A deep dive into cybernetics: the science that made purpose a technical term, gave away the first neural network, tore itself apart over a lie, and then won so completely that almost nobody can name it. No host this time — a narrator takes the whole 45 minutes.
I build control systems for electricity grids. The device at the heart of that work is a loop: hold a target, measure the state, take the difference, act to close it. Every day, in software, at a few hundred milliseconds.
There is a name for the study of that loop. It comes from kybernetes, the Greek word for the man steering a boat, and it is also where we get the word governor — which is what James Watt called the two spinning balls he bolted to a steam engine in 1788. The field that took that name produced feedback control, the first neural network, information theory’s siblings, systems thinking, and the idea that a machine can have a purpose.
Then it stopped existing. Not because it was wrong. Because it was absorbed.
This episode is that story, told properly. Here is some of what is in it.
Two machines
The whole thing turns on a distinction you can see in iron.
In 1712 Thomas Newcomen put an engine next to a coal mine to pump out water. A person stood beside it and worked the taps by hand, every stroke, all day. Within a few years that person was replaced by a plug rod: a rod hung from the great beam with pegs along it, so the engine’s own motion tripped its own valves.
The plug rod measures nothing. It has no target. It cannot be wrong. It does the next thing, and then the next thing, for as long as there is coal.
Watt’s governor is a different kind of object. It measures the thing you care about, compares it against a target built into the weight of the balls, and acts to shrink the difference. It can be wrong, and it can tell.
The plug rod is automation. The governor is control. Two hundred and forty years later, most of what is sold as the second kind is still the first.
There is also a cost in that story that nobody remarked on at the time. Before the plug rod, a person stood at those taps — someone who could notice a thing going wrong that no one had designed for. The first act of automation replaced a human who could notice with a mechanism that could not notice anything at all.
The gun
In 1940, Norbert Wiener and the engineer Julian Bigelow took on the anti-aircraft problem. A shell takes about twenty seconds to reach a bomber; the bomber moves miles in that time; and the pilot is evading. You are not aiming at an aircraft. You are aiming at a guess about a person.
Their predictor never saw combat. A simpler design won the contract. The project failed — and the failure is where the idea came from.
Watching their servos overshoot and swing back and overshoot again, they asked a physiologist, Arturo Rosenblueth, a question no engineer had a reason to ask: is there a disease that looks like this? There is. Damage the cerebellum and a hand reaching for a glass oscillates around it, wider each time. The same mathematics describes a gun turret and a hand.
In January 1943 the three of them published six pages called “Behavior, Purpose and Teleology.” For three centuries science had worked by throwing purpose out. Their move was to let it back in on one condition: stop asking what a thing feels, and look only at what it does. If a system senses the gap between where it is and where it should be and acts to close it, that is what purpose will mean.
Notice what that is. It is not a discovery. It is a definition — and it is the same argument we are having now about whether a model wants anything.
The other paper from 1943
That same year, Warren McCulloch and Walter Pitts published “A Logical Calculus of the Ideas Immanent in Nervous Activity.” A neuron sums its inputs and fires past a threshold, which makes it a switch, which makes a network of them a circuit that can compute anything computable. Von Neumann read it and wrote the report that gave us the architecture of every computer since.
One year, one overlapping circle, two papers: a machine that can want, and a brain that is logic. For about a decade it looked as if the same people would put them together.
They did not. In 1952 Wiener’s wife told him a lie about Pitts, and Wiener cut off everyone in that circle without explanation and never spoke to them again. Pitts — who had left his own family at fifteen, and for whom that circle was the family — burned his dissertation and his notes, and drank himself to death by 1969. Four years after the rupture, a summer workshop at Dartmouth coined the term artificial intelligence, partly, by John McCarthy’s own later account, to avoid having to argue with Wiener. The money followed the new word.
Cybernetics did not lose its ideas in 1956. It lost its title.
Only variety absorbs variety
The one law from the whole field that never stopped being true is Ross Ashby’s, from 1956: only variety can destroy variety. Count the distinct situations the thing you are controlling can get into; count the distinct responses your controller has. If the second number is smaller, you lose — structurally, not occasionally.
The episode spends a while in Santiago in 1971, in a hexagonal room with seven fibreglass chairs, where a British consultant and a young Chilean engineer tried to run a national economy as one control loop. During the truckers’ strike of October 1972 it worked, and it is worth being precise about why: not because the room held more data than the strike, but because its loop was faster than the disturbance. That is the whole lesson, and it is the one people usually skip.
The room was destroyed after the coup on 11 September 1973. But Ashby’s Law had already sentenced it, and the same sentence has been served by every executive dashboard built since. The dashboards have got prettier. The law has not moved.
The fantail years
Here is the part I keep turning over.
Feedback devices are about two thousand years old. Ktesibios built a float regulator around 250 BC — the same device that is in your toilet cistern. Cornelis Drebbel built a working thermostat in the 1620s. In 1745 a blacksmith named Edmund Lee patented the fantail, a small wheel that turns an entire windmill back into the wind. In 1787, one year before Watt, Thomas Mead patented a centrifugal governor for windmills.
Every single time, it stayed a trick. A clever answer to one problem in one machine, with its own name, in its own trade. Nobody saw that they were all the same thing. Building the mechanism was never the hard part. Seeing what kind of thing you had built took twenty-two centuries.
Which is not a story about people in the past being slow.
There are a great many people right now building loops and calling them agents, workflows, pipelines, orchestration, harnesses. Every shop has its own vocabulary. Nobody agrees what kind of thing it is. That is exactly what 1745 looked like — and the people in 1745 did not feel early. They felt like they were living in the age of the windmill. They were right, and they still could not see the shape of the thing they were holding.
Fifty hertz
The episode ends where my own work is. Across Europe the current in the wall is meant to run at fifty cycles a second. It never is. It is at 49.98, then 50.01, then 49.99 — a target the system misses continuously and defends continuously, second by second, for decades without a break. That ordinary, permanent failure is what keeps the lights on.
The steersman does not memorise the sea and never arrives at his heading. He holds it, and corrects, and holds it, and corrects.
Somebody is always at the wheel. Increasingly it is not a person. It was never a mystery, though — only a loop, and a loop can be opened and asked what it was told to want.
Listen above. Runtime 45 minutes.
