Turks All the Way Down

In 1770, at the court of Empress Maria Theresa in Vienna, a Hungarian civil servant named Wolfgang von Kempelen unveiled a machine that played chess. A wooden figure in Ottoman robes sat behind a cabinet. Before each performance Kempelen opened the cabinet's doors one at a time, held a candle behind the gears so the light shone through, and let the audience satisfy itself that there was nothing inside but clockwork. Then the Turk played, and it won.

It won for eighty-four years. It toured Europe and America, beat Benjamin Franklin in Paris, and in 1809 played Napoleon at Schönbrunn. Inside, the whole time, sat a human chess master on a sliding seat, moving out of the candlelight one compartment ahead of the audience's inspection. The trick was never the chess. The trick was the sequence in which the doors were opened.

The most successful machine of the eighteenth century was a cabinet with a person inside. We did not retire the trick. We named our infrastructure after it: when Amazon launched a marketplace for invisible human labour in 2005, it called the service Mechanical Turk, and described the product, with admirable honesty, as artificial artificial intelligence.

The cabinets have been busy lately. Builder.ai raised around $445 million from investors including Microsoft for an AI that assembled apps; when it went insolvent last year, the salient fact was that "Natasha" had mostly been routing the work to roughly 700 engineers, something the Wall Street Journal had flagged as far back as 2019. Amazon's Just Walk Out stores turned out to involve around a thousand workers in India reviewing transactions, near 700 per 1,000 by one report, against an internal target of 50, a framing Amazon disputes. The SEC has charged a public company for the gap between its AI demo and its offshore call floor, and the regulator now has a word for the whole genre: AI-washing. And this year's flagship home humanoid ships with an "Expert Mode", printed right on the order page, in which a human operator in a VR headset sees through the robot's eyes and folds your laundry from a room you will never visit. Figure's Brett Adcock publicly torched that approach as staging "fake videos with human teleoperators in the next room" and calling it a company vision. Two hundred and fifty-six years after Vienna, the industry is still arguing about the sequence in which to open the doors.

In February, Will Manidis wrote an essay called "Tool Shaped Objects" that has been rattling around my head ever since. A tool shaped object looks like a tool, feels like a tool, and produces the unmistakable sensation of work being done. It just doesn't produce work. "The market for feeling productive," he wrote, "is orders of magnitude larger than the market for being productive," and the numbers agree with him: AI is everywhere in consumption and almost nowhere in output. He's right, and I want to concede the full extent of it before I push back.

Because the most respectable cabinet in the industry is not a fraud at all. It is the orchestration layer.

The logic sounds unimpeachable. If one model disappoints, route between five. If the five disagree, let them vote. Put an agent in charge of the agents, a judge model above the worker models, wrap the whole thing in a framework and call it a system. Redundancy is how we built reliable machines out of unreliable parts for a century; it is the most honourable idea in engineering. So this spring we measured it across 67 frontier models: routing, voting, mixtures of agents, every respectable committee. The polite summary is that the ensemble rarely beats its own best member. The models fail together, on the same inputs, in the same ways. A committee of correlated guessers is one guesser in a trench coat.

The reason is not mysterious. The frontier models ate the same internet, chased the same benchmarks, and increasingly learn from each other's outputs. Best-of-N sampling buys you N tickets in the same lottery with the same numbers. Every layer of the modern stack is a bet that many correlated things add up to one independent thing, and the bet quietly fails at every layer: models checking models, agents auditing agents, and at the bottom of more products than anyone likes to admit, a person in Chennai. It is Turks all the way down. The orchestration layer is Kempelen's cabinet rebuilt at planetary scale, and the doors open one at a time: look, no human in this compartment, while the value slides one compartment over.

Redundancy only works when the errors are independent. That is the whole theory of it, in aircraft, in RAID arrays, in juries. So where does independent error come from? It is the question the orchestration business keeps not asking, and the answer is not another head in the cabinet, because the heads agree. It comes from outside, and there is exactly one distribution nobody has finished scraping: the physical world. Every lab has the same internet. Nobody has your contact with matter. An arm that grips, slips, and recovers is drawing samples from a process that is not in anyone's training set and cannot be distilled from a bigger model. Consequence is the data you cannot fake upstream.

Which is why the robotics numbers this summer matter more than the model releases. Figure's humanoids on BMW's Spartanburg line now work, by Figure's own accounting, for about $25 per robot-operating-hour, after a trial that ran ten-hour shifts and loaded ninety thousand parts into real cars. Unitree sells humanoids at e-bike prices and has filed to go public. Optimus is scheduled to attempt mass production this summer. And the people closest to the machines remain the most sober people in the discourse: at the field's own flagship conference last autumn, after a week of watching humanoids walk, dance and box, roughly 90% of the roboticists in the room voted that humanoids will not replace most human workers in the next twenty-five years, and the vote got more skeptical after the debate, not less. Moravec called this sixty years of paradox ago. The dancing is solved. The hands are not.

But look at what the constraint does to the engineering, because this is the part I love. At the edge, every cloud habit dies. A robot runs at batch size one: there is no other traffic to amortise the thinking, the latency you demo is the latency you serve, and the bottleneck turns out to be not bandwidth but the overhead of launching the work. Memory stops being an abstraction: the flash wears out, so every write carries a shadow price, and our policies learned to commit something to memory only when it would change the next action, a discipline I would also like to prescribe to several dashboards. Even redundancy comes back, but honest this time: not five guessers voting on an answer, but one policy acting while a reflex watches its internal state for the wobble that precedes a fumble, ready to hand control to a stronger model just in time. A chatbot that pauses for ten seconds is thinking. An arm that pauses for ten seconds has dropped what it was holding. There is nowhere to put a cabinet at the edge, because there is nothing behind the robot but the robot.

That is the inversion I keep coming back to. The cloud, where nothing is at stake, is where we build the elaborate committees. The edge, where physics audits every millisecond and every write, is where the engineering gets honest. Scarcity did to our stack what abundance never does to anyone's: it made every component justify itself. If the last era's scaling law was more compute per model, the one I am betting on is more world per model. You get there through contact with the world, not by stacking more models on top of each other.

The Turk, for what it is worth, burned in a Philadelphia museum fire in 1854, and the confession came out shortly after. But Kempelen deserves a better epitaph than his cabinet, because he spent his last decades on the opposite kind of machine: a contraption of bellows, reeds and a rubber mouth that could genuinely speak, badly, like a child. He published its full mechanism in 1791, every compartment open at once, and the line of work it started runs through Bell's telephone to the voice models in your pocket. The cabinet made him famous. The bellows made the future.

Build the bellows.

Braindump by Josef Chen