MODULE 01 — FOUNDATIONS · 03 AI Concepts

What is an LLM?

It is not looking up an answer. It is picking the next chunk of text, then doing it again — a few hundred times a paragraph.

What it has written so far

What could come next

The problem it explains

There is no answer stored anywhere

People picture a giant filing cabinet being searched. The machine works the other way round: nothing is retrieved, everything is composed.

A SEARCH ENGINE your query stored page finds returns it unchanged A LANGUAGE MODEL your prompt one chunk next chunk every chunk is fed back in as input …hundreds of times
The loop on the bottom row is the whole difference: the model reads its own output back before every single choice.

How it works

One choice, repeated

every chunk gets a score ≈200,000 of them, ranked
Score everythingEvery chunk in the vocabulary is rated as a continuation.
picked not always the top — that is the randomness
Pick oneA little randomness is what stops every answer being identical.
text so far +1 chunk appended, then scored from scratch
Go againHundreds of times a paragraph, re-reading everything each pass.

What “large” means

Hundreds of billions of dials

Forty-eight dials. A real model has roughly ten billion times as many, and every one was set by training rather than by a person.

Those scores come from an enormous set of numbers — the parameters. Training nudges every one of them until the text it predicts starts to look like the text it was shown.

Nobody sets them by hand and nobody can read them. “Large” is a statement about how many there are, not about how much it knows.

So what

Fluent does not mean informed. The machine is optimised to sound like a good continuation — nothing in it checks whether that continuation is true.