These pages are for the layer before implementation. They make the main idea, the key mechanism, the contrast case, and the carry-away mental model legible before you start tracing code line by line.
The chapter pages are built to answer the same student questions in roughly the same order, so they feel like one family instead of unrelated summaries.
If a page helps you predict what the notebook is about to show you, it is doing its job.
The public surface is intentionally narrow. These pages foreground the chapter’s real mechanism and the mental model a student should carry into the notebook.
The page should name the real conceptual center, not just the sequence of exercises.
Students need one clean story before they need all the details and variants.
The pages try to sharpen intuition by showing the wrong read next to the better one.
Plots, code, hooks, or metrics should feel like evidence for a story the student already understands.
Each one emphasizes a different mechanism family, while preserving the same study-companion feel.
A chapter companion organized around circuit tracing, skip-trigrams, induction heads, and the ladder from pattern to mechanism claim.
A student-facing explainer centered on locality, receptive fields, residual learning, and how BatchNorm and skip connections stabilize deep vision models.
A chapter companion for linear probes, built around the final-token truth setup, MM versus LR, transfer across datasets, and why intervention matters more than readout alone.
The intended flow is simple: orient first, then verify in the notebook while the underlying mechanism stays in view.
These pages try to lower that cold-start cost. They narrow the visual field so the first notebook pass feels like confirmation instead of confusion.