Jul 31, 2026 · 1h 5m · news
The AI Company Simulating the Entire Economy | Simile Co-founder & CEO, Joon Sung Park
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of 20VC, host Harry Stebbings interviews Joon Sung Park, Co-founder and CEO of Simile, about building foundation models of human behavior to simulate economic ecosystems and human decision-making. Park details Simile's journey from Stanford research to rapid enterprise adoption, discussing technical agent architectures, proprietary data sourcing, founder leadership, and high-profile venture funding.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 20.5% of the talking time here. How this is scored →
Harry as informed peer 4.8 Guest teaching 4.8 Guest disagreement 1.6 Harry pushing back 2.1
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 8:05 Joon rejects the foundational importance of prediction
Joon delivers a sharp contrarian take, flatly dismissing the idea that enterprise customers care about raw future predictions, insisting they only value causal mechanisms and counterfactuals.
Hardest push from Harry ▶ 11:10 Harry challenges Joon's dismissal of predictive valueHarry pushes back on Joon's assertion by presenting a tangible business scenario with Starbucks Frappuccino sales and demanding to know what he is missing.
Biggest teaching moment ▶ 17:35 Joon explains the simulation hypothesis flywheelJoon educates Harry on how simulation learning loops diverge from coding models or reinforcement learning like AlphaGo by continuously generating daily testable hypotheses against real-world ground truth.
Harry holds his own ▶ 36:01 Harry defines the contradictory superpowers of world-class CMOsHarry showcases his own operator expertise by articulating how top CMOs uniquely combine rigorous scientific data methodology with artistic imagination.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| The Origin Story: Stanford's Smallville Valentine's Day Simulation | 3 | 5 | 1 | 1 | Host sets up the interview by asking about the Stanford Smallville Valentine's Day experiment. Joon details how early LLMs were combined with explicit memory, planning, and reflection architectures to create autonomous agent interactions. | |
| Solving the Agent Memory Problem via Reflection | 4 | 6 | 1 | 2 | Harry asks technically how agent memory problems are solved. Joon explains using periodic 'reflection' routines to distill raw logs into higher-level beliefs and subjective human biases rather than pure rational intelligence. | |
| Human Behavior, Counterfactuals, and Causal Mechanisms | 4 | 7 | 3 | 2 | Harry asks about the gap between what people say versus do. Joon offers a strong contrarian take: clients do not care about raw future prediction, but rather causal mechanisms and counterfactual intervention through randomized control trials. | |
| Data Strategy and Sourcing Everyday People | 6 | 6 | 4 | 6 | Harry directly challenges Joon's assertion that companies do not want predictions, citing Starbucks inventory planning, and asks if Simile is just a next-gen Qualtrics. Joon clarifies the difference between tool-layer surveys and collective multi-agent simulations. | |
| Ethics, Governance, and Representation at Scale | 5 | 5 | 1 | 2 | Harry inquires about ethical guardrails for government elections and sample sizing. Joon explains the target of population-wide representation at scale to enable dynamic ad-hoc filtering. | |
| Real-World Reward Functions and Daily Hypothesis Testing | 5 | 6 | 2 | 2 | Harry compares simulation data flywheels to AlphaGo self-play. Joon reframes the concept: unlike coding agents with immediate accept/reject signals, simulation uses the unfolding real world to validate millions of daily generated hypotheses. | |
| Total Addressable Market (TAM) & Enterprise Wedge Strategy | 4 | 4 | 1 | 1 | Harry probes Simile's addressable market across consumers versus enterprise. Joon cites Tableau co-founder Pat Hanrahan regarding enterprise monetization as the fastest validation feedback loop. | |
| Finding Product-Market Fit with Synthetic Panels | 4 | 5 | 1 | 1 | Harry asks when Joon realized they had product-market fit and if synthetic panels will overtake human panels. Joon notes achieving 85% accuracy in replicating human attitudes and opening up testing for the 95% of hypotheses previously left unexamined. | |
| Balancing Academic Research with Commercial Product Execution | 5 | 4 | 1 | 2 | Harry asks how Simile balances fundamental research purity against enterprise commercialization and customer success. Joon explains the founding team composition and the direct alignment between model fidelity and customer utility. | |
| Value Extraction, Pricing, and Preventing High-Stakes Mistakes | 6 | 4 | 2 | 4 | Harry drills down on the pricing chasm between creating hundreds of millions in value versus charging nominal fees, and whether Simile cannibalizes prediction markets. Joon emphasizes catastrophe prevention as their main ROI driver. | |
| Team Building Principles and The Figure Painter Analogy | 3 | 6 | 1 | 1 | Harry asks about key hiring and team building lessons. Joon shares his figure painter analogy, the importance of finding common denominators of success across career phases, and seeking contradictory superpowers in leaders. | |
| Managing Contradictory Mindsets: Paranoia vs. Long-Term Faith | 7 | 3 | 1 | 2 | Harry contributes his own operational framework on elite CMOs possessing contradictory traits of rigorous data analytics and pure creative artistry. Joon applies this framework to co-founder Lainey's day-to-day paranoia versus long-term religious conviction. | |
| How Paranoia Drives Founder Success & The AI Talent War | 6 | 4 | 1 | 2 | Harry argues that founder paranoia is essential rather than harmful, and probes the talent war and high researcher compensations. Joon discusses keeping core research cohorts intact across multiple multi-year projects. | |
| Identifying Academic Founders Focused on Real-World Impact | 5 | 4 | 1 | 2 | Harry asks how to differentiate commercially viable academic founders from academic science projects, before discussing Simile's $300M total funding and preempted rounds with Greenoaks and Index. | |
| Capital Strategy and Accelerating AI Compute Inputs | 5 | 4 | 1 | 2 | Harry asks why Joon took an extra $200M when not strictly required. Joon explains controlling research inputs through compute scaling, and shares his evolution in understanding the mentorship value of venture capitalists. | |
| Market Dynamics, Froth, and Variable Compute Costs | 6 | 5 | 1 | 2 | Harry inquires about venture market froth and whether unit economics vary by simulation complexity. Joon outlines a future where single multi-million-dollar inference runs command $100M prices from enterprises. | |
| Bridging Simulation Proof-of-Concept to Production Systems | 4 | 6 | 2 | 1 | Harry plays the time machine game. Joon explains the shift from LLMs as 'CPUs of intelligence' to emergent multi-agent simulations acting as the 'GPUs of intelligence'. | |
| Scalable Human Representation and Financial Market Edge | 6 | 4 | 1 | 3 | Harry asks whether hedge funds represent an obvious customer and if Simile could launch an internal fund that distorts public markets. Joon confirms they have quant researchers on staff and acknowledges potential market shifts under perfect simulation. | |
| Simulating Love, Dating Efficiency, and Human Connection | 4 | 5 | 4 | 3 | Harry pitches simulated speed dating and Married at First Sight efficiency. Joon pushes back from a romantic standpoint, arguing that human connection requires shared lived experiences and co-founding a life together. | |
| Quickfire Round: AI Research, Data Defensibility, and Hardware | 4 | 4 | 1 | 1 | In the quickfire round, Joon highlights data strategy defensibility, chip hardware startups like Etched, and recalls Stanford professor Mary Wootters giving him his initial break in research. |
Statements from this episode (40)
Park: LLMs can simulate realistic human behavior across domains
“One of the early observations that we made was that these models are trained on so much of human behavior data, sentiment data that were expressed on the web. So if you poke at them sort of at the right angle, you could actually extract a lot of realistic huma…”
Park: Stanford simulation first explicitly architected agent memory, planning, reflection
“And then was paired with memory planning and reflection. Really the first times that those concepts Came out to be an explicit part of the architecture in, quote unquote, agentic workflows.”
Park: Reflection Prompts Turn Raw Agent Memory Into Emergent Personality
“So we had this concept of reflection, which basically was every certain interval, it's like a shower thought. You have, you ask agent explicitly to get a bunch of their memory pieces and basically make sense of them. Why did you get omelet so often this week? …”
Park: Simile Builds Foundation Models to Simulate Whole Markets
“We are a company that is creating foundation model of human behavior that can then be used to create simulations of individuals, simulation of sub-populations, and down the line, the simulation of the entire ecosystem and even the market.”
Park: Simile Optimizes for Human Biases Over Super-Rationality
“So the way we see it is if you look at large length model companies today, fundamentally the task they have at hand is to create super rational, intelligent machines that are good at coding, that are good at natural sciences and mathematics. Similarly doesn't …”
Park: Customers do not care about prediction unless trading stocks
“No one really cares about prediction. No one really cares about what's going to happen in the future unless you're trying to predict the stock market. What people actually care about is they want to shape the future.”
Park: Simile trains models on randomized control trials and A/B tests
“So the kind of data that we care deeply about is a lot of randomized control trials. We actually run a lot of A-B testing. We show the models. Imagine people have done this versus that. This is how their behaviors will actually change, and that becomes a core …”
Park: Simile sources data from everyday people, not elite experts
“We don't go after these expert programmers or expert scientists. We go after people like us, like everyday people living their everyday life.”
Park: Simulation extends far beyond traditional survey and interview tools
“Simulation is fundamentally about something different, which is how can you create the most generalizable model of people so that we can represent people's viewpoints at scale? And that goes beyond simply running surveys or interviews.”
Park: AI simulation can help solve wicked problems like climate change
“Simulation I do think can also be a cure for many of what we call quote unquote wicked problems.”
Park: Simulation and AGI are the twin pillars of advanced technology
“The way I see it, simulation as a piece of technology is one of the twin pillars of technology. I'm a fan of science fiction. You read any advanced science fiction, there's always two pillars. One is some form of AGI that always shows up. The other is simulati…”
Park: Simile Models Improve by Comparing Simulation Outputs to Ground Truth
“You know, I think that certainly is the case because, right, there is the data flywheel. There is the learning that occurs as we get more and more simulated results and see what happens in the ground truth. That absolutely yes.”
Park: Coding agents improved rapidly due to clear accept/reject reward functions
“The reason why coding agents have had such massive improvement over the years was because their learning, their reward function was extremely clear. If you make a suggestion and your user says accept, fantastic. If they say reject, also very useful. You very q…”
Park: Simulation models learn by validating daily real-world hypotheses
“I actually think simulation has even better mechanism, which is 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 is mapped onto an end…”
Park: Simile cut production simulation model run costs by roughly 100x
“Right now, we have a model that's been in production. This model used to cost about a hundred times more to run than it does now.”
Simile uses enterprise research as wedge before expanding to consumers
“So enterprise market, market research right now is a wedge that we found that actually have significant budget that have immediate product market fit. But down the line, I do want this technology to be used by the rest of our society, because fundamentally wha…”
Park: Simile models predict human behavior with 85% accuracy
“We show that we can actually predict people's behaviors and attitudes 85% as accurately as people replicate their own.”
Park: Synthetic panels will become larger than the human panel market
“The way I see it, synthetic panels will be larger than our, 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.”
Park: Simile co-founder Percy Liang coined the term foundation model
“Percy was the person who literally coined the term foundation model.”
Park: Simile Replicated Multi-Month Consulting Studies in Two Minutes
“One of the ways we actually got some of our first customers was in the first call, They actually had a finding from, you know, large consulting companies, and they basically queried our system. Hey, if we were to rerun this, what would the system say? And we p…”
Park: Simile Customers Use Simulations to Avoid Multi-Hundred-Million-Dollar Blunders
“One of the core premise and one of the ways that our customers are actually finding value and simile is actually avoiding really damaging decisions that could have costed them hundreds of millions of dollars.”
Park: Simile Differentiates From Prediction Markets by Modeling Causality
“Where I see similarly come in is we are a company that is not just interested in what's going to happen, but more on how it's going to happen and why.”
Park: The most compelling hires possess two contradictory superpowers
“Any expert will usually come in with one superpower, or even sometimes multiple superpowers, but they're all correlated. You're an amazing programmer who happens to be amazing in mathematics. Very common. Where I found things to be particularly compelling is i…”
Park: Great company builders balance daily paranoia with long-term religious optimism
“Actually balancing those two at the same time is quite difficult, because if you are short-term paranoid, then you're likely going to be very pessimistic about your future. And you're, you might be amazing at shorting stocks, but not great as a company builder…”
Park: Close AI research colleagues earn tens of millions in compensation
“So the research talent is very sought after today, and I have my closest colleagues and friends whose total com does range in tens of millions.”
Park recruited doctoral advisors Percy Liang and Michael Bernstein to Simile
“And now when I said, Hey, I want to do this thing and build a simile, I was able to somehow convince Michael and Percy who are, they were actually my doctoral advisors to actually come join me.”
Park: Back Academic Founders Driven by Impact and Revenue, Not Just Problems
“The thing I actually would look for is are they married to a problem or are they married to impact? Sometimes researchers are very much focused on a problem, and something about that problem fascinates them. But oftentimes it's just not a good company, or it's…”
Park: Simile raised $200M, reaching $300M total funding in six months
“So we raised two hundred million dollars. So it brings our total funding to be three hundred million dollars raised over the past six months or so.”
Park: Simile raised capital because scaling data and compute accelerates progress
“We were sort of at this moment where, yes, we can actually significantly raise the input, both in terms of data, compute spend. To actually meaningfully accelerate this progress, that's when we thought it actually makes sense.”
Park: Simile raised Seed, Series A, and subsequent round within one year
“Obviously coming in, I had sort of a, you know, in my mental model, okay, well, if we raise seed now, that means we might raise our A in about a year and maybe B in the year after or something like that. All that happened within a year. We raised seed and I th…”
Park: Parts of the AI Market Are Quite Frothy
“I do think there's parts of market that is actually quite frothy.
For sure.
There's a lot of capital going in, there's a lot of excitement.”
Park: OpenAI and Anthropic Grew via Foreseeable Model Progress
“One of the most interesting thing about how OpenAI Anthropics and these companies grew was there were very strong fundamentals they could actually map out.
They could actually see, oh, the models are getting better at this rate.
Oh, and there's this kind of de…”
Park: Within three years, single AI simulations will cost $20M
“I actually do think simulation can actually be the next frontier of that, where in my vision, I think there's a world in which in about two, three years, we're running a single simulation session, and that's going to take 10, twenty million dollars to run a si…”
Park: PoC simulations are possible for anything; productionization is the real bottleneck
“I actually do think everything that we want to simulate, we can actually create the initial proof of concept. However, as we all know, one of the core challenges of AI is actually bridging the proof of concept with real value productionizable technology.”
Park: LLMs are the CPU of intelligence; simulation is the GPU
“What I see today that's prominent in AI space is what I consider to be the CPU of intelligence unit. You have this one language model that's really large, that's very smart, that can do very complex reasoning tasks. That's like CPU. What I see coming and what …”
Park: Simile may own a small hedge fund down the line
“Maybe similarly, we'll actually own a small hedge fund down the line.”
Park: AI simulation can replace proxy representatives with direct societal modeling
“I actually don't think this is a limitation we have to suffer through in the future. I think there's a world in which we can truly create a layer that becomes a representation layer of our society and of our collective intelligence.”
Park: AI Labs Without Clear Impact Visions Risk Failing as Companies
“I do think Neo Labs, without a clear vision for how they're going to impact the world, I do genuinely think there is some risk that they will turn out to be interesting research project, but not a viable company.”