In 1988 the roboticist Hans Moravec put down an observation that has carried his name since: it is comparatively easy to make a machine perform at adult level on intelligence tests or at checkers, and difficult or impossible to give it the skills of a one-year-old at perception and mobility. On its own that sounds like a curiosity; the explanation is the interesting part. Seeing, grasping, balancing, recognising are the product of a billion years of selection — machinery tuned so thoroughly that we do not notice it running, precisely because it is tuned. Arithmetic, logic and chess are a few thousand years old, and the brain executes them slowly, consciously and with mistakes. We call hard whatever evolution has not had time to make easy.
H. Moravec, "Mind Children: The Future of Robot and Human Intelligence", Harvard University Press, 1988, p. 15. The same argument in M. Minsky, "The Society of Mind", 1986, and R. Brooks, "Elephants Don't Play Chess", Robotics and Autonomous Systems 6, 1990.The consequence that matters: the border between human and machine ability is not a line but a jagged edge. A machine's triumph at chess says nothing about its progress at walking, and the reverse. What follows is a battery of paired tasks. In each pair, one task is trivial for you and brutal for a program, and its mirror is trivial for a program and brutal for you.
A ball flies in from the left and vanishes behind the curtain. Click the point where it will hit the ground. Five launches; the score is your average error in pixels.
You recalled no formulas: your visual system extended the trajectory on its own. Infants a few months old are surprised when an object that went behind a screen fails to come out the other side — the expectation that hidden things keep moving is installed before speech is. Classical AI could do the same only by explicit model: mass, velocity, equations of motion, entered by hand. The current working hypothesis in cognitive science is that the brain carries something like a game physics engine and runs it forward without effort.
R. Baillargeon, E. Spelke, S. Wasserman, Cognition 20, 1985 — infants' expectations about hidden objects; P. Battaglia, J. Hamrick, J. Tenenbaum, "Simulation as an engine of physical scene understanding", PNAS 110 (45), 2013 — the internal physics engine. Both little games on this page are illustrations, not measurements: no norms, no controlled conditions. The result is computed in your browser, stays on this page and is sent nowhere.A 20-digit number will appear for three seconds. Hold as much of it as you can and type it back; the score is how many digits you reproduced correctly in a row from the start.
A machine keeps gigabytes verbatim, indefinitely, and would not need the three seconds. Your working memory holds about four items — not digits, even, but the chunks you managed to group the digits into. Seven or eight in a row is the usual ceiling; twenty is outside the hardware. This is not a defect but a budget — the same bottleneck is measured on the page about the limits of a human.
G. Miller, "The Magical Number Seven, Plus or Minus Two", Psychological Review 63, 1956; N. Cowan, Behavioral and Brain Sciences 24, 2001 — the revision to 4±1 chunks. Both little games on this page are illustrations, not measurements: no norms, no controlled conditions. The result is computed in your browser, stays on this page and is sent nowhere.Reverse CAPTCHA is the side where you win: digits assembled from blobs with no contour to detect, illusory squares, shapes running behind bars. A person reads them at a glance and without preparation; a program that has not been trained on exactly this has nothing to hold on to. That is perception — the line item evolution spent its billion years on.
Mental arithmetic is the machine's side: sixty seconds of sums against a processor that does about three billion of them per second. The contest is announced as fair and is not fair for a single millisecond — arithmetic is the thing the machine was built for.
A few pairs for the record. In each line, the left half is what a person does without thinking; the right half is what machines mastered before they could do the left.
One reservation, without which this page would be dishonest in 2026: the edge has moved again. Language models write coherent essays, pass professional exams and do things this site measures elsewhere — and simple physical common sense still comes harder to them than to a child: what will spill, what will fall, what will jam. And the robots still fold laundry slowly and carefully. The Moravec paradox has not been repealed; it has shifted, and the border it draws is as jagged as before. This page was written in 2026 — check what has already gone stale.
Deep Blue — Kasparov, New York, May 1997; DARPA Robotics Challenge Finals, Pomona, June 2015; V. Allis, "A Knowledge-Based Approach of Connect-Four", Vrije Universiteit Amsterdam, 1988 (independently J. D. Allen, 1988); J. Maitin-Shepard, M. Cusumano-Towner, J. Lei, P. Abbeel, IEEE ICRA 2010 — towel folding at about 25 minutes apiece; AlphaGo — Lee Sedol, Seoul, March 2016, 4:1; the phrase "jagged frontier" — F. Dell'Acqua et al., Harvard Business School Working Paper 24-013, 2023.