On Intelligence: The Brain as a Prediction Machine

A phone designer looked at the neocortex and saw something colder and more useful than mystery: a layered memory system that keeps guessing what comes next. That is the sharpest gift of Jeff Hawkins and Sandra Blakeslee’s On Intelligence. The book is not mainly about IQ tests, chess, or the tricks that make a machine look clever for five minutes. It is about the thing happening before behavior: the brain’s private model of the world.

Hawkins’ central move is simple enough to feel obvious after you hear it. Intelligence is not a list of outputs. It is not the ability to win a game, answer a question, or imitate a human voice. Those are performances. Intelligence is the ability to use memory to make predictions, notice when the predictions fail, and update the model that produced them.

That sounds quiet. In the book, it becomes radical.

The book’s core idea

On Intelligence argues for the “memory-prediction framework.” The neocortex, Hawkins says, is not a blank processor waiting for commands from the senses. It is constantly predicting what it is about to see, hear, touch, or think. Perception is not passive reception. It is a negotiation between incoming signals and stored patterns.

You recognize a coffee cup even when the handle is turned away. You know a song after three notes. You can walk through your apartment in the dark because your body is not merely reacting to the room; it carries a model of the room. The brain’s power comes from that model. It stores sequences, builds stable representations from messy inputs, and uses those representations to anticipate the next moment.

The valuable point is not “the brain predicts.” Plenty of people now say that. The valuable point is how Hawkins ties prediction to hierarchy. The neocortex is layered. Lower areas deal with fast sensory detail. Higher areas combine details into broader, more stable patterns. A line becomes an edge. Edges become a face. A face becomes a person. A person becomes a story about what is likely to happen next.

That hierarchy matters because raw sensation changes constantly. Intelligence needs invariance. Your friend’s face is still your friend’s face in morning light, on a bad camera, turned sideways, half-hidden by a scarf. The brain does not memorize every pixel. It learns patterns that remain useful despite change.

Why behavior is the wrong starting point

One of the book’s best punches lands against behaviorism and shallow machine intelligence. If you judge intelligence only by external behavior, you can be fooled by a clever script. A machine can produce the right answer without understanding the world that answer belongs to. A person can also perform well on a narrow test while having a brittle model outside the test.

Hawkins wants the definition moved upstream. The important question is not “What did it do?” but “What internal model made that action possible?” A thermostat changes behavior when the room is cold, but nobody thinks it understands winter. A real intelligent system should form predictions across many levels of abstraction, use those predictions to interpret noisy input, and revise itself when reality disagrees.

This distinction is still useful in the age of large language models. A fluent answer is behavior. The deeper question is whether the system has a durable, grounded, self-correcting model of the world it talks about. On Intelligence gives you a way to ask that question without getting hypnotized by surface performance.

Memory is not a warehouse

The book also cleans up a common mistake: treating memory as storage. In ordinary language, memory sounds like a box where facts sit until called. Hawkins treats memory more like a living map of sequences. The brain remembers not only objects, but transitions: what usually follows what, what belongs with what, what changes when you move your eyes or your hand.

That is why time matters. Intelligence is not just classification. It is sequence learning. If you hear footsteps behind you on an empty street, the next second matters. If a child hears “Once upon a…,” the missing word is already warming up in the mind. The brain is always ahead of the present by a small amount. That tiny lead is where understanding begins.

The strongest lessons for AI

• Build models, not just reactions.A system that only maps inputs to outputs may look useful, but it remains fragile when the world shifts. Real intelligence needs internal structure that can survive novelty.
• Prediction is a test of understanding.If a system understands a domain, it should anticipate what is likely, what is impossible, and what would be surprising. Surprise is not noise. It is a training signal.
• Hierarchy compresses the world.Useful intelligence turns floods of detail into stable concepts. Without hierarchy, every new scene is too expensive to interpret.
• Embodiment and movement are not side issues.The brain learns by acting in the world. Seeing is partly knowing what would change if you moved. Touch is partly knowing what your hand did. Hawkins’ framework keeps action tied to perception.
• Intelligence is not the same as consciousness.The book is careful to separate machine intelligence from the whole mystery of human experience. That makes the engineering problem cleaner: start with prediction, memory, hierarchy, and world modeling.

Where the book is most valuable

The book’s best value is not that every neuroscientific detail must be treated as final. Neuroscience has moved since 2004. AI has moved even faster. The value is the frame. Hawkins gives readers a durable suspicion of spectacle. Whenever a system looks smart, ask what it predicts. Ask how it represents the world. Ask whether it can notice a wrong expectation and repair the model.

That frame travels well. It helps with AI design, product thinking, education, and even self-understanding. A good learner is not someone who hoards facts. A good learner builds a model, exposes it to reality, feels the error, and adjusts. That is uncomfortable. It is also the only way intelligence grows.

The book leaves you with a clean image: the mind is not sitting at the back of the theater watching a film called reality. It is backstage, holding a rough script, whispering the next line, and correcting itself when the actor on stage says something else.

Sources

• Jeff Hawkins and Sandra Blakeslee, On Intelligence: How a New Understanding of the Brain will Lead to the Creation of Truly Intelligent Machines, 2004.
• Wikipedia summary and bibliographic page for On Intelligence.
• Numenta public materials on Hawkins’ work and books.