Gary Marcus

In the [YOU] on AI Field Guide

[YOU] on AI asks what these machines mean for us; Marcus presses the prior question the industry would rather skip—what, precisely, are they, and what are they not? You cannot answer the first honestly without answering the second, and he has spent twenty years insisting on the order of operations. His value to the cycle is calibration. In an industry that runs on belief, he is the standing demand that the emperor's tailoring be inspected before the crown is placed.

His central distinction—between producing impressive outputs and understanding—is the same one the cycle draws between performance and comprehension, and he draws it with a cognitive scientist's precision. A child who says goed behaves "incorrectly," but the error reveals a correct underlying mechanism: a rule, generalized to a case the data never supplied. A system that has memorized a million examples and cannot handle the million-and-first has not learned the rule. Generalization, for Marcus, is close to the essence of intelligence, and it is exactly what the curve-fitters lack.

He also sharpens the cycle's account of the decorrelation of fluency and authority. Marcus has been blunt that you can get "intelligent-looking behavior" for trivial reasons—including simply repeating things memorized from a vast database. The impressiveness then tells you about the size of the database, not the depth of the intelligence, and the trick is to find the question whose answer was not in the database. There, the performance reveals itself. This is the same hazard the cycle warns of, named with falsifiable specificity.

Origin

Marcus was born in 1970 in Baltimore, a precocious programmer who found his way to cognitive science at Hampshire College and then to MIT. His doctoral work concerned over-regularization errors in children—the goeds and breakeds—and what they reveal about the architecture of the mind. From it came his first major book, The Algebraic Mind (2001), which argued that the human mind manipulates symbols and variables the way algebra manipulates x and y, and that the neural networks then fashionable could not, in their standard form, do this. It was a technical book with a long fuse; two decades later it would read like prophecy.

The prophecy detonated in 2022 with "Deep Learning Is Hitting a Wall." The argument was narrower and sharper than the meme suggested: not that progress would stop, but that scaling alone would not solve compositionality, reasoning, and reliable abstraction, and that pouring more resources into the same architecture was approaching diminishing returns. The systems got bigger and still hallucinated; they got more fluent and still failed at arithmetic a child could do; they passed bar exams and then cited cases that did not exist. He had not predicted that AI would stall. He had predicted that one specific bet about how to reach intelligence would come up short.

He has since moved his frame from the laboratory to the legislature without abandoning it. On May 16, 2023, Marcus testified before the United States Senate Judiciary subcommittee, beside the chief executive of OpenAI, telling lawmakers: "We have to stop letting them set all the rules." He called for an independent oversight body modeled on the Food and Drug Administration—with authority to review systems before deployment and recall those that prove dangerous—and for the transparency without which oversight is theater.

Key Ideas

The cognitive scientist's eye. An engineer asks: does it work? Marcus asks the harder question—how does it work, and is the how the same as the how in us? When a model produces a fluent paragraph he wants to know whether the fluency reflects understanding or merely the statistical shadow of understanding cast by billions of human sentences. The discipline that trained him treats behavior as a clue, not a verdict.

Interpolation versus extrapolation. Symbolic systems generalize beyond their training distribution because they encode rules that apply to cases never seen; deep-learning systems excel within their distribution and grow brittle outside it. They are superb interpolators and poor extrapolators—and the real world, of self-driving cars and medical diagnoses, is relentlessly out of distribution. This is the same brittleness Pearl locates with the do-operator, arrived at from the study of minds.

The case for hybrids. Marcus's constructive program is neurosymbolic AI—the marriage of the connectionist tradition, which learns statistical patterns from data, with the symbolic tradition, which reasons over explicit rules and structured knowledge. "We actually need both approaches," he insists. His 2020 paper "The Next Decade in AI" laid out four prerequisites for robust intelligence: a hybrid architecture, large-scale structured knowledge, reasoning mechanisms, and rich cognitive models of the world. None, he argues, falls out of scale automatically.

Reliability over capability. The word that recurs in his work is not "intelligence" but "trust." A capability that works ninety-five percent of the time and fails unpredictably is, for many purposes, a hazard dressed as a tool. Hallucination is structural, not incidental: a system that generates the statistically likely next token has no mechanism for distinguishing true from false, only likely from unlikely. And because pattern-matching has no concept of its own limits, the system produces an answer for the unprecedented case with the same fluency as for the routine one—it does not know what it does not know.

Causation is the line scaling cannot cross. Marcus holds that neural networks, on their own, will never genuinely understand causal relationships—why things happen as they do—no matter the data or compute. Correlation is what statistics learns; causation is what understanding requires. The convergence with Pearl is exact, and Marcus has noted with dry satisfaction that some celebrated recent systems, by bolting reasoning steps and tool use onto neural cores, "accidentally vindicated neurosymbolic AI."