For several years, we have kept returning to the same paper at Turing Post.

It was published in April 2023 under the title Generative Agents: Interactive Simulacra of Human Behavior, although most people remember it simply as Smallville. The researchers placed 25 AI characters inside a pixelated town, gave them occupations, relationships, memories, and daily routines, then observed what happened.

We have cited it in our articles on agent memory, profiling, multi-agent systems, and agent social networks. A simulated town called Smallville has practically been living in our archive. It gave a concrete form to an idea that was still hard to picture in 2023: an LLM could become more than a chatbot when it was given a persistent history and placed inside an environment with other agents.

Now this town has grown into Simile AI – a company with a $2 billion valuation in less than six months.

And the company wants to simulate all eight billion people on Earth.

Simile CEO describes the long-term ambition as a "CERN of human society" – bank runs, climate cooperation, democratic collapse – and imagines a single foundational simulation that might one day cost $100 million and take months to run.

How are they going to do that? What exactly are they building? Why are investors putting so much money into simulation? And what changed with generative AI that suddenly made this one of the most interesting categories to watch? Let’s explore. It’s a fascinating story of how academic research is becoming the foundation of a new human-simulation industry.

Not interested in this topic? Watch this video about how the inference race is shifting from installing more chips to extracting more useful work from them.

Follow us on

Before Smallville, there was a simulated subreddit

The Simile story did not begin with a plan to replace surveys or create digital customers. It began with a question about social platforms.

Joon Sung Park started working on simulations at Stanford around 2020, just as GPT-3 was demonstrating that a sufficiently large language model could perform tasks it had not been trained to execute in a narrow, traditional sense. Park was interested in a particular property of these models: they appeared to contain many small patterns of human behavior collected from text, conversations, social media, and other parts of their training data.

In 2022, Park and his collaborators – his advisors Percy Liang and Michael Bernstein were among the co-authors – published Social Simulacra, a precursor to the Smallville work. The system allowed designers to populate a hypothetical online community with thousands of generated personas, specify its purpose and moderation rules, then observe the conversations and social patterns that appeared.

There was a straightforward practical motivation. A social network can test whether a button works before launch, but it cannot easily test how thousands or millions of people will interact once the product is released. Poor design choices are often discovered through public deployment, after real users have already experienced their effects. Park’s group wanted to know whether some of those outcomes could be explored earlier inside a simulation.

Smallville (2023) added time and continuity. The agents no longer appeared only when prompted to comment on a hypothetical discussion. It was absolutely fascinating to watch them interact, make decisions, and justify their actions. They accumulated experiences, remembered previous interactions, adjusted their plans, and maintained relationships over several simulated days.

But it was still research, with no plan behind it to start a company. That changed when the paper attracted attention from two very different groups that rarely ask researchers for the same product:

Park describes research as a vehicle for breadth-first exploration. A lab allows many researchers to pursue small pieces of a larger thesis, but researchers are rarely responsible for carrying every promising idea into the real world. A company, in his formulation, is a machine for depth-first search: you develop conviction around one area, assemble the people and resources, and pursue a single direction without hesitation.

That conviction arrived about six months after the Generative Agents paper. The inbound interest showed Park a direct path from the research to something people already wanted to use.

He began discussing what would become Simile with Percy Liang and Michael Bernstein, who had advised his Stanford work. Liang brought deep experience in foundation models and evaluation, while Bernstein’s work spans human-computer interaction and social computing. This was a founding group built around years of shared research, rather than assembled around a newly fashionable category. That explains why investors moved so quickly.

But before they could sell the system, they needed to answer a more difficult question.

From fictional residents to 1,052 real people

Smallville showed that agents could behave coherently enough to seem alive. It did not establish that an agent modeled on a real person could predict what that person would say or do.

To examine that question, in 2025, Park and a large group of collaborators recruited a diverse sample of 1,052 Americans. Every participant completed a two-hour semi-structured interview covering their experiences, beliefs, relationships, work, politics, and life history. The researchers also collected structured survey data.

They then created agents grounded in three different sources:

The agents were later tested on General Social Survey questions, personality measures, economic games, and replicated social-science experiments. A simpler baseline received only demographic information about the person.

This was the important change. The system was no longer asked to invent a plausible 35-year-old teacher in Ohio. It was given evidence from a particular person and asked to reproduce that person’s responses across different tasks.

The research continued with SocSci210, a dataset containing 2.9 million individual responses from 400,491 participants (!) across 210 social-science experiments. Fine-tuning a 14-billion-parameter Qwen model on these responses improved its agreement with human distributions on completely unseen studies by 26% relative to the base model and by 13% relative to GPT-4o, according to the paper.

You could say the idea behind the company was already brewing, because the publications that followed Generative Agents read almost like a product roadmap:

This progression is one reason the company is so interesting. The commercial product did not suddenly appear when synthetic users became a popular startup category. The researchers kept asking a harder version of the same question until a company became a plausible way to continue the work.

What does Simile actually sell

and could it eventually build simulations substantial enough to justify $100 million runs?

Continue reading this fascinating story and learn more about where the industry is heading

Below, we are going to discuss:

Join Premium members from top companies like Microsoft, Nvidia, Google, Hugging Face, OpenAI, a16z, plus AI labs such as Ai2, MIT, Berkeley, .gov, and thousands of others to really understand what’s going on in AI.

⬅️ #7: AI Saved Time. Where Did the Value Go?