Quick Take
- Narration: Michelle Peitz delivers clean, accessible narration suited to the popular-science register, conversational without being casual, which matches the book’s tone well.
- Themes: How AI learns from data, neural network architecture for general readers, the human role in AI systems
- Mood: Curious and demystifying, pitched at the genuinely interested non-specialist
- Verdict: A well-organized introduction to AI fundamentals that treats its reader as intelligent without requiring prior technical background, the ten-secrets structure gives the book a satisfying sense of progressive revelation.
I put this on during a Sunday walk, which turned out to be the right context entirely. How AI Really Thinks by Julian Vexley is the kind of book that works best when you’re moving, not because it’s background noise, but because the ideas are dense enough to sit with while your body is doing something else. This is a book about how AI actually works, stripped of the mysticism that surrounds it in most popular coverage, and written for the person who has watched AI generate images and write emails and defeat chess grandmasters and genuinely wants to understand what is actually happening under the surface.
Vexley organizes the book around ten core secrets of modern artificial intelligence. The structure is explicitly revelatory, each chapter uncovering a fundamental aspect of how AI systems operate. The first few secrets establish the foundations: data as AI’s fuel, algorithms as its rulebook, and the way machines learn from examples in a process that parallels but doesn’t replicate how humans learn from experience. The middle chapters move into the mechanisms that make modern AI work at scale: training methods for image recognition and speech processing, and the neural network architecture loosely inspired by the human brain that allows AI to detect patterns too complex for direct human analysis. The final chapters address prediction, feedback-based improvement, and the differentiation between AI types suited to different tasks.
The Human Role as the Final Revelation
The book’s final secret is also its most important: no matter how impressive AI becomes, it still relies heavily on humans for guidance, correction, and direction. That’s not a controversial claim in the technical community, but it is consistently lost in popular coverage that alternates between AI-as-miracle and AI-as-threat. Vexley lands the point well, not as reassurance but as structural reality. AI systems are built on data that humans curate, trained on objectives that humans define, corrected by feedback that humans provide, and deployed in contexts that humans design. Understanding that dependency doesn’t diminish what AI can do. It clarifies what the human role remains and must remain.
The Ten-Secrets Structure as a Listening Feature
The organizational decision to frame the content as ten secrets works particularly well in audio. Each chapter functions as a discrete revelation with its own satisfying architecture: establish the question, explain the mechanism, show the real-world example. The examples are consistently well-chosen. The explanation of how Netflix recommendations work is more illuminating than the standard description because Vexley connects it explicitly to the training process rather than treating it as a black-box output. The section on how AI recognizes faces connects to the neural network chapter in a way that builds cumulative understanding rather than isolated facts. At two hours and seventeen minutes, the book is short enough to complete in a single extended listening session, and the structure supports that unbroken engagement.
What the Book Doesn’t Cover
Vexley’s scope is intentionally limited to the fundamentals. Readers looking for coverage of generative AI, large language models, or the current wave of tools built on transformer architecture will find those topics absent or only lightly referenced. The ten-secrets framework focuses on the underlying learning mechanisms rather than their current applications. That’s actually an asset for the book’s durability, since the fundamentals of how neural networks learn from labeled data haven’t changed even as the applications built on them have proliferated dramatically. But listeners who came to this book specifically to understand ChatGPT or Midjourney will need to supplement it with more current material.
Who should listen: curious non-specialists who want a structured, accurate introduction to how AI systems actually function, the person who has been asking how does this thing really work for the past two years and hasn’t found a satisfying answer. Who should skip: anyone with a technical background who is looking for depth beyond fundamentals, or readers who want coverage of generative AI and LLMs specifically.
Frequently Asked Questions
Do the ten secrets in How AI Really Thinks build on each other, or can chapters be listened to independently?
They build deliberately. Vexley structures the book so each chapter’s concepts support the ones that follow, the data chapter sets up the algorithm chapter, which sets up the training chapter, and so on. Listening in sequence produces a more coherent understanding than jumping between chapters.
Does the book cover generative AI, ChatGPT, or image generation tools like Midjourney?
The book focuses on foundational AI mechanisms, how systems learn from data, how neural networks detect patterns, how feedback improves performance, rather than the current wave of generative AI applications. Those tools are products of the mechanisms the book explains, but they’re not the primary subject.
Is this book more about the philosophy of AI or the technical mechanics of how it works?
Primarily the mechanics, explained in accessible terms. Vexley explains training methods, neural network architecture, prediction systems, and feedback loops using relatable analogies and real-world examples. The philosophical implications appear at the end, particularly around the ongoing human role in AI systems.
At just over two hours, is How AI Really Thinks comprehensive enough to be useful?
For its intended purpose, yes. The book aims to replace mystery with understanding at a foundational level, not to produce AI practitioners. Listeners who finish it will have a reliable mental model for how AI learns, why it makes mistakes, and what human oversight actually means in technical terms, which is more than most popular coverage provides.
