Simile Raises $300M to Make AI Think Like Humans—Biases and All

Source: 20VC with Harry Stebbings | Published: 2026-08-01T14:00:20Z

While OpenAI works to eliminate human bias from AI, Simile is spending $300M to precisely replicate it—building models that make the same mistakes people do.

That simulated town, called Smallville, was a research project Joon worked on during his time at Stanford. It was also the seed of Simile, which has now raised $300 million and is doing something that sounds slightly insane: building a foundation model of human behavior that lets companies run decisions through a simulated world before executing them in the real one.

The Valentine's Party, and Why Agents Need Memory

The core insight from the Smallville experiment came from a surprisingly simple observation: put two language models together, and every time they meet, they say "Nice to meet you."

The models of that era had no concept of memory. Joon realized that for multiple agents to operate together, they had to remember each other — not in the sense of storing a log, but in the human sense: distilling scattered experiences into a higher-level understanding of themselves and others.

Their solution was called "reflection." Think of the kind of sudden clarity you get in the shower — periodically, the system prompts an agent to surface a set of memory fragments and ask itself: why have you been going to the library every day this week? Do you actually care about this research topic? How does it connect to how you grew up?

The agents started forming their own worldviews. That's what made the spontaneous Valentine's party possible.

This memory-planning-reflection architecture was the first time these three concepts were explicitly written into an agentic workflow design. The paper came out in 2023, built on GPT-3.5 — the research predated the ChatGPT era.

Simile Doesn't Care How Smart the Model Is

The most important line for understanding Simile's positioning comes from Joon himself:

"What the large language model companies are building are super-rational intelligent machines — great at coding, math, natural science. Simile doesn't care about any of that. We care that if a person in a given situation would make a particular mistake, our model makes the same mistake. We want our models to have human biases."

This is a fundamentally different strategic direction. OpenAI and Anthropic are trying to remove human biases. Simile is trying to faithfully replicate them — the biases, preferences, and value judgments that Joon calls the "subjective half of the human brain."

Simile runs on top of OpenAI and Anthropic, borrowing their foundational capabilities. But its own mission is entirely different: not to create smarter machines, but to create more human ones.

What People Say vs. What People Do

The central problem with language model training data is that models learn what humans say, not what humans do.

Internet data is, at its core, a massive record of speech. Simile collects behavioral data: transaction data, observational data, and scenario-specific data gathered in partnership with clients.

But Joon's thinking goes further — large-scale behavioral observation data is best at predicting correlations, and what enterprises actually need isn't prediction. It's the ability to change outcomes.

He uses Starbucks as an example. If you tell a Starbucks executive that Frappuccino sales will drop in two quarters, their first reaction isn't "Good to know." It's "So what do we do about it?" Prediction is the starting point. Intervention is the goal.

Answering "what happens if we do X" requires not a correlational model but a causal one. That's where Simile invests its data collection effort: heavy A/B testing, randomized controlled experiments — teaching the model how human behavior shifts when a variable changes, not just which two things tend to appear together.

The Data Moat: Finding the Right People, Asking the Right Questions

Joon's view is that the core competitive advantage for this generation of AI companies isn't algorithms — it's data strategy. Do you have data others can't access? Do you know how to collect data that's genuinely hard to collect?

Simile's data acquisition has two distinctive qualities.

First, it doesn't recruit expert programmers or scientists. It wants ordinary people — but representative ordinary people. Simile needs to ensure its dataset matches the actual demographic composition of the real world. Do you have data from people of this age group, this income level, this region? That's a harder question to answer than "how much data do we have?"

The second is how Simile questions those people. At the start of data collection, Simile has people tell their own stories: Where did you grow up? What have you been through? What are the hardest decisions you've ever made? This isn't a preference survey — it's about understanding who a person is and what has shaped their judgment.

These two things together are what make the simulated people feel real, rather than statistical averages in human form.

85%, and What That Number Actually Means

Before declaring product-market fit, Simile spent nearly a year validating one thing: could the model accurately predict human behavior?

In late 2024, they published research showing Simile's model could reproduce a person's own behavior with 85% accuracy.

What that means in practice: ask real people the same question, have them reproduce their own answers, and 85% of the time it matches the simulation.

That number changed things. Synthetic panels emerged as a real market. Fortune 500 executives saw the Smallville demo at Stanford and realized this wasn't just interesting research — if you could simulate a market like this, it would fundamentally change how decisions get made.

Three-Month Sales Cycles: Enterprises Are Moving Faster Than Expected

Joon initially expected enterprises to need another year or two to warm up to "simulation" as a concept — with aggressive go-to-market activity not really starting until late 2026.

Reality looked nothing like that.

Shri, CVS's VP of Insights, moved at a pace rare inside any large company after seeing Simile. They closed in three months. The reason: the pain they faced every day was too real — experiments took too long, budgets were too tight, too many decisions had to be made on gut feel. When something comes along and says "that consulting report you just paid for? I ran those conclusions in two minutes" — nobody hesitates.

One of Simile's acquisition tactics is to pull up a client's existing research in the first demo and re-run it through simulation. When the results match, there's nothing to explain. The data speaks.

Current pricing is still in the range of a million dollars per engagement — Joon openly admits that's a fraction of the value Simile creates or protects. But he expects the market to shift toward value-based pricing, with "we stopped a half-billion-dollar mistake" being the clearest attribution story.

The Logic Behind a $100 Million Single Simulation

Joon floated a projection: within two to three years, single simulation runs could cost $10–20 million to compute — with clients willing to pay $100 million for them.

This isn't arbitrary number inflation. Consider how reasoning models today can spend hours of compute on a complex task, and "token limits" are no longer a simple concept. The next frontier for simulation is running models involving millions of agents over week-long horizons — how an entire market ecosystem evolves after a product launch, or how different stakeholder groups negotiate after a policy takes effect. The computational scale and value density are in a different category than anything running today.

The target clients are governments, the world's largest enterprises, and anyone where a single wrong call can cost them billions.

Paranoia and Conviction Must Coexist

When Joon describes his team's philosophy, he comes back most often to his co-founder Laney: she is intensely paranoid in the short term and deeply confident in the long term.

Short-term paranoia means she puts every risk on the table and operates as if not going all-in today means losing. Long-term conviction means she believes at a fundamental level that this will work, regardless of what happens in the near term.

These two traits usually cancel each other out. The paranoid tend to be pessimistic about the future. The confident tend to get complacent. People who hold both at once are, in some sense, wired slightly differently — they use today's paranoia to drive action, but that paranoia doesn't come from despair about the future. It comes from obsession with it.

Joon also mentioned another quality he looks for when hiring: two superpowers that shouldn't coexist, existing in the same person. The best CMOs he's seen have a scientist's rigor with data and an artist's intuition and creativity. These two modes of thinking are normally in opposition. When they coexist in one person, that person tends to do things others can't.

His board member Shardul is the same way — deeply analytical, but relies on instinct to make investment decisions.

Researchers Who Start Companies: Married to a Problem, or Married to Impact

When asked how he evaluates whether academics-turned-founders can succeed, Joon's answer was direct:

Are they married to a specific problem, or married to having an impact on the world?

There's no shortage of researchers obsessed with a problem. But that problem doesn't necessarily become a company. The researchers who actually build companies are fundamentally driven by impact — they actively seek out problems that can reach users, problems that can generate revenue.

Simile's four co-founders are Joon (agent simulation), Michael Bernstein — a leading authority at the intersection of HCI and AI and co-author of the Smallville paper — Percy Liang, who coined the term "foundation model," and Laney, who leads business and go-to-market. Three of four are researchers. But what drives all of them isn't publishing papers. It's watching the thing actually happen in the world.

From a Stanford Garage to $300 Million

Joon ended with a story about where it all started.

After graduating college, he moved to Palo Alto, slept in a friend's garage, and worked on a startup that didn't go anywhere. During that time he could sense the AI wave coming. He wanted to be part of it — but he had no research background. He'd done almost no research as an undergrad, which is nearly disqualifying for PhD applications.

He reached out to many people. A Stanford professor named Mary Ward, who worked in theoretical CS, wrote back — maybe because they went to the same undergraduate school, maybe just out of kindness. She'd said she had half an hour. It turned into a full morning. She helped him think through how to approach AI research, and introduced him to his earliest collaborators.

Joon says he still doesn't know why she replied. But that was the first door into this field.

Three hundred million dollars later, he still remembers.