Key Takeaways

  • Joon Sung Park and co-founder Percy Liang, who coined the term foundation model at Stanford, argue that public web data cannot accurately simulate real human behavior.
  • Frontier language models reflect self-reported opinions rather than choices people make under actual financial or social stakes.
  • Simile AI achieves 85 percent fidelity in digital twins by training models on deep interviews, transaction data, and randomized controlled trials.
  • When Park built the 2023 Generative Agents ("Smallville") simulation, he threw out complex knowledge graphs in favor of simple markdown files.
  • Simulating human behavior requires fine-tuning models on the social physics of real action instead of relying on prompt engineering.

The Internet Is Performative, Not Predictive

Most founders building AI agents assume GPT-4 can already predict how customers act. They write a prompt describing a persona, ask the model if it would buy a product, and treat the response like market research. Park learned at Stanford that this approach fails immediately.

Frontier language models absorb the public internet. The public internet records what people write when they want an audience to perceive them in a certain way. It does not record the hidden incentives, quiet compromises, or messy emotional calculus that dictate real human choices. As Park explained, these public records are “fundamentally the selfexposed attitudinal data with some behavioral data that sprinkle around here and there. And it has yet to learn really deep behavioral nature of people. Not just what people say they do online, but they what they actually do in real life.”

If you want a model to predict human choices, prompting will not bridge the gap. You have to train the model on actual behavioral ground truth. Park notes: “My intuition behind the actual when do you train or even post- train a model versus just a model is if the motor has to learn the underlying physics of the world that it's operating in. So it has to learn new social physics.” Simile AI bridges this gap by feeding models proprietary data from one-on-one interviews, verified transaction histories, and randomized controlled trials.

Why Smallville Ran on Markdown Files

When Park wrote the landmark Generative Agents paper in 2022, his team faced an engineering fork in the road. They needed 25 autonomous characters to remember their interactions, plan birthday parties, and spread town gossip inside an interactive environment.

The research team initially considered building complex knowledge graphs or training bespoke retrieval models to handle agent memory. Park pushed the team to discard that complexity entirely. As Park recalled: “what we decided to do was no no just forget about all this these language models are actually quite good at modeling text and understanding and reasoning about text so just put everything in a markdown file or text file you're done.”

Founders today routinely overcomplicate agent architectures with heavy vector databases and custom graph pipelines before verifying whether the model needs them. Language models excel at text reasoning. Storing an agent's rolling memory stream in plain text files keeps latency low, prevents graph synchronization bugs, and lets the model reason across the raw context directly.

What to Do With This

Audit any synthetic persona or simulation testing in your product pipeline this week. If you are validating pricing, feature adoption, or user churn by prompting off-the-shelf frontier models, stop trusting the outputs. Replace those prompt-based personas with observed transaction logs or real customer interview transcripts stored directly in plain markdown context windows.