The Prediction Puzzle

Imagine typing: “Can you make this warmer?”

You might mean a chilly room, a color palette, a piece of writing that feels distant, or yesterday’s soup. A human friend can draw on the room you are standing in, the conversation you just had, and the look on your face. The sentence arrives inside a world of shared experience.

An LLM receives no such world automatically. It receives the prompt as input data. Before it can produce a reply, the words must be handled in a form the machine can calculate with. The model does not see letters glowing on a screen, or hear a voice in the sentence. It does not feel cold, and it has no soup to reheat. Its starting point is a stream of text supplied by the system around it.

That distinction is easy to lose because the reply may sound so natural. Ask a chatbot to make a paragraph warmer, and it might soften the phrasing. Ask it about a thermostat, and it might suggest raising the temperature. In either case, the answer can fit the situation. But fitting a situation is not the same as experiencing it as a person would.

The prompt is more like a set of clues than a message received by a hidden human reader. The words give the model patterns to work with; the surrounding conversation, if there is one, gives more. “Make this warmer” after a paragraph points one way. The same request after a discussion of a thermostat points another. The prompt supplies the current clues, but it does not automatically resolve ambiguity or provide information that is absent. The system’s response depends on the information available in the input—not on an unspoken awareness of the room or the person typing.

This is not a reason to shrug at what the model can do. Making a useful distinction between two meanings requires more than matching one word to one canned answer. A language model can use the prompt’s structure and context to produce a response that appears tailored, sometimes in ways that are genuinely helpful. To understand how, we need to follow what happens between the text entering the system and an answer coming out.

Nor does the opposite conclusion follow: that fluent language proves a little person is hiding inside the machine. A convincing reply is evidence that the system has learned powerful patterns in language and context. By itself, it is not evidence of human-like experience.

Our task is to keep both facts in view. The model is not a person reading your words from the inside; neither is it a trivial phrase-matching gadget. It transforms input and produces language through mechanisms we can describe. The first step is to see what the machine is actually given—and what it must do with it before “warmer” can lead anywhere at all.

Once a prompt has entered the system, the model’s immediate task is simpler to state than the machinery behind it: continue what is already there.

Imagine this unfinished sentence:

The picnic was canceled because the sky turned dark and the rain began to…

A person might complete it with fall, pour, or come down. The model faces a related choice. It estimates which next piece of text is most likely, given the words so far. Perhaps fall receives a high likelihood, pour a somewhat lower one, and sing a very low one. These are not little words waiting in a queue. They are possible continuations with numerical scores attached to them.

The model’s job is not to select the only continuation that makes sense. Several could work. It weighs possibilities, then one is chosen—often the likeliest, though the exact selection can depend on settings that allow more or less variation. After that, the chosen piece becomes part of the text the model has to continue. It repeats the process, one piece at a time. An answer unfolds one token at a time, and each choice helps shape what can come next.

The technical name for a piece is a token. A token may be a whole short word, part of a longer word, punctuation, or another small unit. So the model is commonly described as predicting the next token, not necessarily the next whole word. For now, think of the sentence as a row of pieces, with the model repeatedly estimating what piece could come next.

If this were all the model learned, it might be an unusually elaborate phrase-completion machine. It could finish familiar lines and imitate common turns of phrase, but it would struggle whenever the right continuation depended on something more than a nearby word.

Consider: “Maya put the glass on the table because it was…” What does it refer to—the glass or the table? The next word might be empty, wobbly, dirty, or broken, and the sentence alone may not settle the question. Add a few lines about Maya clearing dishes or steadying a table, and some possibilities become more plausible than others. To continue well, the model must use clues across the passage, not simply reach for the word that most often follows was.

Now consider a request: “Explain why the answer is 12.” A useful continuation depends on what answer? If the prompt includes a small arithmetic problem, the model must track the numbers and their relationships. If it includes a paragraph about a character, “the answer” may mean something entirely different. The same stretch of text can call for different continuations in different contexts.

So the prediction game reaches beyond familiar phrases. To guess well across stories, instructions, arguments, recipes, jokes, and code, a model has to pick up patterns in how language carries information. It must learn that names refer to people, that pronouns point backward, that questions invite certain kinds of replies, that a recipe has an order, and that a claim may be followed by evidence. It can also learn regularities in the subjects people write about: what tends to happen in a story, how a simple calculation is expressed, or what a line of code is likely to require.

That does not mean the model stores each idea as a tidy fact card, or that every plausible-sounding answer is correct. It means that useful continuation often depends on a rich web of learned relationships. The model’s basic objective remains prediction; the structure it learns to make prediction work is far more than a list of stock phrases.

The game is also played repeatedly on enormous amounts of text. During training, the model sees a stretch of text and practices predicting what comes next. When its predictions miss, its internal settings are adjusted so that, over time, it becomes better at the task. Training builds the learned settings that an ordinary request later uses; producing a response applies them rather than updating them. We will look more closely at how those settings are shaped shortly. For now, keep the central idea in view: an answer is built as a series of guesses, each made in light of the prompt and everything already generated.

This is why “just autocomplete” is both understandable and misleading. It names the basic move: continue the text. But autocomplete on your phone may offer the next likely word in a familiar message. A large language model has learned to continue an immense variety of material, where the right next piece may depend on meaning, distant context, conventions, or a chain of steps. The simple game can demand complicated pattern recognition.

Prediction alone does not guarantee understanding in the human sense, and it does not guarantee truth. A sentence can be a very likely continuation and still be wrong about the world. But prediction across varied situations is a powerful training signal: it rewards the model for capturing patterns that help language hold together and ideas move from one sentence to the next.

The puzzle is beginning to take shape. If the machine learns by guessing what comes next, what has to change inside it for those guesses to become so remarkably flexible?

That question is not just how a machine gets better at finishing sentences. It is how a system trained on language can answer a question it has never seen in quite that form, explain its answer, or solve a puzzle whose solution takes several steps.

Imagine asking: “A jar holds 12 buttons. Maya gives away half, then finds 5 more. How many are in the jar?” A useful answer requires keeping track of the starting amount and the two changes: half of 12 is 6, and 6 plus 5 is 11. The model’s answer can be evaluated against that result, and its intermediate steps can show where its approach succeeds or goes astray.

This raises a practical question: what should count as reasoning? Rather than infer a hidden human-like mental process, we can examine observable performance. Did the system identify the relevant details? Did its steps lead to the answer? Does the result hold up when the wording changes? These tests let us discuss reasoning as a pattern of behavior without treating the model as a person having an inner experience.

A model trained on vast amounts of language encounters examples of how explanations, calculations, code, and arguments fit together. Some systems can also spend additional computation on a difficult task, working through intermediate possibilities before giving a final response. That extra work does not replace the underlying language-model machinery; it gives the system more opportunity to organize its approach to a problem.

Success on the button puzzle would demonstrate a useful capability, not a guarantee of reliable performance on every variation. The important question is how the system handles new combinations of information and whether its steps and conclusions stand up to checking.

That is the puzzle this book will follow: how prediction, supported by learned structure and additional computation, can produce behavior that looks like explanation and reasoning. We can trace the path from prompt to answer by asking what the model has learned, what clues the prompt supplies, and how the response develops. To see how the machinery manages it, we first need to look at what the machine receives. Before it can predict a piece of language, the words themselves must become something it can work with.