The machine that guesses what comes next

We have seen that the machine does not think, and that it arranges our words on an immense map. But a map is silent. It tells no stories. How do we get from a catalogue of mute addresses to those long paragraphs that appear on your screen with the ease of a professional writer?

The answer lies in a mechanism so simple it is almost insulting. You use it every day without giving it a thought.

Take out your phone. Open a new message and type “I am running”. Look just above the keyboard. It suggests three words to continue the sentence: “late”, “out”, “behind”.

How does it do that? It cannot read your mind. It applies a statistical rule. It has observed that after “I am running”, people very often type “late”, quite often “out” and sometimes “behind”. It offers you the likeliest words in order.

Generative AI does the same thing. On a far, far more impressive scale, but fundamentally the same thing.

Generative AI does nothing more than guess the next word.

That is its one and only function. There is no other function hidden behind it. When you ask a question, it does not look up the answer in a secret encyclopaedia. It takes the words in your question, locates them on its map and calculates which word is most likely to come next.

Then it displays that word. Then it starts again. It takes your question, adds the word it has just written and calculates the next one. Word after word, fraction of a second after fraction of a second.

Ask it for the capital of France. It calculates that the likeliest word to begin with is “The”. After “The” comes “capital”, then “of”, then “France”, then “is”, then “Paris”. It never knew it was going to answer Paris. It discovered that along the way, with the final word, at the same moment you did.

That is why the text appears on your screen in little bursts, giving the impression that someone is typing behind it. This is not a stylistic flourish designed to make the machine feel alive. It is the machine working in real time.

There is one problem. If it always chose the number-one word, the most probable of all, it would be deathly dull. It would write lifeless prose and, worse still, it would begin to stammer. You would get sentences such as: “I am on my way home to eat at home to eat at home.” The likeliest word summons the likeliest word, and the sentence becomes trapped in a loop.

The engineers added one indispensable ingredient: a pinch of randomness.

When the machine searches for the next word, it does not select a single candidate. It draws up a list, together with each candidate’s chances. To complete “The cat drinks…”, the result looks roughly like this.

The cat drinks… milk 91 % · water 5 % · juice 3 % · syrup 1 %

Instead of choosing “milk” every time, it draws from the best candidates while respecting their probabilities. Nine times out of ten, it will write “milk”. Every so often, however, it will choose “water”, and the sentence will set off down a different path.

That tiny grain of mathematical madness changes everything. It, and it alone, gives the machine its air of creativity. It is the reason why asking the same question twice never produces exactly the same answer: the route forks at the first slightly lucky draw, changing everything that follows. The machine improvises, in the strictest sense of the word.

This setting can be adjusted. Turned right down, the assistant becomes studious and predictable, perfect for code or calculations. Turned too high, it goes off the rails and eventually produces sentences that look like English without quite being English.

Temperature

The control that adjusts this element of randomness. Cold, the machine repeats the most likely choice. Hot, it takes risks, then loses its way.

Context window

Everything the machine keeps in view while answering you: your conversation from the beginning, plus any documents you have given it.

Autoregressive

The technical name for this mechanism. It simply means that the machine feeds on what it has just written itself.

At this point, a question should be occurring to you. If it improvises one word at a time, how does it avoid losing the thread?

That is exactly the right question. Try telling a joke by improvising each word from the one before it, and you will never reach the punchline. How can this machine write three pages about the French Revolution, complete with an introduction, a developed argument and a conclusion that answers the introduction, if all it does is guess the next word?

By rereading. With every new word, it looks at more than the word immediately before it. It rereads everything.

Imagine that it has reached the thousandth word of its answer. To find that word, it rereads your original question, followed by the 999 words it has just written. It weighs the whole, locates everything on its map and calculates the thousandth word. Then, to find word number 1,001, it rereads everything again: your question, the 999 words, plus the one it has just added.

It is back-breaking work, repeated dozens of times a second. And this obsessive rereading is what allows it to stay on course. It does not plan its conclusion in advance. At every step, it looks back over the entire path travelled before deciding where to place its next foot. It does not lose the thread because it keeps the whole of it in its hand at all times.

You can immediately see the consequence. That hand is not infinitely large. This is the famous context window: beyond a certain length, the beginning of the conversation slips out of view and disappears. Hence the moment when an assistant you have been talking to for an hour seems to have forgotten what you originally asked. It has not forgotten in the human sense, leaving behind a memory that gradually fades. The reality is more absolute: those words are simply no longer before its eyes. They no longer exist.

This is also why a long conversation grows slower and costs more than three short exchanges. Every additional word requires the machine to reread slightly more text than the word before it. When a discussion gets bogged down, starting again on a blank page with a clear summary is often more effective than persisting.

The time has come to put a crack in the phone metaphor. It has done its job and is beginning to mislead us.

Your keyboard looks only at the two or three preceding words, no more. It knows nothing about what you wrote at the top of the message. The machine, by contrast, weighs every word in its window at once and, more importantly, evaluates which ones truly matter to the word it is currently seeking. In the sentence “The report you sent me last week, the one about pensions, is incomplete”, the word “is” relates to “report”, twelve words earlier, not to “pensions”, even though “pensions” comes immediately before it. Your phone would come unstuck. This is precisely what the machine can do, and it was the invention of 2017 that set everything in motion.

So keep the predictive keyboard for what it explains well: the next-word principle and the complete absence of intention. Just do not ask it to explain how a line of reasoning can be sustained across three pages.

One phenomenon remains to be explained, something that has probably intrigued you if you have been using these tools for the past few months. Sometimes the machine displays “Thinking” and makes you wait ten or thirty seconds before answering. These assistants are known as reasoning models.

What is happening during that time? Has it finally begun to think for real? No. It is using exactly the same mechanism, word after word, but now it uses it to write a draft.

Instead of giving you its answer immediately, it first generates dozens of sentences for its own use: restating your question, proposing a hypothesis, testing it, spotting an inconsistency, starting again. Some assistants show you a summary of this process on screen, where you may see phrases such as “wait, that calculation is wrong, let me start again”. Once the draft is complete, the machine uses it as additional context when composing its final answer.

This is remarkably effective for a logic problem, a calculation or a piece of code. For writing a birthday message, it is wasted time. Above all, it changes nothing about what we have just seen: the machine still does not think. It has simply learned to reread its work before speaking.

The entire mechanism has a very plain name in the field: a language model. In science, a model is a mathematical representation of reality. Meteorologists use models to forecast tomorrow’s weather from the wind and atmospheric pressure. Engineers have done the same thing with our language: they have built a model capable of predicting tomorrow’s word from today’s words. A reasoning model, the kind we have just discussed, is nothing more than a language model taught to work things out in a draft.

That is why the media talk about LLMs. The initials stand for large language model. Large, because the system contains billions of internal settings. Language, because it manipulates words. Model, because it calculates probabilities.

One final observation, and its consequences are profound. Run through this chapter again in your mind. We have talked about probabilities, random selection, rereading and the next word. We have not mentioned truth once.

That is no oversight. The machine does not seek what is true; it seeks what is probable. It was designed to be coherent, not accurate. In Chapter 5, we will see everything this small difference causes, and why the problem will never disappear entirely.

Before that, however, we need to answer a question this chapter has carefully avoided. I have told you that the machine calculates probabilities. I have not told you where they come from. Who taught it that “The cat drinks…” is followed by “milk” nine times out of ten? Nobody entered those figures by hand. There are far too many of them.

It learned them by itself. And the way it did so has much more in common with training an animal than attending school.