Why build a simulated SF? To train agents' social intelligence and alignment

davidfromkansas · x · 2026-08-21

The author's motivation for SFSIM: virtual game environments acting as sandboxes for simulating the social intelligence of human-like agents with their own needs and goals. He argues such simulations could be key to scaling AI alignment — training agents to adopt beneficial behaviors while observing how they respond to scenarios hard to replicate physically. The experiment builds on the social simulacra research by @joonspk (of Stanford's Generative Agents), which derived the foundational techniques for generative agents that simulate human behavior.

Related event: Dev Builds SFSIM, a Live 3D Digital Twin of San Francisco with 1,000+ AI Citizens(6 posts)→

Original post →

More from Embodied

Embodied channel →