This Neural Network Runs Entirely on DOOM's Game Mechanics

I built a neural network using DOOM games

A developer compiled neural networks into DOOM maps, where voodoo dolls on conveyor belts perform matrix multiplies, lava lines add weights to sector light levels, and doors act as ablations. The MNIST network uses 1,355 dolls and 158,550 sectors, reading 966 of 1,000 sampled test digits correctly, with every value checked bit-exact against a NumPy integer reference. A transformer even writes DOOM level names like DEMON CANYON, at roughly 70,000 times slower than NumPy.