Joon Sung Park, founder and CEO of Simile, discusses the inspiration behind building large-scale human behavior simulations, drawing from science fiction visions of advanced societies guided by artificial general intelligence (AGI) and simulations. He recounts the Smallville project at Stanford in 2023, where generative AI agents with memory, planning, and reflection capabilities were created to simulate a small town of 25 agents. These agents exhibited complex, emergent social behaviors such as organizing a Valentine’s Day party, demonstrating the potential for AI to model nuanced human interactions and societal dynamics.
The genesis of Simile’s work traces back to early research around 2020 with GPT-3 and foundational models, where Park and his team realized that large language models encode rich human behavioral data. This insight led to experiments like Social Similacra, simulating online communities to understand social dynamics. Over time, improvements in AI models enabled the creation of more sophisticated agents capable of longitudinal reasoning and complex social interactions. However, Park notes current models plateau in simulating human irrationality and diversity of values, highlighting the need for new modeling paradigms that better capture subjective human behavior.
Simile’s platform partners with organizations like CVS to create simulations grounded in real-world data collected through surveys and interviews, including behavioral data from randomized control trials. These simulations enable clients to test concepts, marketing strategies, and product launches at scale, far beyond traditional polling or A/B testing. The platform allows users to explore not only immediate reactions but also long-term, multi-agent interactions and second-order effects, providing deeper insights into market dynamics and consumer behavior that are otherwise difficult or costly to obtain.
Park emphasizes the importance of rigorous evaluation metrics to ensure simulation accuracy, aiming for predictive performance close to how individuals replicate their own behavior. He explains that some simulations converge reliably while others diverge due to inherent randomness or complex social phenomena, and that understanding these dynamics is a key research frontier. The company is also exploring how to responsibly incorporate proprietary customer data to enhance simulation fidelity and tailor models to specific populations, balancing the use of foundational AI models with custom-built components designed to capture human diversity.
Looking ahead, Park envisions simulations becoming a transformative tool for understanding and guiding human society, akin to how scientific instruments like the Hubble telescope revolutionized astronomy. Beyond commercial applications, he sees potential for simulations to address grand challenges in economics, politics, climate change, and social sciences by modeling complex human behaviors at scale. Simile aims to pioneer this frontier, building toward a future where simulations serve as a critical infrastructure for evidence-based decision-making and societal progress.
