// series · ongoing

AI from the inside: from tokens to agents

How a language model works, piece by piece and with code you can run: tokens, embeddings, attention, text generation and agents.

Parts of the series

AI · Part 1 · 13 min

What an AI actually reads: tokens, explained with code

A language model doesn't read letters or words: it reads tokens. What they are, how BPE builds them (coded from scratch in Python), and why they explain its trouble with letters and numbers, what each language costs and a trick to slip past filters.

coming soon

Part 2

Embeddings: how a number comes to mean something

coming soon

Part 3

Attention: how the model decides what to look at

coming soon

Part 4

Generating text: probabilities, temperature and hallucinations

coming soon

Part 5

Agents: the loop that turns a model into something that acts, and its risks