The Stanford Paper That Became a $2B Company in 5 Months

01 /The Concept

The 2023 Stanford paper that became Simile's foundation — "Generative Agents: Interactive Simulacra of Human Behavior," Best Paper at UIST 2023 — depicted AI agents living out 24-hour schedules in a simulated town: making breakfast, attending meetings, organizing parties, spreading gossip, without explicit instructions governing any of it. The paper's contribution was not that AI could simulate leisure. It was that sufficiently detailed agent design produces emergent social behavior that resembles how communities actually function. Joon Sung Park, the paper's first author and now Simile's CEO, turned that observation into a prediction machine for individual consumer behavior.

The product is a library of synthetic user populations. A brand manager at CVS Health does not recruit a panel of 500 consumers to test a loyalty program concept. She specifies the demographic profile, configures the scenario, and receives behavioral predictions within hours. The confidence model Simile has trained — which scores the reliability of each simulation against real-world ground truth — achieves 85% accuracy on the General Social Survey, the closest thing consumer research has to a standardized benchmark. That number converted institutional buyers. Deloitte, Gallup, Wealthfront, and CVS Health are all current clients; CVS Health Ventures joined the B round.

Five months after emerging from stealth with a $100 million Series A led by Index Ventures, Simile closed a $200 million Series B at a $2 billion valuation, led by Greenoaks with Bain Capital Ventures, Hanabi, and CVS Health Ventures participating. Revenue grew 5x in that interval. Tens of millions of simulations have run. The company has 50 employees.

02 /Why It Matters

Market research is a $90 billion industry organized around a supply chain that is now entirely optional. Panel recruitment, fieldwork management, qualitative coding, statistical analysis — each step was necessary when the only way to learn what consumers think was to ask them. The research firm that charges $50,000 for a six-week quantitative study cannot compete on speed with a platform that runs the equivalent study in four hours, even if its sampling methodology is more rigorous.

The more structurally important shift is not cost reduction but hypothesis volume. A product team at a consumer technology company might generate 50 questions about a feature per quarter and have the budget to commission research on three. The constraint is not curiosity; it is the cost and duration of running a study. When the cost of testing a hypothesis drops by two orders of magnitude, the 47 studies that were never run become economically feasible. Simile's growth path is not replacing the $50,000 study — it is creating the market for the research that never happened.

Gallup's participation as a paying client is the most significant signal. Gallup has been surveying people for 85 years; it is the category incumbent in public opinion research. That a firm whose entire business is asking real people what they think is paying Simile to simulate what those people would say suggests the methodology has cleared a bar that internal skepticism normally kills. The research firm that is most exposed to displacement by synthetic respondents is the one that has already adopted them.

03 /Opportunities

Longitudinal simulation is the capability Simile's architecture makes available but has not announced. Current deployments answer a static question: how will this population respond to this stimulus today? The richer question — how does a population's behavior evolve over 12 months if stimulus A is introduced in month 3 and stimulus B in month 7? — requires a traditional research firm to maintain a live panel for a year at costs that limit it to the largest research budgets. Simile could run the same study in days. That creates a product category — dynamic market simulation over time — that has no prior analog and a clear buyer in every brand manager planning a multi-phase campaign or product launch sequence.

The integration that changes when research happens in the development cycle is the second opening. Product teams using AI agents to draft specs, write copy, and design features currently have no way to get synthetic user feedback in the same workflow. A Simile integration that sits inside a product management tool — triggered by a draft feature spec, returning predicted user reactions within the same session — brings qualitative signal to zero marginal cost per hypothesis and changes the decision calculus for what gets tested before it gets built. Teams that test earlier ship better products, and they will pay for the capability regardless of whether they also run standalone studies.