Developer builds 14-byte neural network that solves 2D mazes with 96.5% accuracy

A software developer created a miniature AI model, just 14 bytes in size, capable of navigating 2D mazes it has never seen before. The model achieves a 96.5% solve rate despite having no access to coordinates, map data, or external memory, relying solely on observations of its immediate surroundings. Built over several weeks as a side project following a failed game launch, the AI was refined across 46 training phases and thousands of model iterations. The developer publicly exposed the best-performing model from each phase, allowing viewers to watch how the AI's maze-solving ability improved over time. When the model fails, it typically gets stuck in a repetitive loop, highlighting the challenge of spatial navigation with such minimal computational resources.