Why your to-do list needs a digital pet, not just a better prompt
For makers and artists tired of rigid productivity tools, the dream is often a digital companion that gamifies your workflow. The concept is simple: an AI-driven entity that generates adventures to keep you moving, disguised as a sophisticated to-do list. The reality, however, is a messy crash course in the limits of current generative models.
The Nemotron 30b experiment
The project, originally titled Amazing Digital Pet Dentures, was inspired by the animated series The Amazing Digital Circus. It features a character named Caine, an AI pair of dentures living in a virtual circus alongside digital clones of humans. The goal was to replicate this by creating a digital pet that sends users on daily adventures, effectively turning work into a game.
The journey began with the Nemotron 30b model. The initial strategy relied on simple, lengthy prompts explaining exactly how to build a game. This approach failed immediately; the model frequently generated code that did not function, resulting in broken experiences.
When context windows blow up
Next, the developer attempted to inject specific engineering capabilities by adding skill cards from the GitHub Copilot repository. These cards were intended to guide the model in building a game engine. The result was a context window explosion that wasted compute resources. Increasing the context window to accommodate this data did not resolve the underlying issues.
The final attempt involved using Codex to distill those skills into a single text file, followed by a RAG (Retrieval-Augmented Generation) setup. While this method worked better than the previous iterations, the output remained flawed. The generated games consistently contained errors, often rendering as blank screens rather than playable applications.
The pivot to a simple HTML toy maker
After abandoning the ambitious goal of generating full games, the project was repurposed. It now exists as a basic HTML toy maker capable of creating simple web components in a single shot. The model successfully generates clocks, to-do lists, and classic games like Snake and Breakout.
However, complexity remains a hard ceiling. Attempting to build anything more involved, such as Tetris, causes the model to fail. The project is now hosted at https://huggingface.co/spaces/build-small-hackathon/AmazingDigitalPetDentures.
The creator is currently looking to pivot to a new idea and welcomes suggestions from the community.
Key takeaways
- Generating functional games with LLMs like Nemotron 30b is currently unreliable, often resulting in broken code or blank screens regardless of prompt engineering.
- Adding external knowledge via RAG or skill cards can improve performance but introduces significant context window challenges that may not be worth the compute cost.
- Current models excel at simple, single-shot HTML generation but struggle with complex logic required for games like Tetris.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
