Building AI Twins From Real Human Interviews

Joon Sung Park, co-founder and CEO of Simile, joins Mario on The Generalist for a 70-minute conversation about simulating human behavior.

Interview people before simulating them

• Simile combines interviews, observed behavior, and experiments to model individuals and populations.
• Park says current collection typically takes 15–30 minutes per person, compared with two-hour interviews in earlier research.
• Broad life-story questions aim to capture motivations that transfer beyond one survey or product category.
• He describes consented panels, including a Gallup partnership; the interview supplies no independent privacy audit.

From a synthetic panel to a simulated society

• Customers choose a population and an environment: a survey, product interaction, policy, or conversation.
• Park describes access to individual static, individual dynamic, and multi-agent simulations.
• Whole-market and societal simulation remain the larger ambition behind the headline about everyone getting an AI twin.
• His artistic analogy is a portrait: a selective representation that can reveal something about its subject.

The accuracy claim needs its denominator

• The interview’s “85% plus accuracy” shorthand refers back to research on interview-grounded agents.
• The original paper’s abstract reports survey responses replicated 85% as accurately as people repeated their own answers two weeks later.
• That is a result relative to human self-consistency on specified questions, rather than 85% accuracy on arbitrary future actions.
• Park says customers test simulations against studies with known outcomes, while a separate model estimates likely prediction error.

Observing a purchase does not explain why it happened

• Transaction records show choices; interviews and experiments help distinguish motivations and plausible alternatives.
• Park’s example: a missed coffee purchase could reflect changed tastes or a disrupted journey.
• He describes randomized studies with real consequences, such as actually sending participants the products they choose.

Test the whole objective

• An EV-marketing simulation reportedly improved electric-vehicle demand while reducing demand elsewhere in the manufacturer’s business.
• Park presents that as a warning about optimizing one product line while harming the overall result.
• Simile also used simulation to inform its headquarters move; internal Slack “siblings” rehearse discussions among simulated colleagues.
• These are company anecdotes and simulation outputs, without a demonstrated causal comparison against real-world alternatives.

People still choose the question

• Schelling’s segregation model illustrates how simple individual rules can produce surprising collective patterns.
• Park’s long-term vision resembles science fiction’s goal-directed search across possible futures.
• Competitive advantage, he argues, would still depend on human framing, institution-specific data, and available computing resources.
• Investing, legal strategy, relationship rehearsal, and climate policy are discussed as possibilities with very different evidence requirements.

“It will always be humans who ask the goal and who ask the right questions.” — Joon Sung Park, 59:06

Caveat: this founder interview describes a product and a research agenda. It does not demonstrate a general-purpose system that reliably predicts whole societies.