The $2B AI Startup Simulating Human Behavior with 85% Accuracy

Summary

• Simile’s “85%-plus accuracy” claim comes from a study of 1,000 real people whose agents predicted responses to surveys, behavioral studies and stimuli. Participants supplied life-story interviews, then returned two weeks later for testing; Park dates the result to the end of 2024. His operating principle is that “the simulation has to be grounded” and evaluated alongside people.
• The commercial wedge is research that becomes fast enough to run whenever a decision arises. Park says concept testing, conjoint analysis and MaxDiff studies costing hundreds of thousands or millions of dollars over years become “just a query away.” Customers include CVS and Banco Itaú, with applications spanning store layouts, messaging and customer understanding.
• Predictive value depends on modeling alternatives and causation, including the gap between stated intentions and action. Someone may consistently rate likely gym attendance five or six out of seven, yet choose another hour of sleep instead. Randomized experiments matter because customers ultimately ask, “What do we need to do now to shape the future?”
• Simile’s panel partnerships reach over 40 million people globally, with new population needs usually met within 24 hours. Park says the operation talks to tens of thousands, if not hundreds of thousands, weekly. Participation requires informed consent and the ability to withdraw; excluding underage people leaves “one blind spot that we own.”
• Simile pairs predictions with confidence estimates to help customers calibrate consequential decisions. A companion model produces a zero-to-100 estimate; Park says customers find confidence above “90% or usually 85%” robust enough for high-stakes decisions. Deployment includes helping users determine when to validate findings or collect more data.
• With $300 million raised at a reported $2 billion valuation, Park’s defensibility thesis combines behavior models with close customer deployment. He places simulation in its “GPT-4, GPT-4.5 era,” useful today and changing monthly. He expects the industry to reach today’s coding-agent maturity in about a year and a half to two years, with customer relationships becoming “stickier” through forward deployment.
• Park’s longer-term ambition includes potentially using simulations to evaluate societal interventions. Climate policy involves competing incentives; if AGI is here or imminent, he asks whether simulations could explore where UBI works without waiting for a five-year study. Asked about a $15 billion acquisition offer, he stays conditional: if only Simile can pursue an ambitious idea, “then we’re going to pursue that.”

Deep dive

1. Smallville made human simulation plausible

• Park’s opening “I do think we live in a simulation” rests on a civilizational argument: societies want to understand themselves, and “the best way to understand something is to recreate it.” He calls simulation a societal portrait. Construction starts with individual tastes and behaviors; put those people together, and complex outcomes emerge through their interactions.
• The 2022–2023 Smallville work began with GPT-3 and GPT-3.5, when models were “still fairly janky.” Researchers populated a Pokémon/Sims-like town with 25 agents, each with a name, characteristics and life story, plus memory, planning and reflection. Set the day before Valentine’s Day, the simulation produced party organizing and romantic invitations without preprogramming those interactions.
• The first test was believability: researchers interviewed agents about their lives, relationships and gift choices. Human evaluators compared responses with the agents’ experiences and human-written alternatives. An agent should want to see someone it cared about, Park argues, rather than whoever it passed most frequently. That was the “illusion of life”; the next challenge was faithfulness to actual people.

2. Real-person validation produced the accuracy claim

• Researchers recruited 1,000 people sampled to represent the US across gender, race, ethnicity and other characteristics. Voice-to-voice AI interviews elicited life stories, difficult decisions and important relationships, supplying data for agent twins. After two weeks, participants returned for surveys, behavioral studies and responses to stimuli; agents predicted their answers. Park reports “85%-plus accuracy” at the end of 2024, the initial foundation for Simile.
• Predictability requires translating stated attitudes into behavior. Park describes risk tolerance as “not really known to change,” while intentions and actions follow different curves: respondents may rate likely gym attendance five or six on a seven-point scale every week, but actually exercising requires greater “activation energy.” Learning that gap requires behavioral knowledge, including irrational tendencies.
• O’Shea presses on what makes someone continue a pattern or act differently. Park’s answer is “the existence of optionality”: exercising competes with sleeping another hour or taking another meeting. An abstract survey’s confined answer set misses the alternatives and opportunity costs governing real decisions.
• Interviews, transaction data and mobility data help, but Park emphasizes randomized controlled trials that vary a feature or message and observe behavior change. He was surprised that customers cared less about forecasting than intervention. Told their market will crash in two quarters, they ask what to do now. “What do we need to do now to shape the future?” makes causal mechanisms central to the product.

3. Enterprise research is the initial commercial market

• Concept testing, conjoint analysis and MaxDiff analysis are the starting applications. Park says studies costing hundreds of thousands or millions of dollars over years become “just a query away.” On Simile’s SaaS interface, users specify whom they want to talk to and what they want to ask; question counts currently have no limit. The gain is faster research with fewer resources.
• Retail is one of the larger verticals. CVS explores store layout, messaging and decor; Brazil’s Banco Itaú examines policies and messaging that communicate its values. Entertainment, CPG and major financial institutions also participate. Internal applications include employee retention and recruiting communication, such as appealing to Stanford or MIT students, with informed consent and autonomy from participating employees.
• Purchase decisions require modeling the environment. Park’s water example: a sole source could charge $1 million just to provide access to water, but widely available alternatives—and packaging and branding—change the context. Agents need exposure to competing options before answering what they would pay. “It really is about modeling people and their behavior, but it’s also about modeling the environment.”
• Asked about an expensive mistake avoided, Park cites an automobile study explicitly conducted before Simile. An electric-vehicle launch message looked like “a clear winner,” yet the simulation showed it also shifted perceptions of non-electric vehicles and lowered sales across the market. For Park, the value is examining “the entire field holistically”: useful simulations surface consequences that are “counterintuitive” and “surprising.”

4. Population coverage and confidence depend on data discipline

• Vertical-specific “agent bank” panels give customers a starting population; bespoke needs trigger additional collection and training. Partnerships reach over 40 million real people globally, with “tens of thousands, if not hundreds of thousands” contacted weekly. New panel needs can usually be met within 24 hours. Park highlights Gallup’s panel of hundreds of thousands and relationships spanning decades; representative sampling is the critical discipline.
• O’Shea’s pushback is whether habitual respondents become biased or tell pollsters what they think is useful. Park acknowledges fatigue and gaming for payment: someone may give “some bullshit answer” to participate. Controls include spacing participation, comparing fresh respondents with experienced ones, and attention checks. Asked to recall the interview’s first question, an attentive participant should remember the simulation discussion.
• O’Shea says the discussion makes her nervous, “like you can read my mind”; Park had simulated the interview and anticipated her opening simulation question. For panel collection, his boundary is informed consent and the ability to withdraw. Underage people are excluded, “one blind spot that we own,” as are others unable to consent. Otherwise, Simile aims to represent all populations, including “the good behaviors but also bad behaviors of people.”
• Alongside its foundation model, Simile trains a model assigning responses a zero-to-100 confidence estimate based on training data and the question. Park likens its decision role to the inferential-statistics convention of p < 0.05; he says customers find confidence above “90% or usually 85%” sufficient for high-stakes decisions. Deployment includes helping users decide when to trust results, collect more data and validate findings.

5. Funding supports both model development and customer deployment

• O’Shea puts total funding at $300 million and the reported valuation at $2 billion, naming Greenoaks, Index, Fei-Fei Li and Andrej Karpathy. Park traces the investor base partly to Stanford relationships: co-founder Micah Bernstein worked closely with Fei-Fei; Percy Liang served on Karpathy’s dissertation committee; Karpathy amplified the 2023 generative-agents paper. The underlying pitch is a technology serving an innate need to understand people.
• Simile has about 100 people. Roughly 20% are lab mates or “academic relatives”; science, modeling, evaluation and methodology account for about 20%–25% of the company. Social-science methodologists work alongside AI researchers, engineers, product staff and deployment teams. Studying people is part of the technical task.
• Park places simulation in its “GPT-4, GPT-4.5 era”: useful now, changing monthly. He expects it to reach today’s coding-agent industry’s maturity in about a year and a half to two years. Forward deployment is central: coding companies with close customer relationships became “stickier” and “much more defensible over time,” a pattern he believes will repeat in simulation.

6. Park wants the company to keep earning its next problem

• Inspired by Thomas Schelling’s segregation modeling in the 1970s and 1980s, Park wants to study generational coordination problems. Climate policy involves actors with competing incentives whose interactions create difficult equilibria; repeated simulations could potentially help explore policy outcomes. If AGI is here or imminent, he asks whether simulations could show where UBI works and fails without waiting for a five-year study.
• The career trait he admires is being “the common denominator of success”: taking ownership and repeatedly reinventing oneself. He also prizes seemingly incompatible strengths, exemplified by Michael Bernstein’s mature scientific discipline and childlike excitement about game demos. The durable motivation is the “greatest prize in winning the current game is gaining the right and invitation to play the next game.”
• The failure mode is optimizing for one fundraising, customer win or paper, succeeding, then asking “Now what?” and losing interest. Park qualifies that judgment: it could be “an amazing success,” but becomes failure for someone seeking impact across generations. Asked whether he would sell for $15 billion, his answer remains conditional: if he has reason to believe only Simile can pursue an ambitious idea, “then we’re going to pursue that.”