Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Simile AI raised a $2B Series B backed by GreenOaks, Index Ventures, Fei-Fei Li, and Andrej Karpathy, marking the "Second Summer of Simulation" following the 2023 Generative Agents paper The company builds behavioral foundation models ("social physics") that simulate human behavior at 85-99% accuracy compared to human focus groups, used by Fortune 100 clients like CVS Joon Sung Park's approach combines long-form interviews, observational/transaction data, randomized controlled trials, and post-t
Analysis
TL;DR
- Simile AI raised a $2B Series B backed by GreenOaks, Index Ventures, Fei-Fei Li, and Andrej Karpathy, marking the "Second Summer of Simulation" following the 2023 Generative Agents paper
- The company builds behavioral foundation models ("social physics") that simulate human behavior at 85-99% accuracy compared to human focus groups, used by Fortune 100 clients like CVS
- Joon Sung Park's approach combines long-form interviews, observational/transaction data, randomized controlled trials, and post-training on causal decision mechanisms rather than relying on prompting frontier LLMs
- The technology aims to scale from simulating 1,000 individuals to all 8 billion people, enabling pre-deployment testing of products, policies, UBI, climate strategies, and democratic stability
- Key insight: rational-optimized models fail to capture real human behavior; accurate simulation requires reproducing human biases and irrationality through weight-level training, not just prompting
Why It Matters
This represents a paradigm shift from using LLMs as text generators to using them as behavioral simulators—fundamentally changing how companies and governments can test decisions before acting. For AI practitioners, it signals that the next frontier isn't just bigger models but deeper models of human behavior, with significant commercial validation already underway.
Technical Details
- Data pipeline: Combines long-form interviews, observational data, transaction records, and randomized controlled trials to build population-level and individual-level behavioral models
- Post-training approach: Models are fine-tuned on the causal mechanisms behind human decisions rather than relying on in-context prompting of frontier LLMs; this weight-level training is essential for reproducing irrational behavior and biases
- Evaluation methodology: Digital twins of 1,000 real people achieved 85% behavioral accuracy (how accurately people reproduce their own responses), with Fortune 100 simulations reaching 85-99% accuracy versus human focus groups
- Scaling ambition: Moving from individual-level to population-level to society-scale multi-agent simulations, with the long-term goal of simulating all 8 billion people on Earth—potentially requiring data-center-scale compute
- Architectural lineage: Evolved from the Smallville/Generative Agents project (memory architectures, Markdown-based state, emergent social behaviors) toward foundation models of human behavior
Industry Insight
- The $2B valuation and Fortune 100 adoption signal that synthetic populations will disrupt market research, policy testing, and product development—replacing expensive human panels with scalable simulation
- The distinction between prediction and simulation is strategic: companies should invest in understanding how to shape outcomes through simulation rather than merely forecasting them
- The emphasis on post-training over prompting suggests a new competitive moat: behavioral models trained on proprietary RCT and transaction data will be harder to replicate than general-purpose LLM capabilities
Disclaimer: The above content is generated by AI and is for reference only.