When we first discussed the Summer of Simulative AI in 2024, we knew it would be a short summer. But recently, it has returned with tremendous momentum: first with SimGym in April, and now with Simile AI’s $200 million Series B at a $2 billion valuation. The round was backed by GreenOaks and Index Ventures, with prominent investors including Fei-Fei Li and Andrej Karpathy. Simile AI is running tens of millions of simulations for Fortune Global 100 customers, including CVS, with results achieving 85%–99% accuracy compared with human focus groups.
It’s time to understand why the second summer of simulation is taking off!
Beginning with Smallville—the landmark 2023 Generative Agents paper that demonstrated that AI characters can remember, plan, socialize, and develop emergent behaviors—Joon Sung Park is now attempting to answer a much bigger question: What if we could simulate the world before making decisions? In this episode, the co-founder and CEO of Simile joins us to unpack the journey from generative agents to digital twins, explore why even today’s frontier models still fail to capture how humans actually behave, and discuss what it would take to eventually simulate all 8 billion people on Earth.
We take a deep dive into Simile’s approach to modeling human behavior: long-form interviews, observational and transactional data, randomized controlled trials, population-level and individual-level models, and post-training focused on the causal mechanisms underlying people’s decisions. Joon explains how his research creates digital twins that reproduce human behavior and attitudes with 85% of the accuracy people achieve when reproducing their own answers; why models optimized for rationality may fail to simulate irrational humans effectively; and why understanding “social physics” may require changing model weights rather than merely prompting frontier large language models.
We also explore the broader ambitions behind simulation: testing products and policies before deploying them, finding counterintuitive paths toward desired outcomes, simulating emergent behavior across entire societies, and potentially addressing issues such as climate change, democratic instability, and universal basic income (UBI). Joon also shares his thoughts on scaling laws for simulation, the economics of data-center-scale simulated worlds, the connection to Thomas Schelling and psychohistory, why simulation is surprisingly similar to painting, and whether we might already be living in a simulated world.
We discussed:
- How Smallville and Generative Agents led to Simile
- Why Joon’s team asked: “What if we could directly reconstruct the world we live in?”
- Why useful personal agents need to build deep models of their users
- Memory architectures, Markdown files, and the limitations of prompting
- “Social physics” and foundation models of behavior
- Why web data captures more of what people say than what they actually do
- Interviews, transactional and observational data, and randomized controlled trials
- Why predicting the future matters less than understanding how to shape it
- How Simile creates representative simulated populations
- The difference between simulation and prediction, and its connection to psychohistory in Foundation
- How to evaluate simulations rather than simply stacking up large language model hallucinations
- Creating digital twins for 1,000 real humans and achieving 85% behavioral accuracy
- Why frontier models struggle to reproduce real human behavior
- Why good simulations need to reproduce human biases and mistakes
- Post-training models based on randomized controlled trials
- Population-level and individual-level simulation
- Scaling laws for human simulation
- The long-term goal: simulating all 8 billion people on Earth
- Whether simulation can help solve climate change or detect the collapse of democracy
- Thomas Schelling and the history of agent-based modeling
- Why future simulations may require entire data centers
- Multi-agent simulations, and what happens when simulated people interact with one another
- Replacing expensive human sample groups with synthetic populations
- Why market research is only the starting point for simulation
- Why Joon believes simulation is surprisingly similar to painting
- Using simulation to study issues such as universal basic income (UBI)
- Whether we are already living in a simulated world
- Why artificial general intelligence (AGI) and simulation may be two twin technologies of advanced civilizations
Joon Sung Park
- LinkedIn: https://www.linkedin.com/in/joonspark
- X: https://x.com/joon_s_pk
- Website: https://www.joonsungpark.com
- Simile: https://www.simile.com
Timestamps
00:00:00 From art to AI: Joon’s journey and introduction to the episode
00:01:46 Smallville, Generative Agents, and the origins of simulation
00:05:03 “Let’s just create a world” and the future of personal agents
00:09:53 Social physics and foundation models of behavior
00:14:08 Prediction vs. simulation: How do we shape the future?
00:16:59 How Simile models real humans and populations
00:25:35 Evaluating simulations, digital twins, and 85% accuracy
00:30:23 Using post-training to reproduce human behavior
00:40:04 Scaling laws and simulating 8 billion people
00:43:10 From Schelling to agent-based simulations at societal scale
00:46:13 The cost and economics of simulated worlds
00:52:05 Real-world applications, synthetic populations, and markets
00:57:27 The future of simulation, painting, and universal basic income
01:04:23 Are we already living in a simulated world?
01:06:08 Building Simile and recruiting