模拟整个经济的 AI 公司｜Simile 联合创始人兼 CEO Joon Sung Park

Ready to go? Joon, I'm so excited for this, dude. When Shardul told me that I had to meet you, I'm going to be honest, Shardul does not often tell me that I have to meet someone. So I was like, “Wow, I feel honored. Thank you, Shardul.”

Then we met when I was on holiday with my family, and I remember my grandparents were asleep upstairs, so I was whispering to you. I remember being so excited by what you were building, but also having to be incredibly respectful of the sleeping elderly people next door. Thank you so much for joining me, dude.

1. The Valentine's Day Simulation That Put Joon on the Map

Thank you for having me. Excited to be here.

When I spoke to a lot of your investors and friends beforehand, they all said that I had to start with the very unique background you have. You became particularly well known for a particular project, and it centers around Valentine’s Day and a simulation that happened as a result. Can you explain what happened and how that potentially led to the early days of Simile?

For sure. This was 2023. We had this idea that large language models are often used for simple tasks like classification and simple generation, but we thought that these models actually had a lot more potential. One of the early observations that we made was that these models are trained on so much human behavior data and sentiment data expressed on the web. If you poke at them from sort of the right angle, you could actually extract a lot of realistic human behaviors from them.

I thought that was really interesting, and it was also practically interesting in that it was domain-agnostic. If you look at the literature in computer science for many decades, we've always had the vision of creating agents that are meant to be generalizable, that are meant to really be able to act like a human in any environment. My mind went to, “Well, maybe we have that opportunity here.”

What we ended up doing was asking, “If we were to fast-forward many years into doing this, what would be the most ambitious vision that we might have?” That was creating an entire lived experience of a town. The idea here was that we would make a game town and populate it with 25 NPCs, or non-player characters, except these characters would actually wake up in the morning, do their routines, go to work, have relationships, and do all that. They would remember their interactions and plan their days.

One of the surprising things you end up seeing was that the simulation itself was set the day before Valentine’s Day, and you’d actually see these agents come together, have parties, and self-organize. They would plan parties, decorate the café, and so forth. We thought that was really interesting.

There were 2 fundamental contributions from that work. One was that it was one of the earliest examples of creating agents. This particular set of agents was paired with, back in the day, GPT-3.5 and text-davinci, so we didn’t quite have ChatGPT back then.

2. How Agents Got Memory, Planning & Reflection — The Origin Story

It was paired with memory, planning, and reflection. Those were really the first times that those concepts became an explicit part of the architecture in agentic workflows. The reason why we got that inspiration was that if you had more than 1 agent side by side, you wanted them to remember each other.

Back in the day, language models didn’t really have the concept of memory. So I thought, “Okay, you have to give them memory so that they don’t say, ‘Hey, nice meeting you,’ every time they meet their roommate.” We gave them this concept of memory, planning, and reflection to make sense of a very long-term landscape.

How do you solve that memory problem? Everyone says, “Oh, we have a memory problem today.” How do you solve the memory problem of agents to prevent that from happening?

Back in the day, the initial idea was fairly simple: these language models are actually quite good at processing natural language, so we’d put everything in a Markdown text file. That was it. That sort of worked.

The issue there, however, is that language models have context windows, and even today, even if the context window is getting larger, the kinds of experiences that these agents can have in the small game town are immense. Imagine now if we were to bring this to real life, to a world like the one we live in. The amount of memory that we accumulate is huge.

Imagine you went to get an omelet 5 times in a day. You want to make sense of that rather than just, “Oh, I went to get an omelet 5 times throughout the week,” or something like that. So we had this concept of reflection, which basically was that, at certain intervals—it’s like a shower thought—you’d ask the agent explicitly to get a bunch of their memory pieces and make sense of them.

“Why did you get an omelet so often this week? Were you busy? Do you like omelets? Why are you studying for this test so hard? You were in the library every single day. Does this matter to you?”

They’d actually start formulating ideas that are higher-level than what happens on the ground. Gradually, they start to realize, “Oh, this particular research topic, I’m actually quite invested in it. This might actually have something to do with my childhood or my fundamental memory. This actually shapes who they are as a person.”

That ends up becoming a very useful function in creating these agents that have personality, that actually have a point of view on the world, and that can make sense of a lot of this data. That’s how we did it back in the day.

When we think about simulation models today, for those who don’t know, a simulation model is essentially the creation of agents that then produce a set of activities or actions that show us what a simulated future world might look like. Is that correct?

That’s right.

Got you. When we think about building a simulation model company, would you say Simile is a simulation model company?

Yeah. We are a company that is creating a foundation model of human behavior that can then be used to create simulations of individuals, simulations of subpopulations, and then, ultimately, simulations of the entire ecosystem and even the market.

3. From Researcher to CEO: What Joon Looks for in Academic Founders

Do you sit on top of core foundation models? How do you think about the relationship, for those listening, between an OpenAI or Anthropic frontier model provider and you?

4. Simile vs. Frontier Models

This is a great question. The way we see it is, if you look at large language model companies today, fundamentally the task they have at hand is to create super-rational, intelligent machines that are good at coding and good at natural sciences and mathematics.

Simile doesn’t really care about any of those. What we care about is that if we have a person make a mistake in this context, we want our models to make the same kind of mistake. We want our models to be biased in the same way humans are.

In a way, we want to be a representation of people’s values, preferences, and taste—the subjective half of their brain. That’s what we care about.

5. Say vs Do: Why Behaviour Data Beats Survey Data

I love that. A lot of what people say is different to a lot of what people do. How do you think about the chasm between what people say and what people do, and how that impacts your models?

For sure. If you look at web data, it is fundamentally data of what people have said, not what they have done. Obviously, these models today are trained primarily on this web data.

For us, we do collect a lot of behavior data. We collect transaction data and observational data. We also partner with our customers and our vendors to collect some of this data.

My personal hot take here is that a lot of observational behavior data is amazing at actually helping you create a correlation between the observation and what could happen in the future. That’s good for prediction tasks.

But my take here, after interacting with so many of our customers and also being in research, is that no one really cares about prediction. No one really cares about what’s going to happen in the future unless you’re trying to predict the stock market.

What people actually care about is that they want to shape the future. Imagine you’re Starbucks. It doesn’t really help them to know that their Frappuccino sales are going to tank in 2 quarters. They’ll hear that and be like, “What do we do about that? That’s terrible.”

What they want to know is, “How can we prevent it? What do we need to do now to change the future?” And there, what you really need is a causal mechanism. You need a model that can actually reason about causal mechanisms and counterfactuals.

6. Prediction Is Overrated

So the kind of data that we care deeply about is a lot of randomized controlled trials. We actually run a lot of A/B testing. We show the models, “Imagine people have done this versus that.”

This is how their behaviors will actually change. That becomes a core part of our training asset. This is the data collection that goes beyond the observational data that Simile collects.

Is data collection acquisition the hardest element of building simulation models for you? If you think about the core pillars for traditional models, it might be compute, algorithms, and data. Is data the biggest challenge for you?

Data is an important piece of Simile, for sure. My fundamental thesis here is that, for AI companies of this generation, you need to have an interesting data strategy that's going to be defensible. For us, the data-collection challenge comes from 2 angles.

One is sourcing people. Sourcing people here is a little bit different from what other language model companies might consider to be their people or their population. We don't go after expert programmers or expert scientists; we go after people like us—everyday people living their everyday lives. What we care about is: Are they representative? Do we actually have the same representation of people as we do in the world that we live in?

And then, actually asking the right questions of these people. What are the experiments? What are the questions that actually get at the fundamental core nature of who they are? Some of the questions we ask at the start of our data collection are things like, “Tell us the story of your life. Where did you grow up? What did you experience? What were some of the hardest problems that you had to tackle or decisions you had to make?” Tell us a lot about these people. That's what we try to do.

In terms of people not wanting to predict the future, just so we can drill down on that, I thought they do. If Starbucks can predict that Frappuccino sales will be down in 2 quarters, they can amend their buying cycle. They can change how much they purchase. Isn't that valuable, and what am I missing?

But that's the thing. The reason why they want to know is so they can change their strategy. Certainly, talking about how many resources they actually need to serve this market—that is a kind of change in behavior. But fundamentally, it is about counterfactuals.

We have this market that we want to serve, and we want to maximize our value as a company. What do we need to do to make sure that we react to this dip in the market, whatever it may be? Fundamentally, though, the work that we do is about people. We try to simulate people and represent people's perspectives. The value that we provide is counterfactual in terms of what your consumers, what your population, would do.

7. Is Simile Just a Fancy Qualtrics? The TAM Question

When you look at what can be done for some of the biggest brands—you mentioned CVS there—it's incredibly valuable for surveys, customer feedback, and determining what customers really want moving forward. I didn't know how to say this without being rude: Do you want to just be a next-generation Qualtrics? And how do you prevent that from being the angle?

Yeah. The way we see it is, again, fundamentally, the core primitive we're trying to build is very straightforward. You tell us what population you're interested in, and we'll go model them. So far, the layer of innovation has lived in the tooling layer: How can we create a better survey tool? How can we create a better interview tool?

Simulation is fundamentally about something different, which is: How can you create the most generalizable model of people so that we can represent people's viewpoints at scale? That goes beyond simply running surveys or interviews. Down the line, I see simulation as a field moving into contexts where we can create simulations of many people interacting with each other, so that you can understand all the downstream implications of your decision-making.

Or imagine you have a new product you're about to launch: Can you actually simulate the entire launch—how the audience might react and how the market might shift? This also goes into the scientist part of me. I get quite excited by the vision where simulation—I do think—can also be a cure for many of what we call “wicked problems.”

A good example here might be climate change, which requires collective action across many stakeholders who have different incentives. One of the reasons why such problems are so difficult is that finding the right equilibrium state, where all the different parties come together to make a decision for the global good, is very difficult. Can we actually simulate those decision-making processes? Can we simulate, even in things like, under what conditions does a democracy fail? Can we predict that? These are the kind of questions that simulation ultimately can answer.

Can I ask you a thing about democracies failing and elections? For a government, elections have been an incredibly useful tool. How do you think about who you can and should work with versus who you shouldn't?

This is where the principles matter so much for us. The way I see it, simulation as a piece of technology is 1 of the twin pillars of technology. I'm a fan of science fiction. You read any advanced science fiction, and there are always 2 pillars. One is some form of AGI that always shows up. The other is simulation.

Like with any powerful technology, the potential for misuse is quite real. The way we see it, simulation at its best ought to be representation at scale. People have different viewpoints, different perspectives, and different tastes. Many of their viewpoints are not considered in rooms where important decisions for them are made.

We always want to say, “We listen to our people. We listen to our customers. We listen to our stakeholders.” In practice, it's very difficult. This is a way for us to ensure that, in every decision-making process, we actually listen to people at scale. That's the North Star.

How much data do you need to feel confident that an accurate prediction outcome will be displayed? Is it 100 people? Is it 1,000 people? Is it 1,000,000 people?

8. How Many People Do You Need for Accurate Simulations?

You want to have more people represented so that you can segment down to a specific subpopulation. If you look at any social-scientific literature, if you have a very narrow population of interest, you would usually get statistical significance in the study that you want to run by the time you have 1,000 people.

However, oftentimes, the way that people query our system is they want to come in and say, “Hey, filter down to X with XYZ population.” Those filters are often created on the fly. For us to then be able to simulate people's responses across all those filters, that means we want to represent the entire population. So, that's the journey that we're on.

Is it self-fulfilling? Do you get better and better at predicting over time?

I think that certainly is the case because there is the data flywheel. There is the learning that occurs as we get more and more simulated results and see what happens in the ground truth. That absolutely is the case.

9. The Data Flywheel: How the World Becomes the Ground Truth

This is obviously 1 of the core value propositions for our early partners because they know that, in their business context, Simile is getting better and better and better. Do they have that compounding advantage?

It's kind of like AlphaGo. They just beat the hell out of the model and played it 1,000 times. Every day of activities and outcomes in the world is another game of AlphaGo, where you can correct the model on what was wrong, what you missed, and what didn't happen. After 10,000 days, you should almost be better than the model at the model. Do you know what I mean?

This is actually quite interesting. Let's think about a different example. How does the data flywheel work in simulation, and why would it work?

If I were to take a brief detour and talk about coding, the reason why coding agents have improved so massively over the years was because their reward function was extremely clear. If you make a suggestion and your user says “accept,” fantastic. If they say “reject,” that's also very useful. You very quickly know what is good and what is bad. That was 1 of the core learning mechanisms for these models.

It might be easy to look at simulation as a field and say, “Well, where are you going to get the reward?” Fundamentally, all the things you're trying to predict are happening in the future, so it's going to be hard to validate. That is true.

At the same time, I think simulation has an even better mechanism: The world is our ground truth. We live in the ground-truth world. What we can do is, every single day, generate tens of thousands of hypotheses. Each hypothesis is mapped onto an end state. If this happens, we know whether we can validate the simulation as right or wrong.

We're basically watching the world every day, seeing which of those hypotheses are answerable at what time. We can say, “A month goes by; we generated 1,000,000 hypotheses, and X percentage of them came true.” This is the best way to learn about the world.

Does it take a huge amount of compute to run these simulation environments at scale and well?

Compute is an important piece of simulation. Of course, a lot of the work that we do is to make our simulation more efficient. A lot of our compute initially actually goes into creating the initial breakthroughs in technology. So, it is exploring different ways to train and exploring different kinds of datasets.

Once we have a point of view, we can very quickly make it efficient. Some of the things that I've seen with Simile as we've built this company over the years is that, right now, we have a model that's been in production. This model used to cost about 100 times more to run than it does now.

Some of it does happen because we actually found different ways to model with the same reward model and the same philosophy, but in a way that's much more efficient at inference time. There are these kinds of tricks that we can play and scientific advancements we can make to make things cheaper. A lot of the investment, however, does go into finding that initial point of view.

10. Can Simile Simulate Elections, Democracies & Wicked Problems?

Can I ask you, when you look at serviceable market, or total addressable market—TAM, in venture speak—you obviously have your CVS and your huge enterprises who would absolutely want to work with you. It can also be consumers, like regular consumers wanting to see what happens if they run their own environments. Is this a play for everyone? Is this a play for the biggest companies in the world? How do you think about the TAM for something like Simile?

The start of my career really came from research, obviously, and the job of a researcher is to serve humanity. We do our research for our own enjoyment as well. We love the process of finding new things in the world, but fundamentally, it is a service. It is a belief that if we are able to make scientific breakthroughs, this is going to, down the line, serve everyone in our society.

That is how I see simulation as a field as well. So, right now, we do serve enterprise customers for a couple of reasons. One, obviously—I'll be frank—there's the budget. There is a clear product-market fit that we see today, and that does excite us.

11. Who Should and Shouldn't Have Access to Simulation Technology?

At the same time, it is an amazing way to validate the technology. It is very important to us that we get the feedback loop to be as tight as possible, so we know when our simulations are right, when our simulations are wrong, and we're improving them every single day. There's also a side here that's just as important: when I was at Stanford, I had a colleague whose office was next to mine, Pat Hanrahan. He was one of the founders of Tableau, a graphics professor, and he also won a Turing Award. He's a very well-known person in this landscape.

One piece of advice he actually gave me and some of my colleagues was, “The best way to get feedback is to actually ask people to pay you.” That was the core philosophy at Tableau, and I want to see this here. Getting the best kind of feedback matters a lot.

Enterprise market research right now is a wedge that we found that actually has significant budget and immediate product-market fit. But down the line, I do want this technology to be used by the rest of our society, because fundamentally, what we are trying to do is help people make better decisions.

Before we move on to the expansions that it could be used for, when did you know you had product-market fit? You said you felt that pull. When were you like, “Ah, we got product-market fit here”?

Many of the Fortune 500 board members and their C-suites reached out. In part, they do come to Stanford to see some of the demos that are happening in the lab, and they all saw the Smallville demo after it was released. Everyone thought, “Oh my God, if we can simulate a market like this, this is going to change the way we operate.”

You could immediately sense the product-market fit, and this was really the forcing function for us to then say, “Okay, this is actually quite interesting. We're going to show and validate that our simulation can not just be an interesting demo, but that it's going to be accurate.” So, we spent about a year demonstrating that we can create models of people that are actually amazing at predicting people's behaviors and are validated across surveys, behavior experiments, and real environments.

We showed that we can actually predict people's behaviors and attitudes with 85% of the accuracy with which people replicate their own responses. We put that work out at the end of 2024, and that's really what started the field around synthetic panels and simulations. That's the market that we're seeing today.

Will synthetic panels be larger than human panels in 3 years' time?

The way I see it, synthetic panels will be larger than what we know to be the current human-panel market, in part because this can really raise the ceiling of the kind of questions we can answer. What I see today in the market is actually quite broken.

We have so many questions we want to ask about our market: if we were to release this product, if we were to have this particular strategy, or this particular policy. If you're a scientist, then you want to run this study, or you want to try these macro-scale experiments. You're looking at maybe 5% of those ideas getting answered. The rest of the 95%, we never bother experimenting with because we either don't have the ability to do them—especially if it's something at the emergent scale, where we literally don't have a way to run those experiments—or we don't have the budget and time for them.

A lot of the decisions that we make as a society, we base on our gut instinct. Sometimes they're good, but sometimes they're very biased based on our own narrow experience. So, what simulation will do is unlock that limitation and allow us to actually test every single hypothesis that we have about the world before we launch it into the world.

How do you balance the pursuit of the next dollar and serving customers who pay a lot of money, I'm sure, versus research prioritization and maybe focusing dollars there over building out a customer success team and an FDE team? How do you balance profit maximization with research purity?

Simile is interesting as a company. Simile is a company that has a real product and engineering team, but at the same time, we are a research company. Three of the 4 co-founders are researchers. Our co-founders are myself, Michael Bernstein, Percy, and Laney. Michael, Percy, and I were all researchers at Stanford.

I led research around agents and simulations. Michael was one of the co-authors of ImageNet, which really kickstarted the AI revolution, and he's been a leader in human-centered AI. Percy was the person who literally coined the term “foundation model.”

The vision for this particular area is that we can actually create the next paradigm shift in AI and in the way we view technology and the impact of technology in the form of simulation. The reason why we're able to operate as a research lab, but also have an amazing product and engineering function and go-to-market function led by my counterpart, Laney, is that the alignment between what the technology can do—the promise of the technology—is so close to what our market actually requires.

The better the model gets at representing people, the better simulation we can create. It immediately means a better experience for our users because they'll have much more grounded, much more accurate simulations. It is very difficult to maintain both a lab and a product company if there's not that alignment. But when there is, it can be quite magical, and that's what we're seeing at Simile.

Can I ask you, when we think about pursuing some of the largest companies on Earth—we mentioned CVS, although I'm not sure which customers we're able to say and not say—people always think it's a multi-year, incredibly long sales cycle. Was that something that you experienced, or was it a different experience for you getting to work with some of the biggest companies on the planet?

What's been fascinating to me coming into the field of simulation, especially in this market, was last year, when I started the company. I left Stanford in June of 2025, so it's been exactly 1 year. I actually thought the market would take about 1 or 2 years to warm up to the idea of simulation. We'd basically build the right foundation for this company and for this market, and we'd go aggressive maybe toward the end of 2026. That's what I had in mind.

That's not what we experienced. What we experienced was that our customers were moving extremely fast, also in ways that truly made me change my perspective on corporate America. One of the leaders we work with at, for instance, CVS, is Shree, who's their VP of Insights. She's extremely forward-looking, extremely ambitious, extremely hard-working, and an amazing counterpart to a vision like Simile.

But what I also found was that the pain they were feeling in their day-to-day work was so real. It was way more acute than I could have imagined. When they realized that there is, or there could be, an answer in this market for addressing some of those pains around very slow experimentation, budget, and so forth, they were ready to drop everything and try us out.

So, we actually saw some of the largest customers in the world move at lightning speed for enterprise, where we saw them close deals within 3 months.

3 months. Wow. Okay, that's very different from what people traditionally think. What matters more to them: speed of output—in other words, being able to get results very quickly on their simulations—or accuracy of simulations? Yeah.

It is both. There are so many questions that they are truly relying on their gut decision to answer today. If they can get some form of evidence to at least directionally guide them in the right path, then they're ready to try it.

Then they very quickly realize that this is actually an amazing way to interact with a lot of data. This is an amazing way to gain evidence that is actually quite accurate. One of the ways we actually got some of our first customers was that, in the first call, they had a finding from large consulting companies, and they basically queried our system: “Hey, if we were to rerun this, what would the system say?”

We predicted the outcome of studies that took 3 to 6 months, but did it within 2 minutes.

That's very powerful. It must be so compelling in a customer conversation to be able to say, “You did this campaign. If you had done this campaign, it would have been 12% more effective. Do you want to buy our product?”

[Laughter.]

It is such a good sell. I'm a seller—to be able to have that data is unbelievable. How do you think about value extraction efficiently? What I mean by that is, if you work with a CVS, or you name any of the big companies that you work with, these are massive companies where, if you're able to do your job efficiently, you can move the needle to the tune of hundreds of millions for them, and in some cases billions in revenue. Charging a million bucks feels like a large chasm between value generated and value extracted. How do you think about closing that chasm to be more fair?

Yeah, that's a great question. I see the market moving in this direction. One of the core premises, and one of the ways that our customers are actually finding value in Simile, is by avoiding really damaging decisions that could have cost them hundreds of millions of dollars.

So it's prevention, not optimization?

It's both. But certainly, prevention is a huge, obvious value case, right? That could have been a total disaster. Had we run that, it would have cost us half a billion dollars. We ran a simulation, and that prevented it. That's a no-brainer. This is a true painkiller in their case.

If you do your job efficiently, can Kalshi and Polymarket still exist for a lot of their markets?

It's an interesting question. I do certainly think there is an overlap here, in that we are companies that are fundamentally interested in the future and helping people at least get a glimpse of what the future might be.

Where I see Simile come in is that we are a company that is not just interested in what's going to happen, but more in how it's going to happen and why. In that way—and this is also the value proposition that our customers are most inspired by—it's one thing to simply predict, but can we actually show, “Here are all the steps that your ecosystem is going to take to get to that particular outcome”? This is the way you can prevent that, or you can encourage that. That is ultimately the power.

When it comes to team building, you said something about the craft of team building, and I think it's really interesting because it's an ongoing challenge building the best team. What have been your biggest lessons, coming out of research, in what it takes to build an all-star team at Simile?

A couple of things. One is that the team has to be balanced. There are certain powers that I can bring to the team, but there are also a lot of things that I don't know. I was a researcher; I was not an enterprise seller. I needed Laney to be my co-founder and lead that part of the game.

12. How to Build a World-Class Research Team That Doesn't Quit

Balancing the team and being able to see where your team is lacking, especially as we scale, and the new gaps that are emerging—actually seeing that ahead of time and making sure that we fill those gaps—I do think is a core fundamental of building a great team.

At the same time, I also think it's important that the team remains consistent in its values and rigor. This is more of a painter's analogy. When I was a painter, I was a figure painter, so I worked a lot with human subjects, portraits, and figure studies.

There's sort of this untold secret amongst figure artists, which is that it doesn't matter who you paint: your subject sort of looks like the painters themselves in some ways, or at least they share a similar vibe. I think building a team is actually a lot like that. In the best team, the team that you care deeply about, you really should see yourself in the team.

For me, a couple of things matter the most. One is this: are we the common denominator of success? People live through different stages in their life, and they have different careers and different jobs. At each stage of their life, were they the reason why that thing was successful? If you squint, were they the common denominator?

If the answer is yes, then what that suggests is a number of things: that they have an extreme degree of ownership, that they are the kind of people who come in and say, “It doesn't matter how everything else goes; I will personally make this successful.”

It also shows the ability to reinvent themselves. One of my co-founders, Michael Bernstein, has had a very interesting career as a researcher. During his PhD, 10 years ago, he started in the field of crowdsourcing and collective intelligence. Then, very quickly, during his early years as a faculty member at Stanford, he went into a different area of AI, and then now into generative AI, agents, and simulations.

At each step of the way, you could see in the work he's done that this is very Michael. You could see that this is the person who led a lot of the success. That's an amazing signal.

The second piece for me is a little bit more niche to myself, but I found this to be very true, at least in the way I look at the world. Do my leaders and my team have 2 superpowers that aren't supposed to coexist in 1 person?

13. The Two Contradictory Superpowers That Make Great Founders

Any expert will usually come in with 1 superpower, or even sometimes multiple superpowers, but they're all correlated. You're an amazing programmer who happens to be amazing at mathematics. That's fairly common. Where I find things to be particularly compelling is if people have 2 superpowers that are really contradictory.

The most common one here is actually the greatest CMOs, of which there are very few—I can count them on 1 hand—who are unbelievably data-rigorous, oriented, and scientific in their approach. Then you blend that with this creative artistry and imagination. They are 2 relatively opposing kinds of mental approaches, I think.

Yes.

It's very rare to have that in a CMO, but when you have that, that is the world-class CMO. That's magical.

Yeah. I've actually sometimes used that particular description for 1 of my board members: he's deeply analytical, but he's very intuitive. I think that's how he makes investments that happen to be very successful. Hopefully, we'll continue that success.

In my team, the kind of thing that I also see as an archetype—and I also categorize myself as 1 of these kinds of people—is that, on a day-to-day basis, for instance, Laney, 1 of my co-founders, is paranoid. She's somebody who will come to the table and say, “Unless we put everything on the table today and do everything possible, we'll lose. We'll fall behind. Everything will fail.”

But long term, she's religious. This is somebody who fundamentally believes the world is stacked for her, that no matter how this goes, we will make this successful.

Actually balancing those 2 at the same time is quite difficult, because if you are short-term paranoid, then you're likely going to be very pessimistic about your future. You might be amazing at shorting stocks, but not great as a company builder.

If you're religious, you have the opposite problem, which is that you're complacent. You sort of feel like, “We don't have to put everything on the table today. Things will be okay.”

Balancing those 2 needs somebody who is broken in some ways. Somehow, they found a way to be deeply paranoid, but at the same time ignore all the paranoia of today to believe that the world is going to be amazing.

I think that's exactly me. I think it's actually because you believe that the paranoia you hold today helps that future state be amazing.

I also often interview the world's most successful founders, and I ask, “What do you wish you'd known when you started?” They all say, “I wish I'd known that it would all work out, and I wish I hadn't been so worried.”

I think it's the worst answer you could give me, because the fact that you were so worried meant you did the prep, you put the work in, and you stayed up late to do that presentation. That led to the success.

Without the paranoia, Laney wouldn't have hit the quarter. Liane wouldn't have set the urgency in the sales team. Liane wouldn't have hired those extra people because you didn't know that access to money would be there. The paranoia drives the success. It's a really interesting one.

14. Competing for Research Talent in the Tens of Millions

Can I ask you, Brendan McCord was on the show recently, and he was like, “Honestly, researchers are in the tens of millions of dollars. It is so expensive.” Do you find that to be true, and how do you find this intense war for research talent in the Bay?

Absolutely. Research talent is very sought after today. I have closest colleagues and friends whose total compensation does range in the tens of millions.

When they join Simile, I'm fairly upfront with them. It is not possible, regardless of how many hundreds of millions you raise, to meet them at their base salary. That's tricky.

However, researchers fundamentally care about a couple of things. They care about a vision. If this idea truly comes to fruition, these are people who have literally seen OpenAI go from being the laughingstock in Silicon Valley to becoming a nearly trillion-dollar business. These are people who have seen Anthropic get to that same state within the past 5 years.

These are people who are fundamentally aware that a deeply ambitious vision can actually come to fruition. They care deeply about the vision. They also care deeply about the impact. What is the societal impact of the technology that they'll be working on? And is it actually interesting to them?

Do you worry about the retention problem in the Valley today? You see so many researchers move with such promiscuity, if you can use that word. Do you worry about the retention problem today?

Consistently.

In fact, I view the role of leadership as obviously hiring amazing people, but also providing a platform where individual members can express their superpower to the maximum degree. I do think this part genuinely does matter. Retention can be challenging, but it can be done.

I have a small sense of pride in the way my career has panned out over the past 6 years or so as a researcher. PhD students often go from one project to the next, and their entire co-authorship will change, maybe except for their advisor. I’ve had an interesting career where, in the past 6 years, all my core team members never left. We all moved from one project to the next to the next together.

Can I ask you a really important question for me? I see a lot of amazing people in research and academia who are considering or starting a company in the same way that you did. I always worry that I’m going to finance a science project, because science projects, although great and interesting and intellectually satiating, don’t always make great companies.

You’ve been able to do that incredibly well in the last year to 18 months. If you were an investor analyzing a group coming out of academia, what would you look for that would give you confidence that they would be able to make the leap from research to starting a company?

The thing I actually would look for is: are they married to a problem, or are they married to impact?

Sometimes researchers are very much focused on a problem, and something about that problem fascinates them. But oftentimes, it’s just not a good company, or it’s not a thesis that can really be formed into a company for various reasons.

But there are researchers who are fundamentally driven by the impact that they can have in the world. For them, it means finding a problem that can actually reach people, finding problems that can actually generate revenue, and that’s what drives them. You want to find researchers who are in that category.

So, talk to me. It was Shardul who introduced us. Huge thanks to Shardul for that. There’s a new funding round that’s come to be in the last month or so. Can you talk to me about the funding round, how it came to be, and how you think about it?

For sure. We raised our $100 million round about 5 months ago. Soon after, we were preempted fairly recently by insiders. Shardul at Index led our previous round, and he and some of the other insiders were looking at the market.

Shardul has this comment that he makes every once in a while: he’s seen some of the fastest-growing markets, and his track record does show that he truly has seen different markets. He’s never seen this kind of traction, this kind of pull. When that’s paired with the technological progress that has been made, and also the amount of compute that we can leverage to further accelerate our progress, that’s what prompted our insiders to ask, “Can we actually put in more money now than later?”

15. Raising $300M in Six Months: How the Round Came Together

That’s how the initial round conversation came to be. We were not planning on raising at that particular moment, but there were a couple of teams that I particularly respected in the Valley that, if we were to be raising, I wanted to talk to. I found out that the team at the top of my list was the new enterprise team at Greenoaks.

It turns out their team had been looking deeply into this market and all the players and how the market was going, and they were actually prepared to make the investment. They were looking for the right time to do so. So I reached out and said, “Hey, this is going to be the round. We’re not running a process, so if you’d be interested in joining, we have a few days to make that happen.” They were excited, so the round came together.

We raised $200 million, which brings our total funding to $300 million raised over the past 6 months or so. It gives us meaningful capital to go after this really ambitious modeling challenge and build up this team. It also brings a lot of really exciting people to the team.

Sholto has been a fantastic partner. We actually have a lot of Index connections at Simile. Our seed was led by Mike Volpi, who now runs his own firm, and Sholto, along with Neil and the Greenoaks team, including Patrick, who is the partner.

You didn’t need the money, I take it. You raised $100 million 6 months ago. Was there a consideration of, “We don’t need the money, so why would we take $200 million now?”

There was certainly that consideration. The way we netted out was that the money does buy compute. This is one of those areas where you can actually—here’s a fundamentally interesting part of our research: with research, you really cannot control the outcome necessarily, but what you can control is the input and the process.

We were at a moment where, yes, we could significantly raise the input, both in terms of data and compute spend, to meaningfully accelerate this progress. That’s when we thought it actually made sense.

Totally get that. What did you not know about fundraising coming from a world of academia or research that you now know, having been through 3 rounds?

One thing, coming in, was that I was fairly skeptical about what the role of VCs actually was.

What do they actually do? We all are.

What do they actually do? How do they help? I’ll be honest: I still can’t quite put my finger on it and say, “This is the way they help.” However, if you bring in the right set of people, what I’ve realized is that they can be some of the greatest partners, and they can also be a really strong set of mentors.

I had never run a company, certainly not one like this. This is my first real experience building a company. I have a lot of technical experience doing research, but so much of what I need to do on a day-to-day basis is new. If there’s someone I can trust, then that’s an amazing boost.

Initially, I started to work with Index Ventures and Pear, and we also had another firm, A*, which helped lead our seed. These funds and the team—and we also worked with Shardul very closely there—really became core mentors as I operated in the field. Mike actually introduced me to Laney, who ended up becoming instrumental as I thought about the business. I found a great partner and friend in her, which has also been an amazing part of this experience.

Suddenly, one thing I realized is that VCs can actually help in some magical ways. They’ve seen enough that, if you’re an experienced VC, they can provide advice that founders might not have coming in. That’s one.

Another thing is that things always happen a little sooner than you’d expect. Coming in, I had in my mental model, “Okay, well, if we raise seed now, that means we might raise our Series A in about a year and maybe the next round the year after or something like that.” All of that happened within a year.

We raised seed, and I think our Series A was very soon after. Our next round also came very soon after. I think the market is always moving perhaps one step ahead of where you are in terms of its interest in investing in you. It’s useful to be prepared for those moments. That’s what I’ve learned.

Do you worry that the market is so frothy that it can get ahead of itself? When you announce this fundraise with the people that you have and with the press that you’ll get, you’ll get more interest for the next round. It’s an ongoing cycle, and, bluntly, the hubris is very high right now. Do you think about that?

I do think there are parts of the market that are quite frothy.

Yeah. Do the unit economics vary for you on a per-simulation basis? What I mean by that is, if you look at Anthropic and OpenAI and model routing, some tasks require frontier models, which are much more expensive and much more token-heavy, versus others that are much easier and can use a degraded or older model and a much cheaper model.

Is that the same for simulations? Do different simulations cost different amounts in terms of compute and token usage?

They do. Usually, when you have a simulation that’s trying to answer something much more complex, or something where you want to actually understand all the downstream implications of your decision, or you want to do a market-segmentation study across all of the US, it’s much more expensive.

What I have also seen, however, is that it is in those simulations where we actually get higher ROI for our users. That’s because those decisions are some of the most costly decisions if you fail to make the right one. This is actually interesting for simulation as a field. We’ve all seen, as a community, inference costs going up and up and up, and we now have these thinking models that are thinking for half an hour, a day, and actually start token-maxing and spending a lot of money on just running this process.

I actually do think simulation could be the next frontier of that. In my vision, I think there’s a world in which, in about 2–3 years, we’re running a single simulation session that’s going to take $10–20 million to run, but it’s going to be so valuable that people will pay $100 million for it. That’s where I see it going.

And that would be for the world’s largest enterprises. That’d be for a government or whatever that may be.

Especially on the high end of the spectrum, that’s what it would be.

What cannot be simulated today that you think will be possible in 3 years?

For me, it’s actually a little bit less about what cannot be simulated, because I do think everything that we want to simulate, we can actually create the initial proof of concept for. However, as we all know, one of the core challenges of AI is bridging the proof of concept with real, productionizable value. That’s actually the chasm that I see.

An interesting thing here is that I see the world of simulation going into this world where we are creating that very complex, multi-agent simulation, or we’re running a very long study with many different steps of simulations along the way. We actually started the field with multi-agent simulation when we created Smallville. That was fundamentally the vision.

It also makes sense because we did that because my co-founders, Mike, Percy, and I sometimes sit together and do this exercise called the time machine game. If we were to ride a time machine and go 10 years into the future, what’s going to be the craziest thing we’re going to see, and can we do that now? That was the motivation for running the Smallville experiment.

This can be done, but the question is, can we evaluate the efficacy of these simulations? Can we actually propose this as a scalable, productionizable system that people can actually rely on for making their decisions? That’s the chasm, and that’s the thing that we see getting bridged every day. A huge part of it is also getting models to be better, creating better simulations, and making the system more scalable. All that becomes a part of this.

Can we play a time machine game with me and you?

Let’s do it.

In 10 years’ time, what is the craziest thing that you can see happening?

A lot of things, but one thing I will say is that I am someone who is fascinated by the history of technology and the analogies that we can draw from it. What I see today that’s prominent in the AI space is what I consider to be the CPU of the intelligence unit. You have this one language model that’s really large, very smart, and can do very complex reasoning tasks. That’s like a CPU.

What I see coming, and what I think simulation as a field can offer, is the GPU of the intelligence unit. As I mentioned before, Simile does not care about creating really smart, superintelligent machines. What we care about is creating models that are as smart as we are. I’ve thought about a lot of things. I want to make sure that the model that represents me feels the same way.

The beautiful part about people is that, individually, we have so much diversity and so many different tastes in our world, which makes individuals so interesting. But also, when they come together as a large collective, the emergent phenomena that we’re able to draw out are some of the most wonderful things that we can see in our world.

Creating a society, creating an amazing process that actually allows us to make all these achievements—can we actually replicate that in simulation? I think it’s going to be quite inspiring.

What’s the crazy prediction, then: that every single person will have a replicable twin that acts and behaves like them in a simulated world? I think that’s the vision.

The vision here is, again, representation at scale. We, as a society, have found over the years many different ways to represent our members. Sometimes it’s in the form of government. Sometimes it’s actually companies. Companies are where we, as a society, allocate capital to make sure that they serve society and the needs of people.

But if we can actually create an artifact that, in a much more scalable and granular way, represents all the individuals, what are the new kinds of policies and new kinds of companies that can be created on the basis of it? I actually do think it’s quite interesting.

If I’m a hedge fund, is this not the most obvious buy in the world? If I’m looking for alpha and edge on everyday activity, sign a $1 million contract with you and get unbelievable insight. Yeah.

Maybe Simile will actually own a small hedge fund down the line.

That’s a cool idea. Would you be down to do that?

Well, it turns out we actually do have quants in our firm. Some of the members who have joined actually do have more quant backgrounds. I think right now they’re joining Simile because they actually want to start a quant firm at Simile.

They joined because they see the vision of Simile as very much aligned with their passion and interest, which is to model the world. Down the line, I think it’s actually an interesting idea.

Is there a world—again, I’m saying crazy time machine world—where you are so efficient and so good that stock markets actually become uninvestable because the world is skewed to Simile’s hedge fund or similar providers, and it’s not a fair marketplace?

I think, especially if we were to assume that we’re going to have some form of AGI, and if we were to assume some form of perfect simulator, a lot of the world—a lot of the things that we assume to be true about our world—will change. Certainly, one of these could actually be the stock market.

What else do you think we assume to be true today that you think won’t be true in 5 years’ time?

16. Is Simile the Future of Love and Matchmaking?

What I think we generally assume to be true about the world that we live in is that it is fundamentally impossible to get everyone’s perspective. Therefore, we need representatives of these people to approximate their perspectives. So far, that has worked in some ways. It has failed in other ways.

I actually don’t think this is a limitation we have to suffer through in the future. I think there’s a world in which we can truly create a layer that becomes a representational layer of our society and of our collective intelligence.

Is the future of love not also similar? What I mean by that is, if you were able to create effective simulations of yourself, dating itself could be much more efficient. I feel that if you could date 100 people at the same time—

Yeah.

For the first date, of course, not onwards. You’re going to get in a lot of trouble for that, Joon.

[Laughter]

But if you could, it would be much more effective at finding the one for you who could pass through to the next stage. Do you know what I mean?

I get that. Well, look, I think love comes in different forms, and I think, just as we discussed today, people have such a degree of diversity. Personally, I’m a bit of a romantic, I’ll be honest. This actually goes to the point that I mentioned about long-term relationships.

I do believe that love will, at least in my life, find its way in a more organic way. I actually do think the way I meet the person I personally care a lot about matters. The fact that we have shared a journey is something I personally care a lot about.

I don’t think that piece of humanity will ever change. Actually, experiencing things together and having that shared memory is fundamental to the way we form trust. This is also why we talked about process a lot. For your users using your simulation, can you actually bring them along in this process?

I think finding love is a little bit like co-founding your life with this person. Can you actually find a process that would bring them along in this process of living? I do think that would matter.

Oh, you’re so romantic.

Unfortunately.

Do you know what I’m thinking of, dude? I’m a content person. I’m a content person and an investor. Weird mindset, actually, in both ways. There’s a show called Married at First Sight. You might not know it. It’s where you marry someone on first sight.

I’m just thinking it would be the most phenomenal advert for Simile if you could do the perfect marriage at first sight because simulations have been run before. But I’m just leaving you with pearls of wisdom that I think would be great.

17. Quick Fire Round

We’re going to do a quick-fire round. I say a short statement. Who’s the most underrated AI researcher today?

There are so many, but I actually do really think there are some incredible people who are working at these larger labs whose names are not known because they work at larger labs and they don’t publish. But I think there are some really incredible people in there.

What area of AI do you think is particularly overheated today?

I do think new labs without a clear vision for how they’re going to impact the world genuinely run the risk of turning out to be an interesting research project, but not a viable company.

If you were investing in my seat today, what part of the AI landscape would you say is underinvested and most exciting? You can't say simulation.

I fundamentally believe that for AI companies in the future, you have to have an interesting data strategy. Do you have access to data that no one else has access to? Do you know how to collect data that is very hard to collect? When you see those opportunities, I would invest.

Right now, aside from simulation, robotics is an obvious place where this has become the case. Obviously, there's a lot of money already going into robotics, so I wouldn't say it's underinvested, but I do think it's quite an interesting area.

I also think that, aside from the core robotics or AI space, the inference layer, but also the chip layer—the hardware—is quite interesting. It's a very hard area for people to crack into, but there are a couple of teams that have done an exceptional job in recent months or years. I think they're quite interesting.

Who do you think those are?

Likely Etched recently came out of stealth. I'm quite bullish on their team. I think they're going to be exciting. So that's one.

Final one for you. What's the kindest thing that anyone's ever done for you? I think it's a nice note.

I'm somebody who actually needed a lot of help throughout my career. I didn't come in knowing everything. Certainly, I don't know everything now. But I also didn't come in as someone who was an obvious candidate.

The one person I would quickly call out was when I graduated from college and moved to Palo Alto, living in somebody's garage. I didn't have a job because I was trying to run a startup that didn't really go anywhere. That was a moment where I really felt lost.

I could sense in the air that the AI wave was coming, and I realized that if you want to be a surfer, you need a wave that you can surf. I wanted to make sure that, when the AI wave was here, I would be there to ride it. I wanted to help create the wave in the first place to ensure my seat in it.

But I had no research background. I didn't do research during my undergrad, which is quite rare. If you're a PhD student applicant and have no research experience during your undergrad, unfortunately, it's very hard.

So I messaged a bunch of people, and there was this one professor at Stanford, Mary Wootters. She's a theory professor. I happened to graduate from the same college as her, and she replied.

I still don't know why. I think it was truly out of kindness, and the fact that we're from the same school, she thought, "Well, here's a student who is seeking advice. I'll at least spend half an hour with a student."

She very graciously spent a full morning with me, just talking me through how I should think about the AI space and how I should think about research. She actually connected me with an initial set of people that I started to work with and learn from.

That initial set of people who came together to help give me advice and actually let me have a foot into this area of research—it really was not an obvious choice for them. I really didn't think I deserved it, but that was the bet that they took, truly for their own kindness. I'm very grateful that they did.

Never forget the first believer.

Joon, from quant funds to simulated worlds to love, this has taken many different twists and turns, but thank you so much for joining me.

Thank you for having me.