It answers with
the cable unplugged.
Open source forever. The Ark’s design, described on this site, and its software are free to copy and build, and always will be: the answer engine, the tools that download and verify every file, and the catalog of everything on the drive, published under the Apache 2.0 licence. The book and supported deployment are optional: they fund the free version rather than gate it.
Everything on this site is free to build from the specification. The book is the guided version of that walk — eighteen chapters in which you build a working offline AI on the computer you already own, and finish understanding exactly what it is doing.
Three decisions that shape the whole book.
The laptop you already own
The build in the book is deliberately not the build in the specification. It sits one rung below the Reader — a starter corpus of about 60 GB on hardware you have not bought. The planet-scale maps, the full-precision weights and the solar array are all extensions, never requirements. A book you cannot finish without $8,000 is a book nobody finishes.
Concepts arrive when the build asks for them
Attention is not explained in a chapter called Attention. It is explained the moment you have watched a model finish a sentence and asked how it knew which word mattered. In Chapter 5 you make the model state a confident falsehood on purpose — and everything in Part II exists because of what you saw it do.
The wrong turns are kept in
The dependency that would not install, the file that was not where the repository said it was, the error message that was simply false — all of them, with the fix and the reason. They taught more than the successes did, and they are the reason the build log exists at all.
Four parts. Eighteen chapters.
The structure is the argument. Parts I and II are the system's own two halves — the engine that reasons and the library that remembers →
| Part | Chapters | What you build, and what it teaches |
|---|---|---|
| I — The machine that reasons | 1–7 | A model running locally by Chapter 2. Tokens, attention, quantization, and what your own machine can actually take. It ends by making the model confidently wrong on purpose. |
| II — The library that remembers | 8–13 | The corpus, the embedding model, the overnight index build, and the retrieval layer wired to the model. This is where an answer becomes a sourced answer. |
| III — The human interface | 14–16 | It listens and it speaks, in both languages, and it serves every phone in the room over its own WiFi. |
| IV — The system that survives | 17–18 | Checksums, provenance, cold copies, and the recovery procedure written for somebody who is not you. |
Plus appendices: a parts list per rung, a command reference, a glossary, the manifest schema, a troubleshooting index and further reading. Every figure is drawn to survive greyscale, because the paperback is black and white and e-ink is monochrome.
Being written, in the open.
The outline is frozen and all eighteen chapter briefs are written. The manuscript is being drafted. There is no purchase link yet, and this page will carry one when there is — along with the price, the formats and a sample chapter.
English first. A Spanish edition follows the English launch rather than shipping alongside it, because a technical translation done in parallel with an unfinished manuscript is how both end up wrong.
The book is the guided path, not the gate
The specification, the build scripts and the corpus list are open and stay open, and the book says so in its first chapter. Nothing in it is withheld from the free path — what you are buying is the order to do things in, the reasoning behind each choice, and the mistakes already made for you. Both routes are set out under How to get one →