Watching someone else derive backpropagation feels like learning and is not. The moment you have to produce the gradient yourself, against a checker that knows the answer to four decimal places, the gap between recognizing and knowing closes.
A module is not a video and a worksheet. It is a short concept, the interactive challenge that tests it, a lab where you build the real thing, and a checkpoint that will not let you move on while a gap is still open.
Twenty to sixty minutes of the idea itself, written down. Enough to attempt the challenge and no longer — the explanation is not where the learning happens.
The challenge. Shape Golf, NaN Hunt, Bayes Duel, Reward Hack — the concept is the mechanic, so guessing scores nothing.
Build the real thing on real data: an autograd engine, a transformer from the paper, a RAG system, a fused Triton kernel.
A short assessment that names the gap rather than scoring you. It exists to send you back one lesson, not to award a badge.
Linear Algebra for Machine Learning · Math Foundations
Numerical Computing & Stability · Math Foundations
Instruction Tuning & Alignment · LLMs & Generative AI
