Pages
209
Published
2019
A Guide to Thinking About Machines That Think
Understand what AI can and cannot do — so you can use AI tools with clear eyes and realistic expectations.
Melanie Mitchell cuts through the noise around artificial intelligence to examine what modern AI systems actually do, where they fail, and why the gap between hype and reality persists. Drawing on decades of research, she traces the ideas behind machine learning, neural networks, and computer vision without hiding the hard conceptual problems that remain unsolved. Readers finish with a grounded mental model of AI capabilities — one that holds up under pressure.
Artificial intelligence is everywhere, and so is confusion about it. Autonomous vehicles, image-recognition systems, language models, and game-playing programs are routinely described in terms that suggest human-level understanding. Melanie Mitchell argues, carefully and with evidence, that this framing is wrong in important ways — and that understanding exactly why matters more now than ever.
Mitchell is a professor at the Santa Fe Institute and has spent her career at the intersection of cognitive science and AI research. In this book she does not lecture you about a distant future. She walks you through the concrete mechanisms behind today's most discussed AI systems: how deep neural networks learn from data, what computer vision systems are actually detecting, how natural language processing works at the level of statistics and pattern matching, and where each approach breaks down when pushed outside its training distribution.
Each chapter pairs the history of an idea with its modern implementation and then asks the honest question: does this system understand anything, or is it doing something more brittle? The answer is almost always more complicated than either the optimists or the pessimists claim, and Mitchell is one of the clearest writers working on this material.
If you use AI tools professionally — writing assistants, code completions, image generators, recommendation engines — this book gives you the conceptual vocabulary to reason about what those tools are doing beneath the surface. You will stop being surprised by their failures and start being more deliberate about when to trust their outputs.
This is not a tutorial, and it will not teach you to train a model. What it will do is give you a durable framework for evaluating AI claims — the kind of framework that does not expire when the next benchmark is broken.
Mitchell defines the field, its recurring ambitions, and the persistent gap between what AI researchers promise and what systems deliver. You learn why the question 'can machines think?' is harder to answer than it sounds.
You examine how multilayer neural networks learn representations from data, why the resurgence of deep learning happened when it did, and what the architecture actually computes.
Mitchell traces how machines learn to recognize images and where that recognition breaks down — adversarial examples, texture bias, and the absence of anything resembling visual understanding.
You explore how statistical and neural approaches to language work, what large language models are doing when they produce fluent text, and why fluency is not the same as comprehension.
Mitchell uses chess, Go, and Atari games to explain reinforcement learning and to ask whether a system that defeats world champions has learned anything that transfers beyond the game.
You engage with the hardest open problem in AI: giving machines the background knowledge and analogical reasoning that humans deploy effortlessly, and why current architectures struggle here.
Mitchell examines what it would mean for an AI system to genuinely understand language or images, drawing on cognitive science to clarify what is missing from today's best models.
You work through the landscape of AI risk arguments — from near-term bias and reliability failures to long-term existential claims — and develop a framework for evaluating them with appropriate skepticism.
A science or engineering background helps you get the most from the technical sections, but Mitchell writes for a broad audience and explains core concepts from scratch. You do not need to know how to code.
No. This is a conceptual and critical overview, not a tutorial. If you want hands-on implementation, you will need a separate resource. This book prepares you to think clearly about what you build or use.
The foundational ideas Mitchell covers — how neural networks learn, why generalization is hard, what understanding requires — remain directly relevant to systems released after the book was written. Some specific benchmarks and products will have moved on, but the conceptual framework has not aged.
Mitchell is skeptical of strong claims on both sides. She engages seriously with risk arguments but grounds her analysis in the actual capabilities and limitations of current systems rather than speculative futures.
Yes. Each chapter is self-contained enough to anchor a discussion session, and the absence of heavy mathematics means participants from different roles — engineering, product, policy — can engage on equal footing.
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