Questions readers ask

A very large table of numbers, the weights, together with a fixed calculation that takes a text and returns a chance for every possible next piece of text. Chapter 1.

No. There are no rows, no entries and nothing to look up. What the model knows is spread over its weights as patterns. That is why it can answer questions nobody wrote down, and also why it can be wrong without noticing. Chapter 11.

The calculation is a program, and a short one. The weights are not written by anyone. They were found by training. The behaviour of the model is in the weights, so in the sense that matters, nobody wrote it. Chapter 8.

A weight is one learned number of a matrix. "Parameter" is the common name for all learned numbers, weights and biases. A feature is one of the inputs X and changes with every text. A model has a fixed number of parameters. It has no fixed features. Chapter 3.

It is the brick the model is built from. Between the multiplications there is a bend, the activation function, without which many layers would be worth one. And attention decides which tokens are mixed. Both are made of the same multiplications and additions. Chapter 3 and Chapter 6.

In at least four ways. Its weights are frozen after training, and the brain changes all the time. A weight is one number, and a connection in the brain is a living structure. The model learns from what people have recorded, mostly text, and not from living in the world. And the machines that run it need far more power than the brain, which is estimated to use about 20 watts. What the two share is the idea that knowledge is stored as the strength of connections. Chapter 2 and Chapter 13.

OpenAI has not published the size of GPT-6 Astra. The figure most often reported is 10 trillion, and another report says 4.8 trillion. Neither is confirmed. Estimates for the connections of a human brain run from 100 to 1,000 trillion. We use 200 trillion as a round number. A weight is not a connection, so the two numbers give a sense of scale and no more. Chapter 13.

LLM is large language model. GPT is generative pre-trained transformer: generative because it produces text, pre-trained because it learned from a large amount of text before anyone used it, transformer for the design of its layers. Chapters 1, 8 and 6.

Because a list of all words can never be complete, and a list of pieces can build any text. Common words are one token, and rare ones are put together from several. Chapter 4.

For English, divide the number of words by 0.75: 600 words are about 800 tokens. We measured 0.87 words per token on plain text, so the rule is on the careful side. For code, numbers and other languages, count with a tokenizer. Chapter 4.

The id is a position in the list, a house number. The embedding is the row of numbers the id points to in the embedding table, a place on the map. The model calculates with the embedding and never with the id. Chapter 5.

The idea is the same: a list of numbers as a place on a map of meaning. The job differs. Inside a language model every token has a row. A vector database stores one vector for a whole sentence or page, made by a separate embedding model, so that texts can be found by meaning. Chapter 5, and Thinking in Vectors.

The raw score the model gives to one token, before softmax turns the scores into chances. Only the gaps between logits matter. Chapter 7.

Because the next token is drawn by chance. After the first different token the two texts go separate ways. With temperature 0 the highest token is always taken, and the answer is nearly always the same. Chapter 7.

0 where there is one right answer, 0.3 to 0.7 for answers and summaries, 0.8 to 1.2 for ideas and stories. These ranges are common practice and not a law. In our test a small model wrote broken text at 1.6. Chapter 7.

Because a true answer and an invented one are made by the same steps, and no step checks facts. The confident tone is the tone of the text it learned from. Chapter 11.

The computer is good at it. The model does not use the computer in that way. It guesses the digits of the result as tokens. Give it a calculator as a tool, and the answer is exact. Chapter 12.

Its weights were frozen at the knowledge cutoff. News reaches it only through the prompt, for example when the application searches the web first. Chapter 11.

The model does not. It keeps nothing between requests. If an assistant remembers you, the application around the model stored something and put it into today's prompt. Chapter 9, and the memory book.

Not while you talk to it. A provider may use conversations later as training text for a future model, if its terms allow that. Chapter 8.

An API key, the address of the service, and a request in JSON with the model, the messages and the settings. With the library it is about a dozen lines of Python. Chapter 9.

The first message of a request, written by the developer and not by the user. It holds the standing instructions. Its role is called system, and on OpenAI's newer models developer. Chapter 9.

It ran into max_tokens. The field finish_reason of the response says length. Chapter 9.

Every request resends the whole conversation, and all of it is paid for again. In our example, with a system prompt of 500 tokens and 250 tokens for each turn, 20 turns cost 62,500 tokens and 40 turns cost 225,000. Chapter 10.

The tokens of the prompt can be read together in one pass. Every token of the answer needs a trip of its own. On OpenAI's published price list of September 2026 a token written costs five times a token read: $0.50 against $0.10 per million for gpt-6-luna. Chapter 10.

Retrieval-augmented generation: search your documents first, put the passages that fit into the prompt, and let the model answer from them. Chapter 12.

A program that lets the model ask for tools again and again until a task is done. The model writes the requests, and the program runs them. Chapter 12.

Rarely as a first step. Try a better prompt, examples, retrieval and tools first. They are faster, cheaper and easier to change. Chapter 12.

Yes, a small one. The file must fit into memory with about 20% to spare, by our rule of thumb. At 4 bits per weight a billion weights are 0.5 GB, so a laptop with 16 GB has room for about 26.7 billion at most, and in practice for fewer, because other programs need memory too. Chapter 13.

Each weight is stored with fewer bits than in training, so the file is smaller and the weights a little less exact. Our "4-bit" file weighed 4.5 bits per weight. Chapter 13.

A small model runs without one, as GPT-2 did on our laptop. For a large one the ordinary processor is too slow. The work is a very large number of small multiplications that do not depend on each other, and a graphics card does many of them at the same moment. Chapter 13.

By packing the requests of many users together so that one graphics card works on all of them in the same pass, and by keeping and reusing work already done. This is the subject of a book that is planned for this shelf and not yet written. Chapter 17 gives the plan.