Analysis updated 2026-05-18
Type plain sentences to generate 3D environments in a browser-based world that runs its own living ecosystem simulation.
Study how to run a fine-tuned language model entirely in the browser using WebLLM and WebGPU without sending data to a server.
Explore the architecture of a Rust-to-WebAssembly ecosystem sim that runs in a Web Worker alongside a SvelteKit frontend.
Deploy your own shared-world instance on Cloudflare Workers with D1 persistence for asynchronous multiplayer contributions.
| puruvj/maya | abyo-software/ferro-stash | ariasbruno/glyph | |
|---|---|---|---|
| Stars | 1 | 1 | 1 |
| Language | Rust | Rust | Rust |
| Setup difficulty | hard | moderate | easy |
| Complexity | 5/5 | 4/5 | 2/5 |
| Audience | developer | ops devops | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires Rust toolchain, wasm-pack, and pnpm, AI model weights (~840 MB) download on first run.
World is a 3D browser-based game where you type sentences to build a living environment and then walk around it in first person. You might type something like "a small village with a lake" and a fine-tuned AI model running entirely in your browser transforms that into actual objects placed in the scene: buildings, trees, paths, lamps. No data is sent to a server for the AI part, the model runs locally using your device's graphics hardware. The world does not sit still when you leave. An ecosystem simulation runs continuously in the background. Animals hunt, breed, starve, and age. Settlements grow or fall into ruin. An AI director called Mother Nature makes small adjustments to keep the world alive. The design is inspired by Death Stranding's asynchronous multiplayer: you never see other players at the same time, but you see what they built. Their constructions, the AI's changes, and the simulation's own drift all blend together into a world that feels inhabited. The technical approach separates concerns carefully. The AI model never figures out where to place objects by doing math, instead it emits instructions that follow a grammar, and a separate deterministic renderer converts those instructions into geometry. The ecosystem simulation is written in Rust and compiled to WebAssembly, running in a background thread so it does not slow down the rendering. The visual look of grass, water, sky, and weather is generated by GPU shaders rather than downloaded asset files. The entire world is a pure function of a starting seed and a tick counter, meaning it can be replayed or fast-forwarded exactly. The frontend uses SvelteKit 5 and three.js. The backend, for shared world persistence, runs on Cloudflare Workers with a D1 database. The two fine-tuned AI models that handle text-to-world generation are about 840 MB and 280 MB respectively and are downloaded on first run rather than stored in the repository. Development requires the Rust toolchain, wasm-pack, Node.js with pnpm, and optionally a Cloudflare account for the shared world features. The license is MIT.
A browser-based 3D world where you type sentences to build villages and forests, AI runs locally to place objects, and an ecosystem simulation keeps the world alive with animals and decay after you leave.
Mainly Rust. The stack also includes Rust, WebAssembly, SvelteKit.
MIT license: use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.