Joon Sung Park on Simulating Human Behavior and Building Simile

Joon Sung Park recalls a 2023 research project that asked whether language models could support believable agents in a virtual town. The researchers populated a game environment with 25 characters who had daily routines, remembered interactions and formed relationships. Park describes the results as an early experiment, not a validated model of an entire real society.

The simulation was set before Valentine’s Day. Park says the characters organized a party and decorated a café through their interactions. The episode illustrates the behavior observed in this particular simulated environment; it does not establish that real people would necessarily organize in the same way.

The research paired an early language model with explicit memory, planning and reflection. Park says multiple agents needed memory so that they could recognize one another instead of treating every meeting as a first encounter. These design choices helped the researchers explore longer-running interactions.

Initially, the team stored agents’ experiences as natural-language text. The difficulty was deciding what to retrieve and how to compress a growing history into the model’s limited context. Even a small town can generate far more events than an agent can attend to at once.

Park describes reflection as periodically asking an agent to examine several memories and form higher-level interpretations. Repeated visits to a café, for example, might lead it to infer a preference or a busy schedule. These are generated interpretations within the system, not direct evidence of a human mind.

Stebbings asks what a simulation company does today. Park says Simile aims to build a model of human behavior that can represent individuals, subpopulations and eventually interactions across markets. He presents the broader ecosystem goal as an ambition, rather than claiming it is already solved.

Park contrasts this aim with models optimized for strong performance in coding, mathematics and other rational tasks. His company wants a simulated person to reflect human preferences, biases and even mistakes in the relevant context. Matching a particular population’s behavior remains an empirical question.

Asked about the gap between what people say and do, Park says Simile collects transaction and observational data as well as data through customers and vendors. He argues that written web data contains a great deal of expressed opinion, while actual behavior may reveal something different.

Park says observing behavior can help with prediction, but his customers often want to change an outcome rather than merely anticipate it. His example is a retailer trying to respond to a projected sales decline. A useful tool, in his view, should compare potential decisions and their consequences.

He describes randomized experiments and A/B tests as inputs for learning how behavior changes across conditions. The goal is to reason about counterfactuals rather than rely only on correlations. The interview does not establish that a model can infer causal effects reliably in every new setting.

Data collection has two challenges, Park says: recruiting people who represent the population of interest and asking questions that reveal meaningful context. Simile looks for ordinary participants rather than only technical experts, but representation depends on sampling, coverage and the question being asked.

Some interviews begin with a participant’s life story, including where they grew up and difficult decisions they faced. Park sees this background as useful for modeling a person’s perspective. The discussion does not detail the privacy or consent arrangements for every data source.

Stebbings challenges the claim that prediction itself is unimportant: knowing demand will fall can affect a company’s purchasing plans. Park agrees that forecast information is useful but says the underlying need is to decide what action to take in response.

Asked whether Simile is simply a new survey tool, Park says its ambition is a more general representation of people that can be used across questions. He imagines testing a product launch or interactions among many people, while acknowledging that these are broader future directions.

Park suggests simulations might eventually help people reason about collective-action problems such as climate policy or democratic institutions. He presents this as a research vision, not evidence that the company can solve those problems or accurately forecast political outcomes today.

Stebbings asks about access and potential misuse, especially by governments. Park says powerful simulation technology carries genuine misuse risks and describes a goal of representing perspectives often absent from decision-making. The interview does not specify a complete access policy or demonstrate that a simulation can substitute for direct participation.

On sample size, Park says more participants can help researchers examine narrower subpopulations. He gives roughly 1,000 people as an illustrative social-science figure for some focused studies, but the number required depends on sampling, effect size, design and the population; it is not a general accuracy guarantee.

Park expects a feedback loop in which simulations are compared with outcomes and used to improve later versions. He describes generating hypotheses, mapping them to observable statements and checking what happens. Whether that improves performance depends on the quality and representativeness of the tests.

He compares the problem with feedback in coding systems, where accepting or rejecting a suggestion can offer a relatively clear signal. For simulations, many outcomes take time and not every counterfactual can be observed. Park proposes using the real world as a source of ground truth for answerable hypotheses.

Compute matters as well. Park says Simile’s production model became about 100 times cheaper to run than an earlier version after changes to its modeling approach. This is a company-reported comparison; the interview does not provide enough detail to independently assess the methods or equivalent quality.

Stebbings asks whether the addressable market is limited to large enterprises or could eventually include consumers. Park says enterprise customers currently offer budgets and a fast feedback loop. His broader ambition is to make simulation useful to more of society.

Park recalls advice from Tableau co-founder Pat Hanrahan that asking customers to pay can generate useful feedback. He describes market research as Simile’s initial enterprise entry point, while saying that reaching a wider audience is a future objective.

Park says interest from executives following the virtual-town demo encouraged the team to test accuracy. He cites research reported in 2024 in which models reproduced participants’ responses at about 85% of the consistency with which people reproduced their own responses. That comparison is specific to the study and its tasks, not a universal prediction accuracy.

Asked whether synthetic panels will exceed the current human-panel market, Park predicts they could expand the number of questions organizations can afford to explore. He argues many studies are never run because of time, budget or practical limitations. The proposed market growth remains his forecast.

Park says simulation might help teams consider hypotheses before making decisions in the world. The value of such testing depends on whether the simulated population and outcomes are validated for the decision at hand; an appealing simulation alone is not proof of a real-world effect.

Stebbings asks how Simile balances research and commercial demands. Park says the company combines researchers with engineering and go-to-market leadership, and believes improved models can serve both the research agenda and customers. That alignment is his explanation of the company’s strategy.

Park describes a faster enterprise sales process than he initially expected after leaving Stanford. He says some large customers were willing to try the product within months because existing research methods were slow or expensive. These are his account of the company’s experience, not a universal sales-cycle benchmark.

According to Park, customers want both speed and accuracy. He recounts cases in which a customer compared a simulation with findings from an existing study and received an answer in minutes. The interview does not disclose the underlying studies, evaluation protocol or independent accuracy results for those examples.

The two discuss how a company should price a tool if it can affect high-stakes decisions. Park says customers may value both optimization and avoiding damaging decisions. His examples of potential losses avoided are illustrations of the proposed value, not independently verified savings.

Stebbings compares simulation with prediction markets. Park says his ambition is to explain potential pathways and how interventions could change them, rather than only estimate an outcome. Such explanations are hypotheses that still require evidence and testing.

On team building, Park says complementary skills are essential. He identifies enterprise selling as an area where he needed his co-founder’s experience, while he brought research expertise. As an organization grows, he looks for new gaps in capabilities before they become bottlenecks.

He also wants shared values and rigor. Drawing on his experience as a figure painter, Park says a carefully built team often reflects something of its leader. He looks for people who take ownership and have contributed to success across different stages of their careers.

Park describes his co-founder Michael Bernstein’s movement between research areas as an example of reinvention. He values the ability to carry a recognizable standard of work into new problems, while the specific interpretation of an individual’s contribution remains his own.

Another hiring signal is the combination of abilities that do not always appear together. Stebbings gives the example of a marketing leader who combines rigorous analysis with creative imagination. Park says complementary strengths can make a person particularly effective.

Park describes one of his co-founders as urgent and cautious about immediate risks, yet deeply convinced of the long-term mission. He believes holding both attitudes can counter complacency without losing conviction. This is a leadership philosophy, not a rule that applies to every founder.

Stebbings responds that short-term worry can itself drive preparation and improve the chance of a good outcome. The two agree that urgency and optimism can reinforce each other when a team translates concern into useful work.

Asked about expensive AI research talent, Park says some colleagues have compensation packages in the tens of millions, but a startup cannot always match large labs’ cash offers. He emphasizes the appeal of an ambitious research vision and a chance to work on problems researchers care about.

Retention remains a concern. Park sees leadership as creating an environment where people can use their strengths and trust one another. He says members of his core research group stayed together across multiple projects, and that his former doctoral advisers joined him as co-founders.

Stebbings asks how investors can distinguish an academic company from an interesting research project without a viable business. Park looks for researchers motivated by real-world impact as well as the intellectual problem, including a willingness to find users and revenue.

Park recounts the company’s fundraising: a prior $100 million round followed by a $200 million round, bringing total funding to $300 million in roughly six months, according to his account. He says existing investors initiated the later conversation and Greenoaks joined.

Although the company had recently raised money, Park says more capital could fund data collection, compute and recruiting. He distinguishes inputs that a team can increase from research outcomes it cannot guarantee. The funding figures are company-reported, and added spending does not ensure a breakthrough.

Park says he initially questioned what venture investors contributed beyond capital. As a first-time company builder, he came to value trusted mentors and introductions, including an introduction to a co-founder. His account reflects his experience with particular investors.

He also learned that financing opportunities can arrive sooner than a founder expects. Stebbings asks whether the market has become overheated. Park agrees that parts are frothy and says he focuses on customer demand, technological progress and other fundamentals.

The cost of a simulation varies with complexity, Park says. A broad segmentation study or a chain of downstream scenarios may require more computation, although he believes those questions can be more valuable to customers. The interview does not independently establish the return on those simulations.

Park speculates that in two or three years an especially complex simulation could cost $10–20 million to run and be sold for $100 million. This is an illustrative prediction about a possible future high-end market, not a current price, actual sale or demonstrated customer return.

Asked what will become possible in three years, Park says the key gap is not merely building a proof of concept but evaluating it and making it reliable at scale. The earlier virtual town showed one form of multi-agent simulation; demonstrating decision-grade performance is a separate challenge.

Park describes a thought exercise with colleagues: imagine a surprising technology ten years ahead and ask what can be built now. The exercise helped motivate the town experiment. He says improved models, evaluation and scalable infrastructure are all needed to move from demonstrations to production.

Using a CPU-versus-GPU analogy, Park contrasts a single highly capable reasoning system with many models intended to represent diverse people. He is interested in what might emerge when simulated agents interact, but frames the idea as a vision rather than a demonstrated replica of society.

Stebbings asks whether everyone might eventually have a digital twin. Park instead emphasizes representation at scale: he imagines using models of diverse perspectives to inform decisions. The discussion does not establish that a simulated person can faithfully represent an individual or confer democratic legitimacy.

The host asks whether financial firms could use simulations for an investing advantage. Park floats the possibility of a future quantitative fund, noting that some team members have quantitative backgrounds. He presents it as an idea, not an announcement or evidence of trading performance.

In a thought experiment about a perfect simulator and advanced AI, Park says many assumptions about markets and representation might change. The scenario explicitly depends on capabilities that do not currently exist in the discussion and should not be read as a prediction of inevitable outcomes.

Stebbings asks about simulated matchmaking. Park says he personally values meeting someone organically, shared experiences and the trust that develops through a journey together. He sees the process of building a relationship as important in its own right.

In a quick-fire round, Park says many capable researchers at large labs are not publicly recognized. He worries that new labs without a clear vision for impact may become interesting research projects without sustainable businesses. These are his assessments, not a judgment about any specific lab.

Park again emphasizes a distinctive data strategy when considering AI investments. He identifies robotics, inference infrastructure and hardware as interesting areas, while noting that robotics has already attracted substantial capital. His comments are opinions rather than investment advice.

Asked about kindness, Park recalls moving to Palo Alto after college while an early startup was struggling. He wanted to enter AI research but lacked research experience. Stanford professor Mary Wootters spent time advising him and introduced him to people who helped him get started.

Park says those early mentors took a chance on someone who did not look like an obvious candidate. He remains grateful for their advice and introductions. Stebbings closes by emphasizing the importance of remembering the first people who believed in a project.