What is an LLM?

By the end of this module you will be able to explain, to someone else, what a large language model actually does — without hand-waving, and without needing to know any mathematics.

Imagine someone who has read almost everything ever written, and remembers none of it.

They cannot tell you which book a sentence came from. They have no filing cabinet to check. But they have read so much that they have developed an extraordinarily good instinct for what tends to come next — which word usually follows which, in what context, in what tone.

Ask them to finish the line “the capital of France is” and they will say “Paris” — not because they looked it up, but because in everything they have read, that is overwhelmingly what came next.

That is a large language model. It is the most useful and the most dangerous thing to understand about it, and almost everything else in this course follows from it.

Why “large”, and why “language”

Large refers to two things at once: the amount of text it learned from, and the number of internal adjustable values — called parameters — it uses to store what it learned. Modern models have hundreds of billions of them.

You do not need to know what a parameter is to use one of these systems well. What matters is the consequence: scale is why the behaviour feels like understanding. A small model predicting the next word produces gibberish. A large one produces something that argues, explains, and writes convincingly — while doing the same underlying thing.

Language is the part people underestimate. The model works on text, so anything you can express as text is in scope: an email, a legal argument, a recipe, a spreadsheet formula, a paragraph of Python. That is why one tool appears to do a hundred unrelated jobs.

What actually happens when you type

The reply appears word by word because it is being produced word by word. There is no finished answer waiting somewhere to be revealed.

Step 3 explains something people find unsettling: ask the same question twice and you get different answers. That is deliberate. Always choosing the single most likely token produces flat, repetitive text, so a controlled amount of randomness is introduced. You will meet the setting that governs it — temperature — in Module 05.

The consequence nobody warns you about

If the model is predicting plausible text rather than retrieving facts, then a confident wrong answer and a confident right answer are produced by exactly the same process.

The model is not lying when it invents a citation. It has no concept of a citation being real. It has produced text that looks overwhelmingly like what a citation looks like in that position — which is precisely what it was built to do.

What it does not have

Three absences worth holding on to:

No memory between conversations. Start a new chat and it knows nothing of the last one. Where a product appears to remember you, something outside the model is storing and re-sending that information.

No live access to the world. On its own, a model cannot check today’s date, look at a website, or read your files. Where a tool does those things, it has been connected to them — and that connection is the subject of the second course.

No beliefs. It does not know whether what it just said is true. It cannot want to help you, and it cannot want to deceive you.

Try it — 10 minutes

Step 3 is the one that matters. Most people have to see a fabricated citation once, in a subject they know, before the point genuinely lands.

Check yourself