· 11 min read
Turks All the Way Down
The most successful machine of the eighteenth century was a cabinet with a person inside. Much of the AI stack still is.

In 1770, the Hungarian civil servant Wolfgang von Kempelen exhibited an automaton chess player at the court of Empress Maria Theresa in Vienna. During exhibitions the cabinet's doors would be opened one at a time and its innards, an intricate assembly of clockwork, inspected by candlelight, supposedly to show that no person could be hiding inside. The contraption, a wooden figure in Ottoman robes seated behind the cabinet, became known as the Turk.
A precursor was Jacques de Vaucanson's Digesting Duck, which amazed Paris on 30 May 1739. The duck appeared to eat grain, digest it and later defecate. In reality the grain dropped into one secret compartment, while pre-loaded pellets fell out of another. Vaucanson also made a mechanical flute player, a tabor player and an automated silk loom for the French state.
Kempelen built the Turk in just six months, after claiming he could do better than a conjuror who had entertained Maria Theresa. He seems to have considered it a nuisance and toured it only grudgingly, spending the rest of his life on more practical work: hydraulics, steam turbines, a speaking machine and a typewriter for a blind pianist.
It was a trick. There was a person hidden inside the Turk all along, on a sliding seat. Magnets dangling under the board let the operator follow the game on a pegboard, and a pantograph let him move the arm.
When audiences inspected the Turk by candlelight they really were seeing actual gears; they just weren't attached to anything. Kempelen was a master of misdirection and made sure spectators always peered into the wrong compartment.
The machine went on to win chess games across Europe and America for eighty-four years. Among its victims was Benjamin Franklin, in Paris. At Schönbrunn in 1809 it played Napoleon, who is said to have tried three illegal moves in a row: the first time the Turk put the piece back, the second time it took the piece out of the game, and the third time it knocked every piece off the board with its arm. Napoleon was reportedly thrilled.
After 1804 the Turk belonged to Johann Nepomuk Maelzel, the man who brought the metronome's mechanical tick into the concert hall. He worked with Beethoven too: he made ear trumpets for him, Beethoven wrote him a canon, and then sued him over a battle symphony. Maelzel's Turk exhibit also had a mechanical band of trumpeters and a diorama of the burning of Moscow.
Maelzel took the Turk to America in 1826. It first appeared on 13 April in the assembly rooms of the National Hotel at 112 Broadway in New York, then made its way to Boston, Philadelphia and Baltimore, where it stayed for months and lost to Charles Carroll, a signer of the Declaration of Independence. After that it ranged through the South, Canada and as far west as St Louis. Maelzel's problem was finding a chess player to work inside it. He had trained a woman in France, but eventually William Schlumberger was fetched from Alsace to cross the Atlantic and climb back inside.
The operators over those decades included Johann Allgaier, Boncourt, Aaron Alexandre, William Lewis, William Schlumberger and Jacques Mouret, who sold an account of how it worked for the price of a drink. It ran anonymously in Le Magasin Pittoresque in 1834 as "An attempt to analyze the automaton chess-player of M. Kempelen", and almost nobody believed it. Edgar Allan Poe saw the Turk in Richmond in 1835 and the next year published "Maelzel's Chess-Player", reasoning that there had to be a man inside because a real machine would never lose. The Turk lost from time to time.
In 1837 Schlumberger travelled with the automaton to Havana and died there of yellow fever, leaving Maelzel with a Turk and nobody to sit in it. Maelzel died at sea the following year on the voyage home. The Turk was sold, ended up at the Chinese Museum in Philadelphia, and burned there in July 1854.
In February, Will Manidis published "Tool Shaped Objects". "The market for feeling productive," he wrote, "is orders of magnitude larger than the market for being productive."
This spring, across 67 frontier models, ensembles rarely outperformed their own best member; the models all failed on the same inputs, in the same ways. Last autumn roughly 90% of roboticists at the field's flagship conference voted that humanoids won't replace most human workers within twenty-five years, and after hearing the debate they became even more sceptical. Figure's humanoids do work ten-hour shifts on BMW's Spartanburg line, loading ninety thousand parts into real cars, at about $25 per robot-operating-hour.
In 1819, Charles Babbage played the Turk in London, twice. And twice, he lost. He figured out it was a fraud and said so, but the idea of a thinking machine stayed with him. He dedicated the rest of his days to the Difference Engine and the Analytical Engine, and in 1864, decades after those games, he was still describing how he would build a machine that could play chess. Babbage wasn't fooled by the Turk; it defined the problem for him.
Poe worked the same problem seventeen years later and put his finger on why the honest machine was the harder one to build:
Arithmetical or algebraical calculations are, from their very nature, fixed and determinate. Certain data being given, certain results necessarily and inevitably follow. But the case is widely different with the Chess-Player. With him there is no determinate progression. No one move in chess necessarily follows upon any one other. From no particular disposition of the men at one period of a game can we predicate their disposition at a different period.
— Edgar Allan Poe, "Maelzel's Chess-Player", 1836
Determinate work can be welded into a mechanism. Contingent judgement cannot, which is why there had to be somebody in the box.
The cabinet had successors. Ajeeb, built by the cabinet maker Charles Hooper in 1868, went toe to toe with Harry Houdini, Theodore Roosevelt and O. Henry, drawing tens of thousands of spectators. Its secret: there was a man in there the whole time. In fact, there were several over its long career, including Constant Ferdinand Burille, Charles Moehle, Charles Francis Barker and the former US chess champion Albert Hodges, who in 1898 showed Ajeeb's ropes to Harry Nelson Pillsbury, one of the strongest players alive. Pillsbury operated Ajeeb on and off until 1904, spending countless hours inside while spectators decided he was gears.
Mephisto is the interesting one. Charles Godfrey Gumpel, an Alsatian maker of artificial limbs, built it over six or seven years and first showed it at his home in Leicester Square in 1878. Unlike Ajeeb, Mephisto could actually be searched, with not a human to be found in its compartments. Its secret was electromechanical transmission: somebody was operating it from another room, usually the chess master Isidor Gunsberg. It still wasn't playing chess.
Amazon's Mechanical Turk, a marketplace of invisible humans doing invisible work, has been around since 2005; Amazon's own description was "artificial artificial intelligence". In 2018, Google demoed Duplex phoning to book a haircut, and later clarified that humans would make some of the calls.
In 2019, the Wall Street Journal reported that an app-building service called Engineer.ai relied largely on human engineers. Rebranded Builder.ai, it raised around $445 million from investors including Microsoft, and its AI assistant "Natasha" turned out to be little more than a front for about 700 human engineers. Its administrators filed in 2025.
In 2024 Amazon retired its cashier-less Just Walk Out technology from its Fresh grocery stores, after a report that some 1,000 workers in India were reviewing close to 700 of every 1,000 transactions, against an internal target of 50 (Amazon disputes that characterisation). The SEC has since charged a public company over the gap between its AI demo and its offshore call floor. The practice is common enough that there's a term for it: AI-washing.
This year, the humanoid-robotics startup 1X opened pre-orders for its home robot Neo. On the order page is an option called "Expert Mode", in which a human wearing a VR headset looks through the robot's eyes to do your laundry. Sound familiar? Brett Adcock, of the rival humanoid company Figure, has called such videos "fake videos with human teleoperators in the next room". Mephisto got there in 1878.
The Turk got its honest ending eventually. In 1989 the illusion designer John Gaughan finished a working reconstruction and debuted it that November at a history of magic conference, with a performance eerily similar to Kempelen's. Except the place where the chess master used to hide now held a computer running a chess program. It took two hundred and nineteen years for the Turk to live up to its billing, and by then it wasn't particularly amazing.
Kempelen spent the last decades of his life on a machine that spoke: a contraption of bellows and reeds and a rubber mouth that talked, badly, like a small child. In 1791 he published its full mechanism, every compartment open at once. That line of work stretches through Bell's telephone to the one in your pocket. The chess player made him famous. The speaking machine paved the way.
The regulator arrived in stages. In March 2024 the SEC announced its first two "AI-washing" enforcement actions: against Delphia, an investment firm that told clients it used machine learning on their personal data, and Global Predictions, which called itself the first regulated AI financial adviser. Neither had the capability it advertised. Delphia agreed to pay a $225,000 penalty and Global Predictions $175,000. In January 2025 the SEC announced its first AI-washing case against a publicly traded company, Presto, a restaurant-technology business that went public in a SPAC merger in September 2022 and was Nasdaq-listed until September 2024.
The robotics evidence arrived in the same period and points the other way. BMW's trial of two Figure 02s at its Spartanburg plant ran for eleven months and ended in 2025. Working Monday-to-Friday ten-hour shifts, the robots contributed to over 30,000 BMW X3s, loading over 90,000 components in some 1,250 hours of accumulated runtime. BMW now uses the newer Figure 03 to sequence parts in logistics at Spartanburg, has opened a Center of Competence for Physical AI, and is starting another humanoid pilot at Leipzig. On 20 March, Unitree filed to list on Shanghai's STAR Market, hoping to raise 4.2 billion yuan; retail investors oversubscribed its shares thousands of times over, and its first close was 460% above the issue price.
At an IROS debate in Zhejiang in the autumn of 2025, roboticists were asked whether humanoids would soon replace most human workers. Before the discussion, 80 per cent voted no. Afterwards, 90 per cent did. Unitree's Wang Xingxing was in that audience. Humanoids still embody Hans Moravec's 1988 paradox, thirty-eight years on. The dancing is solved. The hands are not.
For redundancy to help, the errors have to be independent: think aeroplanes, RAID arrays, juries. Adding more members from the same group won't help, because they'll just agree with each other, and right now all the AI labs are training on the same corpus, the internet. Real independence requires interaction with the world: a robot arm grasping an object, feeling it slip, and learning to grip harder. That is data nobody has seen before and that can't be distilled out of a larger model. Consequence is the data you can't fake upstream.
Look at what that constraint does to the engineering, because this is the part I love. At the edge, so much of what we take for granted stops applying. A robot has a batch size of one, so overhead can't be amortised over other requests, and the latency you demo is the latency you serve. The bottleneck is not bandwidth but the overhead of launching the work. Flash wears out, so every write carries a shadow price, and policies learn to commit something to memory only when it would change the next action (I wish my analytics stack could say the same). Redundancy comes back too: not five predictors that will make the same mistake, but one policy acting while a reflex watches its internal state for the wobble that precedes a fumble, ready to hand control to a larger model. A chatbot gets a free ten seconds to think; a robot arm drops what it's holding. At the edge there's no cabinet. There's just the robot.
The orchestration layer is Kempelen's cabinet at planetary scale, with its doors opening one after another. An ensemble of correlated predictors is one predictor in a trench coat. It's Turks all the way down.
The last era of scaling laws was more compute per model. I'm betting on more world per model.
Build the bellows.
Braindump by Josef Chen