Most AI tutorials teach you to copy: clone, run, done. The Workshop is the teaching bench of the Forge, and it works the other way around. Every project here is real, runnable code — small enough to read line by line, documented decision by decision. Because architecture is like the diagram in a geometry proof: it is half the solution before a single line of code runs.
No cloud accounts, no GPU farms. Everything on this bench trains and runs on an ordinary CPU.
Battleship AI — a neural network you can train over coffee
A classic 10×10 Battleship game where the opponent is a convolutional network that learns by playing. It has about 38,000 parameters — thousands, not billions — and goes from random shooting to competent play in five hundred games: under a minute of training on a plain desktop CPU. The network never sees ship positions, only the consequences of its own shots.
The repository ships no pretrained model, on purpose: creating one is the point. The README walks through every place where the “default” answer from a tutorial would have been wrong — why there is no pooling on a 10×10 board, why the input is a four-channel fog of war, why hard rules are masked out instead of learned, and how four reward numbers define the player’s character.
A companion article — how this architecture was designed, and what the network can and cannot learn — is in the works. More projects will join the bench.