Joon Sung Park: How Simile Simulates Human Behavior With 85% Accuracy

Joon Sung Park tells Molly O’Shea that Simile can predict how real people will respond with 85% plus accuracy, and that its customers care less about forecasts than about what to change. The Simile co-founder and CEO, whose Stanford research on generative agents became the company, joined Sourcery at Index Ventures in San Francisco to explain Smallville, the 1,000 person study behind that headline number, the 40 million person panel Simile draws on, how it keeps bad survey data out, and why he thinks simulation is in its GPT-4 era.

TLDW

Simile, reported at a $2 billion valuation after raising $300 million, is building a foundation model of human behavior that takes a description of a person or group and simulates what they would do in a given context. It grew out of Smallville, a 2023 game town of 25 AI agents, and a later study where interview-trained twins of 1,000 representative Americans predicted their own survey and behavioral responses. Park says the hard part is modeling the gap between what people say and what they do, which he ties to optionality and opportunity cost, and that every simulated answer ships with a 0 to 100 confidence score. Customers such as CVS and Banco Itaú use it for concept tests, conjoint studies and messaging, but Park says the real demand is causal: not what will happen, but what to do now to change it.

Thoughts

Simile is not really selling prediction, and Park says so himself near the end of the interview. He admits it surprised him that customers are “not that interested in predicting the future.” Tell a company its market will crash in two quarters and its next question is what to do now to stop it. That reframes the whole business. A forecast is a product you check once and judge by whether it came true. A causal model is a lab you return to every week to test a new price, a new message or a new feature. Park says this is why Simile prizes randomized controlled trial data, where one variable changes and the behavior shift is measured, over observational data like transactions and mobility. It is also where the risk sits. A simulated intervention is most trustworthy when it resembles experiments already in the training data, and least trustworthy for the genuinely new idea a company most wants to test. If the confidence score drops sharply for novel interventions, the system is working as it should. If it does not, customers are getting fluent guesses dressed as evidence.

That confidence score may matter more to Simile’s future than the 85% headline. Park describes a second model, trained jointly with the simulator, whose only job is to output a 0 to 100 probability that a given answer is right, and he compares it to a p-value. Customers, he says, treat scores above 85% or 90% as robust enough for high-stakes decisions. The comparison is loose. A p-value tells you how surprising data would be under a null hypothesis, while this is a model grading another model’s output, and its worth depends entirely on calibration that buyers cannot see. The headline number needs the same care. The underlying Stanford paper measures accuracy against participants’ own consistency when they retake survey items two weeks later. Agents scored between 82% and 86% of that ceiling, while agents built from demographics alone still reached 74%. The interviews earn real gains, but the jump over a census profile is smaller than “85% accuracy” suggests on its own.

The electric vehicle story is the best argument for simulation, and also a reminder of how little proof the episode offers. O’Shea asks for the most expensive mistake Simile helped a customer avoid. Park answers with work from before the company existed: a carmaker’s EV messaging that tested as a clear winner, but in simulation shifted how people saw non-electric vehicles and lowered sales across the whole market. That kind of second-order effect is exactly what a focus group misses, because it looks at one message in isolation rather than at a market reacting. Park’s broader point, that the value is in modeling the environment and the alternatives people have and not just the people, holds up. But the client is unnamed, the result predates Simile, and it ran alongside conventional studies. For a $2 billion company with CVS, Banco Itaú and Fortune 100 customers, a named, measured, post-launch outcome would be far more persuasive than a research anecdote.

The most practical idea for anyone who runs surveys is Park’s account of why people’s answers and actions diverge. Stated intentions, he says, follow something like a normal distribution: asked whether he will go to the gym this week, he answers five or six on a seven-point scale every week. Actual behavior follows a much steeper curve, because the activation energy to really go is higher than anyone expects. What drives the gap is optionality. Surveys present a confined set of choices, while real life offers alternatives with opportunity costs, such as one more hour of sleep or one more meeting. His water example makes the same point at market scale: a sole water source could charge a fortune, and only the presence of alternatives keeps the price down. If Simile can learn that translation layer, from what people say to what they do, it will have built something traditional polling cannot, and it explains why the company models the environment as well as the person.

The episode’s best moment of self-demonstration is also its smallest. Park simulated the interview before walking in, and the model warned him O’Shea might open with whether we live in a simulation, which she did. She points out she tweets at Elon Musk a lot, which makes the prediction less magical, but it shows the product in the kind of setting where it is most useful: preparation, not prophecy. The closing exchange fits the same pattern. Asked what he would do with a $15 billion offer, Park says Simile will pursue any ambitious idea only it can pursue, which is a polite way of saying no without saying it. His answer on careers, that the people who last see winning one game as an invitation to the next, is the clearest statement of what he is trying to build.

Key Takeaways

Simile is building a foundation model of human behavior that takes a description of a person or group and simulates what they would do in different contexts.
The company has raised $300 million in total at a reported $2 billion valuation, with Greenoaks and Index Ventures among investors and Fei-Fei Li and Andrej Karpathy as angels.
Smallville, run in 2022 and 2023 on GPT-3 and GPT-3.5, put 25 agents with memory, planning and reflection into a game town, where they organized a Valentine’s Day party without being programmed to.
The follow-up study recruited 1,000 representative Americans, interviewed each by voice with an AI agent, trained a twin of each, and two weeks later had the twins predict their surveys and behavioral studies with 85% plus accuracy.
Park says stated intentions follow a normal distribution while real behavior follows a much steeper curve, and the gap is driven by optionality and opportunity cost.
Some traits, such as risk tolerance, are not known to change, which makes them highly predictable once learned.
Simile runs concept tests, conjoint analysis and MaxDiff studies in simulation that would traditionally cost hundreds of thousands to millions of dollars.
Customers span retail, finance, entertainment and consumer goods, including CVS for store layout and messaging and Banco Itaú in Brazil.
Simile keeps vertical-specific agent banks ready on day one and can add new panel members within 24 hours through partnerships covering over 40 million real people, including Gallup.
Panel quality is protected by spacing out how often people are surveyed to avoid fatigue and gaming, and by attention checks.
The only excluded group is people who cannot give informed consent, including minors, which Park calls a blind spot the company owns.
Every simulated answer comes with a 0 to 100 confidence score from a jointly trained model, and customers treat 85% to 90% as robust enough for high-stakes decisions.
Simile has about 100 employees, with 20% to 25% on the science side, and runs a forward-deployed model Park borrows from the coding agent cycle.
Park says simulation is in its GPT-4 or GPT-4.5 era and expects it to reach where coding agents are today in about 18 months to two years.

Chapters

0:44 Simulation Theory and Smallville

O’Shea opens with Elon Musk’s view that we may live in a simulation, and Park says he thinks we do, because any society will want to understand itself and the best way to understand something is to recreate it. He describes societies as built from individuals with stable traits, whose interactions produce complex and chaotic outcomes. Smallville was his first attempt at that idea: a game town in the style of Pokémon or The Sims, populated with 25 agents that woke up, went to work and formed relationships. Set the day before Valentine’s Day, it produced a party and invitations nobody had scripted, and introduced the agent architecture of memory, planning and reflection.

6:46 From Believability to 85% Accuracy

Smallville’s agents were tested by interviewing them about their lives, who they wanted to spend time with and what gift they would buy, then having human evaluators judge whether the answers seemed humanlike. Park calls that stage believability, or the illusion of life. The next study aimed for accuracy: 1,000 representative Americans were interviewed voice to voice by an AI agent about their life stories, and their answers trained digital twins. Two weeks later the people completed surveys and behavioral studies, and the twins predicted their responses with 85% plus accuracy. Park stresses that simulation has to be grounded in and evaluated against real people.

11:09 Is Human Behavior Predictable?

Training data includes behavioral knowledge from economic games and randomized controlled trials, which covers many irrational tendencies. Park says some traits, like risk tolerance, barely change, so the modeling challenge is translating observed attitudes and behaviors into fundamental tendencies. He illustrates the say and do gap with the gym: his stated intention is steady, but the activation energy to actually go is much higher. Asked what drives decisions, he points to optionality, the alternatives and opportunity costs people face in real life that surveys leave out.

15:45 How Simile Simulates Customers and Markets

Simile’s simpler use cases are surveys, concept tests, conjoint analysis and MaxDiff studies that can now run almost instantly instead of over months at high cost. Customers then ask for what studies could not do, which requires modeling the environment as well as the people. Park’s example is water: a single source could charge a million dollars, and only the alternatives available keep the price low. Simile models those alternatives and asks agents what they would pay given the full context.

20:05 Customers, Panels and Keeping Bad Data Out

Retail is Simile’s largest vertical, led by CVS, alongside large financial institutions, Banco Itaú in Brazil, entertainment and consumer goods companies. Simile maintains vertical-specific agent banks and collects more data when a customer needs a niche population, drawing on panel partnerships with over 40 million people, including Gallup’s long-running panel. Park says it fights survey fatigue by spacing out questions and uses attention checks, and O’Shea’s suggestion of asking how tall a giraffe is turns out to be a fair one. The only people excluded are those who cannot give informed consent, including minors, because Simile wants to capture good and bad behaviors across all of society.

29:29 Simulating This Episode and the Confidence Score

Before the interview, Simile tested titles and topics for Sourcery’s audience of investors and tech executives. The top simulated title was about the $2 billion company building AI simulations of real people, and digital twins, simulated customer testing and privacy ranked closely as topics. Park explains that customers use a SaaS platform where they choose who to ask and what to ask, with no limit on questions, plus hands-on help deciding when to trust the results. Each answer carries a 0 to 100 confidence score from a jointly trained model, which he likens to a p-value.

33:34 The EV Campaign That Looked Perfect Until It Wasn’t

Asked for the most expensive mistake avoided, Park describes work from before Simile with several large automakers weighing electric vehicles. A messaging strategy tested as a clear winner, but simulation showed it shifted perceptions of non-electric vehicles and lowered sales across the whole market. He says the findings he values most are these holistic, counterintuitive ones that look past a single narrow question.

37:02 Climate, UBI and Simulating the Interviewer

Park’s research hero is Thomas Schelling, whose model of segregation used simulation to study a generational problem. Today he would simulate climate change, which he sees as a coordination problem among actors with competing incentives, and universal basic income, which could be simulated rather than tested in five-year studies. He also simulates his own conversations, and before this one the model warned that O’Shea might ask whether we live in a simulation.

40:13 Investors, Data and a Team for Simulation’s GPT-4 Era

Many angels were colleagues: Fei-Fei Li works closely with co-founder Michael Bernstein, and co-founder Percy Liang sat on Andrej Karpathy’s dissertation committee. Park says the interview method came from his own podcast appearances, where good hosts learned a great deal about him in an hour or two, and life stories reveal the most. Simile has about 100 employees, a science group of 20% to 25%, and a forward-deployed team led by co-founder Lainie Yallen. It also helps companies with retention and recruiting, and Park expects simulation to reach today’s coding-agent maturity in 18 months to two years.

49:43 Success, Failure and the $15B Question

Park admires people who are the common denominator of success and who combine superpowers that should not coexist, naming his co-founders and MIT’s Sandy Pentland, who reinvented himself as a physicist, AI researcher and computational social scientist. Failure, he says, often comes from optimizing for one fundraise, customer or paper, winning it and then asking “now what?” Offered $15 billion, he says Simile would keep pursuing any ambitious idea only it can pursue.

Notable Quotes

“The best way to understand something is to recreate it.”
Joon Sung Park, on why he thinks we live in a simulation

“And none of this was pre-programmed.”
Joon Sung Park, on Smallville’s Valentine’s Day party

“We believe strongly that the simulation has to be grounded and it has to be evaluated alongside people.”
Joon Sung Park, on the data behind Simile’s models

“Optionality and the type of optionality often really governs people’s behaviors.”
Joon Sung Park, on what drives human decisions

“So there’s actually no population that we should avoid. We actually want to know every single population that’s out there.”
Joon Sung Park, on who Simile collects data from

“People are actually not that interested in predicting the future.”
Joon Sung Park, on why Simile focuses on causation

“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?”
Joon Sung Park, on long-term thinking

Watch the full conversation here.