What's inside
Sixteen short lessons in three parts: how it works (1–7), how it learns (8–12), and how to use it (13–16) — then an optional "going further" chapter and a plain-English glossary. Do them in order for the full story, or jump to whatever you're curious about. Use the Next / Back buttons — or your keyboard's ← → arrow keys.
01
What is an LLM?
The whole thing in one sentence: a very good next-word guesser.
02
Tokens
How your words get chopped into pieces the model can handle.
03
Meaning as numbers
How a computer turns "cat" into something it can work with.
04
Context & attention
How it works out which earlier words matter right now.
05
Inside the model
The neural network: billions of tiny dials, stacked in layers.
06
Predicting the next word
The word lottery — and the "creativity" dial.
07
The full loop
Watch all the pieces run together, for every word.
08
The training data
Walk a real web page through the actual data pipeline.
09
How it learns
Guess, measure, nudge: the training loop, made visible.
10
What it learns on its own
Grammar, facts and logic — picked up to win the guessing game, never taught.
11
Scaling up
Why it takes thousands of GPUs — and what size unlocks.
12
Becoming an assistant
Turning a raw autocomplete into a helpful chatbot.
13
At chat time
Context, memory, tools, and "thinking" when you actually use it.
14
Why it gets things wrong
The honest limitations everyone should know.
15
Using it well
Practical prompting that gets far better answers.
16
Agents
How an LLM gets hands: tools, a loop, and why it's needed.
17
Going further optional
Beyond text (multimodal) and chatting with your own files (RAG) — both with demos.
18
Glossary & recap
Every bit of jargon in one line, plus the big takeaways.
Companion course · now with code
Want to build one yourself? This guide is the no-code tour of how LLMs work. When you're ready to see every gear turn, Build an LLM from Scratch walks you through writing one — you implement attention, autograd, and the training loop by hand (runnable right in your browser), then assemble and train a real GPT in PyTorch. Every line of code and every formula is explained from zero.