GPT you can "touch with your hands"

All explanations of how GPT works boil down to the same thing: attention, embeddings, softmax. The words are familiar, but what actually happens inside is unclear.

A Korean developer created ko-microgpt - a tiny GPT model that generates Korean names and shows every step right in the browser. You scroll through the page and see how a token goes through embeddings, normalization, attention, MLP. Live numbers and weights on real data.

The project grew out of microgpt by Andrej Karpathy - the same GPT in ~200 lines of Python. Only here an interactive interface was attached to the model, and it turned out to be a super visual explanation of transformers.

👩‍💻 Data Flow