The cycle that began with [YOU] on AI asks what it means to take the orange pill—to see the machine clearly, without the narcotic of hype. Hebb is the guide for looking at the machine from the bottom up: from the single connection to the distributed pattern to the sequential activation to the question of what, if anything, learning is. He invented a phrase to keep himself honest about the difference between a model of the brain and the brain itself: the conceptual nervous system, an explicit reminder that his theory was a useful simplification, a map rather than the territory. That discipline—naming the idealization so as not to forget its limits—is exactly what the present discourse around AI lacks and most needs. The word "neural" in "neural network" is doing exactly what Hebb's word "conceptual" was meant to flag: marking a relationship to biology that is real in its functional principles and radically attenuated in its substance.
The most practically consequential thing Hebb offers the cycle is a precise account of what the popular paraphrase of his rule forgets. "Fire together, wire together" is the motto of a correlational learner—a system that absorbs co-occurrence statistics without any notion of contribution or direction. Hebb's actual postulate, with its requirement that A take part in firing B, contains the rudiments of a causal learning rule, and the difference between correlation and causation is the deepest known limitation of current AI. A large language model trained to predict the next token learns what tends to occur with what, with extraordinary power and precision. It does not thereby learn what causes what. It knows that rain and wet ground co-occur; it does not know that rain causes wet ground rather than the reverse. The field traded Hebb's causal subtlety for correlational scale, and the trade bought spectacular fluency at the cost of a persistent inability to reliably distinguish cause from coincidence.
Hebb's two-trace theory of memory—short-term memory as reverberating activity, long-term memory as structural change—names a split that the field is still trying to close. A deployed model's long-term memory is its weights, frozen after training, exactly analogous to Hebb's consolidated synaptic change. Its short-term memory is the activity in its context window, transient and session-bounded, exactly analogous to Hebb's reverberating circuit. The difficulty that the field struggles with—that a model cannot easily form new long-term memories from its short-term experience, cannot consolidate what happens in a conversation into lasting learning—is, in Hebb's terms, the failure of the bridge between the two traces. He named the architecture of memory that AI is still trying to complete.
Donald Olding Hebb was born in 1904 in Chester, Nova Scotia, to two physician parents. His first ambition was literary; he took his bachelor's degree at Dalhousie in 1925 intending to be a writer, and drifted toward psychology almost by accident, through Freud and William James. After a master's degree at McGill in 1932, he went to study under Karl Lashley, first at Chicago and then at Harvard, where he completed his doctorate in 1936. Lashley was the great skeptic of his generation—the man who had spent years cutting pieces out of rat brains in search of the physical location of memory and had failed to find it anywhere in particular, concluding that memory was distributed across the cortex rather than stored in a single spot. This failure was the gift Lashley handed Hebb. If memory is not in one place, it must be in the pattern, in the organization of many cells acting together. The whole of Hebb's later theory is an answer to the puzzle Lashley posed and could not solve.
His book, published in 1949 after roughly seventeen years of thinking and rewriting, is unusually well written for a work of theoretical neuroscience—argued in prose rather than equations, built like a long essay following one question relentlessly to its end. The novelist's instinct for a through-line survived the conversion to science. Hebb spent most of his subsequent career at McGill, where he rose to chair the psychology department, and he died in 1985, living just long enough to see the neural network revival of the 1980s begin to vindicate the principles he had proposed in the age before computers had enough power to run them. He would have been characteristically dry about the vindication, and characteristically careful about the overreach.
His legacy divides cleanly between the empirical and the conceptual. On the empirical side, his enriched-environment experiments—raising rats in his home rather than in laboratory cages and finding lasting gains in problem-solving ability—became part of the intellectual foundation for early-childhood intervention programs and remain standard references in developmental psychology. On the conceptual side, the Hebbian postulate, the cell assembly, and the reverberating circuit have exercised a persistent influence on neuroscience and AI that has outlasted every subsequent account of the same phenomena, because the ideas were right in the ways that matter even when they were imprecise in the ways that neuroscience requires.
The Hebbian postulate. When a cell repeatedly takes part in firing another, some physical change grows between them that makes it better at that task. The key word is "takes part in firing"—not "fires simultaneously with." The directionality is the substance. Cell A must come before and contribute to B's firing; simultaneity is not enough. This distinction between contribution and coincidence is the distinction between learning causal structure and learning mere correlation, and it is the fault line that runs through all of modern AI. The popular paraphrase erases the distinction; reading Hebb correctly restores it. Neural networks trained by backpropagation are not strictly Hebbian—they use error signals rather than correlation—but the founding concept of learning by connection adjustment is his.
The cell assembly. No single neuron means anything; meaning lives in groups. Repeated experience welds certain sets of neurons into tightly interconnected assemblies that tend to fire as a unit. Activate enough members and the whole ignites. A thought is the activation of an assembly; a train of thought is a phase sequence of assemblies igniting one another. This is the direct conceptual ancestor of the distributed representation, the embedding, that lies at the core of every modern neural network: the meaning of a word or concept encoded not in a single unit but as a pattern spread across thousands. The properties that make distributed representation powerful—graceful degradation, natural generalization, vast capacity—are precisely the properties Hebb attributed to the cell assembly.
The reverberating circuit and two-trace memory. Activity persists because loops let it return to its starting point. Short-term memory is this ongoing electrical process; long-term memory is the structural change it induces if it continues long enough. The split between dynamic, activity-based short-term memory and durable, structure-based long-term memory is now textbook neuroscience. Its AI mapping is the split between the context window (transient activity, session-bounded) and the weights (frozen structural residue of training). The model cannot consolidate context into weights the way a brain consolidates reverberation into structure, and Hebb's dual-trace picture names precisely what is missing.
Nature and nurture as jointly necessary. Heredity and environment determine behavior the way length and width determine the area of a field—not as competing percentages that sum to a hundred, but as jointly necessary factors neither of which produces anything alone. The question "which matters more" is malformed, a category error. In AI terms: architecture plus training, not architecture versus training. An architecture without data is random noise; data without an architecture to absorb it is an inert pile. The question of whether a model's capabilities come from its design or its training repeats the nature-nurture confusion Hebb corrected in 1949.
The conceptual nervous system. Hebb coined this phrase to keep himself honest about the difference between his theory and the brain itself. His model captured functional principles while omitting the staggering complexity of actual tissue; the naming was an act of intellectual integrity, a reminder built into the vocabulary that the theory is a map and not the territory. The word "neural" in "neural network" is doing exactly what "conceptual" was meant to flag: a relationship to biology that is real in its functional principles and radically attenuated in its substance. Confusing the two directions—inferring that neural networks are brain-like in properties they do not have, or inferring that they cannot think because they are merely physical—is the error Hebb's phrase was designed to prevent.