How to build your 2026 AI summer reading list

Most AI reading lists you’ll see this summer are four books that already agree with each other. Read one and you’ve read them all.

A list built that way just tells you what you already suspect. If you’re a founder or a VP of Engineering deciding what to build in the next two quarters, what you want from a reading list is calibration, and calibration comes from people who would tell you you’re wrong. Which makes the useful question a structural one: what does a list have to contain before it can move you?

Four criteria, and the candidates each one turns up.

It needs the machinery, before it needs any opinions

It’s hard to judge an argument about AI without knowing what a gradient is. One item on your list should teach you how the systems work.

• Why Machines Learn: The Elegant Math Behind Modern AI, Anil Ananthaswamy. The math, honestly presented. Goodreads has it at 3.5, which reflects how demanding it is. Kirkus called it “an illuminating overview of how machine learning works.” Bring calculus.
• These Strange New Minds, Christopher Summerfield. An Oxford neuroscientist argues these systems do something worth calling thinking, grounded in how human reasoning works. The most rigorous statement of the optimistic position.
• AI Engineering, Chip Huyen. The practitioner’s book, and the most-read title on O’Reilly’s platform after its January 2025 release. If your team builds on foundation models, this one will change what you do Monday.

It needs the bull case in the builders’ own words

Optimism filtered through a journalist is worth less than optimism from the people spending the capital. Go to the primary sources and judge the reasoning yourself.

• The Scaling Era: An Oral History of AI, 2019 to 2025, Dwarkesh Patel and Gavin Leech (Stripe Press). Dario Amodei, Demis Hassabis, and Ilya Sutskever in their own words. Treat it as a primary document. These are the people placing the bets, describing the bets.
• Co-Intelligence, Ethan Mollick. A 3.94 on Goodreads and useful as a working manual. Mollick actually uses the tools and reports what happened.

It needs a skeptic, and the skeptic shelf is uneven

The critical books vary a lot in quality. Favor the critics who make technical distinctions over the ones who make noise.

• AI Snake Oil, Arvind Narayanan and Sayash Kapoor. The strongest of them. The Princeton authors separate predictive AI (mostly broken, often harmful) from generative AI (real, oversold), and that distinction is the most useful thing in the genre. It’s also what decides whether a feature needs grounded retrieval or shouldn’t be built with a model at all.
• The AI Con, Emily Bender and Alex Hanna. Sharper prose, thinner argument. Reviewers comparing the two have generally landed on AI Snake Oil as the more convincing case. Second in line, if at all.
• Empire of AI, Karen Hao. Built on more than 260 interviews. An institutional critique: who is making these decisions, and how. A legitimate question, and a separate one from the technical case.
• If Anyone Builds It, Everyone Dies, Eliezer Yudkowsky and Nate Soares. The doom pole. Publishers Weekly called it an “urgent clarion call”; The Atlantic’s Adam Becker called it “tendentious and rambling” and said the authors “fail to make an evidence-based scientific case.” Worth reading for the strongest form of the argument, and for how much of it rests on assertion.

It needs papers, and it should weight them above the books

A book about AI is roughly 18 months stale the day it prints. The research is where your list earns its keep, and these three should reach your Q3 planning.

• METR’s developer productivity RCT (July 2025). Sixteen experienced open-source developers, 246 real tasks. With AI tools available they were 19% slower. Afterward, those same developers estimated AI had made them 20% faster. Then read METR’s February 2026 update, where they call the finding historical and redesign the study. Watching researchers undercut their own viral result is a good education in how to weigh evidence.
• Apple’s “The Illusion of Thinking” (June 2025) and Alex Lawsen’s rebuttal, “The Illusion of the Illusion of Thinking.” Apple claimed reasoning models collapse past a complexity threshold. The rebuttal showed many of those failures were output-token limits, not reasoning limits. Read them back to back; the disagreement teaches more than either paper alone.
• DORA’s 2025 State of AI-assisted Software Development. Adoption hit 90%. AI raised delivery throughput and instability at the same time. Its conclusion: AI is an amplifier, magnifying whatever discipline your team already has, or doesn’t.

The obvious objection

You don’t have time for nine books. Nobody does.

So don’t build a nine-book list. Take one title from the bull case, one from the skeptics, and all three papers. Two books that disagree will do more for your judgment than five that nod along. If the list gets cut to a single item, make it the METR study read next to METR’s own walk-back. That pair shows how to hold a belief: strongly, with a number attached, and loosely enough to drop it when the evidence turns.

The DORA finding is the one we keep hitting on real engagements. A team without tests, review, and a deployment pipeline they trust will use AI to generate broken code faster. The teams pulling real value from these tools already had the discipline in place. That’s why we open an engagement with the Tarmac 10 before any model selection debate, and why shipping production AI is a different job from shipping a demo.

Build the list that can change your mind. It’s cheaper than finding out in production.