The condor Steve Jobs remembered: the real bicycle for the mind chart

The study exists. S.S. Wilson's article "Bicycle Technology" ran in Scientific American in March 1973, and it contains the chart Jobs was describing: the energy each traveler spends to move a unit of its weight one kilometer, animals and machines plotted side by side. Wilson's article was, of all things, an argument for taking bicycle engineering seriously, and his numbers say what Jobs said they say. The walker is mediocre. The rider beats everything.

Among the entries are salmon, horses, a pigeon, cars, a jet transport, a man walking, and a man on a bicycle.

There is no condor on it.

The most famous bird in the history of computing was added in the telling. It is not on Wilson's chart, and when Scientific American redrew the chart from the original data in 2025, it still was not there. The bird flew in from somewhere else, another paper or another memory. Jobs took a data table and improved it into a fable, told it on camera in 1990 for the documentary Memory and Imagination, and the fable worked so well that people have spent half a century citing a chart they never looked at. Wilson's numbers hold, and the bicycle really wins. Jobs misremembered the loser, not the winner.

Stare at Wilson's chart long enough and it gets stranger. The salmon swims. The horse runs. The pigeon flies and the jet burns kerosene. Every entry moves under its own power, except one. The winner of the entire comparison is the single machine on the list that cannot move at all. A bicycle on its own is furniture. Lean it against a wall and it will wait there until the end of the world.

Jobs was not praising the bicycle. He was praising the rider.

That was the deal behind every tool we ever made. The hammer waited for a hand, and the spreadsheet waited for a question. Even the compiler, the most self-important tool in the building, waited for someone to write the code. No intent, no motion. For the entire history of technology, everything we built was a bicycle leaning against a wall, waiting for a person with somewhere to go.

AI broke that deal. It is the first tool that supplies its own motion. Hand an agent half a sentence and it is gone.

We have watched an agent spend an hour carefully refactoring code nobody asked it to touch, like a houseguest reorganizing your kitchen. We have watched one invent a requirement, implement it well, and then write a tidy test suite proving its own invention correct. These are the same machines that return a week of work in thirty minutes.

Autopilots and thermostats move without us too, but they hold a course somebody set. An agent picks its own route, and it will ride into a wall with the same enthusiasm it rides everywhere else, because it cannot tell the difference. The walls have been measured. In DryRun Security's March 2026 study, 87%[4] of AI-agent pull requests introduced at least one vulnerability. Veracode came at it from the other side in 2025, testing more than 100 models on a fixed set of coding tasks, and AI-generated code introduced a security flaw in 45%[5] of them.

The industry's answer to the wandering, so far, is more pedaling. Bigger models, longer autonomy, agents supervising agents, and every benchmark on earth measuring some fresh version of energy per kilometer. All of it pointed at the one column of the chart the bicycle settled in 1973.

A machine that pedals itself is not short on power. It is short on everything a rider brings: direction, brakes, an odometer that cannot be wound back, and a name that answers when it takes the wrong turn.

This is the problem an AI software factory exists to solve. The frontier labs build better bicycles every quarter, and coding assistants make individual riders faster. Good. The factory is not competing with the bicycle. It is the rider's kit for machines that already move on their own.

Direction comes first, and it stays human. You do not prompt the factory. The factory prompts you: it interviews the people who own the intent, reads the policies, and studies the target system until it is confident. Confidence is not the gate. A person reviews and approves each artifact that comes out of that interview before the first line of code exists.

The brakes are the pace. A run takes hours, sometimes days, on purpose, because in regulated software, where the auditor arrives with DORA or NIS2 in hand, the gate matters more than the sprint. And the odometer is the point of the whole machine: nine artifacts[8] every run, one published bar of 22 checks[7]. The factory does not promise identical code each time. No honest vendor does. It promises the same deliverable set, the same bar, and every step on the record.

Wilson's chart measured one number, the cost of moving a kilometer, and nobody ever asked which kilometer. Nobody had to. The rider decided, so for half a century the question never came up. Now the pedaling is close to free and getting cheaper, and when motion costs nothing, direction becomes the scarce thing.

Jobs was right about the bicycle. The interesting part took fifty years to show up. The miracle was never the frame or the gears. The miracle was a walker, unimpressive on any chart ever drawn, climbing onto the right machine and out-traveling everything with wings, because the walker knew where he was going.

The condor, wherever Jobs found it, never did. It just flew.