Show HN: Watch 14-Byte AI "brains" attempt to solve a 2D maze (Its hard)

Category: ai-ml

Tags: neural-networks, maze-solving, model-compression, reinforcement-learning, interactive-visualization

Score: 5.7/10 (Innovation: 7, Technical: 6, Documentation: 3, Utility: 4)

A visual demonstration of tiny neural networks (14 bytes total) solving 2D mazes using only local observations, trained through 46 iterative phases achieving 96.5% solve rate. The project explores extreme model compression for reinforcement learning and offers an interactive web viewer to watch each phase's winning model navigate mazes. It's interesting as a educational experiment in minimal AI and evolutionary training.

Target audience: AI/ML enthusiasts, researchers, and developers interested in tiny models and reinforcement learning

Repository: https://con-dog.github.io/MINIMIO-PUBLIC-FRONTEND/

View on Hacker News