The simulation company. A foundation model that pretends to be you, so the companies you deal with can practice on a copy first.
It is mid-2026. Somewhere in a Palo Alto office, an AI version of a CVS shopper is deciding whether to buy the off-brand allergy pills. A copy of an equities analyst is drafting a question for next week's earnings call. None of them are real. All of their answers will matter.
Simile does not look like other AI companies. There is no chatbot in a sidebar, no glossy demo of a model writing a poem about ducks. What it sells is closer to a wind tunnel. You bring a decision; Simile populates a room with synthetic humans modeled on real ones, then watches them react.
The lab was founded in 2025 by four people with overlapping CVs and unusually intersecting research interests. Joon Sung Park is the chief executive. Before he wrote his dissertation he painted in oils. Michael Bernstein is a Stanford HCI professor and a co-author of ImageNet, which is a polite way of saying he helped set the table for modern AI. Percy Liang runs Stanford's Center for Research on Foundation Models. Lainie Yallen rounds out the founding team. They are friends, collaborators, and now colleagues at a company that, until February, almost nobody knew existed.
The story that gets repeated - and it is the right one to repeat - is the story of Smallville. In 2023, the team published a paper about a tiny simulated town inhabited by 25 generative agents. The agents remembered things. They planned their days. One of them decided to throw a Valentine's Day party, and the others, unprompted, organized around it. The paper went viral inside research circles and well beyond them. It suggested something previous agent demos had not: that large language models, with the right memory and reflection scaffolding, could produce behavior that looked uncannily social.
The leap from a paper to a company was not obvious to everyone. Simulation, as a category, has been quietly unfashionable in tech for years. Agent-based models lived in academic political science and supply-chain consultancies. Synthetic respondents had a brief, hyped, and slightly embarrassing moment when ChatGPT first opened the door. The objection was always the same: simulated humans are not real humans, and the difference is where the money lives.
Simile's bet is that the gap is closing fast, and that the right way to close it is to treat behavior itself as a foundation model. The team has been training on something stranger than the usual diet of internet text: transcripts from interviews with hundreds of real people about their actual lives, historical transaction data, and decades of behavioral-science literature. The output is meant to be a model whose primary skill is not language - that comes for free - but plausible decisions.
A snapshot of where Simile sits in mid-2026.
Simile sells access to a population, not a chatbot. Here is what enterprise customers are doing with that population today.
CVS Health uses Simile to model how different customer segments respond to inventory and product placement choices before stores actually rearrange anything.
Public companies simulate analyst panels with agents modeled on real coverage to anticipate questions. One reported run hit eight of ten.
Teams stress-test policy changes, pricing tweaks, and litigation scenarios against synthetic populations before live rollout.
Run a concept past a synthetic focus group of thousands overnight. Cheaper than a recruit, faster than a panel, blunter than your designer.
Drop simulated users into a prototype to find friction your beta testers will hit next week.
Telstra, Australia's largest telco, is using Simile to model customer reactions across scenarios.
An approximate look at what Simile's behavior model has been fed. These are illustrative proportions, not exact figures.
Stanford lineage, unusual second careers, and one shared obsession.
A $100 million round that surprised people mostly with how unsurprised they were.
Simile is not alone in trying to make AI feel more human. It is, however, taking the longer route.
A growing crowd of startups - Aaru, Syntheticusers.com, Fairly AI, and others - is pitching a faster, cheaper alternative to focus groups. Most rely on prompting general-purpose LLMs to roleplay personas.
Simile's pitch is that behavior is itself worth a foundation model - trained on interviews, transactions, and behavioral-science literature, not just internet text. Slower to build. Harder to copy.
If you want to hear Joon Sung Park explain the idea in his own words.
Share this page, or follow Simile at the source.
The synthetic CVS shopper has made up her mind about the allergy pills. The copy of the equities analyst has filed his question. Somewhere a real product manager is reading the simulation log and changing her plan.
None of these conversations happened. They will, though, soon enough, and when they do, the people running them will have already practiced. That is what Simile sells - not a prediction, exactly, but a rehearsal hall.
The room with 25 agents has gotten very crowded.