Most explanations of large language models go one of two ways. Either they reach for the brain metaphor, which is wrong in ways that matter, or they reach for linear algebra, which is correct and helps nobody. What follows is a third option: a picture that is simplified but not misleading, and that actually changes how you use the thing.
The one sentence version
A language model reads everything before the cursor and produces a ranked guess at what comes next. Then it does it again, with its own guess now part of what came before. Then again. An answer of three hundred words is that loop run several hundred times.
That is genuinely most of it. The sophistication is not in the loop, which is simple. It is in how good the guess is, and the guess got good by being adjusted against an enormous quantity of written text until its predictions started matching what people actually wrote next.
Why "prediction" does not mean "shallow"
The obvious objection is that this sounds like a party trick. Predicting the next word seems like it should produce the kind of thing your phone's keyboard produces, which is drivel.
The reason it does not is worth sitting with. To predict the next word really well across every kind of text that exists, you have to end up encoding a great deal about how the world is described. To finish the sentence "the surgeon put down the scalpel and" you need something about surgery. To finish "she was furious, so when he apologised she" you need something about how people behave. To finish a line of half written code you need something about that programming language.
None of that was taught as facts. It fell out of the pressure to predict well. That is why the systems are simultaneously impressive and unreliable: the knowledge is real in the sense that it is genuinely there, and it is fragile in the sense that nothing in the process ever checked it for truth. It only ever checked it for plausibility.
Everything is a continuation
The single most useful consequence of the picture above is this: the model is not answering your question. It is continuing your document.
Your message becomes a piece of text. The system's job is to write what plausibly follows it. The whole illusion of conversation is built on top of that by adding labels like "user" and "assistant" to the text and letting the continuation carry on from there.
This explains behaviour that otherwise looks arbitrary.
- Sloppy input gets sloppy output, because sloppy writing is most plausibly followed by more sloppy writing.
- Telling it who it is genuinely works. "You are a district nurse explaining this to a new patient" changes what text plausibly follows, so it changes the answer.
- Giving one worked example is worth a paragraph of instruction, because an example sets a pattern and the system is a pattern continuation engine.
- If it starts badly it often continues badly, since its own poor opening is now part of what it is continuing. Starting a fresh conversation beats arguing.
The window, and why it forgets
The model can only look at so much text at once. That budget is called the context window, and everything has to fit in it: your question, whatever the system was told before the conversation began, any document you pasted, and the entire back and forth so far.
Modern windows are large, but they are not infinite, and when a long conversation runs past the edge the earliest part falls out. The system does not announce this. It simply stops knowing something it knew an hour ago, which reads like absent mindedness and is in fact a hard limit being hit.
There is also no memory between conversations unless the product has deliberately added one. A fresh chat starts genuinely blank. The thing that felt like rapport yesterday was in the text of that conversation, and the text is gone.
Training and use are different events
This one causes more confusion than any other, and it is simple once separated.
Training happened once, in the past, over months, at great expense. It is where the model's sense of language and the world was formed. It ended on a particular date.
Use is what happens when you type. Nothing is being learned. The model is not updated by your conversation. Weights that took months to set are not being adjusted by your message about a window cleaning round.
Two things follow. The model does not know about events after its training ended unless the product around it can look things up for you, which some can and some cannot. And correcting it teaches it nothing beyond the current conversation. The correction helps for the next few messages because it is now part of the text being continued, and then it is gone.
Four things the picture explains
- Confident wrong answers. Plausible and true were never separated during training. A fabricated citation is formatted like a real one, so it is exactly as plausible.
- Different answers to the same question. The ranked guess is sampled from rather than always taking the top option. That deliberate variation is what stops the output being wooden.
- Weakness at arithmetic. Long multiplication is not a pattern in text, it is a procedure. Systems that get this right are usually running a calculator underneath.
- The flat corporate voice. Left unsteered it lands on the average of an enormous amount of writing, and the average of everything reads like a brochure.
So is it thinking
The honest answer is that the question is doing more work than it looks. If thinking means holding a goal, weighing it against the world and being moved by evidence, then no: nothing in the loop does that. If it means producing behaviour that requires having represented something about the world, then something in that direction is clearly going on, because you cannot predict that well without it.
What is worth resisting is the leap from "it produced a sentence a thoughtful person might produce" to "it thought about it". The system has no stake in being right. It has no way to notice that it does not know. Those two absences are the entire source of the failure modes, and they are not going to be argued away by a more confident sounding answer.
Holding the picture loosely is the practical position. Treat it as an extremely well read colleague who has never once been embarrassed by being wrong, and who will answer every question at the same volume whether or not they have any idea. You would still find that colleague useful. You would simply check their references.