I was curious what would happen if you replaced git's object store with a neural network. Turns out it works surprisingly well for small files. The network memorizes your code byte-by-byte, I also added a correction layer using a small classifier. The model makes mistakes that look like real typos, and the classifier fixes them