Simile AI, led by co‑founder and CEO Joon Sung Park, focuses on behaviorally grounded models and creating "digital twins" of people. In 2024 the company closed a $2 billion Series B round led by GreenOaks and Index Ventures, with notable backers including Fei‑Fei Li and Andrej Karpathy. Simile counts Fortune 100 clients among its customers, including CVS.
The company reports that in use cases its simulations and digital twins can achieve roughly 85–99% fidelity compared with traditional human focus groups or panels; the exact figure depends on the application and target population.
Research roots: Smallville and generative agents
Park’s work builds on the 2023 Generative Agents paper (often called the "Smallville" paper), which demonstrated that AI agents can remember, plan, socialize and produce emergent behaviors. That line of research motivated the question Simile pursues: what if we could simulate the world before making decisions in it?
Method: data types and model design
Simile organizes its data into three buckets:
- Interview data: long, qualitative interviews in which participants tell life stories and motivations; this provides texture about individuals.
- Observational and transaction data: behavioral events such as purchases or scraped web activity that supply baseline statistics on what people actually do.
- Causal data from randomized controlled trials (RCTs): experiments that reveal why people make certain choices and enable modeling of causal mechanisms.
The company develops both population‑level and individual‑level models: population models represent subgroups’ statistics, while individual models produce personalized digital twins.
Result: simulating 1,000 people and ~85% reproduction
In the study "Generative Agent Simulations of 1000 People" (a follow‑up research effort tied to late‑2024 work), Simile recruited 1,000 US‑representative participants into a virtual lab. After roughly two hours of data collection (interviews, behavioral measures, etc.), participants left and returned two weeks later to complete surveys, behavioral tasks and experiments. Researchers used the initial data to create digital twins and asked those twins to predict how the source individuals would respond in the later tasks.
The digital twins reproduced participants’ attitudes and behaviors at about 85% accuracy — meaning the models approximated how well people reproduce their own responses in these controlled assessments.
Why frontier LLMs fall short on human behavior
Park and Simile argue that contemporary frontier large language models (e.g., ChatGPT, Claude) are excellent at language and reasoning but did not learn the full mapping of humanity’s "social physics." Web‑trained models primarily reflect attitudinal and self‑expressed data rather than real, stake‑based behavior and causal mechanisms. Simile’s goal is to model human biases, mistakes and mundane routines as well as rational choices — sometimes requiring post‑training of model weights rather than only prompt engineering.
Use cases, clients and ethical guardrails
Simile’s customers use simulations to better understand target populations. Typical workflows start by defining the population of interest (for example: US general population or a specific demographic slice), recruiting consented participants, collecting data, building models and then running queries such as surveys, behavioral experiments or A/B tests. Common early use cases include concept testing, focus groups and product or UX testing; the company has worked with customers including CVS and Wealthfront, and has a strategic partnership with Gallup.
Simile emphasizes consent and data governance: participants knowingly share data and are incentivized for participation. The company says it will be cautious about deployment in politically sensitive domains and build guardrails around how simulation capabilities are used.
Scaling, costs and the long‑term ambition
Simile collects tens of thousands of participants’ data weekly and has panel partnerships reaching tens of millions globally. For many enterprise use cases, models built on thousands to hundreds of thousands of participants suffice; the long‑term technical ambition is to simulate up to the world’s population of roughly 8 billion people.
Park reports early evidence of a scaling law for simulation: more human data and compute yield steady, predictable improvements in simulation performance. But running multi‑agent, society‑scale simulations could demand very large compute and storage — potentially data‑center‑scale resources and substantial costs — even as the economic value of better decisions can justify large investments.
From product testing to societal questions
Beyond market research replacement, Simile positions simulation as a way to test products and policies before deployment, to find counterintuitive intervention paths, and to study emergent social behaviors. Park highlights that robust simulations could help explore complex societal issues such as climate change coordination, early warning signals of democratic instability, or questions like universal basic income.
He cites Thomas Schelling’s agent‑based modeling legacy (notably segregation models) as intellectual precedent: with generative agents and richer data, modern agent‑based models could become precise and high‑fidelity enough to inform policy.
Company structure and team
Simile blends research lab and product company. Founders include Michael Bernstein, Percy Liang and Joon Sung Park (research backgrounds) and Lainie Yallen (business lead). The company has about 60 employees, with headquarters in San Francisco and a New York office; team members come from labs and companies such as Microsoft Research, OpenAI and others.
Summary
Simile AI aims to build behaviorally grounded digital twins by combining interviews, transactional/observational data and randomized trials, and to run representative multi‑agent simulations for business and policy decisions. Its published research reports that, in controlled studies, digital twins reproduced human responses at roughly 85% fidelity; the startup has raised a $2 billion Series B, serves large enterprise clients, and views simulation as a route from product testing to addressing society‑scale problems.
