14 What an LLM is, in plain words

15 What an LLM is, in plain words

You will work with these tools more confidently if you have a working mental model of what they are. Not a deep one. A useful one. Two pages, no maths.

15.1 What you will get from this chapter

A plain-words description of how a large language model produces text.
Three properties that explain most of the AI behaviour you will see in the book.
A short vocabulary you will reuse: token, context, training, inference.

15.2 The one-paragraph description

A large language model — an LLM — is a very large neural network trained, on a very large amount of text, to predict the next chunk of text that should follow whatever has come before. That is it. “What word is most likely to come next?” Asked over and over, fast. The trick is that, with enough text and enough parameters, the answers stop being chunks of plausible-sounding language and start being chunks of plausible-sounding thought. We do not entirely know why. It works.

15.3 A few words you will need

A short vocabulary. We will use these throughout the book.

Token. A small piece of text. “Manchester” is one token; “unprecedented” is two; common short words are one each. The model sees and produces tokens, not words. Pricing and context limits are measured in tokens. A useful rule of thumb: one token ≈ three quarters of a word in English.
Context. The set of tokens the model can see at once when it answers — your prompt, the system instructions, any files you uploaded, your previous turns, sometimes results from tools the model has called. Models have a maximum context size, measured in tokens. Today’s frontier models hold roughly 200,000 to 1,000,000 tokens of context, which is to say “a couple of books.”
Training. The process that produced the model in the first place. Training happens once, on huge piles of text. After training, the model’s weights are frozen. Your prompts do not change the model.
Inference. What happens when you use the model. The model reads your context and produces tokens, one after another. Inference is what costs money per use.
Cut-off. The point in time after which the model has not seen any new text. After the cut-off, the model does not know about new events unless you tell it. As of this edition, the cut-offs sit somewhere in 2025; check the documentation for your provider.

15.4 Three properties that explain a lot

Three things follow from the description above. They explain most of the behaviour you will encounter.

1. The model is a probability distribution, not a database.

When you ask “who is the manager of Real Madrid?”, the model is not looking up an answer. It is producing the most likely continuation of your question, given everything it ever read. If a name appeared a thousand times in its training data alongside Real Madrid, it will produce that name. If three names competed and one of them stopped being the answer six months before the cut-off, the model will sometimes give you the old answer.

This is why models hallucinate. They are not lying. They are producing what looks like the answer.

2. It does not know what it does not know.

The model produces tokens. It does not produce a confidence alongside them. The same fluent prose that gives you the right answer to a question gives you a wrong answer to a different question. The tone is the same; the content is not. This is the most important property of LLMs to internalise. We come back to it constantly.

3. The same prompt gives different answers.

There is a sampling step in inference — the model picks the next token from a distribution rather than always picking the single most likely one. This is on purpose; “always pick the most likely token” produces flat, repetitive prose. The cost is that you can ask the same question twice and get two different answers. Sometimes meaningfully different.

For analytical work this is a feature and a bug. If two runs disagree on a numerical answer, that is information; you should be more sceptical. If two runs disagree on the style of a paragraph, the disagreement is mostly noise.

15.5 The “jagged frontier”

A useful image. AI’s capabilities are not a smooth curve from “easy tasks” to “hard tasks.” They are jagged. A task that looks hard to a human (write a small Python script that does ten things) is easy for AI. A task that looks easy (correctly count occurrences of a word in a long document) is sometimes hard. You learn the shape of the jaggedness only by trying. There is no shortcut, and the shape changes between models.

The book is, in part, a tour of where the frontier sits as of Spring 2026.

15.6 Centaur and Cyborg, briefly

Two metaphors get used a lot for working with AI. Centaur — you and the AI take turns; clear handoffs; you do the parts you do best, AI does the parts it does best. Cyborg — you and the AI are interleaved at the level of every paragraph and every line of code; the boundary blurs.

Both are real ways to work. The book leans Centaur in the early chapters (you are still learning where to hand off) and Cyborg later (once you have built the taste for what each side is good at). Neither is wrong.

15.7 Where AI helps · Where AI bluffs

Helps. Explaining what a model is doing on a specific prompt. Translating “I want X” into a prompt that has a chance of getting X.

Bluffs. Telling you, with great confidence, why it produced a particular wrong answer. The model does not have introspective access to its own weights. Ask it to defend an answer and it will produce a fluent defence — of a wrong answer.

15.8 Keep this with you, not the AI

The notion that the model does not know what it does not know.
The decision to ask the same question twice when the stakes are non-trivial.
The notion that the cut-off is a real limit, not a soft one. Recent events are not in the model unless you put them there.

15.9 Try this

Open your chat AI. Ask: “Who is the current chancellor of Germany?” Note the answer. Open a second tab. Ask the same question. Note the answer. Compare with reality (any reputable news source).

If you get the same answer twice, that is one signal. If you get the right answer twice, that is another. If you get two different answers, you have just experienced property number three live.

15.10 AI and me

How did AI support me here?
How did AI fail me?
How did AI extend me?

15.11 Where to go next

Which AI model should I use? — the practical pick.

--- title: "What an LLM is, in plain words" --- # What an LLM is, in plain words You will work with these tools more confidently if you have a working mental model of what they are. Not a deep one. A useful one. Two pages, no maths. ## What you will get from this chapter - A plain-words description of how a large language model produces text. - Three properties that explain most of the AI behaviour you will see in the book. - A short vocabulary you will reuse: token, context, training, inference. ## The one-paragraph description A large language model — an LLM — is a very large neural network trained, on a very large amount of text, to predict the next chunk of text that should follow whatever has come before. That is it. *"What word is most likely to come next?"* Asked over and over, fast. The trick is that, with enough text and enough parameters, the answers stop being chunks of plausible-sounding language and start being chunks of plausible-sounding *thought*. We do not entirely know why. It works. ## A few words you will need A short vocabulary. We will use these throughout the book. - **Token.** A small piece of text. *"Manchester"* is one token; *"unprecedented"* is two; common short words are one each. The model sees and produces tokens, not words. Pricing and context limits are measured in tokens. A useful rule of thumb: **one token ≈ three quarters of a word in English.** - **Context.** The set of tokens the model can see at once when it answers — your prompt, the system instructions, any files you uploaded, your previous turns, sometimes results from tools the model has called. Models have a maximum context size, measured in tokens. Today's frontier models hold roughly 200,000 to 1,000,000 tokens of context, which is to say "a couple of books." - **Training.** The process that produced the model in the first place. Training happens once, on huge piles of text. After training, the model's weights are frozen. *Your* prompts do not change the model. - **Inference.** What happens when you use the model. The model reads your context and produces tokens, one after another. Inference is what costs money per use. - **Cut-off.** The point in time after which the model has not seen any new text. After the cut-off, the model does not know about new events unless you tell it. As of this edition, the cut-offs sit somewhere in 2025; check the documentation for your provider. ## Three properties that explain a lot Three things follow from the description above. They explain most of the behaviour you will encounter. **1. The model is a probability distribution, not a database.** When you ask *"who is the manager of Real Madrid?"*, the model is not looking up an answer. It is producing the most likely continuation of your question, given everything it ever read. If a name appeared a thousand times in its training data alongside *Real Madrid*, it will produce that name. If three names competed and one of them stopped being the answer six months before the cut-off, the model will sometimes give you the old answer. This is why models hallucinate. They are not lying. They are producing what *looks* like the answer. **2. It does not know what it does not know.** The model produces tokens. It does not produce a *confidence* alongside them. The same fluent prose that gives you the right answer to a question gives you a wrong answer to a different question. The tone is the same; the content is not. This is the most important property of LLMs to internalise. We come back to it constantly. **3. The same prompt gives different answers.** There is a sampling step in inference — the model picks the next token from a distribution rather than always picking the single most likely one. This is on purpose; "always pick the most likely token" produces flat, repetitive prose. The cost is that you can ask the same question twice and get two different answers. Sometimes meaningfully different. For analytical work this is a feature *and* a bug. If two runs disagree on a numerical answer, that is information; you should be more sceptical. If two runs disagree on the *style* of a paragraph, the disagreement is mostly noise. ## The "jagged frontier" A useful image. AI's capabilities are not a smooth curve from "easy tasks" to "hard tasks." They are jagged. A task that looks hard to a human (write a small Python script that does ten things) is easy for AI. A task that looks easy (correctly count occurrences of a word in a long document) is sometimes hard. You learn the shape of the jaggedness only by trying. There is no shortcut, and the shape changes between models. The book is, in part, a tour of where the frontier sits as of Spring 2026. ## Centaur and Cyborg, briefly Two metaphors get used a lot for working with AI. **Centaur** — you and the AI take turns; clear handoffs; you do the parts you do best, AI does the parts it does best. **Cyborg** — you and the AI are interleaved at the level of every paragraph and every line of code; the boundary blurs. Both are real ways to work. The book leans Centaur in the early chapters (you are still learning where to hand off) and Cyborg later (once you have built the taste for what each side is good at). Neither is wrong. ## Where AI helps · Where AI bluffs **Helps.** Explaining what a model is doing on a specific prompt. Translating "I want X" into a prompt that has a chance of getting X. **Bluffs.** Telling you, with great confidence, *why* it produced a particular wrong answer. The model does not have introspective access to its own weights. Ask it to defend an answer and it will produce a fluent defence — of a wrong answer. ## Keep this with you, not the AI - The notion that the model does not know what it does not know. - The decision to ask the same question twice when the stakes are non-trivial. - The notion that the cut-off is a real limit, not a soft one. Recent events are not in the model unless you put them there. ## Try this Open your chat AI. Ask: *"Who is the current chancellor of Germany?"* Note the answer. Open a second tab. Ask the same question. Note the answer. Compare with reality (any reputable news source). If you get the same answer twice, that is one signal. If you get the right answer twice, that is another. If you get two different answers, you have just experienced property number three live. ## AI and me 1. How did AI **support** me here? 2. How did AI **fail** me? 3. How did AI **extend** me? ## Where to go next [Which AI model should I use?](03-which-ai.qmd) — the practical pick.