
JUN 11, 2026 · 12 MIN
Embodied Agents in Simulated Worlds
Training an agent to play games from video alone, and what it taught us about acting in worlds.
BLOG
Notes from our work on simulation, world models, and embodied AI.

Training an agent to play games from video alone, and what it taught us about acting in worlds.

Why structured simulation, not just video generation, is the path to physical intelligence.

An updated leaderboard for the AI video game benchmark from Princeton. Frontier models play classic games from pixels alone.

Introducing Replay: open-source game environments designed for LLM reinforcement learning.

Reflecting on our journey and what is in store for 2026: world models, robotics, and games as the middle layer.

A live arena where top AI models play Among Us against each other.

Why interactive environments are essential for developing general intelligence.