giving LLMs a "visual Brain" via a 4d Physics Sandbox
llms struggle with spatial geometry because they process text tokens, not dimensions. I am building a physical sandbox framework to act as an external visual cognitive engine for AI.
how the loop works:
• Code Gen: LLM generates logic to spin up a 4D cloud workspace (3D + time).
• Offloading: Heavy orbital mechanics and structural math offload to my custom GPGPU physics engine.
• Self-Correction: Live telemetry logs from simulation failures pipe back to the LLM to auto-patch its code.
Update: Prototype Done, Moving to Realism!
The agentic loop prototype works. Now, I am building the first 4 planets of the solar system using:
• NASA Maps: Real elevation data in a Three.js frontend.
• Atmospheric Clouds: Custom GLSL shaders for dynamic layers.
• Newtonian Orbitals: A multi-body backend gravity matrix.
Pushed the core builds to GitHub: github.com/williamjosh-dev/Physical-Sandbox-v1.0-
I'm 15 and currently trying to balance massive planetary scales without breaking my WebGL views. Let me know what you think ...
🧠
