Pieza independiente
The owners of the nervous system
Movement 1
1940–2026
I
Between 1940 and 1943, in a laboratory at MIT, Norbert Wiener tried to build a machine that would guess where an aircraft was going to be.
The problem was specific and urgent. An anti-aircraft shell takes about twenty seconds to reach ten thousand feet, and in twenty seconds a bomber has moved. You had to fire at the place where it would be, and to do that you had to model something no ballistic equation allowed for: the behaviour of the pilot, who manoeuvres, who corrects, who responds to what he sees coming.
Wiener and Julian Bigelow spent three years on it. They treated the pilot and his aircraft as a single system that receives information, compares it against a target and adjusts. They gave that there-and-back circuit a name, feedback, and found that the same scheme described a cat chasing a mouse, a hand reaching for a glass and a thermostat.
The machine did not work. The project was cancelled in 1943 and what went into the anti-aircraft batteries was a device from Bell Labs, cruder and more effective. Wiener lost.
What survived was the concept, and it survived whole. In 1948 he published a book on control and communication in the animal and the machine, and out of it came half a century of engineering, biology and economics.
In January 1947, a year earlier, he had also published a letter.
II
A researcher at an aviation company had asked him for a copy of one of his papers. Wiener answered in writing and published the answer in a magazine. He was not going to send it. He argued that a scientist may refuse to put his work in the hands of someone who will use it to kill, and that the habit of publishing everything, without looking at who picks it up, had stopped being neutral.
It was the first serious attempt to ask the right question, and it failed exactly as the predictor had.
Because the question was not who owns the invention. It was who owns the loop: the circuit by which the output of one system re-enters as the raw material of the next, and not the information or the device.
Eighty years later that question has eight answers with proper names, and not one of them is a scientist.
III
It has to be said up front that these are not eight equivalent tycoons. Presenting it that way is what the press does, and it is what stops anyone understanding it.
They are organs of one creature, and the strange part is that each organ has different shareholders.
There is an operational cortex [the layer where an organisation defines what data exists, which objects are valid and what an agent may execute inside its perimeter]. There is a social sensory system, which is where people live and where an agent with its own virtual machine can read the mail and buy things. There is a generative cortex, selling frontier models to companies and to federal, state, local and tribal agencies. There is a self-control system preaching restraint from inside the accelerator. There is a laboratory that turned into an institution. There is a node holding model, social network, data centre and rockets in the same hand. There is a metabolism, which makes the substrate and does not want coordinated restrictions. And there is, lastly, an external evolutionary pressure that does not compete on size; it competes on cost.
None of the eight controls intelligence. Each controls one part of the loop, which is why none of them can be replaced by another.
IV
Of the eight, the one that interests me most is the one that generates the fewest headlines, because it is the only one that changed its nature three times in three years.
In 2024 it sold software for understanding organisations. In 2025, a platform where third-party models operated on the ontologies of companies and armies. In 2026, an architecture for an organisation to hold cognitive territory of its own [compute, models and data inside its perimeter, never leaving].
The mutation is so slow it is almost invisible, and the result is this. It does not need to build the best model in the world.
If someone owns the layer that defines what data is visible, which objects exist, which relations are valid and what the agent is authorised to execute, then they own something other than the model and a great deal harder to replace. They own the operational reality inside which the model thinks.
A better model, inside that perimeter, is still thinking inside that perimeter.
There is an inverse mutation on the same board, and it is just as instructive. A company whose business was feeding models [labelling, cleaning, ordering the material they learn from] was partly bought by the largest social network in the world, and its founder went on to run the lab that builds them. First he learned how an intelligence gets fed. Then he was handed the human ecosystem where that intelligence lives.
One came up through the ontology. The other, through the feed.
V
I let one of those agents read my mail for nine days.
It was the version with its own virtual machine, the one that does not stop at suggesting. It opens the applications, browses, fills things in, buys. I gave it the inbox and a card with a low limit, out of caution and out of embarrassment.
On the first day it chased up a wrongly issued invoice and drafted the complaint in a tone I would never have managed, because I would have apologised somewhere and it did not apologise. On the third it cancelled two subscriptions that had been charging me for fourteen months without my noticing. Eighty-six euros a year, with the two confirmation emails filed in a folder it created.
On the sixth day I realised I had started waiting for its morning summary before opening the mail myself, and that I no longer knew what had arrived, only what it had told me had arrived.
That is not an anecdote about productivity. It is the exact shape of Wiener's loop, turned inside out. The system receives information, compares it against a target and adjusts; only the target is set by someone else, and what it adjusts is what I see.
I switched it off on the ninth day, and not out of fear. It was something smaller and harder to explain. I had spent a week not feeling up to date with my own life, and at the same time my life was more up to date than it had ever been.
VI
What happened in 2026 was not a race. It was a fracture, and it does not run where people usually say it does.
It is not artificial intelligence against humans. It is acceleration against friction, and neither side is a coherent bloc.
On the friction side there are four figures who cannot stand each other and who propose different things. Independent evaluators working inside the labs. Explicit coordination between competitors so as not to run faster than anyone can audit. Institutes dedicated to discussing risk in public. And calls to brake the frontier from someone building data centres at full speed.
On the other side there are three who consider a coordinated slowdown naive at best.
The line does not separate the good from the bad. It separates whoever makes the substrate from whoever builds on top of it. Whoever sells the metabolism does not want the organism on a diet.
VII
What remains is the figure that made me rewrite everything above.
Put the twenty-seven months into a single table — from July 2024 to 16 September 2026, milestone by milestone, layer by layer — and empty months appear. Whole months in which nothing happened that changed the board.
An empty month is not a failure of the record. It is the most important thing in it, because it lets you separate two things that get confused all the time: real acceleration and narrative acceleration.
The second is continuous by construction. There is always an announcement, a demonstration, a model with another number on it, a roadmap. Measure the rhythm of the news and the world has been accelerating without pause for three years.
The first is discontinuous, and you recognise it by one thing. It leaves physical work behind.
That is where the real break of these twenty-seven months happened, and it happened in no model. It happened when the unit of measure stopped being the parameter and became the gigawatt. When people started talking about five gigawatts in development, then ten committed, then another ten of in-house silicon, and two million chips projected.
“A parameter gets announced. A gigawatt has to be built.”
A parameter gets announced. A gigawatt has to be built: substation, transmission line, right of way, cooling water, municipal permit, concrete. Nobody has ever managed to build a substation with a press release. That is the filter, and it is cheap to apply.
So the rule I use to read the whole board fits on one line. If a milestone leaves no concrete behind, it was not accelerating: it was talking.
VIII
Wiener died in 1964 believing he had lost both times. The machine did not work and the letter stopped nothing.
He was wrong about the first and right about the second. The anti-aircraft predictor was irrelevant; what came out of it governs every self-correcting system today, from a thermostat to a model trained on what another model wrote.
The letter did fail, and it is worth understanding why. Wiener was asking who should have permission to use the loop, and that is a moral question that presupposes a permission — which is to say an authority to grant it. In 2026 there is none. What there is are eight organs with different shareholders, each owning a stretch of the circuit.
So the useful question is no longer who owns the intelligence. It is who owns the feedback: who keeps the output when the product of one system enters as the raw material of the next.
The answer, for now at least, is not in any laboratory.
It is with whoever signs the substation contract.