Library · book

Artificial Intelligence: A Guide for Thinking Humans

Melanie Mitchell
2019·Farrar, Straus and Giroux

Source: https://melaniemitchell.me/aibook/

The book that calibrates judgment.

Mitchell explains how AI actually works and then, without cynicism or hype, marks the line between impressive performance and genuine understanding — the line that decides where these systems can be trusted and where they quietly fail.

For directing AI product, that discipline is the whole game: the costly mistakes come from mistaking fluency for comprehension.

Written before the LLM era yet sharpened by it, it pairs naturally with the stochastic-parrots debate as a durable check against overclaiming.

Central argument

A lucid, honest survey of AI's methods and their limits by a researcher who refuses both hype and doom. Mitchell walks through vision, deep learning, reinforcement learning and language, and repeatedly returns to the central puzzle: these systems perform impressively while lacking anything we would recognize as understanding, common sense, or meaning — and that gap has practical consequences.

Critique

Written before large language models reset the field, its concrete examples now feel dated, and it necessarily says nothing about transformers at scale, RLHF or the frontier/open dynamics. It is a conceptual and critical foundation, not a current-state report.

Why it matters for product

For a CTO/CPO, this is the book that keeps expectations calibrated: it trains the instinct to ask 'does it understand, or does it pattern-match?' and to distrust benchmark theater — the same instinct that separates good AI product judgment from credulous adoption. It sits beside the stochastic-parrots debate as ballast against overclaiming.