Not a book: a three-page document I wrote to stop collecting titles without a criterion. It organises reading in layers — the architecture base, AI literacy, patterns for systems built around models, data, and governance — and says in one line what each book is for.
What I took from it
The selection rule, which is the only part that does not age. Frameworks and tools turn over every season; books that teach design, systems and governance still work five years out. When I am torn between two titles, the winner is the one that will still make sense once the tool it describes is gone.
The decision not to go into training models. That is the hardest thing to write about yourself, because it means giving up something interesting: the books on building a model from scratch are excellent and belong to another trade. My value is architecture, integration and governance, and a reading path that excludes nothing is not a path.
And the finding that reordered everything: in governance the real canon is free. The text of the European regulation, the American risk management framework and the international management-system standard are worth more than the 2026 books that summarise them. Most governance books translate those sources into controls.
What I left out
A flaw of my own, worth admitting: a reading list is not a plan. It is comfortable to mistake having read for being able to do, and this path has not a single deliverable attached to it. The next version should say, next to each layer, what has to be built to prove it was understood. Without that, a reading path is an elegant way of postponing.