Somebody Taught the Robot. Nobody Taught the Kid.

In September 2025, a roboticist named Benjie Holson put on a silver bodysuit and filmed himself doing silly things. He cracked an egg into a frying pan. He wiped a window with paper towels. He wrestled a dog-poop bag off a roll.

He was throwing a gauntlet. Holson had just watched the World Humanoid Robot Games in Beijing and come away flat. As fun as rock-em-sock-em is, he wrote, what people actually care about is robots doing their chores. So he posted his own Olympics: doors, laundry, tool use, fingertips, wet work. Real medals, mailed to the winners. He figured the easy ones might fall by December and the hardest might take a year and a half.

Within months, a robot from Physical Intelligence had taken eleven of the fifteen.

Every story written about this has been about speed. I care about how the machines learned, and the answer sits in Holson’s own post, in a sentence nobody quoted.

Explaining why robots still can’t handle force at the wrist, Holson names the real constraint: the robot can only ever perform as well as the human teleoperating it. That’s the ceiling.

The method that beat his Olympics is learning-from-demonstration. A person straps into a rig, does the task with their own hands, and does it again, and again, hundreds of times, while the machine watches. No code says this is how you spread peanut butter. There is only a formed human being, repeating skilled work in front of an apprentice that never tires and never forgets.

Silicon Valley did not invent a technique. It rediscovered a very old one, spent hundreds of millions on it, and called it a data pipeline. What they built is an apprenticeship. They just built the apprentice out of aluminum, and paid for the Meister’s time without once using the word. We pay for that apprenticeship without blinking when the apprentice is a machine. When the apprentice is a kid on a factory floor, we won’t. The machine’s training goes on the books as an asset. The kid’s goes on as a cost.

That’s the argument.

When we launched Lansing Grand River, I helped build the Operator Support Center. We brought in prototype cars and simulated the trim and chassis lines — bodies you could push by hand, a few stations at a time. Then we walked them through. Each team leader did the installs on those cars himself, writing the steps down as he went and taking pictures of what he had just done. Then he put the steps and the pictures together on a job element sheet.

Wally Vinton had the wiring harness on the CTS. Wally and I laughed constantly about his penmanship. He would protest, and dig out the sheets of James, whose writing was worse. But we got there together, to sheets that worked, sheets you could train a new person with.

Set that next to Physical Intelligence and it’s the same loop. A rig that moves the work past a human at human speed. A skilled person performing the task over and over. Capture. Then a model. We did it with a car you pushed by hand and a camera.

Their demonstrator finishes his hundreds of repetitions exactly as skilled as he started. Everything he transferred went into the model. Wally finished his repetitions as the man who knew the harness — and the job element sheet he wrote was the smaller half of what he’d gained. The document held the sequence. It could not hold what his hands had learned about tension, or what a connector sounds like seated versus almost seated, or which run fights you on a hot day when the vinyl has gone soft. That stayed in Wally. Which is why, for years afterward, when that station went down, nobody went and got the job element sheet. They went and got Wally.

Same loop, opposite direction. One deposits into a machine. The other deposits into a person.

Holson is honest about what his robots still can’t do. He lists four limits: no force feedback at the wrist, fingers that amount to open and close, no sense of touch, and precision of one to three centimeters. Set that list next to the machine already standing on your floor. Call it the H-1. It knows where every one of its two hundred-plus joints is without being told. Each hand has twenty-seven degrees of freedom; the best robotic hand ever built reaches twenty-four. Each hand carries some seventeen thousand touch sensors, reading pressure, texture, temperature and slip. It threads a needle and swings a sledgehammer on the same hardware.

Its eighty-six billion neurons run at roughly an exaFLOP. Frontier, the supercomputer, needs twenty-one megawatts for that. The H-1 does it on twenty watts, a dim light bulb. It has walked reliably for two hundred thousand years. A hundred and fifty million units are operational here.

Every constraint the field is spending billions to engineer around is already solved and already on the payroll.

And we don’t form it.

Here is what forming it costs.

Lukas is the standard: the fully-formed H-1, good for business and good for society. Germany produces him on a ladder eight hundred years old. Roughly seventy-five hundred hours — plant and Berufsschule together, under a certified Ausbilder who passed his own federal examination in how to teach — and Lukas is a Geselle at twenty-three. Then the deepening years. At twenty-five thousand hours he is a Meister at thirty: diagnoses a bearing by sound, feels a fault through a housing, forms the next one.

Twenty-five thousand hours. Set that against the American number, which is about forty hours of onboarding and a wish.

I watched that number land at Lansing Delta Township, when we hired for the third-shift launch. The new hires got the standard forty hours, taught by trainers rather than gray beards off the floor, much of it from material years out of date. Then they were dropped on the line at random. One young woman, very short, landed on the glass line. She could not work the suction handles to carry the glass from the robot’s end effector to the body. She