How LLMs Work
Book Labs
All labs
26 Book labs · Sampling

Watch it speak

From Chapter 27 of Large Language Models from the Ground Up. This is not a mock-up: it is the book's actual 824,897-parameter GPT — the one you train in notebooks 08–09 — with its int8-quantized weights downloaded into this page and its forward pass re-implemented in plain JavaScript. Prompt it, set the dials, and watch it roll Chapter 27's loaded dice one character at a time.

Downloading the trained model (0.9 MB) and verifying its forward pass…
What to notice

Everything on this page is machinery from the book, running for real. The weights are the ones notebook 09 trains (Chapters 23–25) — verified on load by replaying the forward pass over "ROMEO:" and matching the exported golden logits. They travel int8-quantized, one scale per matrix row (Chapter 36), and generation runs on a KV-cache — the prompt is encoded once, then each new character appends a single key/value row per head instead of re-deriving them all (exactly Chapter 28's trick; watch the context counter). The dials are Chapter 26's: hold the seed and sweep temperature — around 0.1–0.5 the model digs into a rut ("the world of the world…", degeneration's mild cousin), at 0.8–1.0 it speaks its honest Shakespeare-flavored voice, and by 1.5+ spelling dissolves into noise. Same 824,897 knobs every time — only the dice change.

← Back to all labs