Our mission is to simulate all eight billion people in the world, accurately.
Our AI-driven simulations show how and why customers, employees, or populations respond to change.
Every Simile population starts with real people. Our proprietary algorithms design large-scale studies daily, combining the resulting data with proprietary human behavior datasets to train AI models that know human behavior better than anyone on earth.
We validate against real humans weekly: over 7,000 evaluations across subpopulations and real enterprise use cases, covering distributional distance, effect size comparisons, and accuracy against observed behavior. Our validations train an AI confidence model that tags every result with a predicted accuracy level so you know where to trust it.
We train continuously on new behavioral, macro, pricing, and policy data. Agents consume media like the people they represent, encountering news in real time. New populations, including hard-to-reach ones, on demand.
Loyalty data, balance histories, telemetry: your data, opt-in and customer-governed, trains your custom model and closes the action loop, proving customers behave as predicted.
“With Simile, we don’t need to fail fast in front of customers, we can fail safely in a controlled environment. Teams can test, learn, and refine until they’re confident it’s ready for the real world.”
“Simile expanded our qualitative research scope by 15X without losing depth. Speaking with their simulated customers helps us quickly identify areas of opportunity for ongoing research.”
“Simile is a game changer in how we understand customers — accelerating product while helping teams align faster across the organization.”
“We’re partnering with Simile with the goal to accelerate time to market in our product development life cycle. Simile has been a strong partner in our broader focus on understanding our consumers better.”
“Gallup’s partnership with Simile is built on a shared commitment to methodological rigor, transparency, and technology-driven innovation. Together, we're expanding how deeply we understand people, while staying grounded in what they actually think and feel.”
“In a sea of AI tools, Simile is the first to provide a new way for our private equity team to understand consumers with AI.”
Simile was founded by the 
Joon Sung ParkCo-founder & CEO · Creator of generative agentsStanford PhD who pioneered generative agents and population-scale human simulation, including the landmark “Smallville” study and the first simulation grounded in 1,000 real people. Recipient of Stanford CS’s Arthur Samuel Award for best doctoral dissertation.
Michael BernsteinCo-founder & Chief Data OfficerLeading human-centered AI researcher whose work helped define modern social and interactive computing. Stanford HAI Senior Fellow, Sloan Fellow, Tech for Humanity Prize recipient, and creator of the most-cited research in UIST’s history.
Percy LiangCo-founder & Chief ScientistOne of the world’s foremost AI researchers and founding director of Stanford’s Center for Research on Foundation Models. Creator of SQuAD and HELM, and recipient of the Presidential Early Career Award and IJCAI Computers and Thought Award. behind foundation models, generative agent architectures, and rigorous benchmarking practices.
The report introduced the term “foundation models” and established rigorous evaluation as essential to managing their broad capabilities and risks.
Simile is building a foundation model for human behavior, with evaluation and uncertainty treated as core model capabilities.
Language-model simulations exposed large-scale behavioral paradigms that emerge when social computing systems are populated.
Our platform turns a research question into comparable scenarios and shows how the same population responds as concepts, pricing, or messaging change.
Published landmark "Smallville" paper, built a simulated town of 25 agents, and debuted multi-agent architecture with memory, retrieval, reflection, and planning.
These architectural ideas help simulations preserve context and produce coherent behavior across a study.
Interview-grounded agents reproduced participant responses at 85% of human self-retest accuracy and reduced racial and ideological bias versus demographic prompts.
Simile's foundation model is trained on real people, grounding population-scale predictions in human data.
Finetuning on 2.9M responses from 210 experiments improved alignment on unseen studies by 26% versus the base model, outperforming frontier models.
This is the model direction: predict how anyone behaves in any situation, trained and validated on held-out human responses.
