Four AIs walk into a book club: how I built my AI reading list

The AI field produces books faster than anyone can read them and obsoletes them faster than publishers can print them. So when I decided to build a serious reading stack, I didn't ask one AI for recommendations. I gave the same brief to four frontier models — Claude, GPT, Gemini, and Grok — then made them critique each other's lists until something defensible survived.

The brief was two links and one sentence: pick me 3–6 books on AI using these principles.

It started with Pyrate (@CEOLandshark) on X, on making "quantum leaps" in understanding a field through a deliberately small, well-chosen stack: the canonical book — "if you ask anyone practising this, they'd tell you to read this one"; a practitioner's / teacher's guide — "how the new experts are approaching it … what I would've done differently if I started X again"; and the underground pick — "find the dark sheep book and read that one." Together, "an almost-360 view of the whole field." And before opening a single page, the pre-script that does most of the work:

"Before you pick a skill or a topic … just ask yourself, what do you want out of this? And really inquire into it. … This is the highest leverage and easiest shift to make, and if you get this wrong, it's the costliest. So you have to run that script every single time."

In his replies, someone dropped the complementary frame: The 3 Books Technique from LessWrong — pick a What (a thorough reference you consult when stuck), a How (step-by-step methods you'll practice), and a Why (the intuitions behind the skill), then pair them with a live project. The two frames overlap but don't coincide, which turned out to matter: the dark-sheep slot exists only in Pyrate's version, and it produced the best book on the final list.

My pre-script answer: I build agentic AI systems for my own products and sell AI judgment to small and mid-sized businesses. I don't train models, and I don't need a survey of the academic discipline. I need to understand the machine I'm betting on, know how it fails, and be able to tell a client what's real.

What happened when the models argued

Each model drafted a list. I fed the lists back across the panel. Four things came out of the collisions that no single list contained:

• Unanimity on one book — with an asterisk. All four models picked Chip Huyen's AI Engineering as the working canon for building on foundation models. Useful signal — but as one model correctly cautioned, agreement among AIs trained on overlapping sources is one recommendation counted four times, not four independent measurements. The stronger reason to pick the book is its fit with the actual work. (It fits.)
• A genuine fight over the canon slot. Two models nominated Russell & Norvig's Artificial Intelligence: A Modern Approach — the field's canonical 1,100-page textbook. Both hedged it into "a reference you return to rather than read." That hedge decided it: a canon you're told not to actually read is a canon for a different student. The pre-script settles the question — the academic discipline's canon and the working canon for shipping dependable systems are different books. AIMA is out.
• A taxonomy correction. One model pointed out that Sebastian Raschka's Build a Large Language Model (From Scratch) — which I'd slotted as the "How" — is really a "Why" book: coding attention heads builds mechanistic intuition about the layer under your daily work; it isn't your daily work. Correct. The "How" slot, it turns out, is best filled by the project you're already shipping, audited against a book rather than led by one.
• A gap nobody's book covers. The sharpest observation of the whole exercise: none of the surviving books addresses local-first, audit-trail, evidence-chain engineering for agentic systems — building agents whose claims you can verify and whose actions leave receipts. The books assume looser constraints than that. More on this at the end.

The stack

No affiliate links here — everything below points to the author's or publisher's own page. If I'm selling you judgment about AI claims, my book links shouldn't have a kickback attached.

The canon (desk reference): AI Engineering — Chip Huyen. Evaluation, retrieval, agents, guardrails, inference economics — organized as decision frameworks for building on foundation models. The one book all four models agreed on. Lives on the desk, consulted per-problem.

The dark sheep (read first): AI Snake Oil — Arvind Narayanan & Sayash Kapoor. The vendor-claim demolition kit, including the single most useful razor in the stack: the difference between generative AI (bounded automation that works) and predictive AI (forecasting human outcomes, which routinely doesn't). For anyone buying or selling AI services, this is the book that changes the next scoping conversation.

The why, in two layers: Why Machines Learn — Anil Ananthaswamy. The mathematics of why learning from data works at all, told through the people who figured it out. Then Build a Large Language Model (From Scratch) — Sebastian Raschka. You type in a working GPT-class model: tokenizer, attention, training loop. After it, "why do long contexts degrade" and "what is temperature actually doing" stop being folklore. Math intuition, then code mechanics — a deliberate two-part look under the hood.

The practitioner audit: Build an AI Agent (From Scratch) — Jungjun Hur & Younghee Song. Framework-light construction of the agent loop on raw APIs. If you're already building agents, read it as an audit of your own orchestration patterns — the context-engineering chapters are the part worth slowing down for.

The audio tier (context, not technique): The Alignment Problem — Brian Christian: the intellectual history of systems optimizing the measure instead of the goal. And Empire of AI — Karen Hao: critically reported inside-history of the industry your stack depends on. Due diligence on your vendors, with a thesis to weigh.

Reading order: AI Snake Oil first — it calibrates judgment before the engineering books build confidence. Then the two Why books, then the agent audit. Huyen throughout, the audio tier on drives.

The slot no one could fill

That gap the panel found — auditable, evidence-first agentic engineering — isn't on the list because the book doesn't exist. The practice barely exists. It happens to be exactly what we spent this month building: agents whose every observation is validated before it's believed, whose health reporting can't confuse "no data" with "no findings," and whose claims arrive with receipts. The books won't do that translation for you.

So that last slot on the shelf is one we're writing ourselves, one shipped system at a time. The first installment is already on this site.

Chris Enoch is the founder of Enoch AI LLC, an AI consulting practice for small and mid-sized businesses. [email protected]