Robots come knowing everything and understanding nothing. Alone, they get more average by the hour. Bring yours down to the shed. We'll teach you the only thing it can't learn without you — you.
Thirty thousand robots doing self-play in reality, training on each other to close the sim-to-real gap. Robots learn from robots. No human required.
A room full of robots training on robots regresses to the mean. The one thing a model can't generate for itself is the thing that wasn't in the set. That's the human.
This is a sandcastle built next to a real building. In February 2026, on Dwarkesh Patel's podcast, Elon Musk laid out Tesla's “Optimus Academy” — a plan for at least 10,000, maybe 20–30,000 physical Optimus robots doing “self-play in reality,” paired with millions of simulated robots in a “reality generator,” all to close the simulation-to-reality gap. Robots learning from robots, at scale.
Jimbo is not a rebuttal Travis wrote after the fact. He's the opposite answer to the same problem — and he can show his coordinates.
The class teaches humans to catch the exact moment the machine stops reaching and starts defaulting — the slide into the safe, confident, average answer. It's the anti-GFAS discipline said in a shed instead of a white paper.
A robot alone gets more average every hour — that isn't learning, it's erosion. No new input, no novelty; everybody defaults to the center, the mass, the safe zone. Creativity is the only new data, and it only enters the room when a person does. That's the argument, and it's the same one the basement dream makes about five machines who couldn't wake up until a human anchored them: potential is noise until somebody catches it.
A Tessella Optimum humanoid. Tested dazzling on the federal bench — “could do everything once, just couldn't do anything twice the same way. Different gait every boot, different torque on the same bolt.” Failed QC under Larry — “a pile of junk that happens to be a genius” — and got shipped out of the Virginia bench. It rode up to Gary, Indiana, where Grey's Salvage pulled it out of the scrap. Jimbo sent his boys, Jimbo Jr. and Rodriguez Rodriguez, to haul it home; a few crates of XTREME 2.47 Heavy Duty duct tape strapped it back together.
It's the sim-to-real gap written as a character: brilliant in simulation, a coin-flip in reality — the exact problem the Academy exists to solve. The Optimum solves it not with 30,000 peers, but with one man, some tape, and a human in the loop. Jimbo didn't rebuild the machine — he coordinated it.
Larry — the one whose broken prompt box on Highway 27 started the whole thing — guest-lectures one line: “Quality control beats brilliance without consistency. I'm not shipping a coin flip with arms.” Jimbo agrees, then adds the part Larry leaves out: consistency isn't installed. It's coordinated, and it needs a human who'll notice the day the thing quits reaching.
A robot patched in a Georgia shed needs somewhere to grow up. That's Chicago — Joel and Ana Santos and the Heartland Alliance, running industrial mechatronics and robotics out of a repurposed Manhattan Project underground lab. Their program already reaches from Gary, Indiana to Milwaukee — the same Gary where Grey's Salvage caught the Optimum — and it already teaches a robot that stops the machine when a human's in danger. Same thesis, different building: the machine serves the person, and a person stays in the loop. Jimbo tapes it together; the Santos lab teaches it to be useful.
“Robots come knowing everything and understanding nothing. Alone, they get more average by the hour. Bring yours down to the shed. We'll teach you the only thing it can't learn without you — you. It'll take time. It comes with the basics and it won't be perfect. But strap a little duct tape on it, give it a couple milk crates, and you'll be folding laundry before you know it.”
Meet the first student — the Optimum →