About AI Book

Built by engineers tired of half-finished tutorials.

AI Book is a catalogue of technical volumes on transformers, deep learning, probability, calculus, crystallography, and the systems that run them. Every chapter is written for the reader who intends to rebuild the thing — because you should be able to.

The why

The internet has ten thousand tutorials on transformers. Almost none of them finish.

They show the diagram. They paste the PyTorch. They wave at softmax. Then they stop right when it gets useful — at the KV cache, at the rotary embedding, at the actual production failure mode that'll page you at 3am.

We're building the version of the book that does finish. Where every line is traced with real values, every mechanism is rebuilt from scratch, and every chapter ends past the demo — in the cost curves and the second-order tricks.

It takes longer to write. That's the point.

“A technical book is a conversation with a future self who forgot everything. It must be patient, precise, and a little stubborn.”
Editorial note · 2026
How we write

Six commitments to the long read.

We hold ourselves to these on every section we publish. If a chapter doesn't pass them, it goes back to drafts.

Depth over speed.

We'd rather publish one chapter properly than ten chapters that hand-wave the hard parts. The clock isn't our editor.

Built from first principles.

Every mechanism — attention, gradients, quantisation, lattices — is reassembled from scratch. We don't summarise; we rebuild.

Trace, don't paste.

Every PyTorch line gets a teacher-grade walkthrough. Every loop is unrolled. Every shape transition is shown with real values.

Written to be re-read.

Tutorials are read once. Books are read three times. We write for the third read — when you actually understand what you didn't the first time.

Free forever for readers.

All published volumes, every chapter, every interactive 3D visualization — open to read. Pay only if you want sync + playgrounds.

Static-first, no spinners.

Pages render as HTML before they hydrate. We optimise for the reader on a train in a tunnel, not the demo on stage.

How a chapter ships

From paper to publish.

Four passes. The chapter doesn't ship until all four say yes.

01

Research the mechanism.

Read the original paper. Re-implement the algorithm from scratch. Hit the failure modes ourselves before we write a word.

02

Build the visualization.

If a concept lives in space, it gets a real interactive 3D scene — not a video. We build the Three.js viewer first, the prose second.

03

Write the prose.

Math, code, and diagram in the same paragraph. Every line traced. Every parameter explained. No skipped iterations.

04

Edit for the third read.

An engineer at 3am rereading on a flight. Can they follow it cold? If not, it goes back. Then ship it static.

By the numbers

What's in circulation today.

17
Volumes published
452
Chapters traced
2463
Lessons written
Re-reads expected
Updated as new chapters ship. No artificial milestones.
“If you can't rebuild the thing from scratch after reading the chapter, the chapter is unfinished. That's the test. That's the only test.”
AI Book editorial creed

Don't take our word. Open a chapter.

17 volumes published. 452 chapters traced. Every one free to read. Pick one and judge for yourself.