Moravec's Paradox

Is it really about computation, or about knowledge? Moravec framed it partly in terms of raw processing throughput: the sensorimotor system is doing vastly more operations per second than conscious deliberation. Others locate the difficulty elsewhere.

Marvin Minsky, in The Society of Mind (1986), stressed that the ancient abilities are built from enormous numbers of specialised sub-processes, which is precisely why we cannot introspect on them — there is nothing at the top to introspect on. His conclusion is close to Moravec's but the emphasis is on architecture rather than throughput.

Rodney Brooks made the strongest methodological claim. In "Elephants Don't Play Chess" (1990) and related work he argued that the symbolic AI tradition had mistaken the visible surface of intelligence for its substance, and that the right way forward was to build systems from perception and action upward, with no central world model at all. His subsumption architecture robots walked over rough ground with almost no computation, which was itself a partial argument against the pure-throughput reading.

Steven Pinker, in The Language Instinct (1994), gave the version most people know: the hard problems of AI turned out to be easy and the easy ones hard, and as a result the well-paid analytical jobs are more exposed to automation than gardening, cooking or reception work.

The honest caveat about data. A significant part of the difficulty in robotics may be economic rather than fundamental. Text and images exist in quantities of trillions of tokens, essentially for free. Physical interaction data has to be generated one robot-hour at a time, on hardware that breaks. On this reading, manipulation is not intrinsically harder — it is data-starved, and the gap will close as simulation and fleet learning improve. Several research groups are betting substantial resources on exactly this. It is a live disagreement, and it does not overturn Moravec's evolutionary argument so much as supply a possible route around it.

Where the paradox is weakest. The claim that abstract reasoning is "computationally cheap" has aged awkwardly. Training a frontier language model consumes an amount of computation that would have been unimaginable in 1988, and the tasks it does are exactly the abstract ones. What survives, and survives robustly, is the relative claim: for a fixed budget of engineering effort and compute, you get further on reasoning-shaped tasks than on physical ones.

1986. Marvin Minsky published The Society of Mind, arguing that the mental activities we consider easy draw on the largest and oldest machinery, and are therefore the ones we understand least.

1987–1988. Rodney Brooks's behaviour-based robotics group at MIT built insect-like walking robots that navigated real, unmodified environments using layered reflex-like behaviours and almost no representation — a direct challenge to the assumption that intelligence flows downward from reasoning.

1988. Hans Moravec published Mind Children: The Future of Robot and Human Intelligence with Harvard University Press. It contains the statement usually quoted: that it is comparatively easy to make computers perform at adult level on intelligence tests or at checkers, and difficult or impossible to give them the perception and mobility of a one-year-old. Moravec also gave the evolutionary explanation, and the retina-based estimate of how much computation the visual system represents.

1990. Brooks published "Elephants Don't Play Chess", the manifesto version of his programme.

1994. Steven Pinker restated the point in The Language Instinct, which is where most people encountered it, and where the name "Moravec's paradox" began to attach firmly.

A note on naming. Moravec did not call it a paradox, and did not name it after himself. The label was applied later by others, and the credit is genuinely shared: Moravec, Brooks, Minsky and Pinker each arrived at the same conclusion from robotics, from cognitive architecture and from psycholinguistics respectively, within about a decade.

It redirected a research field. The first three decades of AI concentrated on theorem provers, chess programs and expert systems, on the reasonable assumption that hard-for-humans meant hard-for-machines. Moravec's argument, and Brooks's robots, made the case that perception and control were the substantial problems. The subsequent history — computer vision and robotics absorbing an ever larger share of the field — followed that reweighting.

It predicted which jobs automate, and it got it broadly right. Pinker's corollary has held up remarkably well. Bookkeeping, translation drafting, legal document review, radiological image screening and routine code generation have all been substantially automated or assisted. Plumbing, elder care, cleaning, electrical work and cooking in an unfamiliar kitchen have not. This is not because society values the second group more; it is because the second group is computationally harder.

It explains why introspection is a bad guide to engineering. Asking an expert how they do something gets you the small, recent, verbalisable part of their skill. Michael Polanyi's observation that we know more than we can tell is the same fact from the other side, and it is why decades of effort at building expert systems by interviewing experts produced systems that were brittle in exactly the places the experts had nothing to say.

The last few years have been an unusually clean test, and the result deserves stating carefully.

The abstract side moved enormously. Large language models write competent prose, produce working code, score highly on professional and graduate-level examinations, and handle mathematical competition problems. In 1988 every one of those would have been described as requiring general intelligence. They now run on commodity hardware for fractions of a cent.

The physical side moved much less. General-purpose robotic manipulation in unstructured environments remains unsolved. Robots that fold laundry, load arbitrary dishwashers or tidy a child's bedroom exist as research demonstrations and remain slow, fragile and narrow. Industrial robots are extremely capable inside environments engineered around them, which is precisely Moravec's point — the difficulty is the world's variability, not the motion.

This vindicates Moravec rather than refuting him. It is worth being explicit, because the opposite reading circulates. His claim was never that abstract tasks would remain hard; it was that abstract tasks are the cheap ones and sensorimotor tasks are the expensive ones. A world in which machines write passable legal briefs before they can reliably pick up a sock is the world he described.

Two honest qualifications.

First, the ordering within the physical domain has shifted. Perception — object recognition, segmentation, depth estimation, speech in noise — has largely fallen to deep learning. What remains stubbornly hard is contact-rich manipulation: deformable objects, tight tolerances, force control, and recovering from a grasp that is going wrong. The paradox has narrowed onto that.

Second, the "little computation" half is now false in absolute terms. Frontier model training runs consume vast resources. The defensible modern version is comparative: progress per unit of investment has been far faster on abstract tasks than on embodied ones, and the reason is the one Moravec gave — for abstract tasks we are competing against a few thousand years of human practice, and for embodied ones against half a billion years of evolution.

Practical consequence. Anyone forecasting automation should stop asking how intellectually demanding a job looks and start asking how much of it consists of unstructured physical interaction with an uncooperative world. That single question predicts the last decade better than any measure of educational requirement.