Simile is building a foundation model of human behavior that learns people's values, tastes, biases, and actions in order to simulate individuals, groups, and entire markets. Founder Joon Sung Park explains that the core of simulation goes beyond merely predicting the future to showing "how we can create the future we want." He stresses that becoming an outstanding AI company absolutely requires a unique and defensible data acquisition strategy that others cannot easily copy.

1. Simile Began With a Valentine's Day Simulation

Host Harry Stebbings opens with the unusual project Joon Sung Park unveiled in 2023, the so-called Smallville simulation. At the time large language models were mostly used for classification and simple text generation, but Park and his colleagues saw far greater potential in them.

Language models learn from a vast body of data expressing human behavior, emotion, and opinion on the internet. So the early observation was that if you ask in the right way, you can elicit not just sentence generation but fairly realistic human behavior.

"These models have learned an enormous amount of data about human behavior and emotion. If you poke them from the right angle, you can elicit quite a lot of behavior that looks like a real human."

Park thought the chance had arrived to build the general-purpose agent that computer science had long dreamed of — one capable of behaving like a human in any environment. So he imagined the most ambitious version of that future, and the result was an experiment simulating an entire small town.

Twenty-five NPCs — non-player characters, in game terms — lived in the virtual town. But they were not characters simply repeating fixed lines. They woke up in the morning and went about their day, went to work, formed relationships with other residents, and remembered what happened to them. They also planned their days and adjusted their behavior based on prior experience.

Something interesting happened in particular once the simulation's starting point was set to the day before Valentine's Day. Without anyone instructing them, the residents gathered, planned a party, decorated a café, formed relationships, and organized the town event themselves.

"You could see the agents meeting each other, throwing a party, organizing themselves. They planned the party and decorated the café on their own."

An important contribution of this project was that it explicitly applied the structure of memory, planning, and reflection — now familiar in agent systems — from the very beginning. At the time it used GPT-3.5 and the text-davinci models, before ChatGPT had become widespread.

2. How They Gave Agents Memory and Reflection

For multiple agents to live together, they have to remember each other. Otherwise you get the problem of an agent saying "nice to meet you for the first time" every time it runs into its roommate.

Early on they solved the memory problem in a very simple way. Taking advantage of the fact that language models handle natural language well, they stored all of an agent's experiences in a markdown text file.

"At first we put everything in a markdown text file. That was all of it."

But over time the volume of memories grows too large. Even in a small game town, the events an agent experiences pile up quickly, and to handle the countless people, places, and events of the real world, simply storing every record is not enough.

Suppose, for example, that an agent ate omelets five times in a week. A simple memory stops at "ate omelets five times," but what matters more is the meaning of that behavior. You need to work out why they kept eating omelets — because they were busy, because they like omelets, or because it connects to some particular memory.

So they introduced reflection. After a certain amount of time, an agent rereads its various fragments of memory and organizes higher-level meaning out of them.

"Why did I eat omelets so often this week? Was I busy? Do I like omelets? Why did I go to the library every day? Does that matter to me?"

Through this process an agent moves beyond a simple list of events and begins to form its own personality, perspective, and values. It might realize that it is deeply interested in a certain research topic, and that the interest may connect to a childhood experience.

"I'm actually pretty deeply immersed in this research topic. Maybe it has to do with a childhood experience or some fundamental memory."

This process of accumulating memories, making plans, and reflecting on experience is what gives an agent individuality and a worldview.

3. How Simile Differs From Existing Frontier Models

Park defines Simile not as a simple simulation tool but as a company building a foundation model of human behavior. The model can be used to generate simulations at various scales, from individuals to particular subgroups to entire ecosystems and markets.

Frontier model companies like OpenAI and Anthropic generally aim to build hyper-rational, intelligent machines that surpass humans in areas like coding, mathematics, and the natural sciences. What Simile wants is closer to the opposite.

"If someone makes a mistake in a particular situation, we want our model to make the same kind of mistake. We want our model to be biased in the ways humans are biased."

What Simile is trying to represent is the subjective side of humans rather than their logical capability. It aims to model elements that resist reduction to an objectively correct answer: people's values, preferences, tastes, biases, and emotions.

"We want to build a model that expresses people's values, preferences, and tastes. It's the region corresponding to the subjective part of the human brain."

Simile's model is therefore not aiming to always produce the most rational answer. On the contrary, it has to reproduce the reactions and judgments real people would plausibly show in a given situation — even irrational choices.

4. Why Behavioral Data Matters More Than What People Say

Harry points out that there is a large gap between what people say and what they actually do. Internet data and the training data of existing language models mostly capture what people said. Simile, however, collects data closer to what people actually did.

Simile uses transaction data, observational data, and behavioral data secured in collaboration with customers and partners. But Park does not think observational data alone is sufficient.

Observational data is useful for identifying what behavior occurred and what result followed. That is, it is strong at finding correlations between future outcomes and present observations. But what companies really want to know is not simply what will happen in the future.

"Nobody actually cares about the prediction itself. Unless you're trying to predict the stock market, just knowing what will happen in the future isn't enough."

Even if Starbucks receives a prediction that Frappuccino sales will plunge two quarters from now, that is not good news. What Starbucks wants to know is not the fact that sales will fall, but what to do now to prevent that decline.

This requires causal mechanisms that explain cause and effect, and counterfactual thinking that addresses "what would have happened if a different choice had been made."

That is why Simile places importance on randomized controlled trials and A/B tests.

"They show how behavior differs when people do this versus that. This kind of data becomes our core training asset."

In other words, rather than stopping at observing what people did, they actually present multiple options and measure how behavior changes. This, the explanation goes, makes it possible to build simulations that connect directly to a company's decisions.

5. A Defensible Data Acquisition Strategy

Park says the most important competitive advantage for this generation of AI companies is a unique and defensible data strategy.

"If you're an AI company of this generation, you need to have an interesting, defensible data strategy."

Simile's data acquisition faces two major challenges.

The first is recruiting representative people. Where other AI companies recruit specific expert groups such as outstanding programmers or scientists, Simile targets people living ordinary everyday lives. What matters is not simply gathering a large number of people but how well the sample reflects the composition of the real-world population.

The second is asking people the right questions. Questions that reveal what values and experiences a person holds matter more than asking about surface-level preferences.

"Tell me your life story. Where did you grow up? What have you experienced? What is the hardest problem you've had to solve or decision you've had to make so far?"

Questions like these move beyond recording behavior in fragments and let you grasp the life experience and character behind that behavior.

Over the long run Simile aims for a model capable of representing an entire population. This is because situations keep arising in which only people meeting certain conditions are selected and questioned. Even when a user later filters for a particular group by an unexpected criterion, sufficient representativeness is needed to simulate that group's response.

6. Why Changing the Future Matters More Than Predicting It

Harry pushes back: "If Starbucks knows in advance that sales will fall, it can adjust its purchasing and inventory strategy — so isn't the prediction valuable enough on its own?"

Park acknowledges the point but answers that the reason you want the prediction in the first place is to change your strategy in light of the future.

"The reason you want a prediction is to change your strategy. Ultimately the key is the counterfactual question."

Companies want to know not simply "what will happen?" but:

• What do we need to change to respond to this market?
• If we launch this product, how will consumers react?
• If we implement this policy, how will various groups move?
• What should we do to avoid a negative outcome or reinforce a positive one?

What Simile aims to provide is exactly this kind of actionable simulation. It does not stop at getting the outcome right; it shows the process by which that outcome is produced and the points at which you can intervene at each stage.

7. A Market Beyond the Next-Generation Qualtrics

Harry asks whether Simile amounts to a next-generation Qualtrics that replaces corporate surveys and customer feedback.

Park draws a line: Simile's core is not about building better survey or interview tools.

"What we're building at the core is simple. Tell us the group you're interested in, and we model that group."

Where existing market research tools have improved the instruments of surveys and interviews, simulation focuses on constructing a generalizable model of humans capable of representing people's perspectives at scale.

Going forward, it will be possible to move beyond asking individuals for their reactions one at a time and simulate entire situations in which many people interact with one another. You could test the whole market's reaction before launching a new product, or analyze the chain effects a single decision would have on an ecosystem.

"What if, before launching a new product, you could simulate the entire launch and see how people would react and how the market would move?"

The technology can also be applied to wicked problems like climate change that require cooperation among many stakeholders — simulating problems where a balance point acceptable to everyone is hard to find because each group's interests differ.

It can also be used to study under what conditions democracy weakens or fails, and what outcomes emerge from elections and policy-making processes.

8. Who Should Be Allowed to Use Powerful Simulation Technology

Harry asks who Simile will work with — and who it will withhold the technology from — in sensitive areas like elections and democracy.

Park answers that principles matter enormously here. Simulation is a powerful technology, and its potential for misuse is correspondingly large. The north star he has in mind for the technology is representation at scale.

People hold differing perspectives and tastes, but their voices often fail to be properly reflected in the rooms where important decisions are made. It is easy to say "we listen to the customer's voice" or "we reflect stakeholder opinion," but actually hearing everyone is hard.

"Simulation is a way to hear people's voices at scale in the decision-making process. That's our north star."

Simile therefore holds that it should be a technology that brings more people's perspectives into decision-making — not a tool for manipulating particular groups or covertly moving public opinion.

9. How Many People Are Needed for Accuracy

If you are analyzing a narrow, well-defined group, social science research can achieve statistical significance with roughly 1,000 people. But Simile's users can filter groups by combining new criteria at any time.

They might combine age, region, occupation, income, and particular experiences to ask about the response of an unanticipated subgroup. Responding to such requests requires including a very broad population composition in the model from the start.

"To let people filter to a specific subgroup on the fly, you ultimately have to be representing the entire population."

Simile's goal is not simply to produce a statistically significant result in one survey but to build a digital representation of an entire population capable of responding to a wide range of questions and criteria.

10. A Data Flywheel Where the Real World Is the Answer Key

Simile's model can improve over time, because they can observe simulation results in the real world, check whether the prediction was right or wrong, and keep improving the model.

Harry compares this to AlphaGo, the Go-playing AI. Just as that model corrected its mistakes over countless games, every event that happens in reality each day becomes a new training case that calibrates the simulation model.

A coding agent has a clear reward signal because it immediately knows whether the user accepted or rejected a suggestion. A model simulating the future, by contrast, has the problem that verification is difficult because results occur later.

But Park says simulation actually has a special advantage.

"The world itself is our answer key. We live inside the ground-truth world."

Simile generates tens of thousands of hypotheses a day and links each one to a verifiable outcome. When that outcome occurs in the real world later, they can check whether the simulation was right.

"We can generate a million hypotheses in a month and check what percentage of them actually happened."

Continuously measuring the gap between simulation results and reality creates a data flywheel: run more simulations, obtain more real-world data, improve the model through it, and have the improved model generate more accurate simulations in turn.

Compute cost is also an important issue, but Simile explains that it has already reduced the cost of running the model to roughly one-hundredth of its initial level — raising efficiency by improving training and inference methods while preserving the core technology and philosophy.

11. Product-Market Fit Confirmed in the Enterprise Market

Simile currently focuses its business on large enterprises. Enterprise customers have sufficient budget, and you can quickly verify whether the product solves real operational pain.

Park says he received the following advice from Pat Hanrahan, the Tableau co-founder who was a colleague during his Stanford days:

"The best way to get good feedback is to make people pay."

Simile is using enterprise market research as its initial entry point, because the budget already exists and customers can feel the value immediately. Ultimately, though, the goal is to use the technology to help society as a whole make better decisions.

Product-market fit was confirmed in the reaction of Fortune 500 boards and top executives who saw the Smallville demonstration.

"If we could simulate the market this way, the very way we work would change."

For about a year Simile verified how accurately the model reproduces the behavior and attitudes of real people. It predicted human responses in surveys, behavioral experiments, and real-world settings, achieving roughly 85% accuracy relative to the level at which people reproduce themselves. When these research results were released at the end of 2024, the market for synthetic panels and human behavior simulation began to attract serious attention.

Contracts with large enterprise customers have moved faster than traditional enterprise sales. Some large companies signed within just three months — a sign of how urgent the problem felt to them.

Simile has also shown results comparable to a consulting study that takes three to six months in about two minutes on an initial call.

"We predicted in just two minutes the results of a study that took three to six months."

What customers want is less a perfectly settled prophecy than fast, sufficiently accurate evidence to give direction to decisions that until now relied on gut instinct.

12. Simulation Creates Value Far Exceeding Its Cost

Harry asks whether, if Simile creates millions or billions of dollars in value for a company, charging merely a million dollars leaves a large gap between value and price.

Park explains that Simile's value lies not only in optimizing revenue but also in preventing hundred-million-dollar mistakes.

"If simulation prevented a $500 million disaster, that's an obvious painkiller from the company's point of view."

Value therefore arises in two ways:

1. Optimization — improving performance by choosing better campaigns, products, and policies
2. Prevention — screening out decisions likely to fail before they are made

Simile also differs from a company that simply shows what will happen in the future. The core is showing not only what outcome appears but how and why that outcome is reached.

"Predicting is one thing; showing what steps an ecosystem goes through to arrive at that outcome is another."

Knowing this process lets a company prevent a bad outcome, or intervene at intermediate stages so that the desired outcome occurs.

13. Running a Research Lab and a Product Company at Once

Unlike a typical startup, Simile is a research lab that also runs product, engineering, and sales organizations. The three co-founders come from Stanford research backgrounds. Park worked on agents and simulation; Michael Bernstein is a researcher in ImageNet and human-centered AI; and Percy Liang is the scholar who coined the term foundation model. Another co-founder, Rainy, leads product and go-to-market.

The reason this structure works is that research results and customer value move in nearly the same direction.

"The better the model represents people, the better the simulations we can build — and that means a more accurate, more realistic experience for users."

At a company where research and business goals conflict, it would be hard to maintain both organizations at once. But at Simile, the model getting better is exactly the product getting better, and the direction customers want meets the direction research is advancing.

14. What Makes an Outstanding Team

Because Park came from a research background and lacked enterprise sales experience, he brought in Rainy as a co-founder to fill the gaps in his own skill set. A great team should combine different strengths rather than replicate one person's abilities.

"A team has to be balanced. There are capabilities I can bring, but there's also a lot I don't know."

He says that as a team grows, it is important to identify and fill the newly appearing gaps in advance. At the same time, the whole team must share values and rigor.

Park compares team building to his past experience as a painter of portraits. Just as the painter's own atmosphere comes through in the figure on the canvas no matter who is modeling, a good team reflects the founder's values and attitude.

The criterion he weighs most when assessing talent is whether that person has been the common denominator of success across multiple stages of life and career.

"At each stage that person went through, was the cause of that thing's success the person themselves? Broadly speaking, were they consistently the common denominator?"

Such a person has a high degree of ownership. Whatever way the situation goes, they think "I am personally going to make this succeed," and they can reinvent themselves when necessary.

Another important criterion is whether a single person holds two opposing superpowers that are difficult to coexist. The best marketing leader, for example, is highly data-driven and scientific while also having a creative, artistic sensibility.

Park describes his co-founder Rainy as someone who worries and stays on guard in the extreme in the short term, yet is certain of success in the long term.

"Today I'm deeply anxious, but in the long run I believe the world is on our side and we'll succeed in the end."

Short-term tension gives an organization urgency and drive, while long-term belief lets it endure hard periods. Having only one of the two creates problems: anxiety alone makes you a pessimist, and belief alone makes you complacent.

Harry agrees that this kind of anxiety is what actually produces success.

"It's believing that today's paranoia and worry help make the future a better place."

15. How to Recruit Research Talent Earning Tens of Millions

Competition for AI research talent is fierce. Among Park's close colleagues and friends are people whose total compensation reaches tens of millions of dollars.

No matter how much investment Simile has raised, matching the base salaries of the large labs outright is difficult. What researchers value, however, is not simply compensation.

They care about vision and social impact. Because they watched OpenAI and Anthropic grow from small research organizations into global companies in just a few years, they know well that an ambitious vision can become reality.

"Researchers think hard about what impact the technology they'd be joining will have on society, and whether the work is genuinely interesting to them."

A leader's role in retaining research talent does not end with hiring good people. You have to build a platform where each person can bring their strengths to bear as fully as possible.

Park says the core team members who had been doing research with him stayed together over the past six years, and that he was ultimately able to bring even his doctoral advisors, Michael and Percy, into Simile. Underlying that were the trust built over long collaboration and the conviction a leader conveys to the team.

16. How a Researcher-Founder Makes the Shift to Entrepreneur

When a researcher starts a company, the most important distinction is whether that person is attached to the problem or attached to the impact.

"I look at whether they're bound to the problem or bound to the impact they'll have on the world."

Some researchers are drawn to a particular problem itself, but that problem may not develop into a good business. A researcher who wants to have real impact on the world, by contrast, will find problems that can reach people and generate revenue.

Someone like that has a high chance of converting research into a business. The key is less whether the research itself is interesting than whether it can reach actual users and markets.

17. Raising $300 Million in Six Months

About five months ago, after raising $100 million, Simile ran a preemptive round taking additional investment from its inside investors. Existing investors saw the market response Simile was generating, the pace of its technical progress, and strong customer demand, and judged that they should invest before it was too late.

Park then reached out to investment teams he had long respected, one of which was Greenoaks. Greenoaks had already researched Simile and the surrounding market deeply and was ready to invest.

In the end Simile raised an additional $200 million, for a total of $300 million over roughly six months. The capital goes toward data, compute, and research talent in order to push the ambitious task of modeling human behavior forward faster.

Park says that in research you cannot control the results themselves, but you can control the inputs and the process.

"You can't fully control the outcome of research. But you can control the inputs and the process."

Because putting in more compute, more data, and stronger research talent accelerates technical progress, taking additional investment was rational even with sufficient capital already in hand.

He admits frankly that before founding a company he did not really understand what venture capital was for. But he came to realize through running the business that a good investor can be an experienced advisor and mentor. Because investors have already seen many times the problems a founder is encountering for the first time, they can offer substantive help at the moments it is needed.

Another lesson is that market interest and investment timelines move faster than founders expect.

"Everything happens a little earlier than you expect."

That said, even in an overheated market, Simile focuses on fundamentals such as customers, technical progress, and real market demand. Its position is that rather than the sheer amount of capital flowing in, what must be verified is whether the model's performance is actually improving and whether customers remain willing to keep paying.

18. The Economics of Simulation and the $100 Million Session

The compute cost required varies by simulation. Analyzing the downstream effects of a complex decision — ripple effects across multiple stages — or performing market segmentation across the entire United States is far more expensive than a simple question.

But the more complex the simulation, the higher the return on investment for the customer, because it deals with decisions whose failure is costly.

Park foresees that simulation could become the new frontier of AI inference spending.

"Within two or three years we could see a world where running a single simulation session costs $10–20 million, but the value of that session is so large that the customer pays $100 million."

Simulations like these are likely to target organizations such as global corporations or governments that could suffer enormous losses from a wrong decision.

The current challenge is not whether the simulation itself can be built. Park believes a minimal proof of concept can be built for most ideas. The genuinely hard problem is converting a proof of concept into a reliable production system.

"The real gap is between a proof of concept and productized technology that creates actual value."

19. AI in Ten Years and the Digital Replication of Human Behavior

Park compares today's large language models to the CPU of intelligence: a single enormous model performing complex reasoning. Simulation, he suggests, can play the role of the GPU of intelligence.

Simile is not building a single superintelligent machine but many models reflecting the personalities, tastes, and biases of different people — simulating both the diversity of individual humans and the social phenomena that arise when they interact as groups.

Harry imagines a future in which everyone has a virtual twin that behaves exactly as they do. Park describes it less as replicating individuals than as a new layer representing society at scale.

"What we're trying to build is representation at scale."

Until now, society has represented people's needs indirectly through governments, corporations, and institutions. But if individual perspectives and collective intelligence can be represented in far finer detail, new policies, companies, and decision-making systems could emerge.

They also discuss the possibility of hedge funds using Simile's simulations to find alpha in markets. Park notes that Simile already has people with quant backgrounds, and he does not rule out eventually creating its own quant fund.

If a perfect simulator and AGI arrive, institutions we currently take for granted — the stock market, for instance — could change as well, because if everyone gains the same level of predictive ability, the very way markets function could differ.

The premise most likely to change going forward may be the belief that "obtaining everyone's perspective is impossible."

"I think a world is coming in which we can build a new layer representing society and collective intelligence."

20. Will Simulation Change the Future of Love and Matching?

Harry jokes that if Simile can simulate human behavior well, dating and marriage could become far more efficient — going on virtual dates with several people simultaneously before ever meeting, and passing the compatible ones on to the next stage.

Park acknowledges the possibility but holds a somewhat romantic view of love.

"I'm a bit of a romantic. I believe love will find its way in my life in a more natural fashion."

He says that for him, the way you meet someone and the journey you share matter. This is because people form trust by experiencing things together and building up memories.

"Experiencing something together and having shared memories is, I think, a fundamental element of how we form trust."

He therefore thinks it more fitting for simulation to help people better understand how relationships form and to support the journey of finding someone to live with, rather than replacing love entirely.

21. Quick Questions and Park's Closing Message

In the final questions, Park said there are many excellent researchers inside large labs whose names are not well known — simply because they do not publish papers or because they work inside large organizations.

As an overheated area in the AI industry, he named new research labs launched without a clear vision.

"A lab without a clear vision of what impact it will have on the world risks ending up as an interesting research project. But it may fail to become a viable company."

The criterion he weighs most in AI investing is, again, data strategy.

"Can you access data no one else can? Do you know how to collect data that is very hard to collect?"

He mentioned that robotics, inference infrastructure, semiconductors, and hardware also hold great potential — noting in particular that as AI models advance, the hardware and inference layer that runs them will grow more important.

Asked about the kindest person he has known, he thought of Mary Wootters, a theory professor at Stanford. In the period after college when he was living in a garage in Palo Alto, without a proper job, attempting a startup, Park sensed that a wave of AI was approaching but had no research experience at all.

He sought advice from many people, and Professor Mary Wootters, an alum of the same university, wrote back. She spoke with him for far longer than the thirty minutes originally promised, explaining how to think about AI and research, and connected him with people he could work with.

"I still don't really know why she did that for me. I think it was genuinely out of kindness."

That meeting became an important starting point for Park's entry into research and his later founding of Simile. Harry sums it up in a sentence:

"Never forget the person who believed in you first."

Conclusion

Simile's core vision is not to build a single AI as smart as a human but to represent, at scale, how people actually think and behave. To that end, Simile is gathering not mere web text but real behavioral data, transaction data, observational data, and data from randomized experiments and A/B tests.

Park believes that what matters most in the AI competition to come is not model size alone but a way of acquiring data that no one can easily replicate. And he stresses that simulation can go beyond a technology for guessing the future to become a tool that helps people make better choices and build the future they want.