Key Takeaways
Define intelligence by what a system predicts, not merely what it does.
Behavior is only the surface. Hawkins argues that intelligent behavior alone cannot establish understanding. A program can produce convincing answers by following rules without building any grasp of what those answers mean. His examples include Deep Blue, which beat chess champion Garry Kasparov by searching enormous numbers of moves, and the Chinese Room thought experiment, in which a person follows instructions to answer questions in a language he does not understand.
Prediction offers another test. In Hawkins’s account, a brain understands because it uses memories to anticipate what it will sense next. A familiar room feels intelligible partly because its objects and relationships match expectations; an unexpected change draws attention. Intelligence, on this view, is an internal capacity to model patterns and anticipate them, whether or not the system is visibly acting.
The distinction between performance and understanding remains important in debates about machine intelligence. A system can succeed on a narrow benchmark while failing when context shifts, much as Deep Blue’s chess skill did not generalize into broad knowledge. Prediction is a useful alternative lens because it asks what internal structure supports a response. Still, successful prediction does not by itself settle whether a system understands: a weather model predicts patterns without necessarily having human-like comprehension. The framework sharpens the question, but the answer depends on what counts as understanding.
Treat the cortex as learned memory, not a faster conventional computer.
Speed favors retrieval. Hawkins compares a neuron, which may complete roughly 200 cycles per second, with a computer capable of vastly more operations. He argues that a person can recognize a cat in a picture in about half a second, giving the brain time for a chain of only around 100 neural steps. A computer may need extensive calculation to tackle the same recognition problem.
Experience supplies the shortcut. When someone catches a ball, the brain does not solve a fresh set of physics equations from scratch. It recalls learned sequences of movement and adjusts them to the present throw. Hawkins’s claim is that the cortex mainly retrieves and adapts learned patterns, rather than calculating every answer through a central processor. Its memories are distributed across connections, not stored like files in a conventional computer.
This account highlights a powerful difference between explicit calculation and learned competence. Skilled action often looks effortless because practice has compressed a problem into rapid recognition and response. That idea connects naturally to expertise research, where experienced people notice meaningful patterns that novices miss. Yet the contrast should not be made absolute: brains also calculate, and modern computers can learn patterns rather than rely only on hand-written rules. The book’s lasting challenge is not that computation is useless, but that intelligence cannot be reduced to a fast sequence of explicit instructions.
Remember that brains store sequences and recover wholes from fragments.
Memory unfolds through time. The cortex stores sequences of patterns, Hawkins proposes, because experience arrives in changing streams. A melody is recalled note by note, and a familiar route comes back as one landmark leads to the next. He calls a memory auto-associative when part of a stored pattern can bring back the larger pattern. A few notes can evoke the rest of a song; a partial view of a person can make the whole figure seem present.
Memory abstracts from variation. An invariant representation is a stable memory of what matters across changing inputs. A familiar face remains recognizable as it shifts in size, angle, expression, and lighting. Likewise, a melody can still be recognized when played in a different key. In Hawkins’s account, the brain stores relationships and sequences, not a perfect snapshot of each experience.
These examples make memory feel less like an archive and more like a living process. Reconstruction explains why partial cues can bring a whole scene to mind, and why perception is robust when sensory information is incomplete. It also suggests a source of error: the same machinery that fills gaps can confidently supply details that were never present. Research on memory distortion supports treating recall as useful reconstruction rather than exact playback. The framework’s strength is explaining both flexibility and fallibility through the same mechanism.
Look for one adaptable cortical process beneath many specialized abilities.
A shared structure may do different jobs. Neuroscientist Vernon Mountcastle observed that regions handling vision, hearing, touch, and movement share a broadly similar six-layered structure. Hawkins builds on this observation: perhaps cortical regions perform a common operation, while their connections and incoming patterns determine what each region learns. The hierarchy matters because the world itself contains nested patterns, from notes to melodies, or parts of a face to a whole person.
Experience tunes the system. The book points to striking flexibility: ferret brain tissue connected to visual input can develop visual functions in regions usually associated with hearing, and blind people can recruit visual cortex for processing touch. Hawkins uses such findings to argue that cortex is not born with a separate fixed algorithm for every task. The proposed common cortical algorithm is a theory, not a settled map of every brain function.
The shared-algorithm proposal is an elegant way to turn anatomical resemblance into a research question. It also fits the broad principle of plasticity: neural systems can be shaped by the information they receive. The idea cautions against assuming that a named brain area has one immutable job. At the same time, common structure does not prove identical computation. Specialized connections, local cell types, and the brain’s other regions may matter substantially. The strongest version of the claim is that common principles coexist with specialization, not that every cortical area is interchangeable.
Expect perception to be a negotiated guess, checked against incoming evidence.
The brain anticipates ordinary life. Hawkins describes the brain as continuously comparing incoming patterns with predictions drawn from memory. On a familiar walk through a home, the hand expects the door handle, the body expects the floor, and the eyes expect the room’s layout. A changed door, such as one with a moved handle or unexpected weight, violates these expectations and captures attention. The brain notices the mismatch before it needs to explain it.
Perception combines signal and expectation. The book uses the eye’s blind spot as an example: people do not usually see a hole in the visual field because the brain fills in the missing texture from surrounding patterns. This filling-in is useful, but it also means experience is not a literal copy of sensory input. In Hawkins’s model, errors matter because an unexpected pattern prompts the brain to revise its interpretation.
This account anticipates a central idea in contemporary perception research: the brain does not passively receive a complete picture, but combines sensory evidence with expectations. The altered-door example makes the principle tangible, while the blind spot shows that construction is not an occasional illusion. The model also clarifies attention as a response to surprise. A useful caution is that prediction errors do not always produce conscious correction; familiar assumptions can persist despite contradictory evidence. Perception is an ongoing negotiation, not a guarantee that the best interpretation wins.
Use prediction as the bridge between thought, movement, and imagination.
Action changes what comes next. Looking, reaching, and walking alter the patterns the senses receive. Hawkins therefore treats movement and perception as closely linked: the brain anticipates the sensory consequences of an action, then uses those expectations to guide movement. To walk from one room to another, the brain recalls a learned route and prepares for the sights, sounds, and physical sensations associated with it.
Imagination rehearses consequences. Hawkins proposes that the brain can turn predicted patterns into input for further thought. A chess player can mentally examine a move, picture the changed board, and anticipate an opponent’s response without moving a piece. The same basic predictive process supports planning and mental rehearsal. In this account, imagining is not a separate mental faculty so much as the internal exploration of what a sequence of events might produce.
The link between perception and action is useful well beyond neuroscience. In motor learning, people improve by rehearsing movements and anticipating their consequences; in planning, imagined scenarios help compare possible choices before acting. The account also resembles embodied approaches to cognition, which emphasize that thought is shaped by bodily capacities and interaction with the environment. Hawkins’s claim that prediction can drive movement is ambitious, however. Sensory feedback, motivation, and control systems also contribute. The framework is strongest as an explanation of their coordination, not as a complete account of action.
Build creativity by learning patterns, then testing surprising analogies.
Creativity extends what is known. Hawkins defines creative thought as applying learned patterns to a new situation by analogy. Playing a melody on an unfamiliar instrument works because the brain recognizes a relationship between the new instrument and a familiar one. At a higher level, a mathematician may notice that a difficult equation resembles a previously solved problem and adapt the earlier method.
Practice enlarges the pattern library. Repeated exposure helps people detect subtler relationships, which Hawkins associates with expertise. He also offers a warning: analogy can mislead. Johannes Kepler once tried to explain planetary orbits through nested Platonic solids, a beautiful pattern that did not fit the better measurements available later. Creativity benefits from connecting ideas, but those connections need testing against evidence.
This view makes creativity less mystical and more trainable: a person needs a rich store of experience and the flexibility to compare it with new problems. It also explains why expertise can produce originality, not just efficiency. The caution about Kepler is essential because human pattern-finding readily detects structure, including structure that is not there. In scientific practice, analogy is best treated as a generator of hypotheses rather than proof. The useful habit is to ask both what two situations share and where the comparison breaks.
Treat consciousness as tied to memory, while keeping the hard question open.
Remembering shapes conscious identity. Hawkins suggests that everyday consciousness is closely linked to forming declarative memories, meaning memories a person can later recall and describe. In a thought experiment, someone whose memories of a day are erased afterward might deny having been conscious during it, even though others witnessed their participation. The example separates the experience of an event from the later ability to remember it.
Some questions remain unresolved. Hawkins does not claim to explain qualia, the felt quality of seeing red or hearing music. He speculates that differences in sensory pathways or in the patterns they carry may help explain why senses feel different, but acknowledges that the issue remains difficult. The book also argues that culture and early experience shape the brain’s model of the world, making skepticism important when familiar expectations become stereotypes.
Connecting consciousness to memory helps explain why autobiographical continuity matters to a person’s sense of self. It also fits clinical cases in which memory loss changes how people narrate their lives. But memory cannot be the whole story: an experience can be conscious even when it is not stored for later report, and sleep or anesthesia complicate simple definitions. Hawkins is more persuasive when distinguishing reportable awareness from the unresolved subjective quality of experience. The cultural point adds another caution: a brain’s expectations can feel like reality even when they reflect local learning.
Design intelligent machines around learning and senses, not humanoid imitation.
Build a model of a world. Hawkins’s proposed recipe for intelligent machines is to connect sensors to a hierarchical memory system, then train it through repeated experience rather than programming in every rule. The machine need not have human eyes, ears, or a body. A system monitoring weather sensors, for example, could learn recurring local and large-scale patterns in the same way a brain learns relationships among sensory inputs.
Intelligence need not look human. The book imagines machines that could process information much faster, store more, be copied after training, or use sensors for phenomena humans cannot directly perceive. Hawkins estimates that a cortex-sized memory would have required roughly 80 hard drives available at the time of writing, while noting that the exact numbers were rough. He argues that intelligence alone does not imply human emotions, self-replication, or hostile motives.
The proposal usefully shifts the engineering target from a talking robot to a system that learns a structured environment. This opens applications in fields where people struggle to notice patterns across enormous or unfamiliar data streams. It also separates intelligence from human drives, a distinction that can reduce science-fiction confusion. Still, capability and safety cannot be cleanly separated: a system without emotions may cause harm through flawed objectives, bad data, or excessive trust. The book’s confidence about low danger deserves more caution than its case for nonhuman forms of intelligence.
Judge machine intelligence by useful prediction, while auditing the model's blind spots.
The theory invites experiments. Hawkins offers testable proposals, including that some cortical cells should become active before an expected sensory event, and that unexpected patterns should travel farther up the cortical hierarchy than familiar ones. Such claims make the framework more than a metaphor: experiments could support or weaken particular parts of the proposed mechanism.
Separate principles from machinery. The broad idea is that memory, hierarchy, and prediction help explain perception and learning. The detailed account of cortical layers, cell types, and feedback pathways is more speculative. Hawkins himself presents the book as an incomplete theory and expects specific mechanisms to need revision. For readers evaluating the vision for intelligent machines, this distinction matters: a powerful organizing idea can be worth exploring without every proposed biological detail being established.
A scientific framework earns credibility not from its elegance alone but from making risky predictions that could fail. Hawkins’s appendix gives the theory a valuable empirical orientation, while the book’s own caveat about incomplete mechanisms encourages intellectual restraint. This is a productive way to read ambitious science: preserve the question-generating insight, but do not confuse an explanatory story with confirmed neural circuitry. The same standard applies to machine designs. A system should be judged by its behavior across unfamiliar conditions, not by how closely its architecture resembles a biological diagram.
Analysis
On Intelligence offers a compelling shift in the question asked about minds. Rather than asking what an intelligent system does, Hawkins asks what it has learned to expect. That reframing makes perception, memory, movement, and creativity parts of one explanatory story: a system builds a model from patterned experience, uses it to anticipate what comes next, and notices when reality departs from expectation. The story is memorable because it links mundane acts, such as opening a familiar door, to broad claims about cognition and machine design.
Its strongest contribution is conceptual integration. The book brings time, feedback, and hierarchical organization into a discussion that Hawkins believes was dominated by static inputs and outputs. It also insists that failures of AI are not simply failures of computing power. A program that performs a narrow task may lack the flexible world model needed to interpret new contexts. This remains a useful distinction when evaluating impressive benchmark results.
The central thesis resonates with predictive-processing research, which also treats perception as an interaction between expectation and sensory evidence. That convergence lends the general question weight, but does not validate every proposed circuit. Hawkins’s account of cortical columns, layer-specific functions, sequence labels, and thalamic feedback is a proposed mechanism, often presented as an explanatory model rather than established consensus. The book is most valuable when its claims are kept at two levels: a broad principle that brains exploit learned regularities, and a detailed biological hypothesis that requires testing.
There is also a philosophical limit. Prediction is a powerful measure of a system’s model, but successful prediction does not automatically establish understanding or subjective experience. Conversely, human understanding is itself inferred partly from behavior, even when internal experience is central to the question. The book productively unsettles easy definitions without fully resolving them. Its lasting challenge is practical as much as philosophical: build systems that learn relationships, test them on unfamiliar situations, and distinguish robust understanding from a convincing performance on familiar ground.
Review Summary
On Intelligence presents a novel theory of brain function and cognition, proposing that intelligence stems from the neocortex's ability to make predictions based on stored memories. While many readers found Hawkins' ideas thought-provoking and potentially groundbreaking, some criticized his overconfidence and lack of scientific rigor. The book's accessible writing style and ambitious scope were widely praised, though its philosophical musings on consciousness and creativity received mixed reviews. Overall, readers appreciated Hawkins' attempt to unify neuroscience and artificial intelligence, even if they disagreed with some of his conclusions.
People Also Read
Glossary
memory-prediction framework
Intelligence through learned predictionHawkins’s theory that the neocortex stores patterns from experience, organizes them into memories, and uses those memories to anticipate future sensory and behavioral patterns. Intelligence is framed as the ability to build a model of the world and predict what will happen, rather than as the production of outwardly intelligent behavior alone.
invariant representation
Stable memory across variationA representation of an object or pattern that remains recognizable despite changes in the sensory input. A person can recognize the same face from different angles, distances, or lighting, and identify a melody in different keys. Hawkins proposes that the cortex stores relationships that remain stable across such changes, rather than exact copies of individual experiences.
auto-associative memory
Whole recalled from a fragmentA memory system that can retrieve a stored pattern when given only part of it or a distorted version. In Hawkins’s account, such recall can apply to spatial patterns and to sequences: a fragment of a melody may evoke the rest, just as a partial view can evoke a whole object. He treats this capability as central to cortical memory.
cortical hierarchy
Nested levels of cortical processingThe organization of cortical regions into levels that pass information both upward and downward. Lower levels handle relatively specific, rapidly changing patterns; higher levels combine them into more stable and abstract representations. Hawkins argues that this arrangement reflects the nested structure of the world, such as features forming objects and words forming sentences.
cortical algorithm
Shared cortical learning processHawkins’s proposed common operation across regions of the neocortex. In his account, each region classifies its inputs, learns recurring sequences, forms stable representations, and uses them to predict future patterns. The senses and tasks differ because regions receive different inputs and connections, while the underlying process is broadly shared.
one-hundred-step rule
A limit on neural processing depthHawkins’s thought experiment based on the relatively slow firing rate of neurons. If a person recognizes an image and responds in about half a second, the information can pass through only around 100 sequential neural steps. He uses this to argue that rapid human recognition depends substantially on retrieving learned patterns, rather than calculating a solution from scratch.
declarative memories
Memories that can be reportedMemories that a person can consciously recall and describe, such as where they went or what happened during an event. Hawkins links this ability to the everyday sense of being conscious, distinguishing it from learned skills such as balancing a bicycle, which can be performed without being readily explained in words.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.