Joon Sung Park's company raised $300 million at a $2 billion valuation to build something most AI labs aren't trying to build: not a reasoning machine, but a model of how people actually behave.
Most AI startups chase a chatbot that answers questions well. Joon Sung Park, co-founder and CEO of Simile, told Sourcery with Molly O'Shea that his company does the opposite — it tries to predict what a real person will do, and by the end of 2024 it could already do that with more than 85% accuracy.
"I do think we live in a simulation... at the end of the day, is the experience that we're having actually real? Does it feel real to us? Is this true to who we are? I think that's what matters." — Joon Sung Park
Joon's technology traces back to a 2023 Stanford experiment he ran, called Smallville, that put 25 AI agents in a game town and watched them plan a Valentine's Day party nobody had scripted.
The full interview is covered here so you can skip it. 56 minutes of audio, 8 minutes of reading.
Here are the 9 takeaways that matter.
Sourcery with Molly O'Shea · Oct 26, 2026 · 56m · 48m saved
With Joon Sung Park, Co-Founder and CEO of Simile
Key Takeaways
Simile predicted human behavior with over 85% accuracy by the end of 2024, the technical foundation the company still builds on
The model was trained on 1,000 people's life stories and hard decisions, then tested against their real answers two weeks later
Simile avoided an EV maker's launch mistake that would have lowered sales across the entire market, not just competitors
A messaging strategy that tested as a clear winner for the new EV shifted consumer perception of non-electric cars industry-wide
What people say in a survey and what they actually do follow different curves entirely
Stated intent clusters near the middle; real behavior follows a sharp exponential drop-off because of the "optionality" of competing choices
Simile's panel partnerships, including with Gallup, now cover more than 40 million real people globally
Tens or hundreds of thousands are surveyed in a given week, with new panel segments available within 24 hours
Each prediction carries its own confidence score, used like a p-value
Customers treat a score above roughly 85-90% as robust enough to support high-stakes decisions
Simile is about 100 people, roughly a fifth of them Joon's old Stanford labmates
He models the company's go-to-market on the "forward-deployed" playbook he says made coding-agent companies' customer relationships stickier
1. Models Of Human Behavior
Joon frames Simile as building "foundation models of human behavior" rather than reasoning machines — the opposite design goal of a general chatbot.
People reduce to stable "building blocks" of taste that combine into a simulable individual
"There's always these quantum building blocks of people's taste and behaviors... we build individual building blocks of these larger societies, which is individuals." — Joon Sung Park
Those individual models can then be placed into a shared environment to simulate the emergent, "complex and chaotic" interactions of a whole society — the basis for Simile's $300 million raise from Greenoaks and Index Ventures, with personal backing from Fei-Fei Li and Andrej Karpathy.
