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

1. Key Themes

Simulation as a Foundation-Model Category Distinct from Chatbots and Reasoning Machines

Simile is positioning itself not as another LLM company but as a model of human heterogeneity, trained to capture the full distribution of tastes, values, and even bad behaviors rather than to be a smart assistant. Joon Sung Park says: "Other modeling companies, what they're trying to create is amazing reasoning machine that's great at coding, mathematics, and fundamentally trying to be good chatbot, good assistant. For us, we're trying to create models that can actually model the diverse viewpoints, the values, preferences, and tastes of people." 00:28:44 This implies a different data moat and a different training objective than the frontier labs, and a market the labs are structurally not optimizing for.

Grounded Simulation: The Data Operation Is the Moat

Simile's credibility rests on a large, continuous real-human data operation, not on synthetic data generated out of a vacuum. Park explains: "It's a large data operation, that we do not simply create simulations out of vacuum. We believe strongly that the simulation has to be grounded, and it has to be evaluated alongside people." 00:10:39 The scale: "Simile have partnership with over 40 million real people across the panel partnerships that we have... at any given moment, we're talking to tens of thousands, if not hundreds of thousands, of real people every single week." 00:22:18 New panel needs "can usually be met within 24 hours." 00:22:45

The Say-Do Gap Is the Core Modeling Problem

The technical heart of the company is learning the translation layer between stated attitudes and actual behavior. Park: "Really the modeling challenge here actually is really learning that translational layer. How do you take observed behaviors or attitudes into their fundamental tendencies?" 00:12:00 He gives a concrete mathematical framing: survey answers follow "a normal distribution curve," while real action looks "much more like an exponential curve" because "the activation energy for me to actually go to the gym is actually way higher than what I would expect." 00:13:12 This explains why traditional surveys systematically overstate intent.

Optionality and Opportunity Cost Govern Behavior, So Simulate the Environment, Not Just the Person

Park argues surveys fail because they present abstract, confined choice sets. "Optionality and the type of optionality often really governs people's behaviors." 00:15:18 This is why Simile models both people and their context: "Simulation really is about also creating model of the environment, not just the people." 00:16:54 His water pricing example: a sole source could charge "$1 million," but widespread alternatives cap price, so you must model the competitive environment. 00:17:04

Causal Data (RCTs), Not Prediction, Is What Customers Actually Pay For

A counterintuitive product insight: customers don't primarily want forecasts, they want levers. Park: "People are actually not that interested in predicting the future... what are you going to do if I told you that your market is going to crash in two quarters? They're all going to then ask, 'Well, that's shitty. What do we need to do now to change that?'" 00:44:12 That's why randomized control trial data is prized in training: "It's about understanding the causal mechanism, the causation, because that's ultimately the lever that we have to shift the direction of the future." 00:44:42

Calibrated Confidence as a Product Feature

Simile jointly trains a second model that outputs a 0-100 confidence for each prediction. Park: "Alongside the core foundation model... we are jointly training a predictive model of accuracy... it's a little bit like having p-value." 00:32:02 The adoption threshold: "If confidence is higher than 90% or usually 85%, our customers have found that to be robust evidence that can actually be used to make very high-stake decisions." 00:33:01 This turns a probabilistic model into something enterprises can defensibly act on.

Forward-Deployed Go-to-Market and Change Management as Defensibility

Park draws an explicit analogy to the agentic coding wave: "I look at coding, agentic coding era. I actually have found that companies that were forward-deployed actually found a lot of success. Their relationship became stickier, became much more defensible over time." 00:47:40 He also notes that "this is new way of making decisions," so Simile does change management: "helping them calibrate when to use it, how to use it, where do we need more data, how do we validate it." 00:31:28

Timing Thesis: Simulation Is in Its "GPT-4 Era"

Park gives a specific maturity forecast: "Simulation is basically in its GPT-4, GPT-4.5 era... I really do see the simulation space as one that's going to reach its maturity to a point where, you know, in about a year and a half to two years will be where coding agent as an industry is today." 00:47:08

Academic Lineage as a Talent and Capital Flywheel

The company is an outgrowth of a Stanford research lab, with a cap table and team built from that network. Park: "Fei-Fei, of course, a close colleague at Stanford... Andrej, one of my other co-founders, Percy Liang, was actually on his doctoral dissertation committee." 00:40:36 About "20% of the team actually is our lab mates and our academic relatives," complemented by social-science methodologists. 00:45:30

Evaluation Evolved from Believability to Faithfulness

The validation methodology matured with the field: early "believability" (human evaluators judging whether agents seem human-like) gave way to accuracy against real people. "At the end of 2024, first for the first time really in the field, we could actually show that we can predict people's behavior with 85% plus accuracy." 00:10:13 The test design: 1,000 representative US participants, interviewed voice-to-voice, agents built from the interviews, then compared against what those people answered two weeks later. 00:09:17

2. Contrarian Perspectives

Simulation Is Inevitable Because Civilizations Understand Themselves by Recreating Themselves

Most investors would see behavioral simulation as a research-grade novelty. Park frames it as an innate human drive: "The best way to understand something is to recreate it... it's a little bit like creating a portrait of who we are as a society." 00:01:56 His investor pitch: "I truly do believe that simulation in many ways is inevitable... just like how we write autobiographs or we paint portraits, simulation is a lot like that." 00:41:30 He even concludes any such civilization likely lives in a simulation. 00:02:22

Customers Don't Want Predictions; Accuracy Alone Isn't the Product

Conventional wisdom says the value of a human-behavior model is forecasting accuracy. Park argues the opposite: "People are actually not that interested in predicting the future, which is, to me, this was a little bit surprising." 00:44:12 The value is in counterfactual levers (what to change), making causal experiment data more valuable than observational data.

Good Simulation Should Surface Counterintuitive Findings, Including Harmful Second-Order Effects

Park's example of the most expensive mistake avoided: an automotive client's winning EV messaging "actually shifted people's overall perception of non-electric vehicle, and globally actually lowered the sales within that entire market." 00:34:53 His takeaway: "When it's at its best, it actually should surface insights that's counterintuitive, that's surprising." 00:35:15 This reframes simulation as systems-level exploration rather than A/B optimization of a single metric.

Informed Consent as a Deliberate Coverage Limit, Plus Willingness to Model Bad Behavior

Rather than hide blind spots, Park owns them: "We do not collect data from underage population... that's actually one blind spot that we own." 00:28:15 Yet he insists on modeling the full spectrum: "A part of it is also then learning the good behaviors, but also bad behaviors of people. Simile needs to be able to capture those." 00:29:10 Representativeness plus consent is positioned as an ethical and methodological stance, not a limitation to be engineered away.

Failure Mode Is Winning the Single Goal and Losing the Reason

On personal and company durability, Park flips the usual "optimize for the milestone" advice: "There's that one particular fundraising you want to run... you win it, and you go, 'Now what?' And they lose interest." 00:54:04 His long-horizon test: "Are we doing this so we can have an exit tomorrow, or are we doing it because we want to fundamentally reshape how our society operates for the better?" 00:55:15 He answered the $15B hypothetical with conditionality tied to whether "there's an ambitious idea that only Simile can pursue." 00:55:33

3. Companies Identified

Simile

Foundation model of human behavior company, spun out of Stanford research, reportedly valued at $2 billion after raising $300 million. Mentioned as the subject of the episode. Park: "Simile is a company that is creating foundation model of human behavior, that basically can take as input a description of a person or a group of people, and then simulate what these people might do in different context." 00:00:04

Index Ventures

Venture investor; the interview was recorded at Index's office, and Index is an investor in Simile. Molly: "I know we're here in Index's office, someone who is a big contributor or investor in your company." 00:00:47

Greenoaks

Growth investor named among Simile's backers. Molly: "You have really great names, like I mentioned, Greenoaks, Index." 00:00:20

CVS

Retail customer using Simile for consumer understanding. Park: "CVS. And they have worked with us to create simulations of both how to better understand their consumers, so that they can better serve them, both in terms of the store layout, the messaging, and the overall décor." 00:20:20

Banco Itaú

Brazilian bank and international customer. Park: "International institutions like Banco Itaú in Brazil, where they also have similar needs, but they want to understand, well, what kind of policies or what kind of messaging do we need to put out to better communicate our values to our customers?" 00:20:20

Gallup

Strategic data partner; Simile uses the Gallup Panel for representative sampling. Park: "One of the strategic partnerships that Simile has actually is with Gallup... I would consider Gallup to be one of the most golden-standard polling company." 00:23:47 "They have a Gallup Panel of hundreds of thousands of people with whom they've actually built up relationship over the past decade, two decades." 00:24:13

DoorDash

Source of Simile's head of engineering, praised for AI talent. Molly: "I heard DoorDash has amazing AI talent, like very underrated." Park: "They're killers... they fought in an extremely brutally competitive market, and they won. That speaks volume." 00:46:42

Brex

Episode sponsor; financial platform used by Sourcery and by Vercel, OpenAI, Anthropic, Granola, and Deepgram. Ad narrator: "The companies building what's next, from Vercel, OpenAI, Anthropic, Granola, and Deepgram, all made the same call. They all run on Brex." 00:18:24

Harvey

AI platform for legal work, sponsor. "Trusted by 80% of the Am Law 100 and over 20% of the Fortune 500." 00:19:23

Mercor (APEX)

Sponsor; Artificial AI Productivity Index (APEX) built with partners like Harvey and Cognition to track AI performance on professional work. 00:35:53

Cognition

Named as a partner on the APEX benchmark. 00:35:53

Xone

AI data center campus developer, sponsor. "Xone develops next-generation data center campuses, partnering with AI companies, site developers, and technology leaders to bring compute online faster and at scale." 00:19:39

VCX by Fundrise

Public ticker product giving investors of all sizes exposure to venture capital, sponsor. 00:36:21

Public.com

Multi-asset investing platform with AI agents that automate trading strategies, sponsor. 00:36:21

Deel

Global payroll and hiring platform, sponsor. "Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast." 00:36:46

OpenClaw

AI tool referenced by Molly for ingesting a person's data (photos, emails) and producing a psychological analysis, which shocked Alexander Wang. Molly: "OpenClaw just like ingested everything off of their computers, photos, like emails, everything like that, and gave like a detailed psychological analysis." 00:42:09

OpenAI (GPT-3 / GPT-3.5)

Provider of the early language models used in the Smallville experiment. Park: "Initially we were using GPT-3 and then GPT-3.5, the models were still fairly junky, but you could see glimpse of opportunity." 00:04:10

4. People Identified

Joon Sung Park

Co-founder and CEO of Simile; Stanford researcher and lead author of the Generative Agents (Smallville) work; also a painter. Mentioned as the guest. He led the research establishing 85%+ accuracy in predicting individual human behavior. "At the end of 2024, first for the first time really in the field, we could actually show that we can predict people's behavior with 85% plus accuracy." 00:10:13

Molly O'Shea

Host of Sourcery. Ran a Simile pilot for the episode that simulated the best-performing title ("The $2 Billion Company Building AI Simulations of Real People") for her investor/tech-executive audience. 00:29:55

Michael Bernstein

Simile co-founder and Stanford professor. Praised for combining scientific maturity with childlike excitement. Park: "He is a very mature scientist, understand all about the scientific methodology and so forth, but at the same time, he's a child who gets excited by creating game demo." 00:51:14

Percy Liang

Simile co-founder, prominent Stanford AI researcher and pianist; was on Andrej Karpathy's doctoral committee. Park: "Percy famously within Stanford community, my last co-founder, he's a pianist, but he's also one of the most well-known AI researchers. Amazing combination." 00:52:04

Leni Yellen

Simile co-founder leading go-to-market and deployment. Park: "Leni Yellen, who is deeply paranoid on day-to-day basis that we have to do everything right now... but at the same time, deeply religious that believes that no matter what happens, we will win." 00:51:43 She leads the forward-deployed motion. 00:47:08

Fei-Fei Li

Stanford AI pioneer and Simile angel investor, a close colleague of the founders. Park: "Fei-Fei, of course, a close colleague at Stanford. One of my co-founders, Michael Bernstein, worked closely with her for many years." 00:40:36

Andrej Karpathy

AI researcher and Simile angel; amplified the Generative Agents paper early. Park: "When I first published the generative agents paper back in 2023, Andrej actually was one of the first people to really amplify that particular work." 00:41:01

Thomas Schelling

Behavioral economist whose segregation model Park cites as a research hero for applying simulation to a generational societal problem. Park: "Thomas Schelling, who was behavior economist back in '70s and '80s who introduced what we called the model of segregation." 00:37:20

Sandy Pentland

Former MIT Media Lab professor, an archetype of repeated reinvention (physicist, then AI researcher, then computational social scientist). Park: "In each of the era, you could see that Sandy was such an influential character. That's always an amazing power to observe." 00:53:01

Botan

Simile's head of engineering, recruited from DoorDash. Park: "We actually recently invited our head of eng, Botan, who came initially, actually, from DoorDash. And there he was an amazing leader." 00:46:20

Elon Musk

Referenced for the simulation hypothesis, which opened the episode. Molly: "Elon Musk thinks there could be a probability that we live in a simulation." 00:00:00

Alexander Wang

Mentioned for being shocked by an AI-generated psychological analysis of himself. Molly: "Alexander Wang was shocked by it." 00:42:09

Nat Friedman

Gave Alexander Wang the deep-dive psychological-analysis prompt for OpenClaw. Molly: "From Nat Friedman, the prompt that Nat gave him was to like do a deep dive psychological analysis." 00:42:09

5. Operating Insights

Learn What to Collect from Interviewing Technique: Ask for the Life Story

Simile's data collection design was inspired by being a podcast guest. Park: "After talking to these interviewers for about an hour or two, they know a lot about you, especially the good ones. If they ask the right questions, they know a tremendous amount about both your field and who you are." 00:43:08 The operational payoff: "We found that asking people to just tell us the story of their life actually matters a lot. It reveals a lot of interesting pockets of knowledge about those people." 00:43:08 The tactic generalizes: structured, open-ended biographical interviews are a dense, low-cost signal source for modeling any individual (customers, candidates, employees).

Manage Respondent Fatigue by Spacing Studies and Measuring the Optimal Interval

For anyone running panels, surveys, or user research: fatigue degrades data into gaming. Park: "If you get polled every single week, they start to really get fatigued. And they start to... game the system... So what you do is you actually time it, so that the participation is few in between and sparse enough while being able to still maintain that relationship." 00:24:56 They also "run studies that basically shows we have a new population versus population who had answered X number of questions over the past" to find the right spacing, and they insert attention checks (dummy questions like recalling the first question asked) to filter inattentive participants. 00:25:29

Ship Confidence Scores So Customers Know When to Trust the Output

Pair every AI output with a calibrated confidence number and a defined action threshold. Park's rule of thumb: above 85-90% confidence is "robust evidence that can actually be used to make very high-stake decisions." 00:33:01 This lowers the adoption barrier for novel decision tools by making trust a measurable, legible quantity.

Pre-Simulate High-Stakes Conversations

Park uses the product on his own life: "What I often do is I do simulate these kind of conversations... of course, you bet I simulated the conversations before walking in." 00:39:15 The simulator correctly predicted the host would ask about living in a simulation. 00:39:39 Founders can use persona simulation to rehearse investor, hiring, and partner meetings.

Hire for the "Next Game" Trait and Tension of Opposing Superpowers

Park's heuristic for talent: seek people who are "the common denominator of success" at different career stages and who carry "two superpowers that's really not supposed to coexist in the same person." 00:51:14 He adds that these people treat "the greatest prize in winning the current game... gaining the right and invitation to play the next game." 00:53:31

6. Overlooked Insights

Simulation Becomes Most Valuable When It Models Policy and Coordination Problems, with Internal Stakeholders as an Early Wedge

Mentioned almost as a closing aside, Park lists climate change as a coordination problem ("many actors with competing incentives... the equilibria state they generate is nearly impossible to navigate") and UBI testing ("instead of running five-year study on UBI, can we could we actually have simulated that") as the dream use cases. 00:38:14 Combined with a quietly mentioned HR/retention use case, "it's really about understanding everyone that's in our ecosystem, and many of those important stakeholders are actually internal" 00:49:31, this signals a much larger addressable market than consumer insights: any decision involving heterogeneous agents with competing incentives (policy, labor, recruiting, M&A integration). The current retail/CPG/finance customer set is the wedge, not the ceiling.

The Model Predicted the Interviewer's Question: Simulation Quality Is Already Revealing in Practice

Said with a laugh, but significant: Park ran a simulation of Molly before the interview, and "it did warn me that you might ask whether we live in a simulation... that's not something that every interviewers ask. But for some reason, this is actually a good question that the simulator actually thought, given everything that we know about you, that you might ask." 00:39:32 This is a live, unscripted out-of-sample prediction of a specific individual's behavior from public data alone, a small but real demonstration of the individual-level fidelity the company claims, and the same mechanism that powered the pilot where Simile ranked Sourcery's best title. 00:29:55