Video · 2026-08-01 · 1h 5m · 6 moments

The AI Company Simulating the Entire Economy | Simile Co-founder & CEO, Joon Sung Park

✦ AI generated

timeline · colored by role

Claim

Large language models, when prompted correctly, can extract realistic human behaviors because they are trained on vast amounts of human behavior and sentiment data expressed on the web, and this capability is domain-agnostic.

Joon explains the origin of Simile's approach: LLMs trained on web data can, when probed correctly, produce realistic human behaviors across any domain, leading to the creation of generative agents.

transcript

Joon Sung Park: We had this idea that large language models are often used for simple tasks like classification, simple generation, but we thought that these models actually had a lot more potential. 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 the right angle, you could actually extract a lot of realistic human behaviors out of them. I thought that was really interesting, and it was also practically interesting in that it was domain agnostic.

extends · 1

Claim◆

Simile does not aim to create super-rational machines good at coding or math; it aims to model the subjective half of the human brain — people's values, preferences, taste, and the same biases and mistakes humans make.

Joon distinguishes Simile from frontier LLM companies: Simile models human subjectivity and error, not super-rational intelligence, so its agents make the same mistakes humans would.

transcript

Joon Sung Park: If you look at large language 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 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. In a way, we want to be a representation of people's values, preferences, and taste. Sort of their subjective half of their brain.

explains mechanism · 1

Claim◆

No one really cares about prediction of the future; what people actually want is to shape the future through understanding causal mechanisms and counterfactuals — what they need to do now to change the outcome.

Joon argues that prediction alone is insufficient — companies and policymakers need causal models that show them how to intervene and change the future, which is why Simile focuses on randomized control trials and counterfactual data.

transcript

Joon Sung Park: My personal hot take here is a lot of observational behavior data and what they're amazing at is actually helping you create a correlation of the observation and what could happen in the future. Good for prediction task. But, my take here after interacting with so many of our customers and also being in research, 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. They want to know, imagine you're a Starbucks, doesn't really help them to know that your Frappuccino sales is going to tank in two quarters. They'll hear that and they'll be like, "What do we do about them? That's terrible." What they want to know is how can we prevent it? What do we need to do now to change the future? And there, what you really need is causal mechanism. You need a model that can actually reason about causal mechanisms and counterfactuals.

extends · 1

Prediction

Simulation can be a cure for wicked problems like climate change by modeling collective action among stakeholders with different incentives, revealing equilibrium states and conditions under which systems like democracy succeed or fail.

Joon envisions simulation tackling society's hardest problems — climate change, democracy stability, collective action — by modeling how different stakeholders with conflicting incentives interact and reach equilibrium.

transcript

Joon Sung Park: I also get quite excited by the vision where simulation I do think can also be a cure for many of what we call 'wicked problems.' A good example here might be things like climate change requires collective action across many stakeholders who have different incentives. One of the reasons why such problems are so difficult is actually finding the right equilibrium state where all different parties come together to make decision for global good is very difficult. Can we actually simulate those decision-making processes? Can we actually simulate even in things like in what conditions does a democracy fail? Can we actually predict that? These are the kind of questions that simulation ultimately can answer.

Mechanism

Similation has an even better feedback mechanism than coding agents because the world is the ground truth — every day Simile generates tens of thousands of hypotheses, watches the world, and validates which came true, enabling continuous learning.

Joon describes how Simile's data flywheel works: generating daily hypotheses about future events and validating them against real-world outcomes, creating a self-improving prediction system analogous to AlphaGo.

transcript

Joon Sung Park: 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 statement. If this happens, we know whether we can validate the simulation to be right or wrong. And we're basically watching the world every day seeing which of those hypotheses are answerable at what time. And we can basically say, a month goes by, we generated a million hypotheses, X percentage of them came true. This is a best way to learn about the world.

Prediction

In 2–3 years, Simile will run single simulation sessions costing $10–20 million to operate, but they will be so valuable that customers will pay $100 million for one session.

Joon predicts that as simulation complexity grows — with models thinking for hours or days — a single high-end simulation session could cost tens of millions to run but command $100 million in value for the world's largest enterprises and governments.

transcript

Joon Sung Park: I actually do think simulation could actually be the next frontier of that. Where in my vision, I think there's a world in which in about 2 3 years, we're running a single simulation session that's going to take 10, 20 million dollars to run a single session, but it's going to be so valuable that people will pay 100 million dollars for it.

Highlight slides

Simile models human subjectivity, not super-rationality✦ from: Simile does not aim to create super-rational machines good at coding or math; it aims to model the subjective half of the human brain — people's values, preferences, taste, and the same biases and mistakes humans make.Simile agents reproduce human error and bias✦ from: Simile does not aim to create super-rational machines good at coding or math; it aims to model the subjective half of the human brain — people's values, preferences, taste, and the same biases and mistakes humans make.Prediction Isn't What People Want✦ from: No one really cares about prediction of the future; what people actually want is to shape the future through understanding causal mechanisms and counterfactuals — what they need to do now to change the outcome.The Ask: Causal Intervention✦ from: No one really cares about prediction of the future; what people actually want is to shape the future through understanding causal mechanisms and counterfactuals — what they need to do now to change the outcome.

Related episodes

Audio · Lex Fridman

#464 – Dave Smith: Israel, Ukraine, Epstein, Mossad, Conspiracies & Antisemitism

→

Video · a16z Podcast

Training Robots in Worlds That Don't Exist | World Labs with a16z

→

Video · Syntax

AI Is Stealing From You

→

Large language models, when prompted correctly, can extract realistic human behaviors because they are trained on vast amounts of human behavior and sentiment data expressed on the web, and this capability is domain-agnostic.

Simile does not aim to create super-rational machines good at coding or math; it aims to model the subjective half of the human brain — people's values, preferences, taste, and the same biases and mistakes humans make.

No one really cares about prediction of the future; what people actually want is to shape the future through understanding causal mechanisms and counterfactuals — what they need to do now to change the outcome.

Simulation can be a cure for wicked problems like climate change by modeling collective action among stakeholders with different incentives, revealing equilibrium states and conditions under which systems like democracy succeed or fail.

Similation has an even better feedback mechanism than coding agents because the world is the ground truth — every day Simile generates tens of thousands of hypotheses, watches the world, and validates which came true, enabling continuous learning.

In 2–3 years, Simile will run single simulation sessions costing $10–20 million to operate, but they will be so valuable that customers will pay $100 million for one session.

#464 – Dave Smith: Israel, Ukraine, Epstein, Mossad, Conspiracies &#038; Antisemitism

Training Robots in Worlds That Don't Exist | World Labs with a16z

AI Is Stealing From You

#464 – Dave Smith: Israel, Ukraine, Epstein, Mossad, Conspiracies & Antisemitism