Case Studies

25 AI Villagers Made Him Famous. Now He Wants to Simulate All 8 Billion of Us

Stanford researcher Joon Sung Park's Generative Agents paper captivated the AI world. With his new company Simile, he's turning that research into a business—building digital twins to predict how entire populations will behave.

Font size

In April 2023, a single paper detonated across the AI research community.

Twenty-five LLM-driven villagers were dropped into a pixel-art town called Smallville. They woke up on their own, made breakfast, planned a Valentine's Day party, and gossiped about whether the neighbor was dating someone. Nobody wrote rules. Nobody issued commands. Over two simulated days, these agents spontaneously formed social circles, organized gatherings, and produced the small-town equivalent of tabloid news: Isabella likes Klaus.

The paper was Generative Agents, and its lead author was Stanford PhD student Joon Sung Park. Two years later, Park has graduated, raised funding, and incorporated a company called Simile.

Last month, appearing on the Latent Space podcast, he offered a claim sharp enough to make anyone sit up:

Simulation is the new Scaling Law.

Pretraining scales model parameters. RLHF scales alignment quality. Simulation scales a third thing: the number and complexity of social situations.

His plan is to expand Smallville from 25 inhabitants to 8 billion.

From 25 villagers to an 8-billion-person ambition

The impact of the Generative Agents demo is something I still remember vividly.

It wasn't a product screenshot like ChatGPT, or a real-world video like robotaxis. It showed something emergent: give 25 agents a shared physical space, a common memory stream, and a little reflection mechanism—then step back and watch.

What happens is: a Valentine's Day party actually gets organized. Two agents decide to watch the fireworks together—one of them hesitates because it had previously promised another agent dinner. That kind of social tension never made it into any prompt.

Park put it plainly in the interview: "It was a fun exploration, but the technology itself can do much more serious things."

What are the more serious things? They are what he's building next.

Why simulation is the third scaling law

Everyone knows the pretraining line: Chinchilla tells you how to balance parameters and data so loss keeps dropping. The RLHF line—pioneered by Anthropic and OpenAI—tunes models against human preferences until they're usable.

Park's third path scales something different: the world itself.

The underlying assumption is simple. Now that we have a sufficiently capable LLM, what actually limits AI performance isn't its ability to talk—it's that it lacks enough real-world situations to practice on.

Feed a model 100,000 scenarios of "if Company A raises prices 10%, how will users react on Twitter," and it learns to predict that outcome. Feed it 10 million reactions to different ad campaigns across different demographics, and it learns market response.

The fundamental difference from pretraining:

  • Pretraining stuffs knowledge into model weights
  • RLHF adjusts the model's output style
  • Simulation runs vast numbers of "what-ifs" and teaches the model how humans behave in them

Park's own phrasing: "We're scaling the number and complexity of social situations."

The model is the engine. Simulation is the fuel.

Digital twins aren't chatbots—they're behavior predictors

This is where Simile is most easily misunderstood.

When people hear "digital twin of every human," the first reaction is usually: another Character.AI? Another chatbot companion?

It's not.

Simile's Digital Twin isn't a second self that chats with you. It's a behavior prediction engine. The inputs are:

  • The person's purchase history
  • Survey responses
  • Social media posts
  • Demographic data

The output: how will this person vote, buy, or click when faced with a new policy, product, or ad.

Park sketched a few concrete applications during the interview:

  • Market research: P&G wants to launch a new shampoo. The traditional approach is surveying 5,000 people. With Simile, you spin up 5,000 digital twins, let them "try" the virtual shampoo, and report purchase intent. Speed and cost are an order of magnitude apart.
  • Policy simulation: A city wants to impose driving restrictions. Decision-makers need to know how different income groups will respond—will they buy EVs, or just relocate? Traditional methods see results after the fact; Simile predicts beforehand.
  • Financial modeling: Will retail investors really panic-sell during a market crash? A question behavioral economics has debated for decades can now be tested across a million digital twins running scenarios.

Park's exact words: "It's not a chatbot. It's a behavioral prediction engine."

This positioning matters enormously. Simile isn't competing for Character.AI's users. It's going after Gartner, Nielsen, and Pew Research—the market research incumbents.

From paper author to CEO: the potholes along the way

In this Latent Space episode, Park had several moments where he sounded like a real CEO rather than the NeurIPS demo darling.

Take his description of Simile's deployment model: "We do a hybrid model... We have our own platform... but for our larger enterprise clients, we're happy to deploy on their infrastructure." (See the full Latent Space interview.)

Translated: the client doesn't want behavioral data living on your cloud? Fine, we'll bring the model into their data center.

That's a compliance-driven architectural compromise anyone in B2B has faced. Park almost certainly wasn't discussing these problems in his papers a year ago.

Another telling detail: he mentioned that Simile currently serves "a Fortune 50 healthcare company" and "several financial institutions." He can't name them, but the direction is unmistakable: heavily regulated, data-sensitive industries where individual decisions carry enormous cost. A far cry from Character.AI's target audience of lonely young people seeking conversation.

Park also admitted something candidly: "It's a very different mindset from publishing papers."

Writing papers: "I have an idea, run an experiment, get interesting results, write it up, wait for citations."

Running a company: "The client needs a demo tomorrow—can I get stable outputs on their data? Sales tells me a sector requires SOC 2 compliance—can I pass the audit next quarter?"

This shift from research-driven to revenue-driven thinking, he said, has been "the biggest adjustment."

A serious warning: scale this simulation and the stakes get enormous

At this point in the interview, the mood shifted.

The host asked Park: what you're fundamentally doing is using AI to simulate human behavior. That was the dream of social science seventy years ago. If you can really run digital twins for 8 billion people, how large is the misuse surface?

Park's answer was cautious but honest. He named a few boundaries he considers non-negotiable:

  1. No individual manipulation. You cannot build a customized digital twin of a specific person and then target that person with psychological warfare ads. He draws this line hard.
  2. Predictions are probabilities, not fate. When the model says "this person has a 73% chance of buying," that's a statistical estimate, not a verdict. Downstream users must understand that.
  3. Data provenance must be transparent. Anyone should be able to inspect what data their digital twin was built on. That's basic GDPR-class compliance.

But he conceded these are Simile's current guardrails—not the industry's. If everyone builds digital twins, who sets standards? Who audits?

He cited one scenario that keeps him awake at night:

"If a government wanted to use this to predict civil unrest... that's where it gets really scary."

Simulation itself is the sharpest tool ever developed for studying human behavior. Sharp enough to cut both ways.

Park says he now repeats one sentence to investors, customers, and regulators alike: "The technology is neutral, but the deployment is not."

I've sat with that line for a while. It's the same class of problem that Anthropic's Constitutional AI and OpenAI's Preparedness Framework wrestle with—the more capability, the greater the deployment responsibility. On the simulation scaling law, nobody gets to pretend this question doesn't exist.

Closing

The 25-villager Generative Agents demo was, at its core, a toy.

But the toy contained a real question: can LLMs actually simulate human behavior? If they can, what does it become at scale?

Park's answer: an 8-billion-person prediction engine. New infrastructure for market research, policymaking, and financial modeling.

His concern: it could also become a tool for manipulation, surveillance, and social control.

Both paths are open right now.


Source: The core material comes from Joon Sung Park's appearance on the Latent Space podcast, titled Simulation is the New Scaling Law. The original Generative Agents paper is available at arXiv:2304.03442.

Source: https://www.latent.space/p/simile