The Stanford team that invented generative agents wants to kill the focus group

In April 2023, a Stanford PhD student named Joon Sung Park uploaded a paper that made a virtual town famous. Smallville, population 25, ran on large language models. Its AI residents woke up, cooked breakfast, gossiped over coffee. Then, without a single instruction from Park's team, they threw a Valentine's Day party. One agent mentioned the idea to another. That agent invited a third. Within the simulated day, the entire town was coordinating dates and decorations. Park and his co-authors called these creations "generative agents." The citation count crossed 10,000 before Park had even finished his PhD. On GitHub, 20,600 stars and counting.

On Wednesday, Park's company emerged from stealth. Simile raised $100 million from Index Ventures, Bain Capital Ventures, and a roster of angel investors that reads like a Stanford AI yearbook. The startup says it has built the first foundation model for human behavior, trained not on internet text but on interviews with hundreds of real people about their actual lives. CVS Health and Wealthfront are already testing it.

The market Simile is entering barely existed two years ago. Now it is crowded. Aaru hit a $1 billion headline valuation in December. Artificial Societies, backed by Y Combinator, runs half a million AI personas. The $120 billion market research industry is nervous, and acting like it.

Here is Simile's problem, and it is not a small one. Park's team published the research openly. All of it. Ten thousand citations means ten thousand copies of the blueprints in someone else's hands. The question is whether the people who drew those blueprints can build the company before everyone who has them catches up.

A model trained on lives, not text

Most AI products start with a general-purpose language model and add a layer of specialization on top. Simile started somewhere else.

The company's foundation model draws on structured interviews with hundreds of real people about their daily lives, preferences, and decision-making patterns. It also ingests historical transaction data and text from scientific journals documenting behavioral experiments. The combination produces what Simile calls generative agents: AI systems calibrated not to general human behavior but to the specific preferences of real individuals.

Nobody else has this kind of academic backing. Late in 2024, Park and his co-founders went bigger. The follow-up study scaled Smallville from 25 agents to 1,052. Each agent was paired with a real human participant who had sat for a lengthy qualitative interview. The researchers then tested how well the agents could replicate their human counterparts' responses on the General Social Survey, a standardized instrument used across the social sciences for decades.

The agents matched their human counterparts with 85% of the accuracy those humans achieved when retaking the same survey two weeks later. Nearly as consistent as the original people. The study also showed reduced accuracy biases across racial and ideological groups compared to agents built on demographic data alone. That finding matters for any company trying to model diverse consumer populations without systematic blind spots.

Simile's product translates this research into an enterprise platform. Companies define a question, Simile builds a simulation populated by agents representing a target population, the simulation runs, answers come back. CVS Health uses it to decide what items to stock and display. Wealthfront says it expanded qualitative research scope by 15 times without losing depth. Suntory Beverages & Food is testing it to accelerate product development cycles. A partnership with Gallup provides access to a nationally representative panel for consumer sentiment modeling.

If you run consumer research for a living, the pitch is blunt. A thousand simulated focus groups in the time it takes to schedule one real one.

When your competitors build from your blueprints

Park's team did what academics are supposed to do. They published. They presented at conferences. They posted the code on GitHub. And every competitor in the behavioral AI space has since built on that published work.

San Francisco-based Aaru pulled off a Series A at a reported $1 billion headline valuation last December. Redpoint Ventures led. Accenture Ventures and General Catalyst participated. The company's pitch: AI agents that simulate how consumers and voters behave, delivering market research in minutes instead of months. Aaru called the 2024 US election correctly. For enterprise buyers shopping for behavioral AI, that kind of track record lands harder than academic prestige.

Artificial Societies, a Y Combinator company out of London, took a different path. It raised $5 million in seed funding and built a platform running 500,000 AI personas. Its clients reportedly include Anthropic and the sales automation company 11x. Same target market. Same $120 billion opportunity.

Both reached market earlier and have paying customers. Simile has the research they all cite but not the sales pipeline they already built.

The pattern is familiar. In the 2010s, several teams rushed to commercialize deep learning for image recognition. The researchers who published the original papers, many at Stanford and the University of Toronto, did not always end up building the largest businesses. Google acquired DeepMind. Facebook hired Yann LeCun. The research created the market. The market rewarded execution.

Simile's founders know this history personally. Bernstein co-authored the 2015 ImageNet Challenge paper. That benchmark set off the deep learning revolution, and the people who wrote it did not capture most of the value. Karpathy and Fei-Fei Li were on that same paper. Both are now Simile investors. They watched one set of blueprints transform computer vision and enrich companies that had nothing to do with the original research. Now they are betting the behavioral AI story ends differently.

The inner circle

Walk into Stanford's Gates Computer Science Building on a Tuesday afternoon and you will find roughly half of Simile's brain trust still holding office hours. That is both the company's credibility engine and its most uncomfortable question.

Index Ventures led the $100 million round. The firm backed Figma, Datadog, and Discord, and its participation here signals institutional confidence. But the angel investor list tells you more about what is actually happening.

Fei-Fei Li invested personally. She co-created the original ImageNet dataset at Stanford and later founded World Labs, which pulled in $230 million in 2024. Karpathy wrote a check too. He co-founded OpenAI, ran Tesla's AI division, and apparently likes backing people he has published with. Both were on that same ImageNet paper with Bernstein a decade ago. The Stanford AI establishment is closing ranks around its own, and the mood is less celebratory than protective. They have watched this movie before. They know how it usually ends for the researchers.

Percy Liang, the third academic co-founder, brings a different form of weight. As founding director of Stanford's Center for Research on Foundation Models, Liang co-authored the 2021 paper that introduced the term "foundation model" to the field. He also created HELM, the Holistic Evaluation of Language Models, which became the industry standard for benchmarking large language models across dozens of scenarios. When Simile describes its product as a "foundation model for human behavior," that language comes from the person who defined what the phrase means.

Lainie Yallen is the fourth co-founder, and the odd one out. She studied business at McGill and started at Boston Consulting Group. After a stint investing at Canadian VC firm Inovia Capital, she joined Hebbia AI as employee number eight and helped push revenue up 15x. The ratio tells the story. Three researchers, one businessperson. Which way that tips, company or overpriced research project, comes down to the next twelve months.

Ten people and a hundred million dollars

Ask about Simile's biggest risk and the answer is not competition. It is headcount.

A recruiting page had the numbers. Four co-founders and three researchers, plus two business hires and an admin. Something like ten people. That number will climb fast with $100 million in the bank. But right now, the gap between the money and the bodies tells you exactly how early Simile actually is.

The academic co-founders add a second constraint. Bernstein is still a Stanford professor. He runs the Symbolic Systems program on an interim basis. Liang directs the Center for Research on Foundation Models. Neither has publicly announced a leave of absence. How much of their time Simile actually commands is anyone's guess.

Stanford spin-outs do this all the time. Faculty keep their positions, sometimes taking formal leave, sometimes advising part-time while a CEO handles daily operations. The arrangement works when credibility matters more than speed. It becomes a liability when competitors are already booking revenue.

And competitors are booking revenue. Aaru, emboldened by its headline valuation and election-prediction credentials, is selling to EY for wealth management applications. Artificial Societies is onboarding clients. Traditional market research firms are layering AI onto their existing platforms. Every month Simile spends hiring is a month its competitors spend closing.

There is a counterpoint, though. Simile's product requires deep technical work that cannot be rushed. A behavioral foundation model trained on real human interviews is a different animal from wrapping an existing language model in a survey interface. If the technology gap holds, and the 85% benchmark suggests it does, then the hiring delay costs less than shipping a weaker product early.

What the focus group's death certificate looks like

The market research industry generates roughly $120 billion a year. Much of that money goes to recruiting participants, scheduling sessions, running interviews. AI simulation could compress most of it.

Prices vary wildly. A mid-tier market session costs about $4,000. Run one in New York and the bill hits $12,000. Finding the right participants? That alone eats weeks. Results are qualitative, subjective, and hard to scale. A company testing a new product might run six groups and base a multimillion-dollar launch decision on the opinions of 60 people.

Simile and its competitors are betting that the same insights can come from AI agents calibrated on real behavioral data, delivered in hours rather than weeks, scaled to thousands of simulated participants rather than dozens of real ones.

If you are a CMO at a consumer brand, the implications go further than cheaper focus groups. Simile says its platform can predict what questions analysts will ask on earnings calls, how the public will react to corporate announcements, and how customer segments will respond to product changes. Call that market research if you want. What it actually replaces is the part of corporate decision-making that runs on gut feel and experience.

Skeptics have reasonable ground. AI agents replicate statistical patterns. They cannot account for genuine surprises. A TikTok video goes viral, a celebrity mentions the product, a supply chain breaks. Consumer preferences shift overnight and the simulation never saw it coming. The simulation captures the world as it existed when the interviews happened. What happens next is where it goes blind.

Academic competition is advancing too. Centaur, published in Nature last July by researchers at Helmholtz Munich, represents another approach to behavioral prediction. Be.FM, from the University of Michigan, tackles similar problems. Neither has been commercialized. But the volume of research suggests the underlying capability is becoming commoditized, which means the moat will eventually depend on data quality and distribution rather than the algorithm itself.

The depth bet

Every competitor in this space has chosen scale. Simile chose depth.

Aaru markets speed with hundreds of thousands of agents. Artificial Societies emphasizes its half-million personas. Simile's approach is slower and more labor-intensive: it starts with in-depth human interviews, trains specialized models on the resulting transcripts, and builds agents meant to replicate specific individuals rather than broad population segments.

The bet is that training data quality determines simulation quality. It mirrors a debate across the broader AI industry, where some companies invest in massive datasets scraped from the internet while others focus on smaller, carefully curated corpora. The research supports Simile's position. The 1,052-person study showed that agents trained on detailed interviews outperformed agents given only demographic descriptions, particularly across racial and ideological lines.

But depth is expensive. Interviewing hundreds of people, transcribing their responses, and training specialized models costs more per agent than scraping public data or generating synthetic profiles. The $100 million will buy time, but if the company cannot demonstrate a measurable accuracy advantage in real enterprise deployments, cheaper alternatives win on price.

Here is the test. Simile must convert academic credibility into enterprise contracts at a pace that justifies the funding before competitors close the gap. CVS, Wealthfront, Telstra, and Suntory are named customers, but named customers during a stealth phase often reflect pilot programs rather than committed revenue. The next twelve months will show whether the depth bet produces retention rates and expansion revenue that validate the model, or whether buyers conclude that good-enough simulation at lower cost serves their needs just fine.

Seven months old. Ten employees. A hundred million dollars. And ten thousand citations that cut both ways: proof they drew the blueprints, proof everyone else has a copy. Simile has the intellectual capital. The market will decide whether that converts to the other kind.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]