What Simile Sells Is Calibration, Not Simulation

On July 30, Simile, a startup that simulates human behavior for companies, closed a Series B of more than $200 million led by Greenoaks at a post-money valuation of $2 billion. Five months earlier it had raised a $100 million Series A led by Index Ventures. Total funding is above $300 million, and the angels include Fei-Fei Li, Andrej Karpathy, and Adam D'Angelo. The company has just over 50 people, has been selling for under a year, has grown revenue five-fold, and has run tens of millions of simulations for Fortune 100 clients. Its customers include CVS Health, Gallup, Wealthfront, Deloitte, Itaú, and Suntory. What it does fits in one sentence on its website: simulate human behavior. A company hands over what it wants to test, whether a price increase notice, a new package, a piece of ad copy, or a policy change. Simile simulates the target population and runs how those people might react. CVS used it to run 2.9 million simulated responses to validate research conclusions and test medication-adherence programs. One customer used it to rehearse an earnings call and reportedly hit eight of ten analyst questions. The founding team is the easiest part of the company to tell. CEO Joon Sung Park is the first author of the 2023 generative agents paper, the Smallville experiment in which 25 AI agents lived in a virtual town and which has been cited again and again. Of the two other co-founders, Percy Liang is the founding director of Stanford's Center for Research on Foundation Models, and the term "foundation model" comes from the 2021 report he helped write, and Michael Bernstein is a professor of human-computer interaction. A paper became a company in about two years. Which part of research's cost structure got pressed out To learn how a population will react, the traditional route is surveys, focus groups, and A/B tests. These methods are slow and expensive and have two flaws that can't be fixed: the people willing to answer a survey don't represent everyone, and what people say and what they do are often two different things. The industry calls the first sample bias and the second the say-do gap. Both terms have existed for decades, which shows the problems have too. Simile turns "ask a population one question" into an inference call. The cost structure changes, and the nature of testing changes with it. Testing used to be a step before a decision, on a limited budget, covering only the two or three most important options. Now thousands of variants can be crossed with dozens of subpopulations, and follow-up questions can be asked afterward. CVS's 2.9 million responses would have taken years at traditional research pace. The value at this layer is easy to see, and every company in the field tells the same story. The real disagreement is one layer down: why should the simulated reactions be believed?