A portrait of eight billion people

Before Joon Sung Park https://www.joonsungpark.com/ and I had ever met, he simulated the conversation. What questions would Mario, or a person like Mario, ask him? What would I find most interesting to dig into? I do not know how closely I adhered to my facsimile, though I wish I did. Joon has an advantage in rehearsing these encounters, and thousands of others. He is the Co-founder and CEO of Simile, an applied AI company building a foundation model of human behavior. That involves creating digital twins of real people – like a writer or podcaster – and simulating how they might react to certain circumstances. As you might imagine, that’s an attractive proposition for many companies: CVS Health, for instance, can test a new pricing policy against thousands of simulated patients. A bank can model how customers will react before announcing a fee change. Joon’s ambitions stretch far beyond these cases, though, and Simile has raised more than $300 million, most recently at a $2 billion valuation, to pursue them. — Episode Transcript: https://www.generalist.com/i/222804125/transcript — Every timestamp 00:00 Intro 02:11 What Simile is 03:14 Learning to predict human behavior 04:33 From Schelling’s agent-based models to AI simulations 07:40 What drew Joon to simulations and how the idea took shape 10:32 Joon’s sci-fi inspirations 14:20 The ethical questions raised by simulation 15:34 How Joon got into art 17:58 Art that reveals deeper meaning 22:48 The artist’s framing and its role in simulation 27:04 A Simile customer journey 30:46 How Simile gathers data 37:55 Gaining trust 42:59 Public vs. hidden data and how Simile collects it 45:55 Use cases for simulation, from legal strategy to finding a partner 48:29 SimReddit 51:06 Using simulation internally at Simile 52:43 Surprising results from an EV simulation 54:27 Simile’s “siblings” 56:00 Simulation for investing 58:13 Winning in a simulated world 1:00:58 Joon’s hiring philosophy and his own contradictory superpowers 1:06:22 Final meditations — Resources Simile: https://www.simile.com https://www.simile.com/Schelling’s model of segregation: https://en.wikipedia.org/wiki/Schelling%27s_model_of_segregationWicked problem: https://en.wikipedia.org/wiki/Wicked_problemOn the Opportunities and Risks of Foundation Models: https://arxiv.org/abs/2108.07258Psychohistory: https://asimov.fandom.com/wiki/PsychohistoryGenerative Agents: Interactive Simulacra of Human Behavior: https://arxiv.org/abs/2304.03442Fresh Air: https://www.npr.org/programs/fresh-airWiz: https://www.wiz.io https://www.wiz.io/The Rehearsal on HBO Max: https://www.hbomax.com/show/609d8b4c-f0a6-4a5d-b9d3-bb0f2e207efbRemainder: https://www.amazon.com/Remainder-Tom-McCarthy/dp/0307278352The 36 Questions That Lead to Love: https://www.nytimes.com/2015/01/09/style/no-37-big-wedding-or-small.htmlSocial Simulacra: Creating Populated Prototypes for Social Computing Systems: https://dl.acm.org/doi/fullHtml/10.1145/3526113.3545616Moltbook: https://www.moltbook.com/Triptych: https://www.francis-bacon.com/artworks/paintings/triptych-1— Where to find Joon Sung Park LinkedIn: https://www.linkedin.com/in/joonspark Website: https://www.joonsungpark.com https://www.joonsungpark.com/ — Production and marketing by penname.co https://penname.co/. For inquiries about sponsoring the podcast, email [email protected].

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Mario Gabriele?

  If you can predict the future, then you could be the world's greatest hedge fund ever.

• 0:03Speaker 01

  So it's a really interesting question. Simulay is a applied AI lab that basically allows you to predict people's behavior at the level of individuals and populations. We can predict people's behavior with eighty-five percent plus accuracy.

• 0:16Mario Gabriele?

  And the ultimate goal is to simulate all eight billion humans and predict the future in some sense.

• 0:22Speaker 01

  Yes, it is about representing people in our society and also try to understand the interaction between them. A lot of the companies that we work with, they want to talk to their niche population all the time. But they can maybe, if they're lucky, get to talk to them once. But we create simulations of them once, then they can ask them any question. So this is actually the strength for simulation. So the problems that we would want to go after is really ask questions like, well, how to solve climate change? Can we actually find out the signals for a democracy that's about to collapse? So this is what social scientists would call wicked problems. I often, before interviews, simulate how the interview might go. What are some of the interesting questions that you might ask me today?

• 1:01Mario Gabriele?

  Did you simulate this?

• 1:02Speaker 01

  And.

• 1:11Mario Gabriele?

  A pharmacy chain wants to know why its patients stop taking their pills. A bank is about to change what it charges. A CEO is preparing to walk into an earnings call. None of them know what will happen, but Jun Sung Park just might. Jun is the co-founder and CEO of Simulay. In twenty twenty-three, as a Stanford PhD student, he published a paper that put twenty-five AI agents in a small virtual town and watched them throw a Valentine's party. Simulay is that idea grown up. Digital twins of real people built from hours of interviews that let companies like CVS test a decision on thousands of simulated customers before making a real one.

• 1:52Mario Gabriele?

  That pitch has helped Simulay raise more than three hundred million dollars and reach a two billion dollar valuation. In this episode, we discuss Jun's past as a painter, why the end market for a simulation company might be a hedge fund, and what happens to strategy once everyone can see the future. I'm Mario and this is The Generalist. Jun, I've been super excited about this conversation because I think you're building one of the most interesting companies I've ever imagined. So maybe we can just start very simply. Tell us about Simulay.

• 2:24Speaker 01

  So Simulay is a applied AI lab that is creating a foundation model of human behavior that basically allows you to predict people's behavior at the level of individuals and populations. We then have an amazing product team and engineering team that takes the core foundation model into a product that allows our users to ask any questions they might have, whether it's their customers, at-large institutions, or their environment, and so forth.

• 2:51Mario Gabriele?

  And the ultimate goal is to simulate all eight billion humans and predict the future in some sense.

• 2:58Speaker 01

  Yes, it is about representing people in our society and also try to understand the interaction between them. So when we come together as a group of people, what kind of interaction will we see? And what kind of causal mechanisms can we observe? Those are the core goals of Simulay.

• 3:15Mario Gabriele?

  The implication of of Simulay is, in a sense, that humans are extremely predictable, right? I'm interested in in whether you find that to be true in your own life. Do you find that the people you know are actually surprise you a lot of the time? Or actually, you you you think you could simulate them extremely well?

• 3:31Speaker 01

  What I do think is the case is I do think we can learn about people. What we've noticed over time is that people are made of fundamental building blocks. The core nature that they contain seem to be fairly consistent, so we can learn about them. So for instance, if I have a close friend that I've known since high school, I know what kind of food he likes. But I also know what excites him or what makes him sad. These are the fundamental natures. So one of the core thesis is at the level of individuals, we can pay attention to these kinds of signals that people might have and learn about what would excite them, what is the meaning they seek, and so forth. Now, an exciting part is then you bring in many of these people together in a room,

• 4:15Speaker 01

  and you then see an amazing emergent phenomenon as well. And that's actually much harder to predict in some ways. But this is also what gives us the ability to explore different causal mechanisms. So had we done this, if we had this policy, would it change the world in certain ways? Those are the kind of exploration that we would want to do in simulations.

• 4:34Mario Gabriele?

  When you talk about emergent phenomenons, what does that look like when you start to bring these people together?

• 4:38Speaker 01

  Emergence actually goes back all the way to complex systems and perhaps even the early days of simulation research. So in many ways, the simulation that we're creating with generative AI models is a revived form of the research that had gone in for many decades. But if you look back to nineteen seventies, one of the scholars that I respect the most is Thomas Schelling. He was a behavior economist who created really the field around what we called agent-based models. And he was really one of the first simulation researchers in computation. The model that he became known for was, of course, simulation of the era, really cares deeply about the societal problems of the era. In nineteen seventies, it was the racial segregation.

• 5:22Speaker 01

  The model he created was the model of segregation that basically made this point. Imagine for a moment that you have a checkerboard. On this checkerboard, there are red dots and blue dots. These are, quote-unquote, agents in the system. Now, these agents, because it was back in nineteen seventies, were extremely simple in their primitives. They had one rule, every game iteration, they would look around its corner and see how many of its neighbors are of the same color. If the ratio drops below certain point, then that agent would decide to move to a random location. The point that Thomas Schelling was trying to make was, in our society, we believe that racial segregation happens due to overt racism that people held within them.

• 6:07Speaker 01

  But that's actually not the case. People can have very subtle preference towards living with people of the same color. In no way extreme or overt, but that simple motion would still be enough to generate complete segregation of the society. So that's an instance of emergent phenomenon.

• 6:25Mario Gabriele?

  Yes.

• 6:25Speaker 01

  Where if you look at any individuals within the society, you would not imagine these people to be racist. However, the society stood trends towards being segregated. That was an unexpected finding. Now, of course, that was early days of simulation research. Thomas Schelling goes on to win an award prize. But now the opportunity that we see is, can we create simulation that is of much higher fidelity than what Thomas Schelling used to do? So of course, nineteen seventies, we had the limitation in models, so these were deterministic models. But now that we have generative AI model that actually uncodes a lot of human behaviors, instead of saying red dots and blue dots with these simple rules, can we actually put them in a much more realistic environment?

• 7:10Speaker 01

  So the kind of problems that we would want to go after in the most ambitious side of the scientific research is really ask questions like, well, tell me how to solve climate change. Climate change is a really interesting problem because it has so many actors with competing incentives. So this is what social scientists would call wicked problems. How can we resolve those tensions?

• 7:30Mario Gabriele?

  Yes.

• 7:31Speaker 01

  Can we actually find out the signals for a democracy that's about to collapse? It's these kind of problems that we would be investigating with the next generation of simulation technology.

• 7:42Mario Gabriele?

  When did you first become acquainted with Schelling and simulation work? Like, when did you fall down that rabbit hole?

• 7:48Speaker 01

  So I actually grew up as a big science fiction fan. So, you know, growing up, I watched, read a lot of science fiction. And one of the core things I always resonated with was this idea that in any truly advanced civilization that gets depicted, you really see two pillars of scientific progress. One is some form of AGI and the other is simulation. And I always thought that was super fascinating. And for whatever reason, I was always more drawn to this idea of simulation. And I also think a little bit more philosophically speaking. So I initially got my training as a painter in art.

• 8:26Mario Gabriele?

  Yeah.

• 8:26Speaker 01

  And I think it was perhaps a little bit related to that passion as well. But simulation went on right. It's a little bit like painting a portrait of the society.

• 8:35Mario Gabriele?

  Yes.

• 8:36Speaker 01

  So that we can learn about the society and so forth. So I always had a lot of interest in the topic of simulation. But then during college years, I got acquainted with the topic around agent-based models. So that's when I initially started to learn about this topic and did a very brief stint as a RA in a lab that tried to go deeper into the agent-based model field. And then I sort of forgot about it for a good number of years. And when I started my PhD at Stanford, so this was twenty twenty, GPT-3 was just about to come out. And there was a lot of chatter around the scientific field that this new class of model is going to be a powerful new mechanism for understanding people and also for AI.

• 9:21Speaker 01

  So during that first year, myself and two of my co-founders, Michael Bernstein and Percy Liang, wrote a paper called Opportunities and Risks of Foundation Model. So this was in particular a project that Percy led at Stanford, that coined the term "foundation model." But as we were working on this project, we started to really think about, well, imagine for a moment the foundation model and these generative AI models will be an important part of our technological landscape. What's going to be the killer application of the technology? And we realized that it might be simulation. Because these models have ingested so much of human behavioral data that we noticed that if you poke at models at the right angle, we can actually extract a lot of human behavior.

• 10:05Speaker 01

  In which case, this can actually be the fundamental building block that we had missed in the past for a building simulation. So that, when we had that realization, I sort of connected that with the agent-based models that I used to be fascinated by and said, can we create the next version, the next generation of agent-based models? And maybe even call it generative agent-based models? And that really became sort of the starting point for thinking about this topic and area.

• 10:33Mario Gabriele?

  You mentioned the inspiration from sci-fi. When I, you know, was studying simile, the sort of direct comparison for me feels like Psycho History and Asimov. What were the inspirations for you when you think about what you were reading as a kid or, you know, that you reflect on now?

• 10:49Speaker 01

  Yeah, so Foundation Series is actually a fantastic example here. What really inspires me from the science fiction is not only the existence of this kind of technology, but also shape of this technology. So if you look at simulation technology today, at least the initial building blocks are basically technology that allows you to simulate the results of studies or policies.

• 11:13Mario Gabriele?

  Yes.

• 11:13Speaker 01

  So imagine that we have a behavior study that we want to run, market research we want to run. These are some initial sort of good use cases that people start with on day one. They would come in with a question. But what is interesting is the arc of this technology is much more proactive. And I can describe what that means from the lens of Asimov. So if you look at the Foundation Series, here's a premise for those who may not be familiar with this story. So there's a galactic empire. It's, I was, you know, much more into the future than where we are right now. And a group of scientists realized that this galactic empire is going to collapse in thousand years. And they are basically tasked with figuring out

• 11:59Speaker 01

  what is the way to reduce the following unrest after the collapse of this empire? And they realized that there is going to be 30,000 years of unrest. And they asked, how can we reduce it to thousand years? So they asked their simulator basically this question.

• 12:15Mario Gabriele?

  Yes.

• 12:16Speaker 01

  What is the best way for us to ensure that the social unrest is kept at a thousand years? And if you look at the shape of this technology, the Psycho History, the simulator that they're thinking about, what they're giving this tool is an end state. They're not giving them a question, they're giving them a goal. Here is the end state that we want to get to.

• 12:36Mario Gabriele?

  Yes.

• 12:37Speaker 01

  It's up to the simulator to create the multiverse of possibility and basically suggest, maybe this one, in maybe using the term from Dune, the golden path for getting to that particular end outcome. This is the beauty of simulation. Because this is possible for a simulation because we care deeply about multiverse, creating these different possibilities, exploring into the future, and also understanding the cause and mechanism. Had we done this, how would that have influenced the people's behavior, future? The one last thing I say to this point is the beauty of this way of exploring is sometimes the first steps or even some of the steps that we'll have to take to get to that end outcome that we want can be surprising.

• 13:22Speaker 01

  If you look at this series, they realized that the right first step to mitigate the period of unrest is basically to send the scientist who's been saying that there is going to be a collapse of galactic empire into exile. The most counterintuitive solution, you are about to send the people who are experts in this away to some remote part of this galactic empire, and that's going to be the solve. No one would imagine that. But turns out, at least in this series, of course, it's a science fiction, but at least in this series, the author explores whether that actually is the right first step. And that is the power of simulation at its height.

• 14:00Mario Gabriele?

  I love that example because it also sort of speaks to something we're all grappling with in AI at the moment, which is explainability. There are moves that you can, you know, an AI simulator can make in a chess match that the best players in the world might not see or even might not grok why that's the right choice. And, you know, it's the same sort of idea here around exile. One of the axioms for Psycho History is this idea that you shouldn't reveal what the prediction says. Otherwise, it sort of influences the actions of the population. Do you think that applies for simile at scale? You know, it doesn't maybe matter right now when you're sort of taking small populations and asking, you know, for preference questions, for example. But do you think that that same principle ends up holding?

• 14:45Speaker 01

  I do think this is definitely a part of questions that we do need to wrestle with as similarly. A slightly related and also interesting version of this is, you mentioned down the line, many different parties would have powerful simulator. How do they interact? And if they're based on the same model, how do those interact? Could be the case that in the future, people will want to build their own specialized simulators.

• 15:12Mario Gabriele?

  Yes.

• 15:12Speaker 01

  That is for their own institution, and that is going to be their competitive advantage. That, I actually do think is the case.

• 15:19Mario Gabriele?

  Yes.

• 15:19Speaker 01

  So one of the things that of course simile serves today is building actually more specialized, fine-tuned version of simulator for different institutions that we work with. But I think these are all part of these questions of what happens if people know the results of simulation and vice versa.

• 15:35Mario Gabriele?

  You mentioned that, you know, you trained as a painter and for people who maybe hear that and think, you know, you like to paint. You're like an extraordinary artist, actually. When I was looking at your work on your website, I was like, oh, this is, this is at a very, very high level. You know, this is someone who has clearly devoted a huge amount of time to this and has, you know, remarkable ability. When did that arise for you? Was that something you sort of naturally gravitated towards? Was it something that, I don't know, one of your parents sort of guided you towards?

• 16:07Speaker 01

  So I did always like craft. The act of creating something that people can interact with is something that I was always drawn to. And the way I actually though got to art is actually a little bit more coincidental. So it turns out, when I was in high school, I lived in a dorm. And this dorm had a very interesting doorway, where there was a very quick turn. And one night, when everyone was asleep, I, for whatever reason, thought it would be really funny and just, you know, light prank, that if I can draw really realistic pair of eyes on the wall. And basically, as people go up, they see it and they get slight jump scare and will have good laugh about it. I didn't know how to draw.

• 16:52Speaker 01

  So what I did was I actually that night decided to watch a good number of tutorials on how to draw really realistic pair of eyes from YouTube. And I spent three, four hours and from midnight to maybe 4:00 AM drawing this on the wall. Next day, I would basically hide in the corner and just watch my friend come up to it and interact with it. And I thought that was really fun. I personally got something out of that experience that I have created something that people are reacting to. And in some ways, they're also finding some kind of value and meaning out of it. That particular interaction with the society, with the community, I actually thought was quite compelling to me. And what ended up happening was one of the teachers

• 17:37Speaker 01

  at the school who was one of the proctor at the dorm, saw that. And first he was like, okay, well, you're not supposed to do this. This is graffiti. But then he said, he was an art teacher. Then he said, now you have to come to my class. And we gradually realized that this can actually be a professional career. So that's how I started this journey.

• 17:59Mario Gabriele?

  Wow, that is such a great story. The sort of style that you studied under, as far as I understood, was sort of like, you studied under this very well-known, hyper-realistic painter, which, you know, it's so real. It, you know, you can't imagine how real it looks in a way, but also has this sort of, in some places, sort of surreal elements. You mentioned like there is this sort of connection between simulation and the art. It seems like very clear that you love this process of making something so realistic that you almost can't tell the original from the synthetic. Where does that come from? Like, where do you think that interest stems from in you? Like, what intrigues you about that?

• 18:38Speaker 01

  It's interesting. Of course, my scientific training has really gotten done now in AI and computer science. But for a computer scientist, I am actually somebody who's deeply interested in people. I am intrigued by what motivates them. I'm intrigued by what makes them feel certain emotions. I'm also intrigued by what are the elements that empowers them? And one thing that I realized about the greatest painters, and I worked a lot with figures, and a lot of the best portrait artists, what's unique about them is their paintings are not merely realistic,

• 19:22Speaker 01

  but it says something about the subject. It teaches you something about the subject, something that's not apparent, but something that's hidden away, that gets conveyed in a, at a more emotional level. Those are the portrait artists, those are the painters that I respected the most. So in many ways, simulation, when done right, actually is that. Simulation isn't, of course, reality. It is all within computation. And it is, to a large extent, an approximation of the world that we live in. But when done right, the core ethos here is that we can reveal something that is fundamental about our individuals, our society and community, so that when we look at it,

• 20:06Speaker 01

  we can learn something about ourselves. And that, in the end, can help us teach what is the best way to collectively live in this world.

• 20:16Mario Gabriele?

  When you think about a painting, a portrait, that Does evoke something beyond just, you know, the fidelity of the picture? What comes to mind?

• 20:27Speaker 01

  So when I was a painter, some of the folks that I was heavily inspired by are artists like Gero Richter. So he's a German painter, has had a very interesting career where he originally started as a hyperrealist painter. But he had a very interesting technique where he would take actually photographs. So he, so he's now, he was active throughout his career and his still alive. But he was active for many decades. And when he was making his early career, the idea of photograph and photograph that were getting printed on newspapers and so forth, was an active area of interest for artists. But he had a very interesting tape, which was he would ask his viewers or his friends

• 21:12Speaker 01

  to send him photograph they liked. It can be literally cut out from newspaper. It can be personal photographs that people have taken. And then he would, in a really hyper realistic way, paint that. But he would fade it out in a few different ways. He would fade out the color. So a lot of it actually was black and, uh, black and white, or it had gray gradation, but it didn't have color. And he also did this exercise where after he makes this wonderfully accurate painting, he would smudge it in one direction. It's almost like, you know, something went wrong during the printing process and it got smudged.

• 21:53Speaker 01

  And I always interpreted that as sort of this idea that we want to capture something that's real, that's now as accurately as possible. But as the moment goes on, things are fleeting and certain emotions and memory does split away. And I thought that was really interesting observation where nothing that we capture is permanent, that things are changing. So what are the core essence of that individual that we can really capture? And in many ways, our memories are not perfect. What is the core emotion that you actually still remember after things have faded away? I thought that exploration was actually quite compelling. So a lot of my work actually got inspired by some of the work that Gero Richter has done.

• 22:39Speaker 01

  But of course, I did get a lot of, a lot of influence from American realists who were working on the East Coast. That's where I initially got my training. So it was a mix of those inspirations.

• 22:50Mario Gabriele?

  That Richter anecdote is so interesting, I think, also because it makes me think about the role of the artist in juxtaposition to a machine, which is obviously something we're all wrestling with more and more now, which is, you know, the machine can capture the fidelity even, you know, better than any human can in some sense. The, the role of the artist is maybe not to smudge, but to show that there's a human element also there. You know, I think about probably because of my background with that in writing, right? Right, right now, it's very clear AI is not as good as the best human writers, but that, that will almost certainly happen, right? That seems very possible. And so, you know, what, what are the smudges or, you know, the, the little notes of transience that show up there?

• 23:34Mario Gabriele?

  I imagine, yeah, you, you want sort of the poet's mistake to be in there almost.

• 23:38Speaker 01

  To this point though, in many ways, also the smudge and the artist's role in this ecosystem also comes from the framing. What the artist decide to care about? What they decide to put in their frame, I think matters a tremendous amount.

• 23:52Mario Gabriele?

  Yes, intent.

• 23:54Speaker 01

  Intent, intentionality in art is extremely important.

• 23:57Mario Gabriele?

  Yes.

• 23:57Speaker 01

  Why are you interested in this? Can you express this beyond just the painting itself? This is actually a huge part of the artist's job. As, as I'm sure you are more than familiar, because this is clearly in your sort of straight lean in your profession, finding story that's worth telling, finding elements and phenomena that's worth studying. That's the job of humans.

• 24:18Mario Gabriele?

  Yes.

• 24:19Speaker 01

  Because at the end of the day, we do this because we care about them.

• 24:21Mario Gabriele?

  Yes.

• 24:22Speaker 01

  So as an artist, the kind of things that I would do is before I actually embark on the process of creating paintings, a lot of the work that actually goes into the process is actually building up that emotional rigor in caring about a subject in the first place, that I would actually want to devote my time into actually working on the subject. As a painter, I used to work on a really large scale paintings. For large scale or paintings, each painting could take up to half a year to a year.

• 24:48Mario Gabriele?

  Wow.

• 24:49Speaker 01

  So it's a commitment.

• 24:50Mario Gabriele?

  Wow.

• 24:51Speaker 01

  So when I felt like I was the most successful, there was this deep emotional connection with the thing that I was painting. It was maybe a location that I particularly fell in love with when I was traveling. Or it's a very political subject in everyday life that I found deep interest in, that I decided it's actually worth highlighting in this world. That observation, that framing does come from humans. This is also the case in simulation, because it's at the end of the day, we can simulate anything, anyone.

• 25:23Mario Gabriele?

  Yeah.

• 25:23Speaker 01

  But it is up to the human observers to decide who do we need to simulate? What phenomenon do we need to explore? And what are the solution spaces that are the right thing for us to think about? That's all human researchers.

• 25:37Mario Gabriele?

  Emotional rigor is a phrase I have not heard before, but that, that feels very profound, actually. I love that phrasing. I think you've said before that, you know, in some sense, every portrait is also a portrait of the artist. Does that mean that every simulation becomes in some way a simulation of yourself? Do you, do you recognize that, you know, your hand in these different mini people?

• 26:03Speaker 01

  One of the sort of strange things I realized was portrait artists especially have this secret where no matter who they paint, even if you're a realist painter, the end outcome sort of looks, or at least resembles something about the painter themselves, not their subject. It's one of the sort of secrets of the painting, uh, paintings community. I think this is the case in many ways of the simulation that we create. But it's a, it's much more collective, right? Because when we create simulation, it's not purely about creating these synthetic personas. We actually go talk to real people.

• 26:35Mario Gabriele?

  Yes.

• 26:36Speaker 01

  We get their consent, we get their buy-in. They know how the simulation is going to be leveraged. But it's really a collective process of organizing a large group of people who find this act to be interesting and potentially worthwhile. And when simulation is done right, we represent them. That there is something in this simulation, even though we're trying to simulate many people, there's something about this simulation that does represent or reflect the core individuals that we have talked to in this process. Very much the case.

• 27:05Mario Gabriele?

  This is a good reminder for me that, uh, you know, when I was talking to some folks in preparation for this podcast, one of the, the things they said was, don't forget to really ask about the business, because the philosophy is so interesting here that it can be easy to forget that, you know, there is something that is working extremely well here. And, uh, you know, it seems like you have been through some very rapid growth. So maybe to get more granular of how this sort of tangibly works for folks, maybe you can walk us through a bit of a customer journey. When someone comes to Simile, what do they come for and what does that process look like?

• 27:39Speaker 01

  So in many ways, Simile is a company that is providing simulated panel of people. Today, the day one, the act or the user flow is actually a lot like working with a real panel company. So imagine a customer comes in and they tell us XYZ are the population that we want to deeply understand. What Simile then does is we create the simulation of that population. We load them onto a platform and basically, it's basically a model layer that we generate. And then we make that expose via product to our customers. They can basically query this model any number of times. And the input to this, uh, to this model basically is two things.

• 28:21Speaker 01

  Who, within the population that we have determined, do you want to talk to? So you basically set the filters. And then you give it an environment. The environment can be something as simple as a survey question or behavioral experiment. But then it can be something as complicated or as more involved as things like a real product. Imagine you want these agents to come and use your product for 10, 15 minutes and tell you what they think about it. Or even policies. Imagine we have this new strategy, how will people perceive this? How will they react? And really the arc of simulation, I see this in four stages. It goes from individual level static simulation. This is akin to running surveys, behavior experiments.

• 29:04Speaker 01

  You're asking individuals what they think about X right now. But this then goes into more individual agent in a dynamic setup, where they cannot explore some things like product and so forth in some, you're basically adding the temporal dimension. But then this goes into that multi-agent simulation and even societal simulation. At a multi-agent simulation level, we sometimes simulate your earnings call, often before interviews. And I also simulate how the interview might go. What are some of the interesting questions that, uh, that you might ask me today? And just for me to prepare myself a little bit in terms of what the conversation might look like. And of course, down the line, here we're embedding these multi-agent simulation in a rich environment.

• 29:49Speaker 01

  But if you want to actually simulate the entire market, that is ultimately where this technology goes. Our customers today have access to the first three layers of these simulations.

• 29:58Mario Gabriele?

  Did you simulate this?

• 30:00Speaker 01

  Uh, I did ask our simulator to give me a very brief, uh, preview of what this conversation might look like.

• 30:05Mario Gabriele?

  Was there any question that you were like, huh, I would not have thought to prepare for that?

• 30:09Speaker 01

  It did suggest that we might actually go a little bit deeper into the philosophy and artistic side of the story. And I, I do think it likely has. And I do give as input, uh, the background of our, uh, the person I'd be talking to. And of course, you know, I gave the information that you're a writer, uh, who has worked on these different pieces.

• 30:33Mario Gabriele?

  Yeah.

• 30:33Speaker 01

  And I think it sort of made a connection.

• 30:35Mario Gabriele?

  So there's a mini sort of Mario simulator in there saying, this guy, he's, he's a writer. He's going to ask you about Richter and all these sort of things. That's so fun. Wow. Okay. So interesting. To, to, to take it back to sort of the, the customer side of things, I think the part that maybe people might not realize is that you're actually going out and, and talking to these people in the population, right? You're sort of, you've automated the collection, just like, you know, another panel company might do. And, you know, it, it's really guided by, yeah, real people.

• 31:06Speaker 01

  Yes.

• 31:07Mario Gabriele?

  You know, an example is I'm making this up, but a juice company wants to target a new demographic. They want to, you know, they want to win the 20-year-old audience. Let's say the young, the younger generation. You would then, in an automated fashion, go out, talk to 5,000 people in that demographic, gather that information and then feed that into the model such that the juice company can then go to the model and say, okay, you know, does this marketing campaign work better than this one? Does this pricing strategy work better than this one? Is that sort of roughly the right idea?

• 31:42Speaker 01

  That's roughly the right mental model. But what ends up happening is so across our partnerships with different panel providers, we interestingly have a very fruitful strategic partnership also with Gallup, which is one of the most respected panel companies out there. So through these partnerships, we have access to about 30 to 40 million real human subjects across the globe. On any week, we are talking to tens or hundreds of thousands of real humans.

• 32:12Mario Gabriele?

  Wow.

• 32:13Speaker 01

  Now, the questions we ask them, however, isn't quite so narrow or it's not perhaps even directly related to the question that we want to answer in research. Because in many ways, we want to expose this model to our users who will then ask any questions they want to ask. So in many ways, the goal here is to learn something that's much more fundamental about the people we talk to. For instance, we actually ask people, hey, tell me the story of your life.

• 32:40Mario Gabriele?

  Yes.

• 32:41Speaker 01

  Where did you grow up to? What do you care about? And then we would also go a little bit deeper into certain verticals if we know that we will be serving that vertical. So imagine if it's a CPG company or imagine it's a food and beverage company. We might ask, hey, historically, what kind of beverages did you like? Do you drink? Do you not? We will ask those kinds of questions and we'll also ask you about your experience around that question. What was, when was the time you enjoyed the most when you were, let's say, drinking or eating up with your friend? Those kinds of things. So we want to get a more fundamental understanding of who this person is. And that's ultimately what power simulation.

• 33:21Mario Gabriele?

  How many hours do you need to get to know someone?

• 33:24Speaker 01

  What initially became the foundation for a simile, although the technology actually has moved significantly away from the original papers, was back in 2023, we had a paper called Generative Agents that introduced a small game town with 25 agents living in it. These were basically agents or NPCs in this small game town that were made with GPT 3.5 back in the day, paired with memory planning and reflection architecture. These models were interesting because they were one of the first introduction to the idea of creating agents using generative AI models. But then that's the one that really kickstarted this simulation. But the follow-up to that paper for me was how can we create these agents to be more realistic?

• 34:08Speaker 01

  And the way we went about doing that was actually collecting a lot of interview data. So I would actually recruit about 1,000 or so real human subjects from the US. That's meant to be diverse as our first initial pilot study. Bring them to a lab and we basically did two hours of interview and behavioral data collection. And then we use that to train individual agents that basically became the digital twin of the source individuals. I would then send the people away for a couple of weeks and then bring them back. And when they're back, I would ask them to complete behavior studies, surveys, experiments. And that was the paper that for the first time showed that we can actually predict using these digital twins we created.

• 34:52Speaker 01

  We can predict people's behavior with 85% plus accuracy. Long answer, but here the idea was two hours. Two hour can be very effective. Of course, one of the things that innovation, the simile works on is how can we make this be even more efficient? So today we roughly do about 15 minutes to half an hour of data collection per person. And really the beauty here is actually knowing what questions matter the most.

• 35:17Mario Gabriele?

  Yes.

• 35:18Speaker 01

  And then really fine tuning our process to make sure that we ask the fewest number of questions possible, but still get the most generalized understanding of these people.

• 35:26Mario Gabriele?

  That is so interesting because you're, you're really in a way, yeah, you're doing the work of an interviewer. You know, I was, I was, I've been spending time thinking, how do the best people ask questions? You know, and studying different interviewers. One of the people I was looking at was Terry Gross from Fresh Air. You know, she does these very warm, but also very interesting interviews. And your first question of tell us the story of your life is almost exactly the, the question that she often starts with or recommends as an icebreaker of just like, tell me about yourself. Somehow these are just simple things that actually give people the, the space to tell, tell you a lot.

• 36:05Speaker 01

  It is interesting. Where we got this inspiration for actually running interviews with people actually was through podcasts. So after the original Generative Agents paper back in 2023, I had to go on actually on a lot of podcasts. So I would actually talk to these interviewers. And after an hour, I realized actually these people know a lot about me. This is actually a very effective way of learning about people. And that's what inspired those to say, well, can we want to learn about people? Then let's talk to them.

• 36:36Mario Gabriele?

  How do you deal with situations where the population that you're after is not readily available? Like if you are working with a customer that wants to, you know, know if something will appeal to Fortune 500 CEOs. That's a group that is maybe hard to, to reach.

• 36:50Speaker 01

  So right now, of course, there are populations that are truly hard to reach. And this is where we are actually curating our own bespoke panels of experts. So this is a panel that we actually curate at Simile and form special relationships with individuals. But the ethos here actually is we have large enough a panel set that we can get to individuals that we want to get to most of times. Now, the tricky part, of course, is how frequently can we get to them? You can actually find a way to talk to, let's say, very niche population who have very limited amount of time available to them once. That can happen.

• 37:31Mario Gabriele?

  Yeah.

• 37:32Speaker 01

  But the question is, can you do that frequently? And the, and one of the value proposition for a simulation actually is this is where simulation is actually the most useful in many ways. Because a lot of the companies that we work with, they want to talk to their niche population all the time. But they can maybe if they're lucky, get to talk to them once. But we create simulations of them once, then they can ask them any question. So this is actually the strength for simulation.

• 37:56Mario Gabriele?

  I wonder if this is true, but one of the tricky things I could imagine about the product you're selling is that it might be hard for people to trust the result that you give them. Is that true? How do you sort of convince people that, hey, you know, you, you can act based on the information we're giving you? And, you know, you're, you're probably going to make a good decision here.

• 38:16Speaker 01

  Right. I would actually say trust is one of the most important piece of this industry and this field. It's actually a little bit like how the security industry actually operated too. In security industry, if you look at companies like Wiz, you're fundamentally trying to prevent things that has not happened yet.

• 38:34Mario Gabriele?

  Yes.

• 38:35Speaker 01

  So people have to believe that you are good at doing that to really be a part of the platform. The same thing for simulation. We are fundamentally trying to simulate things that has not happened yet. So what is the way for us to bring them along in the journey? And what we've learned is a couple of things. It's not just about accuracy. How the accuracy is of course the, the table stake. And we're very upfront about here's the validation we run. Here's the benchmark. This is how what we do compared to all the other models. That's how we start the conversation. But what matters perhaps more is people forming a very clear mental model of how this technology works, what data goes into this technology. So even if they do not necessarily know whether this works,

• 39:21Speaker 01

  they, when they put together this mental model of how this works, they can sort of imagine, ah, that does make sense. Of course, that would work.

• 39:28Mario Gabriele?

  Yes.

• 39:28Speaker 01

  So if you think about, let's say consulting, it's very much like they operate very much in this way. Consultants come and they give you an advice that XYZ will happen. Therefore, you should act in these ways. The users do not necessarily think, okay, they predict the future. If they do, they would obviously be in different business. But what they do believe is while these consultants put together a very small group of people have done highly reasonable thing. So that to the extent that any given group of people can have a good insight into what we should do, this is the best evidence we have. And that's how we start the conversation. So for similarly, what does that mean? It means being very transparent about the sources of data. Here is the set of data that we care about.

• 40:12Speaker 01

  This is how we look at curating the right data set. It includes behavior. It includes counterfactual behavioral data. It also includes the interview, like the one that I just mentioned. And then we take this data, process it and put it into a model that can basically reason about this data and provide citations. When we create simulation, it's not just about us creating random predictions, but we actually say real individual set XYZ things. On the basis of this evidence, we believe this is a good simulation. And then the last piece that's also very important is we provide a model of predicted accuracy. But this is a model that we train internally and has been a very exciting project.

• 40:55Speaker 01

  We recently published a blog post that's more technical on this topic. But this is a model that is jointly trained, that is basically looking at the hidden layers of the core foundation model that we train. That is trying to understand, given the data and the question that we have gotten, how confident is our model? And we show very strong correlation between the actual accuracy and the predicted accuracy. This can help people collaborate how much they should rely on this particular prediction.

• 41:23Mario Gabriele?

  Have you ever tried, or is it possible to sort of back test these predictions? Like, you know, could you go back and say, would this population like the first Tesla? And then would this population like the first Cybertruck? And you can sort of say, well, actually, you know, people were very polarized about this, but they really loved this. And that actually, you know, was the case.

• 41:43Speaker 01

  So this is, this happens for almost every deal. So what we do is of course, our customers have run a lot of studies internally. They know the ground truth of the world they live in. So during the first conversation, they'll actually say, hey, here's a study we ran. Can we try this in Simulay and see if we get the same results?

• 42:04Mario Gabriele?

  And so then you can say, okay, it's it's of equivalent accuracy to a traditional panel. But by the way, we can also query this endlessly in a thousand different directions. To what extent is external information sufficiently useful to create a simulation of someone? Like, you know, you, if you were never interviewed by Simulay, would the data that you've shared on a podcast be enough to simulate you well?

• 42:30Speaker 01

  Right. It can be a very good starting point for simulating that person, right? So a lot of writers, a lot of public figures, we can actually get enough data to at least create a persona that does resemble them. But also in many ways, the simulation that we want to create is fundamentally more behavioral. It's not just what people say in public, but it's actually what they do when they're not seen by anyone. That data is often hidden away. So the external data is useful for a step, but it's imperfect.

• 43:00Mario Gabriele?

  How do you get the hidden away data?

• 43:02Speaker 01

  So in few different ways, for people, we actually do collect a lot of behavioral data. And the behavioral data, for instance, that we might collect are things like a transaction data. That's a good starting point. Now, where a transaction data does fall short, however, is it is fundamentally observational data. It tells you what people have purchased. It does not tell you why they have purchased. So if you want to create a counterfactual, imagine this person who used to go to Starbucks every single day, chose not to go to Starbucks today. Why did that person not go? Is it because they got sick of sweet drinks that they were getting? Or was there a car accident? It's very hard to know.

• 43:45Speaker 01

  So if you want to know the counterfactuals, then you actually do need to set up either interview or experiments that does occur naturally. So we do a lot of those activities where what are the natural experiments that we can set up to deeply understand people's reasoning? Or we talk to people. So we do combine this with behavioral data, with interview data. Another thing that we do, of course, is we run randomized controlled trials internally. So this one is with obviously people who have consented to partake in the studies with full understanding of the study's context. But we basically expose people to different conditions of options. And we make the stakes real. Imagine we have a web store that they're going to come and hypothetically buy something.

• 44:29Speaker 01

  At the end of the study, we'll actually send them the item. So the study is real.

• 44:33Mario Gabriele?

  You must have seen the rehearsal. That sounds so similar to something that Nathan Fielder would want to do. How do you, how do you get Nathan Fielder to make season three on Simulay? Because that feels like that would be the most interesting possible version of these, these ideas. Did you get ideas from, from that show or inspiration in some ways?

• 44:54Speaker 01

  You know, we get inspirations from all sources. In fact, after we had written the original generative agents paper, my co-founder, Michael, wrote a follow-up paper. Its title actually was Rehearsal.

• 45:09Mario Gabriele?

  No way.

• 45:10Speaker 01

  It was very much a homage or inspirations from that show.

• 45:17Mario Gabriele?

  There's another book that I always think of when I think about things in this space called Reminders. Have you ever read that?

• 45:22Speaker 01

  Not yet.

• 45:23Mario Gabriele?

  Tom McCarthy, I think is the author's name. And, you know, the sort of premise is a man goes through, I think, a car accident where he gets massive amounts of money as compensation, but has lost huge amounts of his memory. And so he sort of painstakingly recreates in huge detail little things that he can remember or wants to try and recall. And, you know, the set for it becomes sort of increasingly grand, you know, his ambitions grow. Anyway, it's another one that's sort of in that rough zone that is very interesting. When I think about the sort of second, third X number of acts for this company, you can let your imagination spin almost infinitely.

• 46:04Mario Gabriele?

  One of them that I could imagine is, you know, I, in my early years, I wanted to be a lawyer and I did a lot of public speaking and mock trial. I imagine jury selection, for example, would be something that is like extremely applicable here. Have you started to see that?

• 46:19Speaker 01

  We have. In the legal spaces, policy spaces, I think the role the simulation can play is actually quite deep. But this is something that we are obviously quite inspired by, and we do see that opportunity.

• 46:35Mario Gabriele?

  What about in the personal space? Because when I think about the most impactful decisions in your life, picking a partner is one of them. Is there, is there a world in which Simulay becomes sufficiently well, you know, widely spread that people start to simulate potential matches?

• 46:51Speaker 01

  So I think it's possible. Then now what is interesting about this technology is simulation is one of those very few technologies that's very applicable for any groups of people. So right now, of course, we are partnering with enterprises to bring simulation, this innovation to their organizations. But this is something that individuals can also leverage. The use cases they might see are everywhere from, what is the better way for me to, let's say, get better at public speaking? That's a common use case that I've seen. Imagine you have a very difficult conversation that you need to have with your friend or with your colleagues. Is there a way to rehearse the situation so that you're prepared for those conversations?

• 47:35Speaker 01

  So I think these are the places that the simulation could be leveraged. Of course, for the more sort of these dating and romantic side of things, I think this is an obvious use cases that some people might be interested in. Obviously, I personally happen to be a little bit more of a romantic in some ways. I let certain part of my life just fall in naturally. So I think it does depend a little bit on who you are.

• 47:58Mario Gabriele?

  You know, I don't know if you've ever come across it, but there's, I think, a, I think it's the New York Times that has something called like the 21 questions that lead to love. And there are these very good questions that are really thoughtful and do create a way for people to open up. I could imagine someone doing a vast survey of people of those 21 questions and using, you know, that to simulate something pretty interesting. Obviously, that does not feel particularly on strategy for you to do, but it's interesting that the technology is getting to the point where, yeah, you could, you could imagine something happening there.

• 48:29Speaker 01

  I agree.

• 48:30Mario Gabriele?

  One of your earlier projects was called SimReddit, which I thought was really interesting. Maybe you can actually talk a little bit about that.

• 48:36Speaker 01

  So this actually was the project that was the precursor to all simulation effort and also agent effort that we made at Stanford. So this actually predates the Smallville, the generative agents paper. The idea was this, we worked on this back in 2021. The idea was, again, the models have seen so much of human behavioral data that we can actually use these models to simulate people. We wanted to demonstrate that this actually had really good use cases. The interesting part was we actually thought about creating generative agents in Smallville back in 2021. But we couldn't convince ourselves that this was going to be useful. We were like, well, this is interesting, but is it a game? Like, why would we want to do this?

• 49:18Speaker 01

  So we wanted to demonstrate a use case that we were compelled by. And one of the fields that I intersected with was this field called social computing. Social computing is a field that basically has to do with creating computational tools and platforms that enables people to interact with each other and collaborate with each other in a better and pro-social way. And one of the hardest problems in this field for many decades has been the problem around emergence and cold start. So it's one thing for you to run user study and see how five people might use your study, your platform. But what happens when you have millions of people using your platform?

• 49:58Speaker 01

  Or something that we hear often and all the time in many ways is you designed a new social networking service or you created a new online thread. It's very hard to tell what people will do in that space. All the ways they might misbehave, all the ways where they might find opportunities to collaborate. It's really hard to foresee. This is where our policies would also fail. So if you look at Reddit, each subreddit has these moderation strategies. But oftentimes those are said as in a reactive way to something that has happened. Our idea was what if we can create a simulation of subreddits on day one? Show the designers and users every single thing that could happen in this community.

• 50:43Speaker 01

  So that we can, before we put people into this platform, we can put in the strategies and moderation tools and rules. And they will prevent some of the best actors while encouraging some of the really good use cases. So it was 2021, 2022 version of Molebook. That's basically what this was.

• 51:03Mario Gabriele?

  Yes, that's right. I wouldn't have made that connection. That totally makes sense. One of the things that was interesting to me about that is you're sort of simulating the fail states of a company, right? How much of that do you bring into Simulay itself? Like, have you simulated, I don't know, your, your entire company and thought through, I don't know, if XYZ things happen, we should be extra aware of them?

• 51:23Speaker 01

  Right. So we do actually use Simulay to make our own decisions, obviously. We want to believe that this is actually our core advantage as a company.

• 51:32Mario Gabriele?

  Yes.

• 51:32Speaker 01

  So recently we had to decide whether we wanted to move our headquarters. So because we were coming out of Stanford, our first headquarters actually was in Palo Alto. And we asked the simulation, where should we actually have our headquarters? Should it be in San Francisco or can we have it in South Bay? It was very compelling that, ah, yes, we actually would need to want to move our headquarters to San Francisco based on the simulation. And this had to do with our ability to attract amazing colleagues that we want to work with, our ability to be in close proximity to some of our customers. And it has some downstream implications if we stayed in South Bay. So we decided at least for Simulay, it made sense to move.

• 52:16Mario Gabriele?

  Wow.

• 52:16Speaker 01

  These are the kind of simulations that we would actually run.

• 52:18Mario Gabriele?

  And so for, for that simulation, is the sort of input population, current employees, potential employees and customers? Like, how do you sort of make sure you have, you know, the stakeholders you need in that?

• 52:29Speaker 01

  The input to this birth of simulation actually was initially people working in tech in the Bay Area. And of course, that does naturally include people who might already be at this company. But that was the population that we tested against.

• 52:45Mario Gabriele?

  What have been the predictions that have surprised you the most?

• 52:48Speaker 01

  We actually ran a study that was trying to understand how automobile companies are to think about EV, the electric vehicle.

• 52:59Speaker 01

  This actually was a study that we ran for some time, even before we started Simulay. But one of the problems they were faced with was how do we market electric vehicle? And for companies that are not electric vehicle company.

• 53:13Mario Gabriele?

  Yes.

• 53:14Speaker 01

  Marketers and market researchers, what they were focused on was, well, what is the right messaging? That would actually maximize the sales of EV. And we ran that in simulation and we say, hey, you can say XYZ things and this is how your sales will change. Reasonable. But what was surprising, however, was some of the best messages that you could create actually would change people's perception across other lines of business that this set of companies had. That's not EV.

• 53:43Mario Gabriele?

  Wow.

• 53:44Speaker 01

  And what was interesting was it would actually lure the sales of other product lines that wasn't EV. So even though this particular strategy increased the performance on the EV sales, if you take a look at the holistic picture of the business, it was actually in the negative. That actually was really interesting, right? To the point of us finding insight that's counterintuitive. Here, basically counterintuitive thing was you were tasked with optimizing for this micro sort of a segment of the business, but it actually can have negative implication across the board. So maybe it is actually not the right move. Of course, this is something the companies would want to take into consideration holistically and make their own decision. But that was the finding.

• 54:29Mario Gabriele?

  You must have a very good simulation of yourself in a sense. Do you try and increase the fidelity bit by bit over time? Does that happen naturally? Are you sort of, I could imagine a version of it where you're like, I don't know, once a week I give a 30 minute call to my simulation in a sense and let it, yeah, increase its fidelity.

• 54:50Speaker 01

  So what's interesting is what we have, and this isn't internal tools and not a part of our product yet. But internally on Slack channels for the company, we actually have what we call siblings.

• 55:03Mario Gabriele?

  Hmm.

• 55:04Speaker 01

  Siblings can actually model our own team members based on all the messaging that we've been having in the public channels and so forth. So you can actually add your sibling and ask it to talk to another siblings, ask it to talk back. And it's actually very good representation of that person.

• 55:23Mario Gabriele?

  Yeah.

• 55:24Speaker 01

  We, one of my team members actually made this more as a sort of a fun sort of hackathon project.

• 55:29Mario Gabriele?

  Yes.

• 55:30Speaker 01

  But it actually became one of the more viral feature internally. So for instance, when teammates want to, you know, grab dinner and they're deciding what to eat, they would actually do add group of siblings, decide what we want to eat. Or if we have a debate that we want to sort of actually, you know, resolve, we would actually add a bunch of siblings and say, hey, what would these group of people actually talk about? Just simulate the entire debate in a multi-agent fashion. It has been, it has been quite fun.

• 55:58Mario Gabriele?

  Wow, that's so interesting. In many ways, it feels like the, the final state business model of a company like yours should be investing. If you can predict the future, then you, you could be the world's greatest hedge fund ever. Yeah. What do you think about that? How far away might that be?

• 56:16Speaker 01

  Simulay, of course, was founded on the premise that we want to help people and society make the best decisions possible. And that is my personal passion. So beyond obviously the monetary side of the business, you know, we were fundamentally researchers to start. So we do care deeply about the impact that we want to have. But the finance market is actually interesting, less even for the profit. But actually, because this is the way that our society decides to allocate capital for societal decision-making. So I agree that this is actually the final, one of the final lending spot potentially for technology like Simulay. We do have colleagues who are hired directly from leading quant firms.

• 56:59Speaker 01

  So this is of course, something that we are quite fascinated by.

• 57:02Mario Gabriele?

  Have you allowed it to trade small amounts yet?

• 57:04Speaker 01

  Not yet. But I do think it might be an interesting exploration for, for Simulay. We're not going to be able to disclose it.

• 57:12Mario Gabriele?

  Sure.

• 57:12Speaker 01

  When it does happen, at least for some time, but it's an interesting exploration.

• 57:17Mario Gabriele?

  Yeah, you know, I think people have obviously let various agents loose on trading accounts. And it always seems like it goes rather badly so far, but this might be, I could imagine there being a real edge there.

• 57:28Speaker 01

  But the thing that I, that I do think this does raise, that's quite interesting, is more and more agents in the future will be making decisions. And this will be the case in the financial market. This will also be the case when you're trying to even create new UI and UX. Agents will be doing this and they're already doing this. Where they fail is when they do not take into consideration human nature of the world they operate in. Simulation is actually a way to inject our voice into these decision makings. To basically give a sense to these AI models, what people might actually do and feel so that we can have also better control over these models. So that I actually do think is a very important part of contribution that the field of simulation has to make.

• 58:13Mario Gabriele?

  It, it sort of raises an interesting conundrum that I've been trying to think of what it means. I'd be curious how you parse it. In a world in which Simulay or, you know, other such tools, let's say, become truly table stakes for every company and, and kind of every significant organization. You're then in a world in which every company has superior strategy, right? How does someone win in that world? Is it just a sheer research, resource overmatch? Is it the rare moments of human intuition being somehow a little bit actually better than the simulation that make the difference? Because when we think about our world of startups, like strategy is, is kind of one of the key inputs to how

• 58:59Mario Gabriele?

  a small company surpasses a big one eventually.

• 59:02Speaker 01

  I think there is basically three different elements here. One is it will always be humans who ask the goal and who ask the right question. So a part of human intuition that is going to be, that will continue to be important here is basically that ability to put things into the right framing. That I do think will always be there. Then I think there's a question around resources and the resources, I think, come from two different axes. One is, and Simulay does support this today, but these, the ability for these companies to create a simulator that is particularly fine tuned for their population and their environment.

• 59:44Speaker 01

  A part of this would certainly be these companies bringing in their own data to create a forked version of Simulay's model that is going to be institutional, that will be used for their particular institution. The last piece here is also compute. So with Simulay, you can create more multiverses, you can create more experiments. You can look into the future for a longer period of time. And where I actually do see this go is simulation is interesting as AI technology in the way that this is the most extreme version of inference time compute of reasoning. This is the deepest reasoning a model can possibly do.

• 1:00:28Speaker 01

  Because we're literally exploring every possibilities, we're simulating what might happen. And there, another thing that is beautiful in this particular context is there's actually no limit to how much you can simulate. So you can actually see, imagine you want to create a simulation that would solve climate change. I'm sure that's not going to be a $10 session of simulation.

• 1:00:49Mario Gabriele?

  Yes.

• 1:00:50Speaker 01

  So there, I do actually think that differentiation might actually also come purely from the amount of compute and resources that we might have.

• 1:00:58Mario Gabriele?

  I'd be curious to hear how you think about building your team and what you look for in researchers. That's something that some folks I talked to before this said, you have a particular lens of, of what works for you in Simulay.

• 1:01:11Speaker 01

  What I always care about is very consistently two things. Are they the common denominator of success? And do they have two superpowers that contradict each other? And these two things are something that for me, at least for the people that I love and respected the most and that they had most consistently. The former is interesting because it allows us to see whether this is somebody who has extremely high degree of ownership. People who basically wake up in the morning and think, doesn't matter how everything else goes. I would personally make sure that this succeed, that I can be the one to control the success of this particular project or this particular era.

• 1:01:56Speaker 01

  Very powerful, but it also shows that they can reinvent themselves. This is something I personally care a lot about and for myself, for Simulay and the team members. I want Simulay to be relevant, not just in the first era of our company, the era that we live in today. But I want us to be invited to play the next game and the game after and be relevant for many generations to come. That does require us to be able to reinvent ourselves multiple times. And right now we're in the early innings of this journey. So we are now trying to gain the right to win this era and gain access to the next era. But I want the individual team members to also bring in that energy. The winning now doesn't mean our journey is done.

• 1:02:40Speaker 01

  It means we merely gain the right to play the next game. And it's actually the ability to play the game that is the greatest prize. So that I actually care deeply about. Another one here is people with contradicting superpowers. I have now multiple people and in this, who share this trait. But for instance, if you look at someone like Michael, who is one of my co-founders, he actually was the person that I brought and who I had started to discuss the idea of creating a game town. And many of my senior most colleagues thought that was silly. And these were colleagues that I loved and respected the most, but they thought, well, why are we making a game? Is this for entertainment? And I honestly didn't quite have a good answer.

• 1:03:24Speaker 01

  I was like, well, I think it might be just really interesting, but I actually don't know. But Michael was somebody who had this deep analytical insight. So he's an extremely mature scientist, among the very best in my field, if not the best that we've seen. But he's also internally a child. When he heard this idea and we started to talk about it, you sort of would see his eyes light up. And he also couldn't articulate why this was a good idea. But he knew that this was something that he wanted to do. Being able to marry that childlike playfulness and imagination and boldness with the very deep, mature scientific expertise makes this character very compelling. So that I actually care deeply about.

• 1:04:06Speaker 01

  So I see this across my team members and I feel very fortunate to be able to say that. But this is something that I do care about quite a bit.

• 1:04:13Mario Gabriele?

  The loss of playfulness feels like the death of your mind in many ways. As is something I've been thinking about more and more, in part because I have a young child. But it feels like that is the sort of the moment where you've stopped learning in some fundamental way. You can still sort of incrementally learn. But yeah, protecting that is something that's, I think, really hard as you go through the world. I have to ask what your two contradictory superpowers are.

• 1:04:41Speaker 01

  There are two different ways to phrase this. But fundamentally, I am someone who is on a day-to-day basis, very concerned about many different things where we have to put every single thing on the table to make sure that we succeed. But I'm also somebody, long-term, who really do believe that the world is stacked for this team. That no matter what happens, we'll actually succeed. This particular trait I share very closely with one of my co-founders, Lainey Ellen. Lainey is also somebody, I observed, she is paranoid every day. And she's always in the mindset of, we have to do every single thing possible today. Otherwise, we'll totally miss out on this opportunity.

• 1:05:27Speaker 01

  But have this very fundamental faith that no matter what happens, we'll succeed. I personally found this combination to be quite powerful. At least in the way that we were able to operate as a team. And to the point of portrait artists painting their paintings that sort of resemble them. I actually think team building is a lot like that, too. Where as a founder, myself, Lainey, Michael and Percy, we recruit people to accompany and these people bring in their own unicorns that we do not have. But at the same time, it is also us who are recruiting these people, who are finding these people to be compelling.

• 1:06:07Speaker 01

  So in many ways, our belief as a founding team actually does get represented in the team. So in many ways, the team itself is also a self-portrait of who we are.

• 1:06:18Mario Gabriele?

  Yes.

• 1:06:18Speaker 01

  So I do also think this is something that is broadly shared across the team.

• 1:06:22Mario Gabriele?

  If you had unlimited resources and no operational constraints, what is the simulation you'd most like to run?

• 1:06:29Speaker 01

  I am deeply inspired by Thomas Schelling, the work and the career he led. And this was a scientist who had deep insight about the computation and the work of behavioral economics in particular. But also held deep interest in some of the largest societal problems that we faced during those days. The simulation that I would be most interested in are also of that nature. You look at our society today, what are the biggest societal problems that faces our community? It is climate change. It is the stability of our union. Can simulation create insights that would help us better navigate these really difficult wicked problems?

• 1:07:14Speaker 01

  That really does inspire me. And I ultimately want to believe that the final contribution that a company like Simulink can make is one that is societal, that leads our community to be better, to be a more perfect union than we are right now. And that's what excites me.

• 1:07:32Mario Gabriele?

  If you could assign a book to everyone on earth to read, what would you want to give to people? And if you want to throw a painting for everyone to look at, I would, I'd be very interested in that, too.

• 1:07:41Speaker 01

  Okay. I think what book ends up becoming meaningful to the individual is, I actually think, is very personal. So I think everyone should find their own journey. The one that I actually had, that had actually the most impact on me, actually, when I was younger, was The Catcher in the Rye by J.D. Salinger. Without quite realizing, it was very visceral representation of a young mind in the form of Holden Caulfield. And that always stuck with me. And I do think as a young person, I think I first read The Catcher in the Rye when I was, I would say, maybe 14, 15. I do think it actually had a fairly profound impact. And I do believe it also had profound impact to many people.

• 1:08:25Speaker 01

  So not necessarily recommending the book, but that's the book that comes to my mind if I had to pick a title. There is a painting that I have in my apartment, a copy of it, because the real painting, obviously, is very expensive and not available for sale, of course, is The Triptych by Francis Bacon. This is a triptych, but it says, "Three portraits that are painted side by side of fruit."

• 1:08:56Speaker 01

  It's one of the canonical pieces of Francis Bacon. It's very much in Francis Bacon's style. He used to create these figure paintings, but he would distort the shape of the figures. He would smudge it, the eyes don't quite align the right places. It's very mysterious. But that actually was the painting that, when I was an artist, of course, the stylistically, I was actually fairly far removed from Francis Bacon. But it was one of those pieces that I always admired. I'm like, if I can paint something like this in the future, this would be an amazing contribution.

• 1:09:35Mario Gabriele?

  June, this was so much fun. Thank you so much.

• 1:09:38Speaker 01

  Likewise. Thank you for having me.

• 1:09:40Mario Gabriele?

  That's it. Thank you for listening to this episode of The Generalist Podcast. Please subscribe on Apple Podcasts, Spotify, or your preferred podcast app. Ratings and reviews help others discover these discussions. So if you enjoyed the conversation, I'd be grateful if you could take a moment to leave one. For all past episodes and more, visit us at degeneralist.substack.com. See you next time as we continue to explore the future.