Mostly Harmless AI
by Alejandro Piad Morffis
$12.50
Mostly Harmless AI is the working researcher's account of what artificial intelligence is, how it actually works, and how to get the best out of it — written for a general audience, with no code and no math, just the ideas that matter and the contradictions the field hasn't smoothed over.
Second edition (v2.0, May 2026). 302 pages. Three parts, 16 chapters, plus prologue, epilogue, and a back-of-book glossary.
Part I — Foundations (6 chapters). How the technology actually works, not the marketing line. The seventy-year argument between rationalist and empiricist AI; machine learning; deep learning; large language models; generative AI; agentic systems.
Part II — Applications (7 chapters). One chapter per domain, written for the practitioner already in that field who wants an honest account. Knowledge work; science; coding; education; creative work; policy.
Part III — Dangers (3 chapters). Alignment as a research problem; the limitations of current systems; the real risks — bias, security, labour displacement, environmental cost, concentration of power.
New in this edition. A cross-linked back-of-book glossary with 484 entries — every named system, person, technique, and field-of-art term used in the body, linked from the chapters and back. A rebuilt references apparatus with full citations, footnotes back-linked to the body, and links to the original sources.
This book is an evergreen permanent beta. The field moves too fast for any static document to keep up; any version will be wrong about some specific thing within six months. New editions ship roughly annually. Your purchase gives you every future edition in perpetuity — no subscription, no expiring access.
The book is also free to read online at books.apiad.net. Buying here is a gesture of support, not a paywall — what your support actually funds is the time to keep rewriting the book as the field moves. If the price is a real obstacle, write to apiad@apiad.net. No one who wants to read this book should be unable to.
Ongoing updates and discussion live at the Computist Journal.