A model doesn't know things. It predicts what comes next.

Type a question into a chat tool and an answer appears, fluent and complete, as if it had been waiting there. It is natural to picture a very large filing cabinet behind the screen: the question goes in, the matching answer comes out. That picture is wrong, and almost every confusion about AI comes from it.

There is no cabinet. A language model does not look an answer up. It writes one, one small piece at a time, by predicting which piece is most likely to come next. Here is the whole mechanism, and it is short enough to draw.

Your question goes in as text. The model produces a single next word and adds it to the text. Then the question plus that one word goes in, and it produces the next. It goes round the loop once per word, until the word it predicts is a special one meaning "stop". The fluent paragraph you read was never planned as a paragraph. It was a few hundred turns of a loop, each one choosing a word.

What "predict" actually means

Each turn of the loop is not a guess at one word. It is a score for every word the model knows, all at once. Given the text so far, the model produces a number for each possible next word: how likely that word is to come next. Then one word is chosen, usually the highest-scoring one, but not always, and that is what makes two runs differ (that is Note 03).

Where the scores come from

This is the part that replaces the filing cabinet. The model was made by showing it an enormous amount of written text, a large slice of the public internet, books, code, and conversations, and playing one game over and over: cover the next word, ask the model to score it, see how wrong it was, nudge its internal numbers a little so it would be less wrong next time. Billions of numbers, nudged trillions of times. Training is that game played at a scale no person could.

What comes out is not a copy of the text. It is a compressed record of its patterns: which words tend to follow which, in which situations, at every scale from spelling to grammar to the structure of an argument. The model has no page that says "the sky is blue because of the scattering of sunlight". It has numbers which make "blue" the top score after "The sky is", and "because" a high score after that, and "scattering" a high score a few words later. The fact is in there, but as a tendency, not as an entry.

Why it looks like knowing

If all it does is score next words, why do the answers so often come out right? Because in the text the model learned from, questions are usually followed by correct answers. The most likely continuation of a clearly written question is, most of the time, the thing people actually wrote after that question, and people mostly wrote true things. So "predict the likely next word" quietly turns into "produce what a well-informed writer would probably say". It looks like knowledge because knowledge is common in what it read.

And this is exactly where the picture in your head matters. The same mechanism produces a wrong answer with the same fluency, whenever a wrong continuation happens to be likely: a made-up reference that looks like real ones, a confident date that is one year off, a sum that is nearly right. Nothing in the loop checks truth. It checks likelihood. Section 2 is about the ways that goes wrong; this note is only about seeing the loop clearly.

The practical, and what it showed

You can see the loop from outside in fifteen minutes. Give a model a half-sentence with no single right ending, in five separate new chats, and look at what comes back. That is Practical 01. Here is my run.

Both halves of the note are in that drawing. No two runs were identical, not even in the first word, which is what you expect from a loop that chooses among likely words rather than copying a stored answer. And four of the five said the same thing in different words, which is what you expect when one continuation has by far the highest score: it is not the only path, but it is the most travelled one.

A second model added a twist. Three times out of five it did not complete the sentence at all. It read my line as something I was about to say, and asked me to go on: "I'd love to hear what you were about to say!" That is the loop too. After a message like mine, one high-scoring next word starts a completion; another high-scoring next word starts a friendly question back. The model did not decide which reading was correct; it took a likely one. The full ten results are on the practical page.

the answer is never in the box. only the next word is.

One layer deeper *

It is not really words. The pieces the loop produces are tokens: sometimes a whole word, often a fragment ("enter" + "prise"), sometimes a single character. A typical model has a fixed list of around a hundred thousand of them, and every turn of the loop is a score for each one on the list. Note 02 is about why that detail explains some odd behaviour.

The score is a probability. The hundred thousand scores at each turn are scaled so they add up to one; that is why people say the model "outputs a probability distribution over the next token". Choosing the top one every time is called greedy decoding and makes the output almost identical run to run. Chat tools instead sample: they pick at random, weighted by the scores, so a word with a 30% score is chosen about 30% of the time. That is the whole reason my five runs differed, and the dial that controls how much they differ is temperature (Note 03).

"Knowing" is a fair word, with a caveat. To predict the next word well across trillions of words, a model has to pick up spelling, grammar, facts, and a great deal of how arguments are put together, because all of those change what comes next. So it is not true that the model knows nothing; the patterns encode a great deal. What it lacks is any separate store of facts it can consult, and any step that checks a continuation against the world. It knows things the way a very well-read person knows what a sentence is about to say, not the way an encyclopaedia does.

The idea is old. In 1951 Claude Shannon measured how well people could guess the next letter of English text, and used it to estimate how much information a letter carries. Predicting the next symbol as a way of capturing a language has been the core idea ever since; what changed in the last decade is the size of the model, the amount of text, and the architecture (the transformer, 2017) that made scaling it up work.

Two things from the practical that this note does not explain. Not one of the ten runs kept to "in one line", although every prompt asked for it; the reason is that your instruction was not the only one the model was reading, and that is Note 11. And "the model stops when it predicts a stop token" hides the question of who decided how long an answer should be, which is the same note.

Sources

The loop and the scoring follow the standard account: Vaswani et al., "Attention Is All You Need" (2017), and Andrej Karpathy's "Intro to Large Language Models" (2023) for the plain-language version. Shannon, "Prediction and Entropy of Printed English" (1951). The practical results are my own, run on 7 and 8 September 2026 in two consumer chat tools at default settings; the first sentences are on the practical page and the full runs in my records. Terms used: model, token.