Will Simile Kill Kalshi, Polymarkets & NASDAQ with Co-founder & CEO, Joon Sung Park

Why does it matter? Because prediction tools answer the wrong question for every company paying for them.

Joon Sung Park, founder and CEO of Simile, has built an AI company around a contrarian thesis: businesses don't want to know what's coming. They want to change it. That distinction exposes a structural flaw in almost every forecasting, analytics, and market intelligence product sold today.

• Prediction answers "what will happen" — enterprises need "what do we do about it," which requires causal models, not correlative ones
• Simile deliberately trains models to replicate human mistakes, biases, and irrationality rather than optimizing them away
• The defensible data moat isn't web-scraped text — it's randomized control trial data capturing revealed behavior, not stated preference
• A single simulation session may cost $10–20M to run and command $100M in value within three years, reshaping how enterprise decisions get priced

Businesses don't want to predict the future — they want to change it, and that requires a fundamentally different kind of AI

Telling Starbucks that Frappuccino sales will tank in two quarters is nearly useless. "They'll hear that and they'll be like, what do we do about them? That's terrible," Park says. The knowledge only matters if it triggers action — and action requires understanding why it's happening and which levers exist to shift it.

This distinction kills most of the forecasting industry's value proposition. Correlative models, prediction markets, dashboards that extrapolate trends — all answer the wrong question. "What you really need is causal mechanism. You need a model that can actually reason about causal mechanisms and counterfactuals."

The pricing consequence follows directly. A product that tells you what will happen is worth something. A product that shows you all the downstream steps your ecosystem will take to reach that outcome — and maps what you can do to change them — is worth orders of magnitude more. That's what Simile is building.

Simile deliberately trains its models to fail — behavioral fidelity, not superintelligence, is the product

Frontier labs are racing to build machines that reason better than humans. Simile is building machines that reason exactly like them — including the mistakes.

"Simile doesn't really care about any of those. What we care about is if we have a person make a mistake in this context, we want our models to make the same kind of mistake. We want our models to be biased in the same way humans are."

This isn't a concession. It's the entire thesis. If you're modeling how real consumers respond to a product launch, a superintelligent rational agent is the wrong instrument — real customers are irrational and inconsistent. "In a way, we want to be a representation of people's values, preferences, and taste, sort of their subjective half of their brain." Park goes further: "I fail at a lot of things. I want to make sure that the model that represents me fails the same way."

The race to smarter AI may matter less than the race to more human AI. For most enterprise decisions — product launches, market segmentation, policy modeling — replicating human irrationality isn't a bug to eliminate. It's the deliverable.

Web data is what people say — Simile's defensible moat is data from what people actually do

Every major LLM trains on the same substrate: text from the internet. The internet is a record of stated preferences, not revealed ones. "If you look at the web data, it is fundamentally data of what people have said, not what they have done."

For prediction tasks, observational correlations work well enough. For causal modeling — where the question is what happens when people face an actual choice — stated preference breaks down structurally. People say they'll buy sustainable products; they often don't.

Simile's answer is randomized control trial data, the same methodology underlying clinical trials. "The kind of data that we care deeply about is a lot of randomized control trials. We actually run a lot of AB testing. We show the models: imagine people have done this versus that. This is how their behaviors would actually change. That becomes a core part of our training asset."

Park's frame: "For AI companies of this generation, you need to have an interesting data strategy that's going to be defensible." Compute commoditizes. Algorithms spread. Proprietary behavioral data collected through experiments — not scraped from text — is the only moat that compounds over time and can't be replicated by a competitor with more GPUs.

The world is Simile's ground truth — and it runs a million validation experiments every month for free

Simulation seems to have a fundamental feedback problem: how do you correct a model on predictions about the future before the future arrives? Coding agents had it easy — accept or reject creates an instant training signal. Park thinks simulation actually has a better mechanism.

"The world is our ground truth. We live in the ground truth world. So what we can do is every single day we can be generating tens of thousands of hypotheses." Each hypothesis maps to a real-world outcome with a defined resolution date. A month later, you check. "We generated a million hypothesis, x percentage of them came true. This is the best way to learn about the world."

"It's kind of like AlphaGo — they just beat the out of the model and played it a thousand times. Every day of activities and outcomes in the world is another game of AlphaGo where you can correct the model on what was wrong. And 10,000 days in, you should almost be better than the model at the model."

Every enterprise partner sharing outcome data with Simile is feeding a training flywheel that gets harder to replicate with every month that passes.

$10M to run, $100M to buy — Park's target for a single simulation session in 2028

$100 million. That's what Park expects a single simulation session to be worth in two to three years. "I think there's a world in which in about two to three years we're running a single simulation session that's going to take $10, $20 million to run a single session, but it's going to be so valuable that people will pay $100 million for it. That would be for the world's largest enterprises — that would be for a government or whatever that may be."

The logic is already visible in current contracts. Customers find Simile most valuable for "avoiding really damaging decisions that could have costed them hundreds of millions of dollars." One first-call demonstration replicated findings from a consulting study that took three to six months — "just within 2 minutes." At that leverage ratio, pricing isn't constrained by compute. It's constrained by the cost of the decision being simulated.

Think of high-end simulation less like SaaS software and more like a clinical trial: the cost is large, the stakes are larger, and the buyer is comparing against the alternative of not knowing.

95% of business questions never get answered — simulation's real TAM is the decisions companies currently make with gut instinct

The existing market research industry is smaller than it looks. Companies aren't buying comprehensive intelligence — they're buying the thin slice of questions they can afford to pursue. "You are looking at maybe 5% of those ideas get answered. The rest of the 95% we never bother experimenting with because we either don't have the ability to do them."

The questions that do get answered are already the tractable ones. Surveys fit a known methodology; budgets exist; the study is feasible. The decisions made by gut instinct are the hardest: macro-scale interventions, new market entries, complex policy choices. "A lot of the decisions that we make as a society we base on our gut instinct. Sometimes they're good, but sometimes they're very biased based on our own narrow experience."

Synthetic panels won't simply replace the current human panel market. They'll dwarf it. "Synthetic panels will be larger than what we know to be the current human panel market. In part because this can really raise the ceiling of the kind of questions we can answer."

The total addressable market for simulation isn't market research. It's every major decision currently delegated to intuition.

The best hires have two superpowers that aren't supposed to coexist — contradictory strengths beat correlated depth

Most hiring frameworks reward depth. Park looks for contradiction.

"Where I found things to be particularly compelling is if people have two superpowers that's really contradictory." His example: the world-class CMO who is "unbelievably data-rigorous, oriented, scientific in their approach" while simultaneously carrying "creative artistry imagination" — two mental modes that most people and most hiring processes treat as mutually exclusive.

His co-founder Laney is his archetype for founders: short-term paranoid, long-term religious. "If you are short-term paranoid, then you're likely going to be very pessimistic about your future — you might be amazing at shorting stocks, but not great as a company builder. If you're religious, you have the opposite problem: you're complacent." The person who holds both simultaneously is "broken in some ways" — someone who found a way to be deeply paranoid about today while maintaining complete conviction about the long term. That tension drives the preparation; the conviction sustains the building.

Academic founders who build companies are married to impact — not to a specific problem

Between a science project and a company, the diagnostic is simple: are they married to a problem, or to impact?

Researchers attached to a specific problem will work it until it resolves or stalls — and the market rarely cares about their particular problem on their timeline. Those "fundamentally driven by impact that they can have in the world" will find the problem that reaches people and generates revenue. They'll pivot when required.

The sharpest version of this came from Pat Hanrahan, co-founder of Tableau, whose Stanford office sat next to Park's: "The best way to get feedback is to actually ask people to pay you." Researchers who validate through publication cycles are optimizing for a different audience entirely. Impact-driven founders collapse that gap by design.

Park left Stanford in June 2025. By the time of this conversation, Simile had closed deals with Fortune 500 companies within three months of first contact — remarkable for a first-time founder. The speed came from entering the market with a clear thesis about what enterprises needed, then letting paying customers sharpen it.

The most valuable AI won't be the smartest — it'll be the most human

Every AI roadmap right now assumes intelligence is the axis that matters — smarter, faster, more capable. Simile's bet is that fidelity is more interesting: how accurately a model captures the irrational, contextual mess of actual human behavior. If that bet lands, the frontier model arms race becomes largely irrelevant for most enterprise decisions, and companies that built proprietary behavioral data pipelines early will hold an advantage raw compute can't close. The most valuable AI of the next decade might not be the smartest thing in the room — it might be the most human.

---

Topics: AI, simulation, behavioral modeling, market research, synthetic panels, foundation models, data strategy, enterprise AI, startup building, hiring, prediction markets, causal inference