Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Latent Space
Simile AI raised a $2B Series B to model human behavior. It’s already running millions of simulations for big clients, and says it can match focus groups surprisingly well.
Based on reporting by Latent Space — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Joon Sung Park’s pitch is simple and a little unsettling: before making decisions in the real world, why not run the world first? That idea sits behind Simile AI, which just pulled in a $2B Series B backed by GreenOaks and Index Ventures, with Fei-Fei Li and Andrej Karpathy among the notable backers. The company says it’s already running tens of millions of simulations for Fortune 100 customers like CVS, and claims results that line up with human focus groups at 85% to 99% accuracy.
Park didn’t start there. His earlier work, Smallville, the 2023 Generative Agents paper, showed that AI characters could remember, plan, socialize, and drift into unexpected behavior. That paper became a calling card, but the bigger question quickly changed: not how to build a believable agent in a toy town, but how to model real people well enough to make useful decisions about products, policy, and society.
What Simile is building sounds less like chat and more like measurement. Park says the team uses long interviews, observational data, transaction records, randomized controlled trials, and both population-level and individual-level models. The point is to capture how people actually act, not just what they say. Web data, he argues, is good at the latter and much worse at the former. That’s why the company is post-training around causal mechanisms, not just feeding prompts into a frontier model and hoping for a decent imitation of human messiness.
The company’s internal benchmark is blunt. Park says his team built digital twins of 1,000 real people and got about 85% behavioral accuracy, roughly matching how well people reproduced their own responses. He also argues that models optimized to be rational can be bad at simulating irrational humans. In other words, if the goal is social physics, you may need to change the weights, not the wording.
That puts Simile in a very different place from the usual AI sales pitch. The promise here isn’t just cheaper market research, though that’s part of it. It’s the idea that synthetic populations could test products, policies, and even big social questions before anyone commits in the real world. Park’s longer-term horizon is wildly larger: not a handful of avatars, but something that could one day model billions of people and the emergent behavior between them.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI story that gets more interesting when it stops pretending humans are neat. If a simulation only works when it flatters the average user, it’s just a fancy focus group with better branding. The useful version is the one that catches the bad habits, bias, and irrational detours everyone would rather ignore.
Read more about this at: Latent Space