# Defcon1

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Surya Srivastava](https://cerebralvalley.ai/u/slimshady)
- **GitHub:** https://github.com/suryafk/scenefork
- **Demo video:** https://www.youtube.com/watch?v=2OzcMIMkKjw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/349

Video is the most common way we show the world, but it's read-only. You can watch a clip, but you can't move what's in it or ask "what if?" A student can't take a video of their own world and see what changes under different gravity. A creator can't make a clip that viewers can play with. Doing either today takes a game engine, 3D assets and programming skills, which puts it out of reach for most people.

AI could close that gap by turning plain requests into working behavior, but that creates a second problem: generated code can look right and still be wrong, or do nothing at all. Without a way to check it, you can't safely let it change a running system.

SceneFork solves both problems. It turns any paused video into a physical world you can touch and reprogram by voice, with no code and no 3D assets. Click an object and Meta's models lift it out of the frame: SAM 3.1 outlines it, Muse Image repaints the background, and Muse Spark estimates its mass and material. Then say what you want to change. Presets such as "zero gravity" apply at once, and for anything new, Muse Spark writes a small sandboxed program with its own tests.

Nothing generated goes straight into the scene. A verifier runs the tests in a shadow copy of the world, then repeats them without the plugin, so a rule that changes nothing is rejected. Failed tests go back to the model for up to two repairs, and impossible requests are refused with a reason. In our eval on the demo scene, 48 of 50 generated-rule runs passed.

The AI proposes. The simulation decides.

## More from Meta Global AI Developer Hackathon

- [Maximusinfotech](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/346)
- [flood-triage](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/347)
- [ctzurcanu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/348)
- [MxA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/350)
- [CortexNET](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/351)
- [WalletShield](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/352)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/349. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt