A post-mortem from Hugging Face's "build-small-hackathon" describes how developer VirusDumb attempted to use Nemotron 30B to generate complex 3D browser games in a single pass, building a digital pet assistant called "Amazing Digital Dentures" meant to produce custom Three.js games on demand. The attempt ran into recurring code execution errors and was ultimately scaled back.
Attempted Prompt Engineering and RAG Pipeline
The project initially relied on extended system prompts instructing the model to write full interactive games. When long prompts produced broken code, the developer tried integrating GitHub Copilot game engine "skill cards" for more structured instructions. Those skill cards exceeded the small context window that had been configured to save on compute, so the developer used Codex to condense the skills into a single text file and built a Retrieval-Augmented Generation (RAG) pipeline over that distilled knowledge base.
According to the post-mortem, the RAG pipeline improved instruction retrieval but Nemotron 30B still failed to produce working 3D games in a single pass, frequently outputting code that rendered as a blank screen.
Documented failure points include:
- Direct prompting deficits: long prompts didn't give the model enough structure for valid Three.js code.
- Context window overflows: detailed skill cards exceeded the low-compute context limit.
- RAG limitations: retrieval improved code structure but didn't eliminate spatial and geometric logic bugs.
- Single-pass complexity bounds: the system could generate simple HTML apps like clocks, Snake, and Breakout, but failed on more complex titles like Tetris.
Scaled-Back Result
After these failures, the developer reduced the project's scope from a game creator to a simple HTML toymaker, which can produce single-file web utilities in one shot but can't handle complex multi-file game logic.
Why It Matters
The write-up is a useful data point for developers experimenting with open-weight models for automated code generation: this particular 30B model could reliably produce basic 2D utilities and single-file HTML scripts, but full single-shot 3D game synthesis was out of reach without iterative debugging or multi-step feedback loops. It's one project's experience rather than a broad benchmark, but it illustrates a concrete limitation worth knowing about before attempting similar single-pass pipelines.