AI Can Now Predict 85%+ of What You’ll Do Next | Simile Co-founder & CEO, Joon Sung Park
Joon Sung Park, creator of generative agents and CEO of Simile, explains how foundation models trained on real human behavior can simulate people, from a 25-agent town that organized its own Valentine's Day party to 85%-plus accurate predictions of how real participants respond in surveys, behavioral games, and randomized trials.
Generative agents and a town of 25
Joon Sung Park’s doctoral work started from an observation made in the GPT-2 and GPT-3 era: language models absorb so much human data that, probed from the right angle, they yield realistic human behavior. His 2023 Generative Agents paper turned that idea into a small game world of 25 agents, each with a persona, a daily routine, relationships, and memory, living in a grid-based town built with a map editor. Locations and obstacles were encoded as the map’s navigable grid, agents perceived their surroundings as objects, and those observations were translated into language and fed back into the model’s context. Early prompting was crude. The team coaxed GPT-3.5 into dialogue with HTML-like scaffolding, and agents were interviewed through a command line.
Emergent behavior
Nobody scripted the social life that unfolded. Two agents who had crushes on each other ended up in a romantic thread, one asked the other out, and a cafe owner’s Valentine’s Day party, set to start on February 13, became a gathering the agents organized among themselves. Agents received only a persona and a description, and the rest unfolded on its own. The team grew attached to the characters. A conference room at Simile is named after Isabella, the cafe owner, who occasionally appears as Park’s plus-one when he gives talks.
From game to social science
The path to a research contribution was not obvious. The earlier Social Simulacra work proposed simulating a Reddit-style community to prototype social systems and surface antisocial or disempowering dynamics before a platform launches. The designers still make the decisions, but the implications of those decisions become visible in advance. Park describes a ’time machine’ heuristic: imagine the most ambitious thing possible ten years out, then try to build a version of it now. Beneath the playful surface, he frames the project as a portrait of society. A simulation, like a painting, expresses what its creator perceives to be true about the world, and these agents stand in for real people whose interactions show how they might behave in a given environment.
Judging a simulation: from believability to accuracy
Evaluation proceeded in stages. The 2023 work was judged on believability. Agents were interviewed about who they are and how they spend their days, and their answers were compared with responses written by human authors, a kind of behavioral Turing test. The next step required ground truth. The team recruited real participants, collected life-story interviews and behavioral data, and trained agents on that data. Participants then left for a couple of weeks, returned, and were asked behavioral questions while their agentic counterparts tried to predict their answers.
The benchmark covered four categories: the General Social Survey, Big Five personality tests, economic games such as the prisoner’s dilemma, and randomized controlled trials replicated from published studies in venues such as PNAS and Nature. Agents predicted how real participants would respond, and the team reported over 85% accuracy. Park is careful about what that figure means. Many questions are discrete choices with five or ten options. For studies that measure effect size, the team computed Pearson correlations, and values of 0.7 to 0.75 or higher count as strong. Humans are also not perfectly self-consistent. Under retest, people change their answers roughly 13 to 14% of the time, which puts a practical ceiling of about 86 to 89% on accuracy. Predicting people at 100%, or even 95%, would itself be a red flag.
The result Park considers most important is the ability to predict randomized trial outcomes. Observational prediction is useful, but designing systems that improve the world is a causal problem: what happens when you intervene? Estimating the effect of different conditions, rather than only a baseline, is what makes a simulator usable as an A/B testing tool.
Intuition, scale, and the limits of prediction
Simulation is not simply better than a researcher’s hypothesis, though sometimes it is. Park notes that researchers’ intuitions reflect their own values and behavioral tendencies, which often differ from the populations they study. The main advantage of simulation is scale and representativeness. It bridges a researcher’s sense of how people are to the reality of a diverse population. The same logic applies to personal prediction. In some contexts a model can outperform even a close family member, because it can be more objective, more data-driven, and better at pattern-matching details a relative might miss. The model’s limitation is depth. A profile built from a couple of hours of interaction lacks the years of lived experience that family members have accumulated.
How Simile’s model differs from frontier labs
Simile is training its own foundation model of human behavior, which Park distinguishes sharply from models like ChatGPT or Claude. Those are reasoning machines, trained on expert reasoning traces in domains such as programming and mathematics, where correct answers are objective. Simile’s model does not aim to become superhuman at any task. It is trained on data collected from real people through roughly 40 vendors, including a strategic partnership with Gallup, whose representative panels span demographics and populations. The training objective is to learn the distribution of accurate responses across the population of interest.
Two kinds of models are trained. A population-level model describes a group and its environment and predicts how that group’s distribution of behavior would shift. An individual-level model aims to represent a specific person. Simile began with individual models because individuals are the quantum units that aggregate into populations. An accurate individual model leads fairly naturally to population-level ones. Training begins from open-source base models, followed by mid- and post-training that shapes the parameters to represent people at scale.
Park also addresses the skew of web data, which overrepresents heavily online people and underrepresents many others. Simile’s answer is to collect its own ground truth. Each week, the company surveys tens of thousands, sometimes hundreds of thousands, of real people about their attitudes, values, and preferences. That includes areas that rarely appear in text, such as color, texture, and style. For a visual artist, the relevant questions concern minimalism, formalism, or maximalism. The company first establishes views on these underlying units and then percolates them up to specific products, messages, and artworks.
Product and market testing
In practice, Simile’s work starts with concept and message testing and moves into product testing. Agents can use an application for a few minutes and report where they get stuck, what they want to see, and whether a color scheme resonates or feels tiring. Deeper questions compare variants, such as a different color or different wording. As the work nears business needs, the question becomes whether a change would increase retention or sales.
From Stanford research to a company
Interest in the work came first from social scientists and psychologists who wanted to run experiments they could not otherwise afford, and then from large enterprises whose CEOs and board members noticed it. The 2024 paper convinced Park that the models had reached a point of immediate market impact, comparable to the GPT-4 era. He favors visions whose ceiling is set by execution rather than by the idea, and he argues the technology will keep improving with research and resourcing. He left Stanford in June of last year to start the company.
Simile has grown to roughly 75 to 80 people. Its co-founders include Michael Bernstein and Percy Liang, both Stanford professors and Park’s former doctoral advisers, along with Laney, whom Park recruited to build the enterprise business. Park describes Simile as a hybrid of an AI neolab and a product company, with about a fifth of the staff drawn from academic lineages connected to the founders. The company uses a forward-deployed model, placing its people inside customers’ operations to build tailored use cases and relationships.
Moving from researcher to CEO
Running a company differs from leading a research team. Enterprise customers require clear ROI and a degree of certainty, while research presents a vision and a prototype. Park also sees deep similarities, since both come down to craft and to persuading others to join a vision. He looks for people who take high ownership, reinvent themselves, and combine contradictory strengths. Bernstein is his example. He is an analytically rigorous scientist who still has childlike wonder, and when Park first pitched the idea of a game world, Bernstein responded that he did not know why it was a good idea, but it was compelling.
Park describes his own leadership as holding two tensions at once. There is a sense of urgency and even a degree of paranoia about execution, alongside a near-religious conviction that the technology will be positive when implemented well and that the company will succeed.
Simulating meetings and organizational decisions
Park runs important meetings through simulation first. Conversations with public figures are easy to simulate given their online presence. Internally, Simile built ‘siblings,’ agentic simulations of employees derived from Slack activity. The idea began as a hackathon project. Colleagues can ask a sibling questions such as what June wants for dinner or whether he would approve a budget. The company uses this heavily, though siblings cannot grant approvals, so those still go to the human.
Free will and responsibility
Asked whether this changes his view of free will, Park separates the two questions. For him, free will is less about which action one takes and more about ownership, the willingness to be accountable for consequences and the right to take that responsibility. He believes in free will, and he regards the ability to take responsibility for one’s actions as a privilege of being human.