Show HN: Watch 14-Byte AI "brains" attempt to solve a 2D maze (Its hard)
Hey HackerNews, I built this project over the last few weeks as a palette cleanser from a failed game launch. I wanted to learn a bit about AI/Neural-Networks and naively thought I could build a tiny maze-solving AI in a weekend with a 100% solve rate. Well - I couldn't, but I got pretty close. 14 Bytes total model size, and a 96.5% solve rate on unseen mazes. Trained across 46 phases experimenting with different ideas to improve the model (better performance, smaller size). Its quite fun to watch the model attempt to solve the maze, when they fail its usually due to getting stuck in a loop. The models have no access to coordinates, map-data, or external memory scratches - they must navigate using only immediate local neighbourhood observations. There is a model dropdown and you can see how the model has progressed over each phase, constantly getting smaller and increasing its solve rate. Total trained models number in the thousands - I just expose the winning models from each phase. Overall a fun experiment, with much implementation help from AI agents to scaffold and implement the code (I'm a lazy software dev).
Discussion Highlights (4 comments)
purple-leafy
Facts: - 14 Bytes does not include the runtime! - All that counts towards the 14 Bytes is the model weights - Lack of git history is because my actual repo is private, this is just the frontend for now + models
elil17
I would love to understand what the inputs actually are. Very neat visualization.
iwhalen
If your goal is to get the network as small as possible, you could try using NEAT[1]. A cool example of a small network for a hard task using NEAT is MarI/O[2]. [1]: https://en.wikipedia.org/wiki/Neuroevolution_of_augmenting_t... [2]: https://www.youtube.com/watch?v=qv6UVOQ0F44
vunderba
Neat! Maybe consider doing a longer postmortem/blog post on this. Small bit of feedback: it might help to add an explicit small (i) icon for areas which contain tooltips especially for interfaces (phones) that don't support hover events. I discovered that the models (such as pareto-r1) had additional information by complete accident~