ML Book Stacks, AI Engineering Foundations & The Data Curriculum
Let's kick things off with something practical — because not everyone needs a PhD to get good at this stuff. A former Meta, Netflix, and Airbnb data engineer — who now runs a tool doing a hundred and fifty thousand dollars a month — just made a bold claim: you only need four books to master ML and data engineering.
The list? "Fundamentals of Data Engineering" by Joe Reis. "Designing Data-Intensive Applications" by Martin Kleppmann. And two from Chip Huyen — "AI Engineering" and "Designing Machine Learning Systems." That's it. No massive reading list. No drowning in whitepapers. Just four books, done right.
He pairs those with four leadership reads — "Radical Candor" and "Atomic Habits" are on that side of the stack. The argument is simple: technical depth plus behavioral discipline equals outsized results.
Now — Chip Huyen's "AI Engineering" keeps coming up, and for good reason. It's apparently O'Reilly's most-read book since it dropped. So what's actually in it?
Here's the core mental model it pushes: AI engineers build *on top of* foundation models — they don't build from scratch. That distinction matters more than people realize.
And the adaptation hierarchy is genuinely useful. Start with prompt engineering — zero cost, try it first. If that's not enough, move to RAG. Only after that do you consider fine-tuning. Full fine-tuning needs millions of examples. LoRA? You can get away with hundreds. That alone is worth writing down.
There's also a section on the Chinchilla scaling law — the rule of thumb being roughly twenty times your model's parameter count in training tokens. So a three-billion parameter model ideally trains on sixty billion tokens. It's a useful back-of-the-napkin sanity check.
On evaluation — the book leans into AI-as-judge, but it's honest about the problems. Verbosity bias. Position bias. Self-bias, where a model tends to favor its own outputs. Knowing those failure modes matters if you're actually shipping anything real.
And the diagnostic framing is clean: are you dealing with an information failure — the model doesn't *know* something — or a behavior failure — the model knows but acts wrong? Different problems, different fixes.
Zooming out a bit — there's also a twenty-book curriculum floating around that takes a more product-and-strategy angle alongside the technical stack. On the generative AI side, it pulls in "LLM Design Patterns" by Ken Huang and "Generative AI with LangChain." And for the business layer — books like "The Profitable AI Advantage" and "Your AI Survival Guide." It's a broader sweep, but the through-line is the same: understand the models, then figure out how to build something people actually pay for.
The takeaway from all of this? The reading list consensus is converging. Kleppmann for systems. Huyen for AI engineering and ML design. Reis for data foundations. Everything else is supplementary.
That's your AI digest for 02 May 2026.