Overview
What if decision-makers could test consequential choices on a credible simulation of humanity before acting in the real world? June describes a research and commercial program built around behavior foundation models: models trained not merely on what people publish online, but on interviews, observed actions, and randomized controlled trials that reveal why behavior changes. Simile recruits consenting participants, constructs individual and population-level representations, and validates digital twins against the same surveys, economic games, and experiments later completed by their human counterparts. One study reportedly reproduced behavior and attitudes 85% as accurately as people reproduced their own responses, compared with substantially weaker results from general frontier models in some settings. The immediate applications resemble accelerated human-panel research—concept testing, product evaluation, behavioral experiments, focus groups, and simulated earnings calls—but the thesis is broader. Prediction only says what may happen; simulation should expose the sequence of interventions capable of changing the outcome, including counterintuitive second-order effects. Scaling therefore requires more than a larger language model: it combines behaviorally grounded models, individual agents, multi-agent interaction, and rich environments. June ultimately envisions simulations approaching the scale of eight billion people, capable of investigating societal coordination problems while remaining grounded in consent, empirical validation, and continued real-world verification.
Sections
Higher-Order Implications
Patterns and implications that emerge across the interview.
- Behavioral simulation inverts the usual AI objective: irrationality, inconsistency, and personal bias become features to preserve rather than errors to eliminate.
- Simile's durable competitive advantage would likely come from its consented behavioral and causal data system, not merely from its model architecture, because general web data lacks many of the mechanisms the model must learn.
- The practical value of synthetic populations is initially less about replacing every human study than about representing stakeholders in the many routine decisions for which direct consultation is currently too slow, costly, or burdensome.
- As simulations become more influential, empirical validation and political guardrails become inseparable from product quality: an inaccurate or manipulable model of society could affect the society it claims to represent.
Core Concepts
Terms used to explain behavioral modeling and simulation.
- Behavior foundation model: a model trained to reproduce the attitudes, actions, biases, and causal decision mechanisms of people or populations rather than maximize abstract reasoning performance.
- Social physics: the underlying behavioral regularities and mechanisms governing how people respond to environments, other people, incentives, and interventions.
- Concept testing: evaluating alternative messages, product ideas, or propositions with a target population before committing to them.
- Attitudinal data: information about what people say, believe, or report that they would do, without necessarily observing a consequential action.
- Behavioral data: evidence of actions taken under real stakes, including transactions, platform activity, and consequential experimental choices.
- Randomized controlled trial: an experiment that varies selected conditions while holding the rest of the setup comparable, helping reveal causal mechanisms rather than correlations alone.
- Wicked problem: a complex coordination problem involving many actors with competing incentives and no simple, isolated solution.
Forecasts and Ambitions
Future developments anticipated by the speaker, with certainty represented qualitatively.
- Behavioral simulation models will continue improving through breakthroughs in data, algorithms, and more aggressive scaling.
- Large simulation runs may eventually cost as much as training a foundation model because their social and economic value could justify data-center-scale computation.
- Simulations may progress toward representing all eight billion people in rich multi-agent environments.
- High-fidelity societal simulations could help investigate climate coordination, democratic collapse, monetary systems, and universal basic income.
- The scientific impact of advanced social simulation could eventually merit a Nobel Prize in economics.