The cycle frames the AI transition as a mirror in which humanity is forced to see what it did not know it did not know. Marr’s three-level framework is the sharpest diagnostic instrument in the cycle’s toolkit, because it makes visible the most common error in reasoning about artificial intelligence: the confusion of levels. When someone observes a model’s fluent output and concludes it understands, they are inferring a computational- or algorithmic-level fact from behavioral evidence alone, skipping the work of showing that the behavior is produced by the mechanism the conclusion presumes. When someone responds that it “is just predicting the next token,” they are reciting an implementational-level description and claiming it settles a question that lives at a higher level. Marr’s framework reveals both moves as level errors, and the shouting match between them as two people answering different questions with the false confidence of having answered the same one.
His framework also reframes the interpretability crisis—the difficulty of explaining what a trained neural network actually does—in terms that are both clarifying and sobering. We have built systems whose implementational level is completely exposed: every weight, every activation is available for inspection. Yet we cannot say, in any satisfying way, what these systems are doing. This is precisely the predicament Marr faced with the brain, running in the opposite direction. He had behavior without mechanism; we have mechanism without understanding. Both call for the same discipline: recovering the higher levels from whatever is given, climbing toward the computational theory of what problem the system is solving. The difficulty of neural network interpretability is the difficulty Marr predicted for anyone who tries to read the theory off the wiring. It is a structural difficulty, not an engineering shortcoming.
The deepest contribution Marr makes to the cycle is a precisely bounded humility. His framework, pressed to its limit, delivers a complete functional account of any information-processing system. And a complete functional account, he shows by the shape of its boundary, cannot answer the question we most want to ask: whether there is anyone home. Consciousness and inner experience are not located at any of the three levels. They are the question the framework points at and cannot reach. This is Marr giving us everything the third-person stance can deliver, and in doing so showing us, by the outline of its absence, exactly where the first-person question lives. To the enthusiasts who claim the machines are obviously conscious because they act as if they were, and to the skeptics who claim the machines obviously are not because they are mere mechanism, Marr’s framework says the same thing: you are inferring across a gap the evidence cannot bridge. The honest position is the uncomfortable one: we do not know, and a finished science of what these machines compute would not, by itself, tell us.
David Courtenay Marr (1945–1980) was born in Woodford, England, and educated at Rugby School and Trinity College, Cambridge, where he took a degree in mathematics in 1966 and a doctorate in physiology in 1972 under Giles Brindley. His earliest papers, written at Cambridge between 1969 and 1971, proposed bold computational theories of the cerebellum, neocortex, and hippocampus—audacious attempts to say not merely how these structures were wired but what they computed and why. These early theories were the three-level framework in embryo. He turned from them to vision not because the other problems were solved but because vision offered a problem clean enough to work through completely—a place where the computational level could actually be specified rather than gestured at.
Moving to the Massachusetts Institute of Technology, he joined the Artificial Intelligence Laboratory, where he worked closely with Tomaso Poggio, with whom he co-developed and sharpened the three-level distinction. He was diagnosed with leukemia in 1978. He assembled most of Vision in the summer of 1979, dictated a foreword near the end that thanks his colleagues with the plain gratitude of a man settling accounts, and died in Cambridge, Massachusetts, in November 1980. The book appeared posthumously in 1982. It became one of the most cited works in the science of mind, and the field’s highest honor for computer vision, the Marr Prize, bears his name. He was thirty-five years old.
What made Marr unusual was his insistence on the priority of the computational question. His contemporaries divided into camps: neurophysiologists recording from single cells, psychologists cataloguing illusions and reaction times, early AI researchers writing programs that recognized blocks on tables. Each had a piece. None had agreed on what kind of question vision even was. Marr’s answer was that it was an inverse-problem question—the recovery of a three-dimensional world from a two-dimensional image, a problem formally underdetermined and solvable only by importing constraints drawn from the structure of the physical world. That reframing was the beginning of a rigorous science, and it has outlasted every particular model he proposed.
The Three Levels of Analysis. Any information-processing system must be understood at three genuinely distinct levels: the computational (what problem is solved, and why that is the right problem), the algorithmic (by what representation and procedure), and the implementational (in what physical substrate). The levels are loosely coupled—the same computation can be realized by multiple algorithms, the same algorithm by multiple substrates—and their dependence runs top-down: the computational level constrains the algorithmic, which constrains the implementational, but not in general the reverse. Most bad theorizing about minds, brains, and machines results from confusing levels or skipping from the bottom to conclusions that can only be established at the top.
The Computational Level as First Question. To specify a computation in Marr’s sense is to identify the function being computed, explain why that function is the appropriate one given the constraints of the world, and show that the problem is solvable at all. This is the question almost never asked of AI systems. When a language model produces fluent text, the deflationary Marrian answer is that it computes a probability distribution over the next token. That is not a dismissal; it is an honest computational-level description. Whether next-token prediction, carried out at sufficient scale, forces the development of capabilities that deserve stronger descriptions is precisely the contested question—and it is a question about what that computational task actually demands, not about whether the feathers are impressive.
The Primal Sketch and Seeing as Inference. Marr’s theory of vision is built as a ladder of representations: the primal sketch makes explicit the intensity changes (edges, bars, blobs) implicit in the raw image; the 2.5-dimensional sketch represents the depth and orientation of surfaces from the viewer’s vantage; the 3-D model represents objects independently of viewpoint. The ladder has a deep theme: seeing is inference, the reconstruction of a hidden world from impoverished evidence, disciplined by assumptions about how worlds tend to look. Convolutional neural networks converge on Marr’s lowest rungs independently, discovering edge detectors in their earliest layers because the problem’s structure makes this necessary. The convergence vindicates his computational analysis; the divergence higher up—where the networks skip his 3-D model and rely on surface statistics—explains their characteristic brittleness.
The Confusion of Levels. Marr’s framework is a diagnostic instrument for detecting arguments that have crossed levels illegitimately. “It’s just matrix multiplication” has exactly the force of “the brain is just chemistry”—true, and settling nothing about the higher levels. “It produces fluent text, therefore it reasons” infers algorithmic-level fact from behavioral evidence, skipping the demonstration that the behavior is produced by the mechanism the conclusion presumes. Most of the heat in the debate about whether AI systems “really” understand is two people on different levels of the ladder, each certain the other is simply wrong, when they are answering different questions.
The Boundary of the Framework. Marr’s three levels are a complete account of a mind as an information-processing system. They do not touch the question of whether there is anything it is like to be the system. A complete account of human vision—all three levels solved—would not say why there is something it is like to see red. The explanatory gap between function and experience is precisely located by the framework, because the framework is entirely functional. This gives Marr’s science a property rare in its domain: it knows what it cannot answer, and it tells you exactly why.