Episode Summary
Executive Summary: Jun Sung Park, founder/CEO of Simile, argues AI’s next frontier is not smarter reasoning alone but defensible behavioral data and simulation. He describes Simile’s origin in agent-based “Smallville” research, its enterprise wedge in market research, and a future where simulations become a core decision layer for companies, governments, and possibly markets themselves.
Main Topics: Simile’s origin in agent-based simulation research (Priority: 5/5): Park explains how early experiments with LLM agents, memory, planning, and reflection led to the Smallville demo, where NPCs lived routines and self-organized around Valentine’s Day. Foundational model of human behavior (Priority: 5/5): Simile positions itself not as a frontier LLM company, but as a model of human behavior that reproduces human biases, values, preferences, and mistakes for realistic simulation. Data strategy as moat (Priority: 5/5): Park emphasizes that defensible AI companies need an interesting data strategy. Simile combines behavioral, transactional, observational, and experimental data, especially RCTs and A/B tests, to infer causal mechanisms rather than only predict outcomes. Enterprise market research as the first wedge (Priority: 4/5): The company serves large enterprises first because they have budget, urgent pain, and immediate ROI. Park says customers use Simile to test campaigns, policies, and product decisions before launching. Accuracy, validation, and the data flywheel (Priority: 4/5): Simile claims high predictive accuracy and improves over time as simulated outputs are compared with real-world outcomes. Park frames the world itself as ground truth for learning. Research company and startup balance (Priority: 4/5): Park discusses the tension between scientific purity and product urgency, arguing Simile succeeds because the research mission and customer value are tightly aligned. Future vision: simulation as societal infrastructure (Priority: 5/5): He envisions a future where simulations scale to populations, ecosystems, and even markets, enabling better decisions on climate, democracy, product launches, and potentially finance.
Key Arguments: Defensible AI companies need proprietary data strategies; generic model access is not enough. Human behavior simulation requires data about what people do, not just what they say. Prediction alone is not the end goal; customers want causal understanding and counterfactuals so they can change outcomes. Enterprise customers move faster than expected when the pain is acute and the value is obvious. Simulation can unlock the 95% of hypotheses that companies and researchers never test today due to time, cost, or scale constraints. The world itself provides continuous validation, making simulation a learnable system with a strong feedback loop. The best teams combine contradictory superpowers, such as analytical rigor and creativity, or paranoia and long-term conviction. Simulation may become a new “representational layer” for society, helping decisions reflect people’s perspectives at scale.
Data Points: Initial model cost reduction: 100x cheaper - Park says one production model at Simile used to cost about 100 times more to run than it does now. Prediction accuracy: 85% as accurate as people replicating their own responses - He says Simile demonstrated models of people validated across surveys, behavior experiments, and real environments. Enterprise sales cycle: within 3 months - Park says some of the largest customers in the world moved quickly and closed deals in about three months. Funding raised in latest round: $200 million - He says Simile’s recent round was preempted by insiders and brought total funding to $300 million. Total funding raised: $300 million - Park states the company has raised $300 million over the past six months or so. Prior round size: $100 million - He notes Simile raised a $100 million round about five months earlier. Time to evaluate prior research studies: 3 to 6 months - He contrasts traditional study timelines with Simile’s ability to query outcomes in minutes. Simulation runtime vision: $10M to $20M per single session - Park predicts high-end simulation sessions could cost this much in 2-3 years. Per-session value vision: $100M - He believes enterprises or governments may pay this much for a single high-value simulation session.
Pivotal Quotes: "My fundamental thesis here is for AI companies of this generation, you need to have an interesting data strategy that's going to be defensible." — Jun Sung Park: Park explains his core view on AI moats and why Simile focuses on data. "We want our models to make the same kind of mistake. We want our models to be biased in the same way humans are." — Jun Sung Park: He distinguishes Simile from frontier model companies that optimize for superhuman rationality. "In my vision, I think there's a world in which, in about two, three years, we're running a single simulation session that's going to take 10, 20 million dollars to run. Single session, but it's going to be so valuable that people will pay $100 million for it." — Jun Sung Park: Park outlines the long-term economic upside of simulation as a product category.
Implications: Simulation may become a major AI category for enterprise, policy, and finance, shifting value from prediction to actionable counterfactuals. Winners will likely own proprietary behavioral data and highly validated models of human decisions.