Simulating Humanity: From Generative Agents to 8 Billion Digital Twins
Latent Space · youtube · 2026-08-22
Latent Space interviews Joon Sung Park (Simile AI) about the evolution from the Smallville Generative Agents experiment to building foundation models of human behavior. Key topics include:
- Why Simulate?: To test products and policies before deployment and model emergent societal behaviors.
- Limits of Frontier Models: Current models fail to capture actual human behavior because they are optimized for rationality, whereas real humans are biased and irrational.
- Simile's Approach: Using long-form interviews, observational data, transactions, and RCTs to post-train models on the causal mechanisms behind decision-making.
- Digital Twins: The team created digital twins that reproduced human behavior with 85% accuracy.
- Shaping the Future: The goal is less about predicting the future and more about understanding how to influence it.
The conversation also touches on scaling laws for simulation and the economics of data-center-scale simulated worlds.
More from AGI Musings
- Superintelligence hailed as pinnacle of Western tech to lead humanity across star systems — SydSteyerhart · 2026-08-22
- Opinion: Datacenter Sudo Creds Are the 21st Century Keys to Power — curious_vii · 2026-08-22
- Tech evolution history: From nuclear energy to potential human divergence — Dr_Singularity · 2026-08-22
- Astra discussion shifts to containment and autonomy — VraserX · 2026-08-22
- Jeff Ladish: banning datacenters only works if both the US and China do it — JeffLadish · 2026-08-22
- Public opposition to data centers stems from intuitive sensing — JeffLadish · 2026-08-22