Night School. How AI works, taught on a coast at night.

Start here

New to the maths? Begin with Pylon on the hill above the harbour: Vectors: a list of numbers is a place, then Matrices move space, Gradients, Backpropagation, Optimisers and The held-out set. All six come back to the attention lesson's small language model, so its numbers carry over.

Then down to the dock, where language models start: Tokens: cutting text into crates with Pylon, and Embeddings: every token gets coordinates with Tallow in the lamp room. Then Tallow's wheel: Next-word prediction and the loss, and Decoding: temperature, top-k, top-p. Then the whole machine: The transformer block in Tallow's tower, and The training loop on Pylon's dock, then what feeds it and how far it goes: Data and Scaling. Then down to the small boatyard, where big models are made small: Distillation: a big teacher, a small student, and LoRA and fine-tuning. Then to the prism at the tide line, where Rerun checks the numbers: Error bars: how sure is the number, and When the check can be fooled. Then Evaluating a language model puts both to work, and Pylon opens the cannery, where graphics chips do the sums: Threads, workgroups and memory. Then make a model smaller and faster: Quantisation with Tallow, and Matmul tiling in the cannery. Then attention without its table of scores, Fused attention, and the model in 16-bit numbers, Mixed precision with Tallow. Then two case studies with Rerun on the beach: a real check that failed, The denoiser's tensor-core parity, and A score that carries its receipt. And the algorithms under all of it: Search and sorting, and Dynamic programming. Last, back at the boatyard, the whole method on a real contest, misses and all: Our MMBU Challenge entry.

Attention: how a language model decides what matters. Tallow keeps the lamp at the top of the lighthouse. A word shines its question back along the sentence, and the earlier words that answer it best get the light. In the lab you train a 2-layer model with 20,864 numbers on sentences from the coast and watch its heads learn where to look.

The lessons

30 lessons in six tracks, from vectors to fused attention kernels. "After" names the lessons that help first; nothing is locked. 30 lessons are live, every one of them. To see them all at once, as lights along the guides' shore, open the Lantern Coast.

How it works

The films are drawn by the raw-native engine, live in your browser or rendered to video, from the same scene files. The labs train real models in plain JavaScript on your own computer; nothing is sent anywhere. Every claim in a film has its source listed beside the lesson, and every number about a model comes from the trained model's own file.

There are no points, streaks or badges here, and nothing plays the next lesson on its own. Come when you want, learn a thing, and go.