Level 1 of 14 · about 3 weeks

How LLMs work, for builders

Build a working mental model of tokens, context and sampling, so model behaviour stops being surprising.

Modules

1
Tokens and tokenisation
What a model actually reads, and why it decides your bill
Core 7 parts · about 6 hours
2
Embeddings and attention, by intuition
Meaning as direction, how words look at each other, and why long inputs cost more
Common 6 parts · about 7 hours
3
Training stages
How a text predictor becomes an assistant that follows instructions
Common 7 parts · about 6 hours
4
Sampling and decoding
How a model picks each next token, and the settings that control it
Common 5 parts · about 5 hours
5
Choosing a model
A repeatable way to pick, test and re-check the model behind your agent
Core 5 parts · about 6 hours
6
Cost, latency and limits
Why agent runs get expensive, slow and rate-limited, and what to do about it
Core 7 parts · about 6 hours
7
Failure modes
How language models go wrong, how to detect it, and what to do about it
Core 5 parts · about 6 hours
8
Deep-dive track
Inside a modern transformer: the block, rotary positions, grouped-query attention, mixture of experts, scaling laws
Specialist 6 parts · about 10 hours, optional