Books I keep coming back to.

This was my first actual AI book, not the “let’s re-explain linear regression for the 47th time” kind, and it delivered. What makes AI Engineering stand out is that it treats AI like what it really is: an engineering problem. Not a research paper. Not a hype machine. A messy, real-world system that has to work in production without breaking every five minutes. The book pulls off a rare trick, it’s comprehensive without being exhausting. Chip Huyen covers a ton of ground, but never drowns you in theory you’ll forget in two days. Everything feels intentional, practical, and immediately useful. The parts that really hit for me:

• RAG and agents: finally explained in a way that makes you think, “oh, so that’s how people are actually improving these systems,” instead of vague architectural diagrams that go nowhere.
• Fine-tuning: concepts like quantization and PEFT are usually treated like insider secrets. Here, they’re explained clearly enough that you can actually use them, not just nod along pretending you understood.
• Observability, testing, monitoring: the least sexy part of AI, and therefore the one most teams ignore until everything explodes. This section alone should be mandatory reading for anyone shipping LLM apps.

Also, the LLM application sections are packed with practical “here’s how you build this thing” guidance, which is refreshing in a space where too many books stop at “imagine what you could build.” Bottom line: whether you’re just getting into AI or you’ve already deployed models in production, this book earns its spot. It’s the kind of resource you don’t just read, you keep coming back to when things inevitably break.

Josh Starmer is basically the Gandalf of machine learning—he makes all the scary, math-heavy stuff approachable without turning you into a zombie from endless equations. This is the book you grab if you want a solid grip on ML without drowning in jargon or crying over derivatives.

Sure, it won’t replace a full-on math bootcamp, but it gives you the big picture and shows how the gears turn behind the scenes.

My favorite bit? Watching Josh walk you through turning raw input data into predictions.

• First, he fiddles with weights and biases like a chef seasoning a dish.
• Then, he reveals the “magic” of gradient descent so your model actually learns something.

It’s like watching a wizard at work, but with numbers.

I went in super hyped because I loved Josh Starmer's previous book. And while this one covers a lot of ground, the PyTorch examples felt a bit like Groundhog Day—helpful for building architectures from scratch, but light on the kind of juicy variations that make you truly understand the concepts.

Still, it’s a solid resource if you want a high-level grip on neural networks without drowning in math. Like the last book, this isn’t about deep equations—it’s the big picture view, showing how networks tick.

Where it shines is the peek behind the curtain at the latest AI toys: encoder-only Transformers, decoder-only Transformers, and the many flavors of attention mechanisms. Josh grabs your hand and walks you step by step through the math that magically turns raw words into numbers and then back into words. It’s like watching a wizard translate gibberish into sense, and somehow, it works.

This book doesn’t just talk the talk—it makes your code walk the walk. LangChain Expression Language gets demystified like a magician revealing tricks, with the pipe operator and batch invocations turning messy workflows into neat, readable chains. Preprocessing? Parallel inference? Check. RAG and agentic logic? Covered, with advanced toys like ColBERT, RAPTOR, and Reciprocal Rank Fusion handed to you with practical patterns for reflection and modular subgraphs. Forget just theory: you’ll learn how to build self-correcting agents and slick retrieval routing like a pro. If you want your autonomous systems to actually perform reliably, this is your blueprint.

It all kicked off on YouTube when Salvatore Sanfilippo, Redis’ founding wizard, popped up in my feed. I knew zilch about Redis, but curiosity got the better of me. Then Instagram decided to guilt-trip me with an O'Reilly ad for Managing Memory for AI Agents, a collab between the Redis crew and O'Reilly. The title hit me like a paper cut, so I figured, yeah, probably should read that.

Thankfully, 'should' quickly became 'did'. True to O'Reilly’s efficient style, the book is detailed yet compact—a rare combo. I absolutely geeked out over the authors’ metaphor using System 1 and 2 from Kahneman’s Thinking, Fast and Slow to explain agentic memory types and multimodal architectures. If you know that book, it’s like sliding into a warm, familiar pair of socks.

Bonus points for the section on Collective Intelligence, which was completely new to me but super fascinating. And the part dissecting framework vs custom architecture? Classic Joel Spolsky territory—timeless dilemmas, well served.