Simulating Humans at Scale: Simile's Joon Sung Park

1. Key Themes

From Research Curiosity to Commercial Validation

The genesis of Simile was a years-long research arc at Stanford, beginning with GPT-3-era experiments and culminating in a validated claim: simulations can predict human behavior with meaningful accuracy. The company was only founded after Joon achieved a concrete validation benchmark.

"We went out and actually created simulations of thousand people of the US population. We demonstrated that using our architecture and the models, we can actually predict people's behaviors 85% as accurately as people replicate their own." 00:11:55

The Say-Do Gap Is the Core Product Insight

The foundational problem that Simile is solving isn't just scale — it's that existing LLMs are trained on what people say, not what they do. Closing this gap is Simile's core technical differentiator.

"A lot of the large language models are trained on attitudinal data. Fundamentally, it is the things that people have said online... One of the things that Simile's simulation platform does is actually closing the gap... a lot of the data that we end up collecting by nature are behavioral." 00:16:36

The CPU vs. GPU of Intelligence

Joon draws a sharp architectural distinction between what frontier labs are building (rational, objective, singular) versus what Simile is building (diverse, irrational, parallel). This is not just positioning — it reflects a genuine thesis about what is missing in the current modeling paradigm.

"Imagine the current today's models are akin to the CPU of intelligence unit... Simile's model is much more akin to developing something that is closer to the GPU of the intelligence unit, where the idea here is we don't actually need a model that is superhuman at Simile. In fact, we want a model that's as human as possible." 00:20:36

Convergent vs. Divergent Simulations as a New Scientific Framework

Joon introduces a conceptually important framework for evaluating simulation outputs — some simulations converge to reliable answers regardless of compounding errors, while others diverge and require confidence interval thinking. This is framed as the early days of a new inferential statistics.

"I see simulation as a field as akin to developing your day one of inferential statistics... Similarly, it's working on setting the same kind of threshold and standards for the rest of the field." 00:31:08

Life Story as the Most Efficient Data Signal

The most efficient signal for building faithful human agents is not demographics or attitudes — it is narrative biography. Simile trains an AI interviewer specifically optimized to extract maximum behavioral signal from minimal time with a real human.

"We literally ask a question: tell me the story of your life... if you understand the person's story of your life, the kind of data you get from it is what we consider to be the long tail information about this person... it's an amazing way to build a translational layer between attitudes and behavior." 00:19:14

Second-Order Market Effects as the True Long-Term Value Proposition

The near-term use case is concept testing. The long-term value — and what forward-looking customers are already asking for — is simulating downstream, multi-order consequences of strategic decisions across markets.

"What does that do to the perception of let's say non-electric vehicle? Does it change the market perception? Then what does it mean for the rest of the product line? And how do you balance those kind of second-order impact of your decision in a way that is more evidence-based? Today, there's no way to test for this. You can run this in simulation." 00:25:52

Simulation as a Foundational Scientific Instrument

The deepest ambition expressed is that simulation will do for social sciences what the telescope did for astronomy — enabling entirely new categories of scientific knowledge, including Nobel-worthy breakthroughs.

"You look at the greatest scientific innovation — they often start from an amazing measurement. How about telescope really change the trajectory of how we understand the universe — simulation can be that for human society." 00:36:42

2. Contrarian Perspectives

Bigger, Smarter Models Are Actually Worse at Simulating Humans

The conventional assumption is that larger, more capable frontier models produce better simulations of human behavior. Joon directly contradicts this: as models optimize for rationality and objective correctness, their ability to replicate human irrationality and diversity plateaus or diverges.

"You actually start to see divergence in model size going up and the performance in its ability to predict and simulate human behavior. So we have sort of plateaued with current modeling paradigm, our ability to really simulate humans." 00:09:29

The Most Valuable Question to Ask a Person Is "Tell Me the Story of Your Life"

Conventional research collects structured data — surveys, discrete choices, demographic attributes. Joon argues that unstructured biographical narrative is actually the highest-signal data for predicting behavior — a deeply counter-intuitive claim for market researchers.

"Turns out if you understand the person's story of your life, the kind of data you get from it is what we consider to be the long tail information about this person... it's an amazing way to build a translational layer between attitudes and behavior." 00:17:05

A Simulation That Takes $100 Million and Several Months to Run May Be Worth Every Penny

At a time when people celebrate fast, cheap AI outputs, Joon argues that the most important simulations of the future may be extraordinarily expensive and slow — but solve fundamental societal questions, justifying the cost.

"I do see a future where today this is something not the case. Today, a simulation is quick and fast to run. But what about a simulation that takes actually a hundred million dollars to run once? And could take many months to run. But when we run it, it solves one of the fundamental questions of our society." 00:35:20

Field Testing Your Product on Real Users Is the Wrong Default

The dominant tech industry approach to testing product and social design is to ship and see what happens. Joon argues this imposes real human costs and that simulation is the correct first step — a position that challenges deeply embedded "move fast" culture.

"The only way we test it today is you basically field test it. You release your prototype, see what happens. And sometimes it actually comes at a real cost... if you have a bad design, imagine you have a feed on social media that is more likely to propagate certain emotion that is negative... this now gets tested in the field." 00:05:28

3. Companies Identified

Simile

AI simulation company building a platform that models human behavior using agents grounded in real-world data. Co-founded by Joon Sung Park, Percy Leung, and Michael Bernstein. The platform allows companies to simulate consumer populations, test concepts at scale, run multi-agent simulations including earnings calls, and map second-order market effects. Has a strategic partnership with Gallup and customer engagement with CVS.

"We made this observation that large language models can now encode a lot of human behavior... We basically created generative agents that is paired with generative AI model with memory, planning and reflection to basically create this lived experience of agents living in this small town." 00:01:13

CVS

Major pharmacy and health retailer. Cited as a live, nearly six-month customer engagement for Simile. The internal champion is a Senior VP leading human insights who discovered Simile through the published academic validation paper.

"CVS has been partnering with Simile for the past, I would say, nearly half a year... our main buyer at CVS is the lead, is a senior VP who leads human insights... he basically read my paper that validated the agent simulations and thought we have to bring this to CVS." 00:12:58

Gallup

Global polling and analytics company. Cited as a strategic data partner for Simile, used to reach representative human populations for grounding simulations in real-world behavioral and attitudinal data.

"We have through our partnership with vendors, we have a strategic partnership now with Gallup, for instance, who is a polling and panel company, where we go out, work with our vendors to actually reach out to real humans." 00:14:24

Anthropic

AI lab. Cited as an example of a frontier lab building toward rational, superhuman intelligence rather than human-faithful simulation — used to frame Simile's differentiated thesis.

"If you look at many of the large language model companies, whether it's OpenAI, Anthropic and many of the new labs that are getting formed, the models they are creating are models that I would consider to be... meant to be rational... supposed to be really amazing at tackling problems that have an objective answer." 00:08:37

OpenAI

AI lab. Cited in the same context as Anthropic — as a representative of the rationality-optimizing paradigm that Simile is deliberately not following.

"If you look at many of the large language model companies, whether it's OpenAI, Anthropic and many of the new labs that are getting formed, the models they are creating are models that I would consider to be... meant to be rational." 00:08:37

4. People Identified

Percy Leung

Co-founder of Simile and head of the Center for Foundation Models at Stanford. Lead author on the foundational "Opportunities and Risks of Foundation Models" paper. One of Joon's co-advisors at Stanford. Key intellectual collaborator over six years.

"It was led by one of my co-founders, Percy Leung, who is now the head of the Center for Foundation Model at Stanford... he sometimes says you look at the greatest scientific innovation — they often start from an amazing measurement." 00:04:12

Michael Bernstein

Co-founder of Simile and Joon's Stanford advisor. Described as a researcher in human-computer interaction and social computing.

"Michael Bernstein, was a researcher and my advisor at Stanford. Both of them were actually my advisors. So the three of us have been working together for five years." 00:12:24

Sri (Senior VP, CVS)

Senior Vice President of Human Insights at CVS. The internal champion who discovered Simile's validation paper independently and pursued the partnership. His cousin connected him to Joon.

"Our main buyer at CVS is the lead, is a senior VP who leads human insights... he basically read my paper that validated the agent simulations and thought we have to bring this to CVS because today we are bottlenecked by the number of questions we can field test." 00:12:58

Thomas Schelling

Nobel Prize-winning economist. Cited as the pioneer of agent-based models of segregation — Joon uses Schelling's work as the intellectual ancestor of what Simile is building, but with far richer agents.

"Scholars like Thomas Schelling would actually build agent based models that are extremely simple and rudimentary, but that showed something deep about human macro behaviors. And he, of course, went on to win a Nobel Prize." 00:33:22

5. Operating Insights

Validate Before You Build the Company

Joon did not start Simile on the strength of the Smallville demo alone. He ran a specific, quantified validation study first — and only when he hit a threshold he trusted did he make the leap to company formation. This is a replicable discipline for deep-tech founders: define your validation threshold before fundraising.

"We went out and actually created simulations of thousand people of the US population. We demonstrated that using our architecture and the models, we can actually predict people's behaviors 85% as accurately as people replicate their own. When we saw that, we thought, OK, this is something that we feel comfortable providing to our users." 00:11:55

Start Customers on Concrete Use Cases, Then Expand Their Imagination

Simile's customer journey deliberately anchors on the most familiar and concrete task (concept testing) before introducing the more novel and powerful capabilities (multi-agent simulation, second-order effects). This sequencing reduces friction and lets customers arrive at the bigger vision themselves.

"Our customer journey usually does, however, start with a very concrete use cases and problems they are trying to solve. Concept testing is a big one... they quickly see as well... what does it look like for us to test instantly thousand different ideas across thousand different subpopulations?" 00:22:05

Use a Strategic Partnership with a Data Vendor to Solve the Representativeness Problem

Rather than building proprietary human data collection from scratch, Simile partnered with Gallup to solve the hardest part of the go-to-market: reaching the right populations. This is an asset-light way to solve a problem that would otherwise require enormous infrastructure.

"We have a strategic partnership now with Gallup, for instance, who is a polling and panel company, where we go out, work with our vendors to actually reach out to real humans. So these simulations are grounded in real data." 00:14:24

6. Overlooked Insights

Earnings Call Simulation Is an Unexpectedly Common Enterprise Use Case

Buried in a list of applications, Joon mentions that multiple customers are already asking Simile to simulate their earnings calls — stress-testing CEO messaging against a simulated analyst and investor audience. This was framed as surprising even to Joon, but it points to a recurring, high-stakes, time-sensitive enterprise workflow that has almost no existing software solution. It implies a potential wedge into the investor relations and C-suite communications market that nobody is currently serving.

"Some of our customers very routinely actually ask us to simulate their earnings call. This is actually a use case that both surprised me at first, but this is also surprisingly a common ask. Because, of course, the CEOs and board members always need to think about, hey, how are we going to design our earnings call? How would the audience react?" 00:22:58

Customer First-Party Data Fine-Tuning Is the Hidden Enterprise Moat

Mentioned briefly and almost in passing, Joon describes a model where enterprise customers bring their own proprietary behavioral data (e.g., CVS's 90 million customers) to fine-tune Simile's base model into a customized simulation engine. This is not just a services conversation — it is a structural moat: customers with large behavioral datasets could create proprietary simulation layers that are impossible for competitors to replicate, locking them deeply into the platform and dramatically increasing switching costs.

"Our customers then come in, see that potential, and their mind goes to, wow, we have 90 million customers, let's say here at CVS. How can we leverage this kind of data to create better simulations? So there's also a conversation around how can we in a responsible and ethical way leverage existing data that is also in-house for our customers, then use that to create augmented version of Simile's model." 00:18:18