Joon Sung Park studies human–AI interaction and social computing. His research includes the Smallville generative-agent simulation, experiments that ground agents in people’s interviews and survey responses, and methods for prototyping social systems. He is also the founder and CEO of Simile, where he is exploring behavioral simulation for research and commercial decision-making. — Sequoia Interview.

Part 1: The Pace of Algorithmic Interaction
- A Pause Can Support Reflection: In controlled visual-estimation studies, Park and colleagues found that a slower algorithmic response gave participants more time to assess advice before changing their own answer. — ACM CSCW Research Blog.
- Slowness Can Help Users Recognize Good Advice: When an algorithm’s estimate was accurate in the study’s jelly-bean task, participants moved closer to it when the advice arrived after a delay than when it arrived immediately. — ACM CSCW Research Blog.
- Slowness Can Also Help Resist Bad Advice: With deliberately inaccurate advice in the same task, delayed responses produced smaller shifts toward the algorithm’s estimate; the reported comparison was suggestive rather than a universal guarantee. — ACM CSCW Research Blog.
- Speed Is Not the Only Design Goal: Park’s team argues that optimizing exclusively for speed can crowd out reflection in consequential human–algorithm decisions, though waiting also imposes a cost. — ACM CSCW Research Blog.
- Make Space for Consequential Judgment: Park and coauthors suggest designing more reflective interfaces where algorithmic advice affects important decisions; their experiments establish this only in a simple estimation task, not in bail or healthcare decisions themselves. — ACM CSCW Research Blog.
Part 2: The Genesis of Generative Agents
- A Small Town Demonstrated Agent Interaction: Park and coauthors populated a game-like town with 25 generative agents that followed routines, formed relationships and interacted in an open-ended simulation. — Generative Agents Paper.
- A Single Prompt Can Trigger Coordination: In the Smallville demonstration, one agent’s Valentine’s party intention led others to circulate invitations, arrange dates and gather at the cafe without those steps being scripted individually. — Generative Agents Paper.
- Agents Can Forget Invitations: Park recalls that some simulated residents received invitations yet did not attend; the paper’s end-to-end analysis identifies missed planning and conflicts among the nonattendees. — Generative Agents Paper.
- Unexpected Social Paths Can Emerge: Park recounts an uninvited resident arriving at Smallville’s party as another agent’s date—an example of social coordination not scripted as a fixed event sequence. — Sequoia Interview.
- Language Models Encode Behavioral Patterns: Park’s starting observation was that large language models contain patterns of human behavior from their training data that can be elicited in specific situations, beyond simple text-generation demonstrations. — Sequoia Interview.
- Believable Agents Were a Longstanding HCI Goal: Park places generative agents within a longer effort in human–computer interaction and cognitive modeling to build characters that can sustain routines and relationships without brittle scripts. — Park’s Dissertation.
Part 3: The Architecture of an Agent
- Memory, Reflection and Planning Work Together: The Smallville architecture stores experiences, synthesizes reflections and retrieves relevant memories to guide plans and reactions in a changing environment. — Generative Agents Paper.
- Each Architectural Component Matters: In ablation tests, removing observation, planning or reflection reduced agent believability on the researchers’ evaluation tasks; that result does not prove human-equivalent behavior. — Generative Agents Paper.
- Reflections Synthesize Experience: The agent system records observations in natural language and periodically derives higher-level reflections that can inform later behavior. — Generative Agents Paper.
- Retrieve the Memories Relevant Now: Rather than feeding an agent its entire history, the architecture selects memories relevant to the current situation before generating a response or action. — Generative Agents Paper.
- Specific Context Improves Behavioral Probes: Park found that asking what a person might do in a particular situation elicited more useful microbehaviors from early language models than asking for generic behavior. — Sequoia Interview.
Part 4: Prototyping Social Systems
- Scale Changes Social-System Design: Park argues that platform designers must consider how many users interact, because emergent group behavior is harder to anticipate than an interface’s immediate usability. — Sequoia Interview.
- Field Tests Can Impose Human Costs: Park notes that testing social-platform designs only after public release can expose people to harmful dynamics, such as feeds that amplify negative emotions. — Sequoia Interview.
- Simulate Before Exposing Users: Park’s research explores simulations as a way to probe possible community behavior and moderation choices before a design is launched; such probes are not a guarantee of real-world safety. — Sequoia Interview.
- Prototype Communities with Synthetic Members: Before Smallville, Park’s Social Simulacra project used language models to populate a proposed online community with personas and interactions so designers could explore its dynamics. — Social Simulacra Paper.
Part 5: Grounding Agents in Real Human Data
- Real-Person Simulation Needs More Than a Fictional Bio: Park distinguishes believable fictional agents from research agents intended to approximate actual individuals; the latter need person-specific evidence, not just a short invented persona. — Stanford HAI Interview.
- Agents Can Role-Play Interviewees: The research agents were prompted to answer as particular interview participants using those participants’ own self-reports; this is a predictive method, not proof that an agent becomes that person. — Stanford HAI Interview.
- An AI Interviewer Can Gather Life Stories: For a study of 1,052 U.S. adults, the team used an AI interviewer to conduct roughly two-hour semi-structured conversations with adaptive follow-up questions. — Self-Report Agents Paper.
- Model-Generated Expert Reflections Summarize Interviews: The agent architecture adds short model-generated reflections from four social-science role perspectives to draw out implications in participants’ self-reports; these are not notes written by human experts. — Self-Report Agents Paper.
- Individual Data Can Reduce Stereotyped Prediction: In the study, agents grounded in people’s interviews or surveys generally narrowed racial and ideological accuracy disparities relative to demographic-only prompting, without eliminating them. — Self-Report Agents Paper.
- More Data Can Have Diminishing Returns: The combined interview-and-survey agents performed best on held-out General Social Survey items, but their gains over single-source agents were modest, suggesting diminishing returns after sufficient person-specific evidence. — Self-Report Agents Paper.
Part 6: Accuracy and Behavioral Prediction
- Benchmark Against People’s Own Consistency: The researchers compared agent answers with participants’ answers across two survey waves, using each person’s two-week test–retest consistency as a benchmark rather than treating a single answer as an infallible ground truth. — Self-Report Agents Paper.
- Report Survey Accuracy Relative to Test–Retest: In the revised study, interview-only agents reached 83% and interview-plus-survey agents 86% of participants’ own two-week consistency on held-out General Social Survey items; these are normalized, not raw, accuracies. — Self-Report Agents Paper.
- Treat Economic-Game Results as a Boundary Case: The agents predicted choices in five incentivized economic games, but performance differences among agent types were not statistically significant; the authors treat this as a limitation, not a demonstrated advantage of interview grounding. — Self-Report Agents Paper.
- Experimental Effects Need Careful Interpretation: Across five social-science experiments, agent-estimated effect sizes tracked the direction and relative size of the human sample’s results, but the authors caution that five studies leave the comparison underpowered. — Self-Report Agents Paper.
- Generalize Across Evaluated Outcomes, Not All Behavior: The self-report agents predicted held-out responses across surveys, personality measures and experiments without task-specific training; the authors explicitly say population-level correlational structure remains untested. — Self-Report Agents Paper.
Part 7: Commercial Applications and the Say-Do Gap
- Distinguish Convergent from Divergent Simulations: Park distinguishes questions that tend toward similar aggregate outcomes from those where small differences can yield very different trajectories; the appropriate evaluation depends on which type is being modeled. — Sequoia Interview.
- Treat TVD as a Proposed Decision Threshold: For quantitative population-response comparisons, Park says Simile aims for total variation distance below 0.15 as evidence useful to its decisions; this is his stated threshold, not a universal validation standard. — Sequoia Interview.
- Estimate Uncertainty in Divergent Scenarios: For simulations whose outcomes vary, Park proposes repeated runs and bootstrap-style resampling as ways to examine confidence and possible futures; he presents this as an evaluation question, not a proven requirement. — Sequoia Interview.
- Explore More Concepts Than Traditional Panels: Park describes testing many concepts against many subpopulations as a commercial vision for simulation, beyond the five to ten ideas a customer might conventionally test; the transcript does not validate thousands of forecasts. — Sequoia Interview.
- Model Diverse Preferences, Not Just Rational Answers: Park contrasts frontier models optimized for objective tasks with his aim to model heterogeneous human values, tastes and preferences; his CPU/GPU comparison is a metaphor for that research direction. — Sequoia Interview.
- Address the Say–Do Gap with Behavioral Data: Park argues that online text captures much of what people say, while behavioral and life-story data may help model what they actually do; closing that gap remains Simile’s goal. — Sequoia Interview.
- Study Second-Order Product Effects: Park proposes using simulations to explore whether a new product might change perceptions of an existing product line; the interview presents this as a possible use case, not a validated forecast. — Sequoia Interview.
Part 8: Ethics, Privacy, and the Future of Society
- A Person Should Control Their Agent: Park argues that an agent based on someone’s detailed personal data should belong to and remain under that person’s control, comparing the sensitivity to genomic data. — Stanford HAI Interview.
- Consent Must Be Revocable: Park says people represented by research agents should be able to withdraw consent and have the corresponding agent removed from contexts where it operates. — Stanford HAI Interview.
- Restrict Access to Real-Person Agent Banks: The study’s agents were not released for public use; Park says researchers seeking access should apply and provide privacy assurances. — Stanford HAI Interview.
- Use Simulations to Explore Policy Hypotheses: Park hopes empirically grounded simulations could help probe complex policy questions such as climate action, while acknowledging this is a potential research testbed rather than a proven solution. — Stanford HAI Interview.
- Science Fiction Suggests Two Complementary Tools: Park says science fiction influenced his vision of advanced societies combining powerful AI with simulations that help them reason about collective choices. — Sequoia Interview.
- Current Models Permit a First Attempt: Park believes current models make a first serious attempt at social simulation possible; he explicitly frames more capable societal simulators as future research, not an existing achievement. — Sequoia Interview.
Part 9: From Prototypes to Experimental Science
- Populate a Community Before Launch: Social Simulacra takes proposed community goals, rules and personas and generates synthetic members, posts and replies, including antisocial interactions, for design exploration. — Social Simulacra Paper.
- Compare Design Interventions in Prototype: In the paper’s simulations, generated community behavior changed when designers altered rules or tried member and moderator interventions, enabling what-if exploration before release. — Social Simulacra Paper.
- Use Prototypes to Surface Failure Modes: Designers in the study used populated social-system prototypes to identify possible successes and failures and refine their concepts; these observations are design inputs, not real-world outcome forecasts. — Social Simulacra Paper.
- Build Behavioral Data from Many Experiments: The SocSci210 dataset assembled 2.9 million responses from 400,491 participants across 210 open social-science experiments for behavioral-prediction research. — SocSci210 Paper.
- Fine-Tuning Improved Response Distributions: In the SocSci210 study, fine-tuning improved the strongest model’s alignment with human response distributions in unseen studies relative to its unchanged base model. — SocSci210 Paper.
- Test Generalization on Unseen Studies: The authors evaluated whether learned behavioral patterns transferred to studies absent from training and reported better distributional alignment for their fine-tuned model in that setting. — SocSci210 Paper.
- Partial Studies Can Inform Nearby Conditions: Fine-tuning on some conditions within a study improved predictions for conditions the model had not seen, according to the paper’s evaluation; this is within-study transfer, not arbitrary counterfactual forecasting. — SocSci210 Paper.
- Measure Demographic Prediction Gaps: Using the dataset’s demographic information, the researchers reported a 10.6% reduction in demographic parity difference after fine-tuning on their evaluated tasks. — SocSci210 Paper.
- Use Simulations as Scientific Instruments: Park argues that generative-agent simulations should complement empirical observation by challenging assumptions, revealing dynamics and generating testable hypotheses before real-world deployment. — Park’s Dissertation.
Part 10: General User Models and Proactive Computing
- Represent User Context with Calibrated Propositions: The general-user-model paper converts observations such as screenshots into confidence-weighted propositions about a person’s knowledge, goals and preferences, rather than storing only raw activity. — General User Models Paper.
- Revise a User Model as Evidence Changes: The proposed architecture retrieves related propositions and revises them as new observations arrive, so an earlier inference need not remain fixed. — General User Models Paper.
- Carry Context Across Applications: The paper demonstrates a cross-application user model that can contextualize assistant prompts, notification filtering and agent actions, rather than limiting personalization to one app. — General User Models Paper.
- Prototype Proactive Assistance: In the research prototype, assistants using a general user model identified and performed some useful actions participants had not explicitly requested; this is an evaluated prototype, not a general deployment claim. — General User Models Paper.