4 lessons
Self-attention, explained without the jargon
The core trick that lets a model figure out which words in a sentence matter to which other words — the mechanism that actually reads your context window.
The full math of attention, worked by hand
Scaled dot-product attention end to end — every formula, why the scaling factor exists, and a complete numerical example you can trace by hand.
Why context windows cost quadratically (and what the KV cache does about it)
The precise reason doubling your prompt length more than doubles the work, worked out from the attention matrix itself, plus the caching trick that makes generation bearable.
How a model knows word order at all
Attention alone can't tell "dog bites man" from "man bites dog" — positional encoding is the patch that fixes this, and it's stranger and more elegant than you'd expect.