Man Didn’t Fly by Building a Better Bird

For four centuries the way to build a flying machine was obvious. Watch a bird. Copy the bird.

The results were not encouraging.

Around the year 1000, a monk named Eilmer strapped wings to his hands and feet, launched himself off the tower of Malmesbury Abbey, covered a respectable two hundred yards, and broke both legs. He was lame for the rest of his life. He spent that life explaining to anyone who asked that it would have worked if he had remembered to fit a tail.

Leonardo drew his ornithopter around 1490: a man in a frame, hauling levers to beat a pair of articulated wings. He was not being eccentric. He was being orthodox. For eight hundred years hopeful men climbed towers with flapping contraptions lashed to their arms and jumped. In 1809 a Swiss watchmaker named Jacob Degen announced that he had flown by flapping. He had. His machine was also tethered to a balloon, which was doing most of the flying and possibly all of it.

The error was not mechanical. It was conceptual. Everyone had assumed flight was a behavior, something a bird does, rather than a physical condition, something a shape achieves in moving air.

George Cayley broke that frame in 1799. He separated the problem into lift, propulsion, and control, and then did the thing no flapper had dared: he held the wing still. A wing that stops moving stops being a bird’s wing and becomes an aerodynamic surface. By 1853 he had put a man into a glider and sent him across Brompton Dale.

But gliders fall. Cayley knew exactly what was missing and said so. Sustained flight would have to wait on a lightweight engine.

It waited fifty years. In 1901 the Wright brothers discovered that the lift tables everyone relied on were wrong, so they built a wind tunnel in the back of their bicycle shop and tested some two hundred wing sections until they had numbers they trusted. Flight was solved on paper. And still they could not fly, because no manufacturer in America would sell them an engine that made eight horsepower and weighed under two hundred pounds. So they built their own: aluminum alloy crankcase, hand-finished by their mechanic Charlie Taylor, about a hundred and eighty pounds, no throttle.

The engine was not the insight. The engine was permission to use the insight.

Now hold that shape against the history of AI.

Geoffrey Hinton was not a computer scientist. He read experimental psychology at Cambridge, having passed through natural sciences, art history, and philosophy first, and spent a year apprenticed to a carpenter before going back. His doctoral supervisor favored symbolic AI and regarded neural networks as a dead end.

Neither Wright brother was trained in aerodynamics either. Wilbur never collected his high school diploma. Orville left short of the credits for one. Neither attended a day of college. Wilbur read his way through his father’s library during three years homebound after a hockey injury, and both learned their engineering in a bicycle shop, where an unstable machine stays up only because a rider controls it.

They went looking for control while everyone else was trying to build an aircraft that would fly itself.

The computer scientists were building flapping machines. Their premise was that thinking is a behavior: observable, decomposable, and therefore writable down. Expert systems. Rule bases. Ever more elaborate programs for mimicking the steps a mind appears to take. Each one more intricate than the last, each demonstrated to considerable excitement, none of them able to leave the ground.

Hinton was after something else. Not a description of thought but a mechanism, a way for a system to learn its own representations from data. Right on paper by 1986. Then a quarter century of near-silence, because the machine to run it did not exist. In 2012 the engine arrived. Hinton and two of his students trained a network on a pair of consumer graphics cards and cut the ImageNet error rate from 26 percent to 15.

Now consider what we actually built.

A 747 does not flap. It does not perch, migrate, or land in a tree. It carries four hundred people at six hundred miles an hour, seven miles up, which no bird has ever done and no bird ever will. Twelve men have walked on the moon. Not one of them got there by flapping.

Then run it the other way.

We have been flying for a hundred and twenty years and we still cannot build a bird. Engineers keep trying, and it is worth noticing how it goes. In 2011 Festo built a robotic herring gull that genuinely flies by flapping. A working ornithopter at last, five centuries after Leonardo, whose natural habitat is the trade show floor. A Stanford lab spent years teaching a drone to grip a branch and succeeded, in a laboratory, on a branch selected in advance.

Meanwhile the actual bird lands on a branch it has never seen, in wind, carrying a stick, which it then threads through a gap the size of a coin and weaves into a structure it builds without plans in a tree it chose itself. In 2022 a bar-tailed godwit flew from Alaska to Tasmania: eight thousand four hundred miles, eleven days, no stopping, no food, no engine, and no navigation system anyone has fully explained.

The 747 will never build a nest. The godwit will never reach the moon.

Two capabilities, each extraordinary, neither reducible to the other. Copying either one is how you end up with neither.

That is the whole lesson for human-AI teaming, and it is not a consolation prize for the humans. AI does things people cannot do. It reads ten thousand documents before lunch, holds a hundred variables in view at once, and finds the pattern nobody thought to look for. People do things AI cannot do, and always will. The mistake is not in choosing sides. The mistake is designing either one to imitate the other.

There is a name for why this cannot be fixed by trying harder. Stephen Wolfram calls it computational irreducibility: for many systems there is no shortcut to the outcome, no formula that lets you skip ahead. The only way to know what the system does is to run it, step by step, and watch.

A model’s route to an answer is like that. It is not a tidy procedure sitting inside the machine, waiting to be extracted, written down, and handed back as instructions. You can watch this for yourself. Ask a model to make one small change to an answer it has just given you, and it does not edit that answer, because there is no answer sitting there to edit. It re-derives the whole thing from the beginning with your new requirement folded in. There is no diff to apply and no procedure to patch. There is only the run.

So when you specify the human method and require the machine to follow it, you are assuming a path that compresses into steps. It doesn’t. You are not managing the work. You are fitting it with a tail.

So the question in redesigning work is not which of our tasks AI can perform. Tasks are the flapping. The question is what outcome we actually need, and then letting each kind of intelligence reach it the way that kind of intelligence reaches things.

Man didn’t fly by building a better bird. And no machine has yet built a nest.