What is an LLM?

At its core, an LLM (large language model) does one thing: it predicts the next token, based on the tokens that came before it.

A token is a small piece of text. It might be a whole word, part of a word, a punctuation mark or a few characters. The model picks one token, adds it to the text so far, then uses the longer text to pick the next one. Over and over again.

Under the hood that scoring is a huge amount of matrix maths. You don't need to follow the maths to use AI well. You do need to remember that every answer is built this way: one likely piece at a time. The model doesn't always take the top score. Most apps let it pick among the likely ones, which is why the same question can get different answers.

Predicting the next token sounds simple. But to get good at it across a vast amount of text, a model has to learn a mind-boggling number of patterns: grammar, facts, code, structure, style, and how ideas fit together.

That is why next-token prediction can produce answers that look a lot like thinking. Human pattern recognition and an LLM are not the same thing, of course. But both can produce a sensible-looking response from a learned pattern.

It also explains the famous failure: when the pattern is strong but the facts are missing, the model still produces something fluent and confident. It is completing a pattern, not checking a record.

On its own, an LLM has no goals, no tools, no memory between conversations and no way to check whether its answer is right. Text goes in, text comes out. Everything else is the system built around it, and the rest of this course covers each part:

Prompts

tell it what job it is doing

Context

gives it the information to work with

Tools

let it look things up and take actions

Memory

lets it carry facts from one conversation to the next

Loops

let it check its work and try again

Evals

tell you whether it actually worked

One more distinction you will hear a lot. Open-weight models can generally be downloaded, run on your own computers and adapted, subject to their licence. Closed-weight models are only available through someone else's product or API.

Neither is simply better. The choice affects control, privacy, cost and where the model can run.