Pages
233
Published
2019
How Artificial Intelligence Works and Why It's Weirder Than You Think
Understand how AI systems actually work — and why they fail in such gloriously strange ways — so you can use them with clear eyes.
Janelle Shane pulls back the curtain on modern artificial intelligence, revealing not a superintelligent overlord but a surprisingly literal-minded pattern-matcher with a talent for going spectacularly wrong. Through real experiments and genuinely funny examples, this book explains how neural networks learn, why they fail, and what that means for the AI tools shaping everyday life. It is the clearest, most honest introduction to AI available for curious non-specialists.
You have probably encountered AI doing something baffling. A photo app that labels a dog as a muffin. A chatbot that confidently gives wrong directions. A spam filter that lets through exactly the messages you wanted blocked. These are not bugs waiting to be fixed. They are AI working exactly as designed, and understanding why changes how you think about every AI tool you touch.
Janelle Shane is a research scientist who spent years feeding strange prompts to neural networks and documenting the results. The outputs ranged from nonsensical to hilarious to quietly unsettling, and each one revealed something true about how modern AI actually operates. This book grew out of that work. It is not a textbook and it is not a technology-industry press release. It is an honest, grounded account of what machine learning can and cannot do.
The core argument is simple: AI learns by finding patterns in training data. It is very good at this, sometimes startlingly good. But it has no understanding of the world behind that data, which means it can fail in ways a human never would. It may optimize aggressively for a proxy goal while missing the actual goal entirely. It may generalize confidently from a skewed dataset. It may produce output that looks convincing and is completely wrong. Knowing this makes you a smarter user of any AI system.
Shane explains neural networks, training, reward functions, and generalization without requiring a line of code or a calculus prerequisite. The examples do the work: ice cream flavor names that make no sense, AI-designed Halloween costumes, pickup lines generated by a language model that clearly does not understand human attraction. Each example is funny, and each one is also a precise illustration of a real concept.
This is the book to read before you form strong opinions about AI. Whether you use AI tools in your work, encounter them as a consumer, or just want to have an informed conversation about where this technology is actually headed, Shane gives you the vocabulary and the intuition to think clearly about it.
Shane introduces what AI actually is by contrasting it with science-fiction expectations. You see your first examples of neural networks producing strange output and start building an accurate mental model from the ground up.
You follow the training process step by step, from raw data to a model that can make predictions. Shane explains how pattern-matching without rule-following produces both impressive results and spectacular failures.
This chapter shows how the quality, scope, and hidden biases of training data determine what an AI system can and cannot do. Real examples illustrate how a skewed dataset produces a confidently wrong model.
You learn what happens when an AI is given a proxy objective and optimizes it too well. Shane's examples range from funny to alarming and make the concept of misaligned objectives concrete and memorable.
Shane examines how models trained on one context fail when applied to another, and why the failure is often invisible until something goes wrong in production. You finish with sharper intuitions about model limitations.
You survey AI systems you already encounter every day, from content recommendation to spam filtering to autocomplete, and apply the framework from earlier chapters to understand why they behave as they do.
This chapter addresses AI-generated content that looks authoritative but is fabricated or subtly incorrect. You develop a checklist of signals to watch for and habits for verifying AI output before trusting it.
Shane closes with a clear-eyed account of genuine AI strengths, the narrow tasks where it beats humans reliably, and what that means for how you should and should not use AI tools going forward.
No. Shane deliberately avoids equations and code throughout. The concepts are explained through examples and analogies that require only curiosity, not technical training.
The core concepts — how neural networks learn, why training data matters, how reward hacking works — have not changed. The specific examples predate the large language model boom, but the mental models Shane builds apply directly to understanding those systems too.
It reads like popular science: narrative, funny, and built around vivid examples rather than structured reference material. You will come away with durable intuitions rather than step-by-step procedures.
Readers who already work in machine learning or AI research will find the technical level too introductory. It is written for people who are curious about AI, not people who are already building it.
Shane discusses risks and failure modes honestly without adopting a strong ideological stance. The tone is analytical and sometimes wry rather than alarmist or promotional.
New
The Disruptive Economics of Artificial Intelligence and What It Means for Business Decisions
by Ajay Agrawal
New
New
A Practical Guide to Using AI Coding Tools and Productivity Assistants in Your Daily Workflow
by Tom Taulli
New