← @bgSponge

GitHub - PhiloLabs/fable51-worlds: worlds via code, from fable 5.1

Summary

PhiloLabs uses Claude Fable 5.1 agent swarms to transform text, video, or images into walkable Three.js worlds generated entirely via code.

  1. Code-based worlds enable verification.

    Playwright drives the Three.js app to diff screenshots against real photographs, treating geometric errors as failing tests rather than subjective taste.

    A wrong number is a failing test, not a matter of taste.

  2. Procedural generation ensures editability.

    Modifying a world is a code diff rather than a latent space re-roll, ensuring changes occur only where requested and persist in future renders.

    The world changes exactly where you asked and nowhere else

  3. Agent swarms automate world construction.

    The pipeline uses reconnaissance agents for surveying, asset scripts for kit emission, and independent reviewer agents to file QA reports.

    Builders never grade their own work

  4. Fable 5.1 outperforms GPT-6 Astra.

    Side-by-side unedited comparisons of Union Square and Kyoto show Fable 5.1 results against GPT-6 Astra's filmed walkthroughs.

    left GPT-6 Astra, right Claude Fable 5.1

  5. The roadmap targets agent environments.

    Future development plans to use these worlds as training grounds where code provides both the environment and the reward signal.

    the same code that builds the place also supplies the reward

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →