20VC: How Simile Uses Unique Data to Predict Human Behavior

**SPEAKER_1** (0:01)
When AI gets real enough to model people, the stakes change fast. Today we're breaking down Harry Stebbings' conversation with Jun Park, founder of Simile, a simulation company trying to predict human behavior.
Jun makes this bold claim. If frontier models are the CPU of intelligence, simulation may become the GPU.

**SPEAKER_2** (0:26)
That's a huge claim. So where did this all start for Simile? What was the origin story?

**SPEAKER_1** (0:32)
It began with a Valentine's Day experiment. Jun and his team created a game town and populated it with 25 non-playable characters. Those agents woke up, worked, built relationships, planned parties, and even decorated a cafe.
That project introduced memory, planning, and reflection into agent architecture.

**SPEAKER_2** (0:55)
What exactly does reflection mean in this context?

**SPEAKER_1** (0:59)
Jun calls it a shower thought interval, where the system synthesizes experience into higher level beliefs and personality. But here's where it gets interesting technically. Jun argues that web data only shows what people say, not what they do.
To truly understand behavior, Simile needs transactional and observational data, plus randomized trials and A-B tests.

**SPEAKER_2** (1:24)
So, they're after causality and counterfactuals, not just correlation.

**SPEAKER_1** (1:29)
Exactly. Jun says the real value isn't just prediction, it's shaping outcomes.
A retailer doesn't only want to know sales may dip, they want to know what action will prevent the decline.

**SPEAKER_2** (1:42)
That makes sense. But what about defensibility?
Every AI company is competing on models right now.

**SPEAKER_1** (1:50)
Jun believes the next great AI companies will win through data, not just model size. He says you need an interesting, defensible data strategy.
The hard part is sourcing representative everyday people and asking the right questions, because the best simulations depend on inputs that are hard to copy. As he puts it, the world is our ground truth.

**SPEAKER_2** (2:14)
How's the product market fit looking? Who's actually using this?

**SPEAKER_1** (2:18)
Currently enterprises, because they have budgets and can validate quickly. Some customers have replaced studies that once took three to six months, with results in two minutes.
Jun quotes a Stanford professor who told him, the best way to get feedback is to ask people to pay you. They're claiming 85% accuracy against human replicants.

**SPEAKER_2** (2:40)
That's impressive.
What about the bigger picture? Where does Jun see this going?

**SPEAKER_1** (2:45)
He calls simulation one of the twin pillars of technology alongside AGI.
He believes synthetic panels could eventually outgrow the human panel market entirely, because simulation raises the ceiling on the questions society can ask. For Jun, the future isn't about replacing people, but representing them better at scale, so better decisions can be made before the world pays the price. They recently raised a large round to accelerate with more data and compute.

**SPEAKER_2** (3:15)
So it's not just about prediction. It's about giving us better tools to understand ourselves and make decisions we can't make today.

**SPEAKER_1** (3:24)
Precisely. That's the vision.