Notes

The notes written so far.

In order. The full plan, including notes not yet written, is in the contents.

00
Why this notebook existsOpening · what this notebook is for, and how it is kept.
01
A model doesn't know things. It predicts what comes next.Section 1 · nothing is looked up; every answer is built one word at a time. With Practical 01.
02
Tokens: why the model reads "strawberry" differently from youSection 1 · the model never sees letters, only numbered pieces; why that explains counting, arithmetic, cost and languages. With Practical 02.
03
Why the same question gives different answers: temperature and samplingSection 1 · the dial on the draw, and why it only matters when a question has more than one good answer. With Practical 03.
04
The model is a snapshot: training cutoffsSection 1 · the model has read nothing after one date; the only way in for anything newer is the text in front of it. With Practical 04.
05
Hallucination is the same mechanism that makes it usefulSection 2 · the loop that writes a true sentence writes a false one the same way; in 2026 it mostly answers a nearby question instead of yours. With Practical 05.
06
Why a system that writes poetry struggles to multiplySection 2 · arithmetic by prediction gets the shape right and the middle wrong; in 2026 the product usually runs code for it, and the model's account of which is itself a prediction. With Practical 06.