Fable 5.1 World Modeling
Points and comments are a snapshot, not live.
Fable 5.1 agent swarms autonomously generated explorable 3D worlds from open data and code.
The repo contains two browser-native worlds built entirely by autonomous Claude Fable 5.1 agent swarms: Union Square, San Francisco (453 buildings, 129 storefronts, working traffic lights and cable cars, explorable Apple and Nintendo interiors) and Higashiyama, Kyoto (2.3 km of walking route, hand-painted anime style, 1,938 trees). Each world follows a pipeline of reconnaissance from OSM and elevation data, offline asset generation via Blender scripts, runtime assembly in Three.js, and camera-match QA against real photographs. The Kyoto world corrected six widely-repeated figures using LiDAR survey data.
What commenters are saying
Commenters were impressed by the scope but split on terminology: several objected that "world model" has a different meaning in ML research (a physics-grounded internal simulation, not a 3D scene generator). The author noted a single run cost ~$33 and ~8M tokens over 2 hours via a lengthy prompt with subagent orchestration. Technical discussion covered NPC logic (on-rail nav graphs), topology advantages of code-generated geometry vs. mesh generation, and comparisons to Opus 5 for lower-fidelity tasks.