Slides
Lecture slides for each day
Each deck is published on the morning of its teaching day, so only Day 1 can be downloaded now. The rest of the page shows what is coming. Day 1 runs to 108 pages, from one unit to a checked parameter update and on to what a held-out number is worth. Day 2 runs to 60 pages on convolution. Day 3 runs to 67 pages on attention, language models and retrieval. Day 4 runs to 64 pages on the point where an answer becomes an action.
Inputs and targets; one unit, then a layer; why depth needs a nonlinearity; a complete forward pass; backpropagation, with the softmax cross-entropy gradient derived rather than quoted; and a measured run whose held-out loss turns upward while accuracy stands still.
A dense layer on a 64×64 colour image needs 1,572,992 parameters and assumes nothing about it; a bank of 32 filters of size 3×3 needs 896 and assumes that meaning is local and does not depend on position. The day ends by auditing whether those assumptions held.
Attention computed by hand on a three-token example; the training objective shown to be next-token likelihood rather than truth; and retrieval ranked on similarities where the stale passage scores 0.956 against the current one at 0.984, above an irrelevant page at 0.271.
Tool contracts, state, authority, and the controls between a proposed action and an executed one. A seven-step agent trajectory runs through the day; its third step is a well-formed request that the gate refuses.
