microgpt
Andrej Karpathy distills the algorithmic core of GPT training and inference into a dependency-free, 200-line Python script. It trains a 4,192-parameter GPT-2-like model on 32,033 names, making every essential component—from autograd to attention—inspectable.