Simile AI

Simile AI builds models and agents that simulate human behavior, allowing organizations to test decisions against modeled individuals and populations. Cofounded by Joon Sung Park, Michael Bernstein, Percy Liang and Lainie Yallen, with Park as CEO, the company combines behavioral-model research with products for concept testing, customer experience analysis and other business decisions. Its public launch in February 2026 accompanied a $100 million Series A; in July it announced a Series B of over $200 million at a $2 billion post-money valuation.[1][2][4][3]

Simile’s model accepts a natural-language description of a population, a situation and a set of valid actions, then returns a probability distribution over those actions. Training inputs include grocery-store and delivery-app transactions, AI-led voice interviews and surveys. The distinction matters: transactions record choices people made, while interviews supply context about their lives and reasons for those choices.[5]

Customers begin with a population of interest, such as people in their twenties and thirties living in California. Park describes recruiting participants with consent and incentives, collecting their data and creating models that customers query through population filters and an experimental environment. Those environments can include surveys, behavioral experiments and A/B tests. Simile trains both population-level and individual-level models, each taking a description of the people being modeled and a stimulus.[1]

Park argues that accurate human simulation requires a different objective from producing an ideal reasoning assistant: a simulated person must reproduce the source person’s biases and mistakes. He describes combining life interviews, observational behavior and randomized controlled trials to learn both patterns of action and responses to interventions. In his account, simulation helps decision-makers ask how a change could produce a desired outcome, beyond forecasting what happens without it.[1]

At launch, Simile named earnings-call rehearsal, litigation-outcome modeling and policy testing as applications. Its July announcement named CVS Health, Wealthfront, Deloitte and Gallup as adopters. Park described Wealthfront as an early customer interested in product testing with multimodal inputs, including images, Figma mockups and agents that use websites.[2][3][1]

In its February 24, 2026 CVS Health account, Simile described agents built on 2.9 million consented responses from more than 400,000 participants across over 200 behavioral scenarios. CVS used simulations to examine customer journeys and test reminder cadences, education content and benefit designs for their effects on intent to refill prescriptions or adopt clinical services. Simile reported that simulated results helped prioritize downstream pilots and experiments; these were tests of intent and interventions before deployment, rather than measured improvements in medication adherence.[6]

Simile added question-level confidence estimates to address a practical limitation: strong average performance does not establish that a particular answer is reliable. In Building confidence in Simile, published August 25, 2026, Andrew Wesel, Sarah Chen and Percy Liang explained a model that predicts a simulation’s error. The error measure is total variation distance (TVD) between predicted and observed action distributions: zero means an exact match, while one means the distributions are disjoint. The product displays estimated error when the real-world distribution is unavailable.[5]

The team compared regression on basic output features, semantic embeddings, probes of the simulation model’s hidden states and fine-tuned models. The hidden-state probe reuses activations from the forward pass that produced the simulation, requiring no extra simulation-model call. A continued fine-tuning approach replaced the language-modeling head of a Simile fine-tune of Qwen3.5-27B with a linear head predicting scalar error, trained using root mean squared error loss.[5]

Simile reported evaluating the confidence methods on approximately 8,600 questions roughly representative of production queries, using fivefold cross-validation with each question’s text and entire experimental sample kept in the same fold. Continued fine-tuning achieved an AUROC of 0.736 for identifying simulations below 0.16 TVD, compared with the probe’s 0.730; the authors described that difference as statistically indistinguishable, although fine-tuning improved error and correlation metrics. The 0.16 threshold came from an internal exercise involving approximately 2,750 ratings by fourteen raters. It is Simile’s evolving definition of “decision-quality,” rather than a universal accuracy standard.[5]

These products followed earlier collaborative research by Park, Liang, Bernstein and other researchers. That work supplies the technical context for Simile’s progression from believable agents in a fictional town to models evaluated against real individuals.[1][7][8]

Park recalled a “time machine game” with Bernstein and Liang: imagining which application would matter most ten years later. Their question became, “Well, what if we can just recreate the world that we live in?” He identified Social Simulacra as a precursor to their generative-agent work. Park also recalled considering personalized assistants, but betting that accurate models of users would be a prerequisite for more capable personal agents.[1]

The April 2023 paper Generative Agents: Interactive Simulacra of Human Behavior was authored by Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Liang and Michael S. Bernstein. Its architecture extended an LLM with natural-language records of experience, higher-level reflections and dynamic memory retrieval to plan behavior. The researchers populated a Sims-inspired sandbox with 25 agents. From a single seeded intention to hold a Valentine’s Day party, agents spread invitations over two days, formed acquaintances, arranged dates and coordinated attendance. The experiment evaluated believable individual and emergent social behavior.[7]

The original November 15, 2024 version of Generative Agent Simulations of 1,000 People moved to agents representing 1,052 real individuals through qualitative life interviews. Its authors were Park, Carolyn Q. Zou, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Robb Willer, Liang and Bernstein. They reported that agents reproduced General Social Survey responses 85% as accurately as participants reproduced their own answers two weeks later. This was a result relative to human test–retest consistency on that survey, rather than 85% accuracy across all human behavior.[8]

Simile publicly introduced its business on February 12, 2026. That day, Index Ventures announced leading its $100 million Series A. Simile named Hanabi, A*, Bain Capital Ventures and individual backers including Andrej Karpathy, Fei-Fei Li, Adam D’Angelo, Guillermo Rauch and Scott Belsky as participants.[2][4]

On July 30, Simile announced raising over $200 million at a $2 billion post-money valuation, naming Greenoaks and Index Ventures as co-leads. The company reported fivefold revenue growth since launch, tens of millions of simulations for Fortune 100 enterprises and more than 50 employees. These were company-reported growth figures; no absolute revenue figure accompanied the announcement.[3]

In the August 21 interview, Park reported about 60 employees, headquarters in San Francisco’s Mission Rock and a newly opened smaller New York office. He described weekly data collection involving tens of thousands of people and panel partnerships providing access to tens of millions globally. That panel access was distinct from the populations actually being simulated. Simile’s stated mission to simulate all eight billion people remained a longer-term ambition.[1][3]