loom.

A GPT built from scratch in pure NumPy — every gradient derived by hand — trained on Shakespeare and running live in your browser. Generate text, then inspect what each attention head was looking at while it wrote. Part of a from-scratch LLM stack with mosaic (tokenizer) and nabla (autograd) · source on GitHub.

Generate

loading Pyodide runtime (~10 MB, one time)…

Output

Attention — what the model looked at

Row i shows where the model attended while predicting the token after position i (last 48 characters of the output). Brighter = more attention. The empty upper triangle is causality: the model cannot see the future. Different heads learn different habits — some watch the previous character, some scan for the start of the line or the speaker's name.

Everything you see is computed by the same NumPy code that trained the model — no framework, no server, no API. The model is a deliberately tiny ~350K-parameter, character-level GPT: it writes plausible Shakespeare-shaped scenes, not good ones. That's the point — it's small enough to read, train, and fully understand.