Simulation: the new Scaling Law — Joon Sung Park, Simile AI

Simile is a company building models that imitate how particular groups of people actually behave, including their biases and mistakes. In this interview, co-founder and CEO Joon Sung Park says models trained to be rational, helpful assistants make poor simulations of people, because simulating them is a different data and training objective.

Park describes three kinds of data: long interviews that supply personal context, observational records such as transactions that show what people do, and randomized controlled trials that show how decisions change when one variable changes. Customers name the population they want modeled, Simile recruits consenting participants from it, and the customer then queries the model with surveys or experiments. As a check on fidelity, Park cites a research study in which digital twins built from interviews reproduced real participants' answers about 85% as accurately as those participants reproduced their own two weeks later.