Chess Fell Before Laundry: Why Contact and Uncertainty, Not Hardware, Limit Manipulation — Epoche C2
In 1997 a machine beat the reigning world chess champion. Decades later, no robot folds laundry reliably in an ordinary home. The usual reading of that gap is that the hardware is not yet good enough. Moravec (1988) offered a different reading, and the difference matters, because it changes which engineering problem is worth working on. This essay sets out his account, separates the part of it that carries weight from the part that does not, and then gives the two results — one about contact, one about observation — that actually explain why the towel resists. The claim, and the causal chain Moravec proposed Moravec's observation in Mind Children is that it is comparatively easy to make computers perform at adult level on intelligence tests or at checkers, and hard to give them the perception and mobility of a one-year-old. He wrote it in 1988, nine years before Deep Blue, so the chess half is a prediction rather than a rationalisation. His explanation is evolutionary. Sensorimotor competence — seeing, balancing, reaching, grasping — has been under selection since the first mobile nervous systems, roughly $5\times10^{8}$ years. Explicit symbolic reasoning is recent: at most $10^{5}$ years if dated to behavioural modernity. The ratio $5\times10^{8}/10^{5} = 5\times10^{3}$ is between three and four orders of magnitude, and Moravec reads it as that much more optimisation. The causal chain from the ratio to the felt difficulty runs in four steps, and each is a separate claim. Long optimisation produces dedicated, massively parallel circuitry. Dedicated circuitry does not report its intermediate steps, because reporting them would cost time and serve no function the organism has. What cannot be introspected cannot be written down as rules. Therefore the tasks whose algorithms are hidden feel effortless while the tasks we can state in rules feel hard, and we mistake the absence of felt effort for the absence of computation. The third and fourth steps are the ones with independent support. Felleman and Van Essen (1991) assembled the macaque visual cortex from the anatomical literature into a single map: 32 distinct visual areas, 305 identified pathways among them, arranged into roughly ten hierarchical levels, occupying on the order of half the neocortex. All of it runs while the subject reports seeing a scene and nothing else. That is a measurement of how much machinery introspection does not reach, and it is the closest thing Moravec's argument has to a test. The compute estimate, and how much weight it will bear Moravec attached a number to the hidden machinery, and it is worth reconstructing because every factor in it is a separate assumption. The retina resolves about $10^{6}$ points, which is the order of the optic nerve's fibre count. It delivers roughly 10 detections of edge and motion per second at each point, which is set by the temporal resolution of the retinal circuitry rather than by anything about the image. He costs a single detection at about $10^{2}$ elementary operations, the price of a small convolution and a threshold. The product is $$10^{6}\times 10\times 10^{2} = 10^{9}\ \text{operations per second}$$ for the retina. He then scales to the whole brain by mass: retina about $0.02$ g against a brain of about $1500$ g, a ratio of $7.5\times10^{4}$, giving $7.5\times10^{4}\times10^{9} = 7.5\times10^{13}$, of order $10^{14}$ operations per second. The weak link is that last step, and the evidence against it is quantitative. Azevedo and colleagues (2009) counted the cells of the human brain by isotropic fractionation and found about 86 billion neurons in total, of which about 69 billion sit in the cerebellum, a structure of roughly 154 g. That is about 450 million neurons per gram, against about 13 million per gram in the cerebral cortex, which holds 16 billion neurons in about 1233 g — a factor of 34 between two parts of the same organ. Computation per gram is not remotely uniform, so a mass ratio is not a computation ratio. The figure is an order-of-magnitude gesture, and nothing in what follows depends on it. What the evolutionary explanation buys, and what it does not Before turning to the two tasks, one honest accounting is owed, because the essay's own thesis does not need Moravec's premise and is stronger without it. The evolutionary story predicts a ranking of engineering difficulty by evolutionary age. But the age ranking was read off the difficulty ranking rather than measured against it: nobody dated the capacities first and then predicted which would resist. Its one genuinely independent consequence is the introspective-opacity claim, and that is the one Felleman and Van Essen support. Nor does the time ratio measure what it is asked to measure. Cumulative selection depends on generation counts, population sizes and selection intensities, none of which the ratio of elapsed years captures, and the two capacities differ in all three. Three to four orders of magnitude of time is not three to four orders of magnitude of optimisation . The ranking has also shifted since 1988 in a way that narrows the claim. Perception, which Moravec placed squarely on the hard side, has largely yielded — and it yielded by the route his account predicts, through learning from data rather than through anyone writing the rules down. What has not yielded is a smaller and more specific residue: manipulation in contact with objects whose state is not known. That residue has an explanation which owes nothing to evolutionary history, and the rest of this essay gives it. Moravec's paradox survives as an excellent description of the asymmetry and a poor test of its cause. Why chess yielded Chess is finite, discrete, fully observed and deterministic: given a position and a legal move, the successor position follows with certainty and can be written down exactly. Its combinatorics are enormous — Shannon (1950) estimated about $10^{43}$ legal positions — but they are the kind of enormous a machine ca