Show HN: Commit code into a neural network, see what it remembers (browser demo)

Category: devtools

Tags: neural-network, version-control, browser-demo, lossy-compression, interactive

Score: 5.3/10 (Innovation: 7, Technical: 6, Documentation: 5, Utility: 3)

lgit is a browser-based demo that reimagines version control by storing committed code or images as weights of a small neural network rather than as raw bytes. Each commit trains the network, and checkout reconstructs the data via inference, producing lossy, sometimes hallucinated results. It's an interesting exploration of neural compression and failure modes, but primarily a conceptual toy rather than a practical tool.

Target audience: backend devs, researchers, and curious developers interested in neural networks and version control experiments

Repository: https://nicobrenner.github.io/lossy-git/demo.html

View on Hacker News