Simile AI's Joon Sung Park pitches simulation as the next scaling law, claiming 85% accuracy at reproducing human responses

Latent Space·medium signalXBlueskyLinkedInCopy link

In an August 21 Latent Space interview, the generative-agents researcher turned CEO says behavioral foundation models trained on three data types (interviews, observed behavior, and randomized controlled trials) hit 85% accuracy reproducing human responses, versus 20-30% for frontier LLMs on niche populations. Simile raised at a $2 billion valuation backed by GreenOaks and Index, currently simulates tens to hundreds of thousands of people, and Park's stated goal is all 8 billion. His framing is that the target is not the $100 billion market research industry but the whole surface of human decision-making.