Simulating Humans at Scale: Simile's Joon Sung Park
Joon Sung Park, founder and CEO of Simile and creator of Stanford's "Smallville" generative agents study, explains how Simile is building the "GPU of intelligence" to simulate human society, diverging from frontier models that act as the "CPU of intelligence." He details Simile's approach of grounding simulations with real human behavioral data, its diverse corporate applications, and its long-term vision to create a "CERN of human society" to solve fundamental societal challenges.
Joon Sung Park, founder and CEO of Simile and creator of Stanford’s “Smallville” generative agents study, emphasizes the necessity of advanced simulations alongside AGI for technologically mature societies. He identifies a crucial opportunity to build such simulations today, a conviction developed through deep research.
The “Smallville” Experiment: Unlocking Micro-Behaviors
In April 2023, Park’s team at Stanford conducted the “Smallville” experiment, observing that large language models (LLMs) could encode human behavior from vast training data. By prompting these models correctly, they could elicit specific micro-behaviors. This insight led to the creation of generative agents—LLM-powered entities endowed with memory, planning, and reflection—designed to live out experiences in a simulated small town. The experiment involved 25 agents with distinct personas, who would follow daily routines, interact, form relationships, and exhibit emergent behaviors, such as spontaneously organizing a Valentine’s Day party. This demonstrated the models’ ability to generate complex, human-like social dynamics.
From Research to Company: Identifying a Critical Gap
Park’s journey into this field began around 2020, coinciding with the emergence of GPT-3. His early work at Stanford focused on the generalizability of these new models and their potential to encode human behavior. Coming from a social computing background, he recognized the difficulty of designing social platforms at scale, where traditional field testing is costly and slow. This motivated the 2022 project, “Social Similacra,” which used thousands of “personas” to simulate subreddits, revealing valuable insights into community dynamics and collaboration. This foundational work demonstrated the early potential for applying these simulations.
He argues that current frontier models, often developed by companies like OpenAI and Anthropic, serve as the “CPU of intelligence.” These models are designed for superintelligence, excelling at rational, objective problems with clear-cut answers. However, human society is inherently irrational, driven by subjective values, preferences, and tastes. This creates a divergence: while LLMs provide a foundational level of understanding, they plateau in accurately simulating true human behavior.
Simile, in contrast, aims to build the “GPU of intelligence.” Rather than striving for superhuman rationality, Simile focuses on encoding the diversity of human values, preferences, and tastes, creating models that are “as human as possible.” This recognition of the gap between rational problem-solving and diverse human behavior was a key catalyst for forming Simile as a company.
Simile’s Approach: Grounding Simulations in Real Data
Simile’s transition from research to a company was spurred by significant inbound interest from social scientists and Fortune 500 companies seeking to run experiments and answer market research questions beyond the scope of traditional methods. A pivotal moment was validating their simulations: by creating agents of 1,000 Americans, Simile demonstrated an 85% accuracy in predicting behavior compared to self-reported answers.
Simile’s customer engagement typically begins with understanding the target population. They partner with vendors like Gallup to collect unique behavioral data from real humans. This data helps close the “say-do gap”—the difference between what people say (often reflected in LLMs trained on online text) and what they actually do. They gather “long-tail information” through interviews (e.g., asking for a person’s life story) and use surveys for more discrete choices. This data, including insights from Randomized Control Trials (RCTs), is used to train a “foundation model of human behavior,” which can be further augmented with customer-specific in-house data.
Diverse Use Cases and the Value of Simulation
Simile’s platform, offered as a SaaS product, supports a range of use cases:
- Concept Testing: Rapidly testing new product ideas, market messages, and pricing strategies. Customers can test thousands of concepts across numerous sub-populations, far exceeding the capacity of traditional methods.
- Product Experience Simulation: Agents can “experience” a product over time, providing temporal feedback.
- Multi-Agent Simulations: Simulating complex scenarios like earnings calls to predict audience reactions.
Compared to traditional methods like running Facebook ads, Simile offers significant advantages:
- Scale and Representativeness: Simulations can achieve unlimited scale and ensure a truly representative population, overcoming biases present in online experiments.
- Downstream Implications: Simulations can model second-order impacts of decisions (e.g., how an EV launch affects perception of non-EVs), enabling evidence-based long-term strategic planning that is impossible with current field testing methods.
Model Evaluation and the Future of Simulation
Joon acknowledges the theoretical limit of predicting human behavior due to inherent randomness. However, Simile evaluates its models rigorously:
- Quantitative Metrics: For quantitative questions and RCTs, they use Total Variation Distance (TVD) to compare the distribution of simulated responses to ground truth, aiming for a TVD less than 0.15 for decision-making.
- Convergent vs. Divergent Simulations: Some simulations (e.g., network formation leading to hubs) converge to predictable outcomes despite small errors, as the underlying system dynamics are strong. Other simulations (e.g., election outcomes or historical counterfactuals) are inherently divergent, with outcomes varying based on initial conditions. For these, Simile focuses on confidence estimation (running simulations multiple times to observe outcome frequency) and illustrating the diversity of possible futures and their causal mechanisms. Simile views its work as laying the groundwork for “day one of inferential statistics” for simulation, aiming to establish robust mathematical frameworks and standards for the field.
Park envisions a future where simulations can address profound societal questions, much like Thomas Schelling’s agent-based models for segregation earned a Nobel Prize. These could include understanding bank runs, climate change collective action problems, the signals of a collapsing democracy, or the origins of monetary systems. He imagines a “CERN of human society”—large-scale, potentially multi-million dollar and months-long simulations—that could unlock fundamental insights into humanity and social sciences, akin to how the Hubble telescope revolutionized astronomy. This application of simulation to achieve a deeper understanding of human society and guide its improvement is what truly excites him.