On Intelligence

What is this book about?

Published by Times Books in 2004, written by Jeff Hawkins, inventor of the PalmPilot and a member of the National Academy of Engineering, with the New York Times science writer Sandra Blakeslee. Hawkins was dissatisfied with the dominant approach in artificial intelligence — defining intelligence by behaviour that passes the Turing test — and sceptical of reducing the brain to an input-process-output computer. He proposes the memory-prediction framework: the neocortex is a stack of predictors, each layer forecasting the input it will receive next on the basis of sequences it has seen before. Perception is not passive reception; the brain continuously generates predictions, compares them with reality, and updates only when a prediction fails. The book first asks what intelligence is, then shows how the framework explains memory, perception, creativity and consciousness, and finally discusses how to build machines that genuinely learn.

Why read it?

Because it offers a portable framework for thinking, not merely a claim about the brain. Defining intelligence as prediction changes the questions you ask: instead of asking whether a machine can do X, ask whether it holds an internal model of the world and can anticipate what happens next. That lens still has explanatory force against today's deep learning and amounts to a serious challenge to the purely data-driven route — Hawkins argues that fitting inputs to outputs on vast labelled datasets lacks precisely a model of how the world works. Its position is plainly refutable: cognitive scientists note that predictive coding did not originate with Hawkins and that the framework struggles to explain language and abstract reasoning, while engineers point out that it has not produced a comparable system in two decades. The value of reading it lies not in believing the conclusion but in acquiring a reference point against which to inspect every AI narrative. Evidence of standing: (1) a standalone Wikipedia book article (Times Books, 2004); (2) a professional book review in the IEEE Intelligent Informatics Bulletin in 2006; (3) successive formal Chinese editions from Shaanxi Science and Technology Press (2006), China Overseas Chinese Publishing House (2014) and Zhejiang Education Press (2022); (4) Hawkins was elected to the National Academy of Engineering in 2003.

Core Ideas

• Prediction, not behaviour, is the core of intelligence. Judging intelligence by whether something behaves like a person mistakes mimicry for understanding; the real mark is the ability to anticipate before an event occurs and to correct an internal model by its error.
• The brain is a hierarchical prediction machine. The neocortex is a stack of many structurally similar cortical columns; lower layers predict nearby, short-horizon input, higher layers predict more abstract patterns over longer timescales, and invariant representations let the same object be recognised across angles and senses.
• Memory is not playback but the extraction of sequence regularity. What we store is not frame-by-frame imagery but the structure of what usually happens next; creativity recombines fragments of different sequences and forms new links where prediction fails.
• Genuine machine intelligence must be built on time and sequence. Hawkins argues that static pattern recognition detached from the temporal dimension cannot lead to general intelligence; a machine must, like a brain, continuously predict the next input.

What questions does this book try to answer?

• If intelligence is prediction, what exactly does passing the Turing test prove, and what does it leave out?
• Can a system that only fits inputs to outputs genuinely understand what it processes without a model of the world?
• Can the prediction framework explain language, mathematics and abstract reasoning? If not, where are its limits?

Who should read it?

Readers who keep asking what intelligence actually is; AI practitioners and researchers, especially those wanting to step outside the stack-more-data-and-scale-more-models paradigm and look for an alternative; students of cognitive science and neuroscience; and investors and observers following the debate over paths to AGI. It is not for those seeking concrete algorithms or an engineering tutorial — it is a manifesto, not code.

How to Read It

The first half — the definition of intelligence and the hierarchical structure of the brain — is the skeleton and deserves slow reading; the middle sections on memory, perception and creativity unfold the framework and can be checked against your own experience; the final chapter on routes to machine intelligence reads well against present-day deep-learning practice. It is worth pairing with A Thousand Brains (2021) to see how Hawkins revised his own theory seventeen years later. The most recent Chinese edition is New Machine Intelligence (Zhejiang Education Press, 2022); the earlier Shaanxi Science and Technology edition appeared as The Future of Artificial Intelligence. Read it with scepticism intact: treat prediction as a heuristic, not a dogma.