback loops, and emergence may be ways to study consciousness. The brain is a complex system that a simple set of rules cannot fully define.
- See: Brain modeling from scratch of C. Elgans
What did you think? Let us know.
Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:
- LinkedIn - Episode summaries, shares of cited articles, and more.
- YouTube - Was it something that we said? Good. Share your favorite quotes.
- Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.
The AI Fundamentalists a podcast about the fundamentals of safe and resilient modeling systems behind the AI that impacts our lives and our businesses. Here are your hosts, andrew Clark and Sid Mongolek. Hello everybody, welcome to today's show. We are going to be discussing complex systems and before we do, we've seen some interesting news this month released from the American Statistical Association, which I'm going to put Andrew on the spot to talk a little bit about their announcement and why we do what we do here at the podcast.
Speaker 2:Yeah. So it's really great to see. As a member of American Statistical Association, it's really great to see that they're finally kind of coming around here.
Speaker 2:Originally, when I was going through school, the statistical organization was kind of like, oh, data science and AI and just kind of like poo-pooed it almost a little bit.
Speaker 2:So it's really good to see that they're I think they're a little late to the party, but they now have a working group around and they're kind of reframing some of the messaging and essentially the meat of this post that they had was really really solid read. It's kind of saying that like you really have to use statistical techniques and causal relationships and things like that to really use AI well and to really get the most value from these systems, which is a lot of what as we've tried to coin the term hasn't really caught on too much is bringing stats back as kind of the goal of this podcast. So American Statistical Association is definitely along those same lines of really seeing that to really build safe, performant, resilient systems, you really have to be using statistical best practices in creating that data and curating it, versus just doing, you know, data mining and building systems from there. So it's just good to see that we're having some coalescing around those ideas from multiple different sources.
Speaker 1:And it's a nice segue, because we're going to be talking a lot about complex systems today, and it also sounds like this organization has recognized that there's not going to be one way and modeling systems are only going to get more complex. So the need to work together is very important and for our complex systems topic today, we also have a special guest. We have Rachel Osaka with us today. She is an astrophysicist who's worked with major galaxy formation simulations for many years. She hails from Leiden University and the University of Florida and now works as a senior data scientist where she works on modeling road safety. Welcome, rachel.
Speaker 3:Hi, thank you so much for having me.
Speaker 4:Yeah, we're especially glad to have Rachel on board because she really works with complex systems, and we here in the data science and AI world like to think that we work with them, but we don't always get to work with truly complex systems. So let's hop a little bit into what a complex system is. So a complex system is any structured system that involves multiple agents, actors or elements, and we can think of these complex systems as existing on several levels. Right, this can be in like a really physically small size, like human cells making up the body, maybe a more medium size, which is economies at scale of agents of those economies, or even like some of these larger scale ones, like weather systems or, as Rachel's going to walk us through today, specifically galaxies and astrophysical simulations. And one thing that we're going to note here is that we're calling these complex systems because we can't describe how they work with single rules or reduce how they work to simple explanations or even just one level of explanation. Often, we're going to be dealing with multiple levels of explanation of how these models work and we can't then predict how more complex levels are going to build off of our simple levels. So in the past, when we tried to do research and study these types of complex systems like studying the weather we often had to summarize a lot of their system dynamics and their processes with simple but basically provable in general explanations.
Speaker 4:Since then, we've gotten access to much more powerful computing resources, right? And when I say much more powerful, I'm just talking about basically the 90s right, where we had access to big amounts of data, big amounts of processing, and we can finally start running simulations on huge amounts of data, right. These are not hand calculations anymore. So now researchers can study the complexity of factors involved in a subject without simplification, without major simplification or reduction, and we've seen a lot of techniques used for this. You may have seen stuff used in like agent-based modeling or in MESA layer techniques, but we'll hop a little bit more into a high level here. So I'll pass it over to Rachel to talk about how the early days of physics research on galaxies and otherwise is really fundamentally different from how it is today, and how the shift in complex system studies has changed.
Speaker 3:Yeah, so a galaxy just so we're all on the same page is a group of stars, gas dust and a whole bunch of dark matter.
Speaker 3:If you haven't heard of that, don't worry, we can't see it anyway. But to know that any of this stuff is there and how it's moving and interacting with each other, we had to make a lot of assumptions before we had computational resources to explore the dynamics more in depth. Previously, we kind of figured that galaxies were rotating. We were able to observe this with specialized telescopes that could show us the rotation, and we're able to count the stars and see how many of these stars billions of stars move around in, let's say, a disk galaxy. To actually know how galaxies interact with each other, though, takes place over tens of millions of years, if not hundreds of millions of years. That's something that we can't see happening in just one snapshot of a telescope or in one human lifetime, and so today, we use hydrodynamical simulations and other types of simulations to see how galaxies interact with each other over the course of tens of millions of years, using the basics of physics that we know, such as gravity and hydrodynamics.
Speaker 2:Excellent. This is super exciting and very much in line with a lot of the approaches that we talk about on the podcast and really just these understanding what you're doing with modeling and really modeling something specific right. And, as you mentioned, your, your basis is still in the fundamentals of their physics-based mechanisms right. That you're trying to then simulate to see how those things are to respond based on what you know and then figuring out what you don't know, which is what's really crucial. So, as on the spectrum of the models we've talked about, a lot of models we've kind of talked about is a smaller churn prediction type stuff that a lot of companies are doing in the industry versus this is the far side of like the some of the top flight research going on in the world and the most complex simulation. So it's really good to see that whole gamut Really excited to have Rachel on here.
Speaker 2:And also these types of simulations are used a lot of other places too. And also these types of simulations are used a lot of other places too. This corresponds a lot with the NASA conversations we've had and the systems engineering and the spacecraft as well, as there's a lot of from backgrounds in economics. There's a lot of economic-based modeling going on here at the different layers agent-based modeling, the MESA layer and things like that that Federal Reserve and other people are doing for banking as well, and just financials, like how will the economy respond for easing monetary policy and things like that. But I'm very excited to be digging in today on the physics and astrophysics side specifically.
Speaker 4:Yeah, and so when we're working with these complex systems, there's a lot of challenges and difficulties that come with this that we don't see in other simpler systems. So I'll just rattle off the list here, but then we're going to go into detail on these. So this is going to include things like non-linearity, chaotic systems, collective dynamics and emergence. So let's dig a little bit into this non-linearity piece. And I think all of us in the modeling space are very familiar with linear models right, this is our least squares our GLMs and our linear SVMs.
Speaker 4:These are really powerful methods and they got us really far and they're nice and explainable, but they're not good at modeling all possible functions. For example, if we have, let's say, a decision boundary and in one circle we have group A, a surrounding circle we have group B and a surrounding circle around that we have group C, there's no linear boundary we can draw which can meaningfully separate these groups. We could try and approximate it and throw a lot of lines together, but this is really not the best solution to the problem. So we run into a problem with a lot of generalizability not being accessible to us. This is where we see stuff like neural networks take off, where they're really good at modeling nonlinear situations because of those nonlinear layers that we put into the model, and so theoretically they could then model any function. So I'll hand it to Rachel to talk a little bit about how nonlinearity also exists in the natural world and why we'd be modeling these types of situations that we throw neural networks at, sometimes not thinking about why they're the right thing to use use.
Speaker 3:Yeah, so when we think about in space, two objects interacting via gravity, a planet orbiting a star, it seems pretty straightforward. We have one equation, newton's law or, if you want to get fancy, you can use Einstein's principles to model how that planet is going to orbit the star, model how that planet is going to orbit the star. But now we're throwing in a moon and seven other different planets and plenty of asteroids that can mess up this orbit considerably. This is non-linear. Adding all of these different bodies into the system causes a lot of wobbles, causes some unpredictability. If we're thinking of Earth specifically, it's incredibly minute. But if we have a system for a different star with smaller planets that are closer together, they can influence each other more readily. Now we're dealing with a nonlinear addition of gravitationally bound bodies. They're going to start wibbly wobbling all over in their orbit.
Speaker 4:Totally, and I think that gets us a little bit into this idea of like a feedback loop, right, that one motion pulls one motion pulls another motion. Um, those of us who, you know, have been in the nlp space since before transformers, you know you might remember stuff like recurrent neural networks, right, where the outputs of one node gets passed back to an older node, right. So this idea that, like, there is a feedback between outputs going into inputs, outputs into inputs, which is similar to what Rachel is describing here, with how the way that these things interact is not just linear, they're repetitive and recurrent, even and I think that leads us really nicely into this next step, which is chaos and randomness that these are an inherent part of these complex systems. We often expect that these systems are going to have something that's non-deterministic, something that we can't strictly always say this is what's going to happen next. They may trend towards some central outcome, but there's no definitive cause and effect cycle, and this can cause problems when the systems are measured either intermittently or rarely.
Speaker 2:And we're all got most familiar with the chaos theory from Jurassic Park for the movie fans out there.
Speaker 2:But it's definitely a very key component of all of these different components of what makes a system really complex that non-linearity, chaos, randomness, feedback loops.
Speaker 2:We're going to also talk about emergence and things like that soon and that's where you start seeing today we're focusing specifically on the astrophysics implementation, but these types of complex systems showing up in multiple places and then sometimes that's that's where the neural networks could be helpful.
Speaker 2:But sometimes it's even that much broader than that of Monte Carlo simulations as a component, or really composite modeling systems with many different areas, or you're completely making things up as you go along in some, some instances, um, to really capture how large and complex the problem is. So it's really about this is almost a foundational type use case, but this is where we get. We get into problems and really defining those fundamentals of what is the use case, what are you doing and why, and making your decisions from there, versus picking tools. Because, as Sid's highlighted, for why SVM and why linear or nonlinear systems are like neural networks make sense is because if these use cases don't match, but just going for the deep neural network when you haven't checked to see do you have. Nonlinearity in your data set is the definite overkill.
Speaker 4:haven't checked to see do you have non-linearity in your data set is the definite overkill. Yeah, I think that's absolutely true and you know, use these more complex systems can sometimes help us catch some of these, more like what we call butterfly effects. I don't know how many you've seen the old ashton kutcher movie or just know the original short story that this comes from. But like the story goes that like a butterfly flaps its wings a thousand years ago and then a thousand years in the future that causes a tornado right that changes like the later conditions of the entire world. So it's this idea that small initial changes or inputs to a complex system can result in large and unpredictable or untractable, intractable changes in the output.
Speaker 3:Yeah, so here with our tornado example, it's a great example of chaotic theory applied to climate simulations, which are extremely necessary to consider when trying to predict the weather On a larger scale.
Speaker 3:If we have three bodies of roughly the same mass out in space, so you can think three stars in the system, they all influence each other gravitationally in a similar way.
Speaker 3:They'll start to pull at each other, but within a couple of steps or steps in time, you can think of it as maybe they're 10,000 years apart the influence that they have based on their position changes, and so it's incredibly difficult to predict where these stars are going to be after a certain amount of time and after even more time. It's impossible to predict where they're going to be based on just those initial conditions. It's impossible to predict where they're going to be based on just those initial conditions. Even if you do out the math, there's no formula that will tell you exactly where it's going to be without having to do that formula every single time, step reiterating over and over again to make sure that you have the exact, correct positions of those stars. The same way that with a weathering model, you kind of have to wait for the weather to get closer. You have to wait for that tornado to be closer to know exactly where it's going to be, and even later on, you won't be able to really know where it's going to be two weeks later.
Speaker 4:Yeah, I think that's absolutely applicable and I think that you know once again leads into what we're going to talk about, which is this idea of collective dynamics or hierarchy. When we're talking about these interactions between these agents, it's not like they happen like exactly linearly at all scales, at all levels. So, you know, we don't think about like a neural network just like moving through with unweighted nodes. Don't think about a neural network just moving through with unweighted nodes, and we don't think about the relationships between the first layer of nodes and the last layer of nodes as being a direct interaction. We have to think about these interactions on the scale that they really occur, which is, basically, who are your neighbors? Can you tell us a little bit about how things happen on different scales in like a simulation that, like you, don't just do like agent A to agent B across the universe?
Speaker 3:Absolutely. One principle of gravity in general is that everything with mass affects everything else in the universe that has mass. So, technically, you sitting in a room listening to this podcast, maybe're out on a walk you have mass. Technically, your mass affects the sun and the supermassive black hole at the center of the galaxy and the supermassive black hole at the center of the andromeda galaxy. Does it matter that much? I'm sorry it doesn't, but we have to consider what mass affects each other on the correct scales. So moons orbiting Jupiter will affect each other because that's the scale that they're on. It doesn't affect the sun, which is on a totally different scale and is affected by the galaxy, which is on a totally different scale.
Speaker 3:Our galaxy is affected by our other closest galaxy, andromeda. They affect each other, but you go further out. You go out to a galaxy that's, you know, hundreds of thousands or millions of light years away. They aren't really influenced by Andromeda and the Milky Way interacting with each other. And so we have to think about, especially with gravity and all other physics that we apply to our simulations hydrodynamics, magnetics everything we include is extremely at a scale. For the most part, when we have galaxy simulations, we don't even model individual stars. The stars that we have in our simulation are actually a thousand or a million or 10 million stars all clumped together, because it's only at the scale of tens or hundreds of thousands of stars does it start to interact and affect other things in the area, at the scale of galaxies.
Speaker 2:Yeah, so you're even going past, so you're doing more, almost like the mesolayer modeling would you call it, kind of the groupings of stars, if you will, instead of doing the agent-based layer of individual stars to be able to see kind of the adaptations and emergences.
Speaker 3:Right. So we have to make a lot of assumptions because we're smoothing over all of these stars. In a universe, stars will clump together in clusters. Our sun happens to not be in a cluster, but the majority of stars are in clusters, and so it's kind of easy to just say well, I'm not modeling a star, I'm modeling a stellar cluster, because the properties of the cluster are the things that are going to affect other clusters in the area, or they're going to make certain elements and then spew it into the surrounding gas. But the effects of one individual star are not as important on this scale, and it would be computationally expensive to even attempt to model every single individual star when we're looking at a galaxy, because again, it's billions of stars. And then we also have to think about all the gas, because the gas does behave differently than stars, and so gas particles we also treat as their own thing, and it is roughly a million suns worth of mass that the gas particle is.
Speaker 2:Very, very interesting and part of what we've talked about in this podcast before that really aligns with this is the same sort of technologies we've talked a lot about in computer science and statistics specifically, but the same sort of methodologies are developed in so many different areas.
Speaker 2:Like specifically where Rachel's talking about, is that in economics we have this type of thing too, or that's what we call it, the mesolayer, where you're essentially taking a zoom operation up on agents in an economy.
Speaker 2:It's the same sort of thing to see better those trends and because we can't computationally or practically model all of those areas and part of this with that adaptation just because I think in economic terms I'm just kind of rephrasing some of those in that way is very much like if the Federal Reserve wants to stimulate the demand say the economy's in a recession you're going to increase maybe the money supply so there's a little bit more money out there maybe cause a little bit of inflation, kind of try and stimulate that demand side Well, that feedback loop and adaptation will then affect asset prices and you'll start getting bubbles over there and you're always kind of like whack-a-moling in different places all the time.
Speaker 2:And it's that feedback loop and adaptation is what makes all these things so complex, and I'm sure, um rachel, you can probably speak to some of this, specifically some of those feedback loops. I'd love to hear some of the astrophysics, because this conversation completely new to me. I have a, you know, uh, whatever you learn in middle school and high school level of understanding of astrophysics, so this is very exciting.
Speaker 3:Absolutely so. The way that a group of stars evolves is they will burn hydrogen into helium. We say burn, it's nuclear fusion, crazy thing. And that's going on at an atomic level. We're not modeling the atoms that are taking place inside of stars, because those stars are essentially the size of atoms when we're starting to look at thousands of galaxies. So we have to do a lot of assumptions and smoothing over details about what goes on inside of individual stars and what goes on inside of each stellar cluster. But the stellar cluster will over time release compounds and elements back into space and this feeds into the surrounding gas, so that gas now has more developed elements.
Speaker 3:The universe starts out mostly hydrogen and some helium, but through stars and stellar production, stars are the foundries of all of the other elements, and so those other elements oxygen, carbon, nitrogen, all the ones we love are now being able to populate the surrounding gas. When that gas ends up making more stars, it collapses and condenses and then eventually makes their own stars. Those stars start off with much more oxygen, nitrogen and carbon than the previous generation, because they started off with the remnants from the previous stars. And so there's this feedback loop where the stars are going to release more and more of these elements because they started out with more of them, and so it's not extremely dramatic. The dominating element in the universe is still hydrogen, but we start to see more and more development of oxygen, nitrogen carbon and all of the rest of the periodic table.
Speaker 4:And I think this is going to bring us into what's probably the last complexity with complex systems, and I think this is the one that can feel to a lot of people like magic. Right, this is, this is the idea of like emergence that as you build the system with a bunch of simple, simple rules right, all you're doing is a bunch of matrix, multiplication and the neural network. But as you add these small components together, you then get properties in the system that can't be explained by those simple rules. So can you tell us a little bit about how, like taking basic axioms in math or physics, we put them into these simulations, but then we see behaviors that we didn't code into the system, right, the same way that in a neural network, we didn't code this to tell is this image a hot dog? We just taught it some math and some multiplication. So can you walk us through a little bit of that small steps resulting in results that we couldn't have predicted?
Speaker 3:So we put in information about gravity, about hydrodynamics, which is how hot things interact with each other, fluid dynamics. The gas essentially acts like a fluid. We code for dark matter, so we do account for that. It doesn't just come up from the system like magic, but what does end up happening is the galaxies form into disks, much like the disk galaxies we see in space. You see the spiral galaxies. They have their nice arms going around and they have a nice bulge in the middle.
Speaker 3:We did not code galaxies to form spirals and to have a bulge in the middle and to have a bulge in the middle.
Speaker 3:They naturally will do that when they form at the start of our simulation or the start of the universe. We also see that after these spiral galaxies interact with each other, maybe they smash into each other again, just due to gravity and hydrodynamics and the other types of physics that we implement into the simulation. We see that they make globular and kind of ovally galaxies. We also observe these types of galaxies in space. We didn't really know why certain galaxies were formed as disks and others were blobs, why certain galaxies were formed as disks and others were blobs. But through our simulations we can see, oh, when disk galaxies interact with each other, they can mess up the form that they have and end up making more of a blob. We didn't code those galaxies to make a blob at the end of their interaction or to make a disk at the start of that galaxy's lifetime. That naturally happens through the processes of physics that we put into the simulation.
Speaker 4:I think that's what makes this so exciting and interesting is that, you know, we build these simple systems with small little components and understandable agents and we get these really cool emergent behaviors. That either is modeling the real world or doing something new that we've never done before.
Speaker 2:For sure. It's definitely one of the most interesting areas of modeling and one of the ones I think doesn't get enough press or understanding by a broader community. It's like astrophysics they do. Are you using deep neural networks under the hood, or is it all just more physics-based functional programming with monte carlos or just at a high level kind of how would you go about creating a complex simulation?
Speaker 3:so some of these simulations are more simple, um, and they take a lot more assumptions. They do put in a lot of properties that would otherwise be emergent from the more complex simulations. The simple ones are probably going to be a team of five people coding up a storm for five years, and probably in C or God forbid Fortran they will code that up. They have a huge library that they put together accounting for all of the different physics. These were made before neural networks were really at all accessible, and so and astronomers are pretty slow on the uptake, I have to be honest we're probably not going to use neural networks to make simulations anytime soon. I have seen research done where a colleague of mine used machine learning to try and guess what the initial conditions were based on the results of the simulation, and that was pretty interesting, um, to try and guess what the initial cosmological, or like the initial cosmology setup was. We did that in order to then have that neural network point itself at real images of our own universe, to try and guess what the initial conditions of our universe are, because those are some things that we're still trying to figure out today. But as far as making the simulation itself, the larger ones take dozens of people across many universities, probably eight years, to put together, and then they'll have their papers publishing for the next two decades. That's what's been going on.
Speaker 3:I joined two different teams. I joined the Eagle Simulation at Leiden University and worked with their code for a year. I was mostly analyzing the results of the simulation instead of tweaking the simulation directly. But I was able to tweak the simulation directly with the Illustrious or Illustrious TNG simulation. Yes, tng is just the next generation. The astronomers love Star Trek a lot. I was able to work directly with the illustrious TNG simulation at the University of Florida, where I attempted to improve the resolution of gas particles in certain areas.
Speaker 3:That code itself is written in C and it's extremely complex and there is just an entire directory of directories, of directories of scripts that all write different aspects of the physics that are involved. It's very large as a code base and it does take quite a long time to include all of the physics and run the simulation On the supercomputers that are available at universities like the one at University of Florida. The longest simulation I ran, I believe, took five days on their supercomputers. So we're as efficient as necessary in order to get our results and publish. Our goal is never to be the most efficient, because we just want to be the most accurate. Astronomers are more focused on using the physics that astronomy as a whole grew up, on putting those into a computer to let the computer do the calculations step by step, because we can't, we can't just on paper do emergence and hierarchy and chaos and non-linearity, between hierarchy and chaos and non-linearity.
Speaker 1:We do very much rely on the supercomputers to do that, but we're not quite at the point of using machine learning to take it to the next step. Rachel, a lot of what you're saying is reminding me of two things, and Andrew and Sydney might have thoughts on this as well. One we had a discussion, probably back in December, with Michael Herman about consciousness, and some of the things that you're saying right now were reminding me of some things he was saying then. It's also bringing up some of the arguments of why the current AI systems are they ready to reason, and I think your explanation from a completely different context is super helpful for some of that discussion that's going on out there. Sid and Andrew, I don't know if you have anything any of that struck you at all.
Speaker 2:I think that, just as an industry, you know faster move, move, fast, break things. We need to deliver it in two weeks or whatever has taken hold a little too much of like I know you're, you're very, very open-minded to what data science and machine learning could potentially do in the astrophysics area. But I mean I there's probably some computational efficiencies and things that could be done, but as a whole I think that we have that tendency of that's always oh, we got to go to machine learning, we got to try and make it more efficient. It's like I would argue sometimes, for like this is fundamental research for understanding the galaxy. I wouldn't trust machine learning with that. Machine learning is kind of swaggy a lot of the time. Right Versus like this is literally some of the brightest minds in the world coding in exactly to our current level of understanding and then letting the complex system from the go from there, right. So I think that obviously not everything is near as complicated as astrophysics and understanding the galaxy. However, these types of methodologies and really understanding the principles or the the first principles and like what, how the mechanics of the underlying system work, and then going from there and exploring the use of these systems, a solid team could do something in six months. We're talking about insurance pricing or things like that. Like it doesn't have to be like. This is the one extreme way, more complex than any economy that Rachel's describing.
Speaker 2:Just because it takes that long, for this doesn't mean that complex systems and actually not trying to just do machine learning versus maybe it can be helpful from some of the discovery and fitting, but really understanding the system you're building and going from there. It's kind of like actuarial modeling and they might be able to do some more complex systems, but it's very much. Understand exactly what are those pricing factors and what are we exactly. How do we know what's driving it? Are we comfortable with it? That level is what statistics is as well Understanding those correlations, not just causations and oh sorry, causations, not just correlations.
Speaker 2:I had it backwards and then really understanding what you're building and then use a simulation from there or a model given on it's a models are just representations of a reality and we don't always want to be just turning our brain off and turning ChatGPT on instead. So I really like that. To me, studying astrophysics and how, like experts, really use complex systems is really great and I think honestly this community of machine learning has a lot more to learn from complex systems done in the real world than necessarily they have to learn from us. So this is very interesting and exciting to hear and in either case the cross pollination I think is huge from both sides.
Speaker 4:And you know, speaking of brains, there's some really great researchers out at MIT, CMU, tel Aviv that are trying to do a lot of this AGI stuff and this like brain modeling, the fundamentally, which is crazy. They're actually taking like a diamond knife and taking micron thin slices of brains from nematodes and from rats and building complex systems of how these brains work, neuron by neuron. They're doing the non-linearity, they're doing the hierarchy, they're doing the feedback loops, they're getting the emergent properties and this stuff. It's incredible and I'm not here to say that one is going to be better than the other, but there are researchers out there we don't hear a lot from them because they're not making chat, gpt who are trying to build general intelligence systems from a complex systems perspective interesting.
Speaker 2:I was not aware of that, that is. I mean, if we're going to get to general intelligence systems, I think it's going to be from that, that stream, not from predicting the next word uh, language systems, right. So that's, that's very interesting to hear. And yeah, there's a lot of really great research at like academia and and these other research, like santa fe institute for economics, complexity x and economics. They do a lot of work there. There's also a new england institute and then there's all of the experts around the world and doing astronomy and things. They're not loud with good marketing departments, but that's where a lot of the real innovation is happening. And I think it behooves the whole business industry and machine learning industry to be more in tune with some of these other discipline control theory, astrophysics, economics, all these other areas and trying to understand what are they doing and how are they really understanding their roots, versus just let's get a new, you don't understand it. Add another couple of depths in your deep neural network and now we'll get better. I don't think that's necessarily the solution.
Speaker 4:Totally, and these complex systems are everywhere in our lives, right? We might feel like, oh Totally, and these complex systems are everywhere in our lives, right? We might feel like, oh well, that's just something that scientists are doing. But I think data scientists and computer us that are in the reinforcement space, or even, just you know, boolean networks, which we learned about in intro to AI, classes At large enough scales, start acting like complex systems. These are absolutely around us all the time and we might feel like, oh well, I don't do that, but we often are engaging with these complex systems in a lot of the tools that we use.
Speaker 1:I would like to wrap it up for everybody. We've done a lot of talk. You know, Rachel, thank you for everything that you've discussed with us about these complex systems from an astrophysics perspective and, you know, Sid and Andrew really helping to bring light to that from a modeling perspective. Is there anything that you guys want to say that would really tie up some of the points you made here today?
Speaker 4:with a bow, I think we spent a lot of time today talking about the difficulties of working with complex systems, but I want to, you know, make the note that these complexities are part of what make these models so powerful and so interesting and so cool and so descriptive of the natural world that we live in, in a way that we shouldn't expect simple, linear models to work in. If we want to get these really great and deep understandings of the world that we live in, we should look to complex systems, because that's really how these things work in the real world, and so the model should match that model should match that.
Speaker 3:Yeah, I should emphasize the simulations that we have some of the best ones that we have out there. Yes, they're huge, they run on supercomputers and they don't use machine learning or neural networks. But if you look at the results that come out, we can make a picture from the simulation that looks identical to observations we make with Hubble and James Webb. It is eerily similar and we have plenty of other analysis to prove that our simulations are extremely accurate compared to what we see in space, which is just so amazing how we're able to mimic the entire universe with understanding and respecting all of the different components of complex systems.
Speaker 1:It is fascinating and you know I'm I'm eager to hear from you. Rachel, what did you think of your first podcast episode?
Speaker 3:This was a lot of fun. I hope that not too much of it went over people's heads. I'm very used to talking to other astrophysicists and I hope that we were all able to make some connections between our two different domains. I definitely learned a lot about how complex systems are used in data science and machine learning, and it's given me a lot to think about going forward as well, because I don't really take the time to think about how complex our simulations are.
Speaker 1:For everybody listening. I think you know this discussion also, I would dare to say, pairs well with some of the information we shared before. I mentioned the consciousness episode in our catalog. You're welcome to go check that out. In our podcast we also have some systems engineering episodes that I will link at the bottom of the show notes. That will help round out this discussion from a couple of different perspectives. This discussion from a couple different perspectives. Rachel, on behalf of myself, and Andrew and Syd, thanks for joining us today to guide us through this discussion and bring us a new perspective to help all of our modelers out there. We appreciate you.
Speaker 3:Yes, thank you so much for having me. Thank you.
