Watch a professional outfielder catch a fly ball. In under four seconds they read it off the bat, break into a run, and arrive exactly where it lands. To do the same thing on paper you'd need projectile motion, air resistance, and a differential equation. The player solves none of it. They couldn't write down the math if you asked — and it wouldn't help if they could, because System 2, the slow deliberate mind, is far too slow to catch anything. The catching is done by System 1: fast, automatic, and completely wordless.
This is what the philosopher Michael Polanyi meant when he wrote, "We can know more than we can tell." The deepest knowledge an expert has is tacit — lived in the body, impossible to fully spell out. The catch looks like instinct. It is, in fact, one of the most computationally demanding things a human ever does.
Now run it in reverse
Demis Hassabis, who leads Google DeepMind, has posed a sharp thought experiment: suppose you trained an AI on nothing but written text — no photos, no video, no falling objects. Would it know that a dropped glass shatters? Intuition says of course not; it has never seen anything. Modern language models say otherwise. In their training text, "dropped," "glass," and "floor" sit right next to "shattered," "shards," and "swept up." The statistical shape of our language turns out to mirror the physical shape of the world. The model builds a usable map of reality out of words alone.
That's the baseball player inverted. The outfielder has deep physical knowledge they can't put into words. The text-only model has deep physical knowledge it got only from words. In both cases the shock isn't the skill on the surface — it's how much understanding sits underneath, in a form we never expected it to take. (Whether a system with no body truly understands "shatter" is a real and unsettled debate — but it can clearly reason about it.)
The pattern has a name
Both stories point to the same rule, first noticed by the roboticist Hans Moravec: high-level reasoning is cheap for computers; low-level sensorimotor skill is staggeringly expensive. Chess, bar exams, and calculus fell to machines early. Folding laundry, walking over gravel, and gripping an unfamiliar object remain unsolved. Steven Pinker put it bluntly in 1994: "The main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard."
The reason is evolutionary. We spent hundreds of millions of years perfecting perception and movement, so it feels effortless — the enormous computation is simply hidden from us. Abstract reasoning is a thin, recent layer, so it feels like the hard, "smart" work. Machines invert that ranking entirely. What humans and AI share is a large reserve of unknown knowns: things they can do but cannot explain. The catcher doesn't know the calculus; their body knows the trajectory.
Why this matters for your AI plans
Here's the operational point, and it's not philosophical. If human intuition about what's "hard" for intelligence is demonstrably backwards, then it's a terrible instrument for predicting what a model will and won't do for your business. The task you're certain a model can't handle may be trivial for it. The "simple" one may be where it quietly fails.
So stop trying to win the argument about capability, and start running the test. A one-afternoon pilot on your real, messy data settles a debate that a month of meetings can't. Don't benchmark on toy demos — point the model at the actual work you want changed and measure whether the outcome moves. And re-test on a cadence, because "experts say it's years away" has a habit of aging badly: Go, protein folding, and passing the bar were all supposedly decades off, right up until they weren't.
The hard part of AI in a business is rarely the model. It's replacing confident assumptions with cheap experiments that tell you the truth about where it actually helps.
We built an interactive resource to make the paradox click — watch a fielder catch a ball two ways, quiz a text-only model on physics it never saw, and test your own gut against Moravec's paradox on eight tasks.