这家估值20亿美元、以85%准确率模拟人类行为的AI初创公司

Elon Musk thinks there could be a probability that we live in a simulation.

I do think we live in a simulation. Simile is a company that is creating a foundation model of human behavior in partnership with over 40 million real people. For the first time in the field, we can actually show that we can predict people’s behavior with over 85% accuracy.

You have raised $300 million in total at a reported valuation of $2 billion. You have really great names, like Greenoaks and Index. What happens if someone comes to you and offers you $15 billion to sell your company?

Well—

Joon, welcome to Sourcery.

Thank you for having me.

I know we’re here in Index’s office, whose team has been a big contributor to or investor in your company. Congratulations on the recent big announcements. I’m super excited, because this is an area I don’t quite understand. I don’t really understand where the company is positioned in the market, what you’re replacing, or what you’re making now. I don’t know exactly what simulations are, but I find it fascinating. Elon Musk thinks there could be a probability that we live in a simulation. What does that mean to you?

I do think we live in a simulation.

You do?

I do.

Explain.

Of course, not surprisingly, because of the company and technology we’re building, we get asked this question quite frequently. At some point, I started to just say, “Hey, we live in a simulation.” But the part that matters more to me is: At the end of the day, is the experience we’re having actually real? Does it feel real to us? Is this true to who we are? I think that’s what matters.

Probability-wise, I think it’s a pretty good argument that humans and societies like ours will always have a desire to better understand themselves. What we’ve found is that the best way to understand something is to recreate it. One way I sometimes describe a simulation is that it’s a little bit like creating a portrait of who we are as a society. I think this is an innate desire for humanity, or a civilization like ours, to have. In that case, yes, any civilization will create a simulation, and at some point, it’s likely that we live in a simulation.

How, in the real world, are we in a simulation? What would make that up? Then we’ll get into what you’re building.

Right. Let me say a little more about what makes up the simulation. What do you mean?

How could humans be living in a simulation? We all have the same— we’re all programmed the same way. We all have the same psychological behaviors. It’s like everything is already written. What does it mean?

The way I often think about a society or an experience like the one we’re having right now is that there are always these quantum building blocks of people’s tastes and behaviors. They might be things like my inherent characteristics. People’s risk tolerance doesn’t really change. There are certain aesthetics that I really enjoy. These all come together, and we build larger societies from individual building blocks: people.

When we think about simulations, we think about individuals first. Can we create simulations of individuals that faithfully and accurately represent a diverse persona and the characteristics and behaviors of that person? To create a civilization-wide simulation, we put all these people into the same environment and let them interact with each other. The part that’s always complex and chaotic is what gets born in their interactions. That’s ultimately what building a simulation on a societal scale is going to look like.

And this all started with Smallville?

It did.

Explain what that is, and your whole background behind it.

For sure. Smallville was an experiment that we ran in 2023 and 2022. We initially observed that these large language models—we were using GPT-3 and then GPT-3.5—were still fairly janky, but you could see glimpses of opportunity. These models were trained on such broad human data that, when you poked at the right angle, you could extract a lot of human data that appeared in their training data. This might be from the web, Wikipedia, or social media.

We started asking ourselves, “Can we use this to create a truly new interaction paradigm that goes beyond simply generating content or classifying different tasks? Can we generate truly new behaviors?” We decided the most ambitious version of this would be to create a simulation of an entire person, and then, once we had an entire person simulated, create an entire community. Smallville was an experiment that came out of that effort.

It was a game town. Computer scientists love games because they’re an amazing way to demonstrate new capabilities. We created a game town that looks a lot like the towns you might see in Pokémon or The Sims. Then we populated it with 25 agents. They all had different names, characteristics, and life stories, and they were basically language models paired with what we called an agentic architecture, which had memory, planning, and reflection.

They would wake up in the morning, do their routines, go to work, and have relationships. The simulation takes place the day before Valentine’s Day, so the agents come together to have a party. There’s a character organizing the Valentine’s Day party, and there are characters asking each other out. None of this was preprogrammed, but we could see that we could create simulations of these people with emergent phenomena coming out of the simulation.

That was the experiment we ran in 2023: Smallville. That particular work introduced a few key concepts. One was that we can create simulations of people. The other was the core agentic architecture: that we can create agents using generative AI models. It was the very early days of these agentic workflows, so we could really start to see them becoming possible. Since then, we’ve been building on the technology to create simulations that are more accurate, more diverse, and bigger in scale.

How do you test the validity of this? How do you know if it’s actually working?

Our ability to test has shifted quite a bit alongside technological progress. On day one, when we created Smallville, we interviewed the agents. We pulled individual agents out of the game town, sat them in a hypothetical chair on a command line, and asked them questions that probed certain qualities, similar to how people reason about different tasks in their lives.

A simple question was, “Tell me your name.” Then we’d ask, “Tell us your life story.” We’d ask questions like, “Who’s the one person you want to spend time with?” And then, “If you had to get that person a present, what would you get them?” We asked these questions because, when you think about the human brain and mind, we can answer them because we can remember, plan ahead, and reflect.

Then we generated all these responses and had human evaluators look at them and ask, “Given the agents’ experiences in their game town, are these experiences or insights actually human-like?” We’d—

Hmm.

—we’d have human evaluators compare human-authored responses with the agents’ responses to see which ones were more realistic. We did quite well; these agents were quite realistic.

One way to tell is to imagine asking an agent, “Who do you want to hang out with the most?” It’s a little strange if the agent says whoever they happen to see most on the street. The agent should likely say the person they actually care about. Who do you have a special relationship with? What kind of conversations did you have with people? What stood out? The agent should reflect on that and pick those people. Those are some of the experiments we ran and the signals we were able to get.

Over time, we wanted to go beyond that. The initial evaluation was really about believability: whether we started to see what we call the illusion of life in these agents. As we matured as a field, we started running evaluations focused more on faithfulness and accuracy.

The experiment we ran next used this architecture to create simulations of 1,000 real people. We went out and recruited a large number of real people, sampled representatively from the U.S. It was a representative sample of 1,000 people from the U.S. population across gender, race, ethnicity, and so forth.

Then we collected data from these people. We conducted a human interview. We built an AI agent that could talk with people voice-to-voice. We asked questions similar to the ones we ask agents in Smallville: “Tell me the story of your life. Where did you grow up? What were some of the hard decisions you had to make? Who matters to you?”

We used that data to train AI agents that could be twins of these people. Then we sent people away for 2 weeks, brought them back to a lab, and asked them to complete a battery of surveys and behavioral studies and respond to stimuli, while their agents predicted what they would say. At the end of 2024, for the first time in the field, we could show that we can predict people’s behavior with 85%+ accuracy using these agents.

That’s ultimately what gave the company the foundation for today’s technology. Of course, the technology itself has shifted quite a bit, but that was the initial building block: we could see that these models and agents can predict human behavior, and that’s amazing. At the same time, it’s a large data operation. We don’t simply create simulations out of a vacuum. We strongly believe that the simulation has to be grounded and evaluated alongside people. So it really is both.

I have so many questions. Do you sometimes inject anomalies or things that could create factors that aren’t predictable? And my second question, which I wanted to be my first question, is: Is human behavior predictable?

1. What Makes People Predictable

In the training data, we inject a lot of behavioral knowledge. It’s not just about what people say, but also about what they do, and there are a lot of irrational behaviors in human tendencies. A lot of things we study in economic behavior, games, and randomized controlled trials try to cover those.

At the same time, there are certain things about people that aren’t known to change. Your risk tolerance, as I briefly mentioned, isn’t really known to change in a person. Once we’re able to learn that, it becomes a very predictable trait. So it’s really about understanding the gap between what people think they are and what they actually do in their lives. The modeling challenge is learning that translational layer: How do you take observed behaviors or attitudes and get to their fundamental tendencies?

What’s in that gap? Is it time? Is it events around you? How do you determine the gap that ultimately creates the outcome?

What’s very consistent in our experimentation is that these 2 functions are governed by different functions, or different shapes of the curve. What people say in a survey often follows a normal distribution. If I’m asked, “Am I likely to go to the gym this week?” I might respond according to some normal distribution curve: “Yes,” maybe a 5 or 6 on a 7-point scale every week.

But do I actually go to the gym that week? The activation energy for me to actually go is much higher than I would expect, so that looks much more like an exponential curve, with a much clearer drop-off in terms of how much activation we need to get to the state of, “Oh, I’m actually going to exercise this week.” Those are the kinds of different functions we need to learn, and that’s often where the gap is.

Again, I have so many questions about this. What’s the number one, or most common, indicator of human behavior? What are those traits or things—

Mm-hmm.

—that you’ve noticed create the reactions or events? What causes people to do things? Why would someone make a decision versus not? How do you determine if someone will continue a pattern or not?

It turns out that this sometimes has to do with the existence of optionality. Let’s briefly look at how psychologists would usually study human behavior in a survey. We have an abstract question and a certain set of options presented in multiple-choice questions or on a Likert scale of 1 to 7. That’s a very confined space, but people don’t usually behave in a vacuum or under abstract conditions.

They often behave in the real world. Even with a simple question like, “Where do we want to go eat?” if somebody said, “Hey, let’s go eat X,” the odds are that if I’m really hungry—in fact, if I’m very hungry and there’s no food available—I’d almost certainly say, “Yes, let’s go do that.” But what often ends up happening is that we don’t have access to certain options, while we do have access to others. That’s where people have optionality, so now they have to think about the opportunity cost of taking that option.

That’s what really governs people’s behavior. I can go exercise this week, but if I don’t, I can sleep for one more hour or take one more meeting. Optionality, and the type of optionality, often governs people’s behavior.

Do you pull a lot of data from dating apps?

Not dating apps in particular, no. We don’t.

No? Okay. Taking all of this into consideration, how does this become the business of Simile?

2. Turning Simulation Into A Business

Simile is a company creating a foundation model of human behavior that can take as input a description of a person or a group of people and simulate what they might do in different contexts. Today, simple use cases include surveys and experiments. Many insights teams have studies they were meant to run. Simile can run them at a much faster pace and with far fewer resources.

Those are obvious use cases for concept testing. In more complex cases, we do conjoint analysis and MaxDiff analysis. Traditional analyses that researchers had to run could cost hundreds of thousands of dollars, sometimes millions, and take many years. We can run them in simulation almost instantly. They’re just a query away.

That’s the use case today. But users also come to us and ask, “Okay, it’s great that we can now do these studies. Can we do something our studies couldn’t do? What would people actually do? How would they react to different stimuli?” Simulation is also about creating a model of the environment, not just the people.

Imagine you’re trying to sell water. That’s actually quite interesting as a behavior question. Imagine there’s only one source of water. That particular person could likely charge us $1 million just to have access to water. But the reason that can’t happen is optionality: there’s a context around the water source where water is everywhere. It’s about packaging and, of course, the right brand. At the end of the day, water is widely available, and there are other options in this space.

We want to model that environment. What are the other options? What are the other alternatives? Given that context, can we expose our agents to this decision-making? At what price would you actually buy this water, given all the environmental factors that go into it? It really is about modeling people and their behavior, but it’s also about modeling the environment and everything that surrounds them.

And who are some of your customers? What types of companies are these?

Yeah. We work across a few different verticals today. Right now, one of the larger verticals is retail.

Okay.

One of the customers we’ve been working with is CVS. They’ve worked with us to create simulations to better understand their consumers, so they can better serve them in terms of store layout, messaging, and overall decor.

But we also work with financial companies, including some of the largest financial institutions today, as well as international institutions like Banco Itaú in Brazil. They have similar needs, but want to understand what policies or messaging they should put out to better communicate their values to their customers. We also work with entertainment companies and CPG companies. The verticals are pretty broad, but what unites them is that they need to better understand their user base and are trying to be even more customer-centric.

Today, there’s a real limitation on being able to talk to every single person whenever a company needs to. But if we can create simulations of people and ask them questions at any time, that becomes a superpower for these companies.

For each one of these customers, do you create their own digital twins? How do you create new populations within that, instead of using the same base for everything and getting duplicates?

3. Building Representative Populations

Simile maintains vertical-specific agent banks across key verticals. These are simulated panels that are always available from day one, including when we start a new customer relationship. But customers often have further requests, because no one is interested in the general population; everyone is interested in their own unique population that they’re trying to better understand.

In those cases, Simile collects more data to create and train the model to cover those populations. Today, Simile has partnerships with more than 40 million real people through our panel partnerships. This is across the globe, and at any given moment, we’re talking to tens of thousands, if not hundreds of thousands, of real people every single week. That’s a scalable operation we’re running today. If there’s a new need for panel members, that can usually be met within 24 hours.

Who are these people? I’m always really curious about this, because I don’t understand polling. I don’t understand when people say 53% of Americans think X or Y. Who are these Americans? I don’t know; I’ve never met someone who’s been polled or surveyed. Have you? Well, you have. But who are these people?

They’re just everyday people who choose to answer polls or participate in these data collection processes. The magic of this process, though, is knowing how to sample them. Social scientists have been developing methods for randomized samples and probability sampling to get a broad understanding of a representative set of people.

Some panel companies specialize in this. They try to form real relationships with people across the entire country. One of Simile’s strategic partnerships is with Gallup. Gallup is one of the most well-established polling companies, and I’d consider it one of the gold standards for generating these kinds of insights. They have a Gallup panel of hundreds of thousands of people with whom they’ve built relationships over the past decade, two decades, and so forth.

But wouldn’t those people be biased at some point, or feel entitled to give them what they think would be the best information based on their own circumstances? How do you know that’s valid information?

We control for these things in a couple of ways. Biases can creep in in different ways, so there are different ways to address different kinds of biases. Not surprisingly, this is an entire area of study, with many researchers working on it. One clear bias we often see is fatigue. If you get asked the same questions—or even different questions—but get polled every single week, you start to get fatigued and answer in certain ways, or try to game the system.

Mm.

You might think, “If I participate in this, I get paid more, so I’ll just give some bullshit answer and participate.” You start to see those behaviors. So you make participation sparse enough while still maintaining the relationship. We run studies comparing a new population with a population that has answered X number of questions in the past. We measure the right amount of time between asking different questions to get high-quality responses—the gold-standard quality responses we’re looking for.

Another thing we do in these kinds of studies is use attention checks. These are dummy questions that aren’t part of the core measurement.

Yeah.

You’re basically sneaking in questions to filter out participants who aren’t actually paying attention.

What would one of those questions be?

A question might be something like, “Imagine right after our interview, you asked me what the first question you asked me in today’s conversation was.” If I can’t answer, that’s tricky. You asked me about Elon, “Are we living in a simulation?” If I was paying attention to our conversation, I should know that. That’s an example of an attention check.

I was thinking of something really silly, like checking someone’s attention by asking, “How tall is a giraffe?” or “How many hamburgers can a human eat?”

Technically, those are also good attention checks.

Really?

Yeah.

You can use those if you want.

Thank you. These are actually questions we sometimes ask, too. Some surveys or data collection processes will literally ask people to do simple arithmetic, so if they say 10 plus 10 is 30, we know they’re likely not paying attention.

Or they’re not—

Or they might not be good at math in those instances.

That brings up another question. I’m just in the mood now. Are there certain populations you have to avoid, like drug addicts, mentally ill people, people who might skew the data one way or another? Or is that part of the overall picture? I know that’s an appropriate question, but I’m curious: what are the no-gos in who you survey, and how do you screen for that?

For us, the no-gos are participants who can’t give informed consent. We take informed consent very seriously. Every member of a panel who joins our study and data collection process knows, at least broadly, why the data is being collected, and consents to that process. They also have the opportunity to withdraw their consent.

For instance, underage populations can’t participate. That’s one blind spot we own: we don’t collect data from underage populations. This also applies to other populations that can’t give informed consent. Aside from those populations, Simile is different because we’re trying to create a model that can truly be a representative snapshot of who we are as a society. There’s no population we should avoid; we want to know every single population out there.

That’s also where we’re quite different from companies simply creating a model, like a language model, or other modeling companies. Other modeling companies are trying to create an amazing reasoning machine that’s great at coding and mathematics, and fundamentally trying to make a good chatbot or assistant. We’re trying to create models that can capture the diverse viewpoints, values, preferences, and tastes of people. Part of that is learning people’s good behaviors but also their bad behaviors. Simile needs to be able to capture those.

For part of the prep work for this, we had a Simile project pilot for Sourcery, and I thought it was really interesting because I was like, “Who are these people?”

Yeah.

Who are you interviewing? There were 2 different tests: one on topic appeal and one on titles. The number one title it predicted—or, I don’t know, what are the terms you use?

Simulated.

Simulated. Of course. Duh. The number one title it simulated was “The $2 Billion Company Building AI Simulations of Real People.” That was the strongest title overall with my audience, which is mostly investors and tech executives.

That was interesting, and the topics really went into digital twins. The number one was digital twins, simulated customer testing, and privacy. There was no clear winner among them. I kind of agree with that, because right now the biggest concerns with all this AI stuff are cybersecurity and privacy risks.

I obviously asked a lot about how simulated customer testing works. The digital twins thing is quite interesting. I don't know. How does this work for other people? What is the process by which it outputs this information?

What our users have today is access to an online platform. We basically serve a SaaS platform, so they get access to a web interface where they can come in and define 2 things: who they want to talk to and what questions they want to ask. They can ask anything, and right now they can ask any number of questions. There is no limit to the number of questions they can ask, and we'll generate responses. That's what they get today.

One of the most important things, though, is that this is a new way of making decisions. We haven't quite seen simulation that is so generalizable in a social context that it can inform all human decisions. One of the things we do, working very closely with our users, is change management: helping users take these results and calibrate when to use them, how to use them, where we need more data, and how to validate them.

One of the core functions attached to these reports is this idea of a model of confidence, or predicted accuracy. This model is quite important to us. Simile will generate predictions, simulations of the future. Alongside the core foundation model we're training to make the simulations, we're jointly training a predictive model of accuracy. Its job is to output a probability in the 0-to-100 range for how confident we are in a particular response, on the basis of all the data we use to train our model and the question we got.

For a social scientist, this is a little bit like having a p-value in inferential statistics. In science, when the p-value is less than 0.05, that's the canonical way of saying, “We're ready to make a decision.” We provide that. If confidence is higher than 90%, or usually 85%, our customers have found that to be robust evidence that can be used to make very high-stakes decisions.

So it's the whole package: delivering the insight and also delivering undeniable evidence that, if this is the evidence, then this is a totally reasonable outcome that we have to prepare for. Then we provide this model that gives us a degree of confidence calibration.

It's really interesting. One of the questions I had was, what's the most expensive mistake—

Uh-huh.

—Simile helped a customer avoid?

Yeah.

What is it?

There's an interesting study that we ran. This was technically slightly before Simile, but I'll still mention it because I think it's an interesting example. We worked with a few automobile companies, and all of them were interested in electric vehicles: Should we do it? When should we do it? How should we do it? That was top of mind for every electric vehicle company we met, and these were some of the largest players in the field.

One of the simulations we ran had to do with whether we could predict or simulate how people would react to different messaging for the entry of a new electric vehicle in a particular market, and how we could optimize that messaging to convey our values to our customers. There was a strategy that said, “Okay, this is a clear winner. Fantastic. We're ready to go, and this is good.”

What was interesting, though, is that this had downstream implications. This particular messaging, this particular approach to the market, actually shifted people's overall perception of non-electric vehicles and globally lowered sales within that entire market.

Jeez.

This was run in studies, and it ended up becoming an input. Again, this wasn't part of Simile back then, but these are the kinds of findings I find interesting. They're not narrowly focused on one particular question; they're more generalizable, and you're looking at the entire field holistically.

Simulations allow you to explore that, and at their best, they should surface insights that are counterintuitive and surprising. So that's really the goal and the value.

Do you have any dream scenarios or dream simulations that you want to answer or uncover? What's the wildest thing you're trying to solve, or wish someone would ask you to solve, or that you could solve on your own?

We're running this company and delivering so much value, but a small part of me is still the researcher. As a researcher, I'm inspired by questions that are truly societal and generational. Every generation has a few of those questions.

One of my research heroes is Thomas Schelling, a behavioral economist in the '70s and '80s, who introduced what we call the model of segregation. Racial segregation was a big topic, a generational topic, so he decided to apply simulation to understand its causes and mechanisms.

I do think about what questions we have today. One of them is climate change. Climate change is hard, not just because of the physics of it, but because it's about coordination. There are so many actors with competing incentives, each trying to optimize for their own things. When they come together and clash, the equilibrium state they generate is nearly impossible to navigate.

If we can simulate that over and over again and see what kind of outcome we might get depending on the policy, that could be quite powerful. Another question is: if we believe AGI is here, or imminent, then obviously we as a community, at least in some pockets of this community, have presented a few different ideas. One of them might be UBI. There might be some other ideas. Can we simulate these kinds of findings? Instead of running a five-year study on UBI, could we simulate it and show where it worked and where it didn't? Those are good simulations to run.

Have you simulated who you should be taking money from? I know you just raised a huge round.

Well—

You must have, right?

In fact, I often simulate these kinds of conversations.

Mm-hmm.

Obviously, Molly, you have a very big online presence, so of course I simulated the conversations before walking in.

Is this one of them?

But here's the funny part: it did warn me that you might ask whether we live in a simulation.

Of course.

I should have taken that advice more seriously.

Pull up your results. Let's go through this right now. What else did it simulate?

That was a big one.

Mm-hmm.

I thought that was interesting. It's not something every interviewer asks.

Hmm.

For some reason, this is actually a good question that the simulator thought you might ask, given everything we know about you.

Well, I do tweet at Elon a lot, so that is part of it. I'm curious, though: you've raised $300 million in total at a reported $2 billion valuation, and you have really great names, like I mentioned: Greenoaks, Index Ventures. You also have some really cool individuals, like Fei-Fei Li and Andrej Karpathy. How did you collect such an interesting base of investors and angels?

We were, in a way, fortunate to have worked with many of these people in the past, so many of them were actually our colleagues. Fei-Fei Li is, of course, a close colleague at Stanford. One of my co-founders, Micah Bernstein, worked closely with her for many years, and even today we're colleagues.

Andrej Karpathy was one of my other co-founders. Percy Liang was actually on his doctoral dissertation committee. So we've known Andrej for a long time. In fact, when I first published the “Generative Agents” paper back in 2023, Andrej was one of the first people to really amplify that work.

So in that way, it felt like we were coming full circle. But I think in many ways, we were also fortunate because the idea, the core research, and the core need for this technology are so real that I truly believe simulation is inevitable. There’s a deep, innate desire and need for people to better understand who they are.

Just as we write autobiographies or paint portraits, simulation is a lot like that. In many ways, we’re painting a portrait of who we are as a society. I think many of our angels see that this is an inspiring idea, an idea that needs to exist, and that’s what got them excited.

Alexander Wang was talking about how, when he was first exposed to OpenClaw—I think through Nat Friedman—the prompt Nat gave him was to do a deep dive psychological analysis based on all the information. OpenClaw ingested everything from their computers—photos, emails, everything—and gave a detailed psychological analysis.

Yeah.

Alexander Wang was shocked by it. Is that an angle as well? How do you see how much data you need?

It’s both about how much data, but more about what data you have. It turns out that asking the right questions matters quite a bit. We do interviews with people, just like what we’re doing right now. I got the idea of interviewing people to learn about them back in 2023 and 2024, when I was on a lot of podcasts for the academic work we led at Stanford. I realized that after talking to interviewers for an hour or two, they know a lot about you, especially the good ones. If they ask the right questions, they learn a tremendous amount about both your field and who you are.

That’s where we got the inspiration. We ask questions that are both topical and quite general, and we’ve found that asking people to tell us the story of their life reveals a lot of interesting pockets of knowledge about them. But we also collect a lot of behavioral data. Some of it might be observational, like transaction data or mobility data, which makes a lot of sense. What really matters, though, is the causal mechanism: randomized controlled trials. Imagine you have the same experimental setup, but you vary a particular variable, whether it’s a product feature or messaging, and observe how people’s behavior changes. That data matters a tremendous amount.

What we’ve learned over time is that people aren’t that interested in predicting the future, which was a little surprising to me. Unless you’re trying to beat the stock market, what are you going to do if I tell you that your market is going to crash in 2 quarters? People will ask, “Well, that’s shitty. What do we need to do now to change that?” The second part is what matters more: What do we need to do now to shape the future? For us, it’s not just about finding correlation; it’s about understanding the causal mechanism, the causation, because that’s ultimately the lever we have to shift the direction of the future.

Who makes up the team? What kind of talent are you looking for? Researchers, scientists, engineers? How is it composed?

We’re now a team of about 100 people. Twenty percent of the team are our lab mates and academic relatives. These were people who were my lab mates in Percy’s group, in Michael’s group, and in the computer science department at Stanford. They went on to frontier labs and then came back to work with us to build these foundation models of human behavior.

We also have an amazing group of people in our circles who are more like methodologists. For us, it’s not just about creating models; it’s about studying and understanding people. We have a small but growing group of methodologists with backgrounds in the social sciences.

Mm-hmm.

That’s the science part of Simile: about 20% to 25% of the company. Modeling, evaluation, and the science of conducting these simulations are very important parts of the company. Then we have an amazing group of engineers and people on the product side, because this technology is being handed to many Fortune 100 companies and other users today.

We’re creating this technology with an amazing group of people. We recently brought on our head of engineering, Baton[?], who initially came from DoorDash.

Hmm.

He was an amazing leader there.

I heard— By the way, I heard DoorDash has amazing AI talent. It’s very underrated.

They’re killers.

Yeah.

What I’ve always appreciated about DoorDash and that particular team is that they’re soft-spoken in some ways, but they fought in an extremely brutally competitive market, and they won. That speaks volumes. We have an amazing group of engineers and product talent, matched with an amazing group of people in go-to-market and deployment. We’re a very forward-deployed company. That part is led by one of my co-founders, Lainie Yellen.

Our fundamental thesis is that the data and the model are extremely important. Simulation is basically in its GPT-4, GPT-4.5 era: the technology is here to make an impact, but it’s changing and evolving every month. A lot of this battle is also going to be in the market, in our forward-deployed motion. I look to previous cycles in the recent AI era. For instance, in the agentic coding era, I’ve found that companies that were forward-deployed found a lot of success. Their relationships became stickier and more defensible over time. I think that’s also going to happen in simulation, because I see the space reaching a point of maturity where, in about 1.5 to 2 years, we’ll be where coding agents are as an industry today. We’re learning a lot from that era.

We’re also in an era of spies, people leaking information, and employees who are flight risks. What just happened with—did they have a spy, or was someone bugging them? Do you help HR teams determine who to hire and who not to hire?

Those are some of the use cases we hear about, and in some cases, we collaborate. Of course, we do this with the informed consent and autonomy of the members of our customers’ companies. Many companies, especially as they get large, are deeply interested in retention. How can we make sure our people are excited, and especially that our most high-performing people are committed to the company? That’s something we can help simulate.

Another use case is recruiting. Imagine you’re a company trying to recruit the most talented students from Stanford or MIT. What’s the right way to communicate with that population? That’s also something we can help simulate. It’s not just about understanding consumers or customers; it’s about understanding everyone in our ecosystem. Many of those important stakeholders are internal.

I have to say, the more we talk and the more you answer questions, the more nervous I’m getting. I feel like you can read my mind a little bit. I don’t know if that happens often, but I genuinely feel like, “Shit, you did your research.” What does the simulation say?

As we close out, this is one of my favorite questions. You kind of already answered it, but I’m curious if you might have a different answer. I take this in a personal direction. I think performance is a matter of who you surround yourself with. Some people say it’s your 5 closest friends. You might have a more scientific answer, but it could be a mentor or someone who’s inspired you along the way. I know you’re also a painter, so maybe it’s an artist. Who is that person for you, or who are those people?

The people I respect and love the most in my career had 2 traits. One is: Are they the common denominator of success? Are they people who, at a different point in their career, stood up and basically said, “I will make this successful, no matter how everything else goes. I will be the reason this succeeds”? That has always demonstrated to me their ability to reinvent themselves and take deep ownership.

The other trait is having superpowers—two superpowers that really aren’t supposed to coexist in the same person. I sometimes describe one of my co-founders, Michael Bernstein, this way: he’s a very mature scientist who understands scientific methodology and so forth, but at the same time, he’s a child who gets excited about creating a game demo.

Or Lainie Yellen, who is deeply paranoid on a day-to-day basis that we have to do everything right now and really make sure we do everything we can to make this successful. But at the same time, she’s deeply religious and believes that no matter what happens, we will win and we will succeed. Those kinds of tensions I find particularly beautiful.

And of course, Percy, famously within the Stanford community, is my last co-founder. He’s a pianist, but he’s also one of the most well-known AI researchers. It’s an amazing combination. These people really do inspire me.

For people who are the common denominator of success, you can see it and feel it in their behavior and their careers. I was recently coming back from a workshop, and I happened to sit next to a scholar named Sandy Pentland. He used to be a professor at the MIT Media Lab for a long time and trained many of the most brilliant minds in my field and in computer science more broadly.

He’s had an amazing career, and he’s somebody I’ve always looked up to. There were many generations: an era when he was a physicist, an era when he was an AI researcher and ended up training some of the most well-known AI researchers of that generation, and an era when he was a computational social scientist. In each era, you could see that Sandy was such an influential character. That’s always an amazing thing to observe.

And what makes that up? How do those people become that different person in each era? Is it just extreme commitment and durability? What is it?

I think it’s extreme commitment, stamina, and the ability to continuously get excited. These are also people who are in the game so they can play the next game. Their greatest prize in winning the current game is gaining the right and invitation to play the next game. That’s what makes them special, and it’s something we try to emulate.

If those are the traits of success, what are the traits of failure? What causes people to fail?

I think, as they say, every failure has its own reasons. Every failure can look slightly different. But I think the reverse of what I said can sometimes be a mode of failure: trying to over-optimize for one particular outcome.

One thing I often saw was that there’s that one particular fundraising round you want to run, that one customer you’re trying to win, or that one best paper you want to win. And you win it, and you go, “Now what?” Then they lose interest. I’ve seen so many institutions and individuals achieve incredible things at some point in their careers, then lose interest and disappear over time. Whether that is failure or not, that’s not for me to say.

Mm.

In many of those cases, I think it could actually be seen as an amazing success. But for those who want to play this game for a long time and have an impact across many generations and many eras, that is a failure moment.

So it’s the difference between short-term thinking and long-term thinking. Fascinating.

And why are we doing this?

Yeah.

Are we doing this so we can have an exit tomorrow, or are we doing it because we want to fundamentally reshape how our society operates for the better?

So what happens if someone comes to you and offers you $15 billion to sell your company?

Well, if I have every reason to believe there’s an ambitious idea that only Simile can pursue, then we’re going to pursue that.

Okay. Well, Joon, this was such a fun conversation. Thank you very much, and I appreciate you answering my silly questions.

Thank you for having me.