What an AI model actually is
Underneath every chatbot answer is a single mechanism: the model predicts what word probably comes next, writes it, then asks the same question again. The fluency, the confident wrong answers, and the two different replies to the same prompt all come from that one mechanism.
The mechanism
Section titled “The mechanism”An AI model is a pattern-matching machine over text.
It was shown vast amounts of writing. The internet, books, papers, code. From all of that, it absorbed how language behaves: which words tend to follow which, in what order, in what tone, after what setup. Not by memorising the text. By compressing the shape of how the text works into a set of internal numbers, its weights. Producing those numbers is what training does.
Given a prompt, the model answers one question: based on all of that, what word probably comes next? It picks a word, adds it to the text, and asks the question again. What word comes next now? It picks again. And again, until the answer is finished.
That repeating step is the loop, and it is the whole machine. (Strictly, the model picks tokens rather than words: chunks of text a few characters long. Nothing in this chapter changes if you read “word” throughout.)
See it happen
Section titled “See it happen”The model writes one piece at a time, and each pick is shaped by everything written so far. The first 3 picks here have an obvious favourite. On the 4th, nothing tops 1 in 5, and which word lands is a weighted dice roll over the list.
This is also what the streaming in a ChatGPT reply is: each word appearing as the model commits to it. When a product shows a “thinking” line before the first word, it is running the same loop over its own working first. How AI generates a response traces that step by step.
What this explains
Section titled “What this explains”Smooth writing is the model’s basic skill. It spent training absorbing how good text flows, so it can produce a fluent paragraph on almost anything. The smoothness says nothing about whether the content is true, because the machine that produces the smoothness never checks.
Confident wrong answers come from the same place. The model continues patterns; it does not look facts up. When the most likely next words happen to be untrue, the model writes them exactly as fluently as true ones. This is called hallucination, and it is not a malfunction. It is the ordinary loop on a stretch of text where likely and true differ.
Asking the same question twice gives two different answers because the pick between likely words includes deliberate randomness, the dice roll in the demo above. The content usually lands in roughly the same place. The wording shifts every time, and now and then the substance shifts with it.
And prompts matter because the prompt is what the predictions are based on. A prompt that states the situation, the task and the constraints makes useful next words likely. A vague prompt makes average next words likely.
What it can hold in front of it
Section titled “What it can hold in front of it”The model sees only what is in the current conversation: the messages so far, plus anything pasted in. This window of visible text is called the context window. Every model has a size limit on it. In a long enough conversation the earliest messages fall outside the window, and the model answers as if they never happened, because on this turn they were not in front of it.
The model also keeps nothing between conversations. A new chat starts from only what is put in front of it. When a chat product does seem to remember things about its user, that is the product saving notes and putting them back into the window, not the model remembering.
Where the line sits
Section titled “Where the line sits”Everything the model does is producing text. That covers more work than it sounds: drafting an email, summarising a document, rewording a clause, translating a paragraph, proposing 5 subject lines. All of it is text out of text, and all of it sits inside what the loop does well.
Booking a flight, sending an invoice, updating a record in another system: these are actions, not text, and the model cannot do them. A product can be built so that the model’s text triggers an action through tools, and when several such steps are chained with checks in between, the result is called an agent. In every case the model’s own contribution is still writing.
The same goes for knowing things. The model holds no client history, no record of last quarter’s decisions, no verified facts on demand. Where a product appears to know these, the product fetched them and placed them in the window.
That is the line to carry through the rest of Horizon: the model writes from pattern, and everything else, acting, remembering, looking up, checking, is built around it. Whether a piece of work sits on the writing side of the line or the other side decides most of what the later chapters cover.