The back of the notebook.

Context window

The fixed amount of text a model can have in front of it at one time: your question, the conversation so far, any document you pasted, and the answer it is writing. When it is full, the oldest text drops out. It is a working space, not a memory.Explained properly in Note 09, Section 3.

Model

The trained system that turns text in into text out. It is not the same as the product you use it through: the chat window, the search, the memory features and the safety rules are wrapped around the model by whoever built the product.Explained properly in Note 12, Section 3.

Token

The unit a model actually reads and writes. Roughly a word or a piece of a word: "notebook" may be one token, "strawberry" may be several. Everything the model does, and everything it costs, is counted in tokens.Explained in Note 02.

Sampling

The step after the model has scored every possible next token: one of them is drawn, weighted by its score. Sampling is why two runs of the same question can differ.Explained in Note 03.

Tool call

A line the model writes, in a form the product taught it, asking for something to be done outside the model: a web search, a calculation, a lookup. The product's code runs it and pastes the result back in front of the model. Whether the model writes that line is a prediction like any other.Explained in Note 04; the loop is Section 5.

Training cutoff

The date on which the text used to train a model stops. The model has read nothing from after it, and answers questions about later events from the older world unless the newer text is put in front of it. Providers publish the date; the model's own statement of it is a guess.Explained in Note 04.

Top-p (nucleus sampling)

A cut applied before the draw: keep the smallest set of tokens whose shares add up to p, discard the rest. Stops high temperature from letting nonsense through. Top-k is the blunter version: keep the k highest.Explained in Note 03.

Temperature

The dial on sampling. At 0 the top-scoring token is always taken and the answer is repeatable; higher settings spread the draw across lower-scoring tokens. It changes how often a lower-ranked word is chosen, not how the model ranks them.Explained in Note 03.

1. Queued for Section 3Why do models read the middle of a long document worse than the start and the end?
2. Queued for Sections 3 and 8How much of what feels like a model's personality comes from its training, and how much from the instructions the product puts in front of it?
3. Open in the fieldIs there a reliable way to tell a model's training cutoff from the outside, without being told? Researchers have proposed tests; none is settled.

• 9 Sept 2026 Note 04 said the search bolt-on "decides when to run, not you". The product does not decide; the model does, following instructions the product wrote, and that decision is a prediction like any other. The distinction matters because it explains why "search on" sometimes does nothing. Corrected in place, with two new paragraphs on how the decision is made and what a search costs in tokens, after a question from Pradeep.

a notebook should say how it was made.