Three Pounds Against a State

In 2024, the world’s data centers drew about as much electricity as France. Artificial intelligence is a growing share of that load. The organ that conceived it weighs three pounds and runs on some twenty watts.

We all learned long division by hand at school: the long count, the remainder, the patience. A calculator performs the same operation in a blink and, under the same conditions, returns the same result. And when the calculation grows to crossing millions of records, or testing more combinations than a person would get through in a lifetime, the machine keeps solving it in seconds while we would still be starting. For decades we knew that superiority and it did not frighten us. It occurred to no one to call the calculator intelligent. It was not intelligent, we knew. It was fast. It was powerful. Speed of calculation, however complex beneath the surface, stays in our service: a tool, our tool.

And there was a deeper reason to feel safe. That machine always did the same thing to the same input: same sum, same result. It followed a fixed rule, and every fixed rule is crystallized human insight. When your phone traces a route, it executes layers of geometry others discovered and left as steps, so anyone can apply them without having thought them through. And so the route does not drift by miles, the GPS system adjusts the satellite clocks for Einstein’s relativity, which found that time runs differently up there, from the great speed and the weaker gravity. Ancient geometry and modern relativity, working together from the device in your pocket. An algorithm is the indelible trace of an insight: the understanding someone reached for the first time, set down in instructions a machine executes without understanding. Executing Pythagoras’s insight is not the same as having Pythagoras’s insight; applying Einstein’s equations is not the same as having grasped spacetime. The flash was human and it moved us forward as a civilization; the machine only applies what those insights left behind.

That is where the real magic lives: in the instant someone understands what nobody had understood. Once it becomes a rule, it stops being magic and turns into technique anyone repeats. The frontier moves; further out waits the next mystery, still unnamed. It is an endless spiral: each magic becomes a rule, and each rule clears the ground for the magic that follows.

Artificial intelligence crossed a threshold neither the calculator nor the GPS came near. It crossed, all right, and away from where we think, and away from where we fear.

From certainty to plausibility

The calculator and the GPS determine: they apply an exact rule and deliver, with certainty, the result the rule dictates. Artificial intelligence does something else: it predicts. Rather than executing a fixed rule, it learns, across a mountain of absorbed examples, regularities that let it produce plausible continuations. To the same question it can answer differently on two attempts. From that mechanism it receives no commitment to the truth.

There it loses a certainty the calculator had. It can be right, and the mechanism that produced the answer guarantees nothing about that rightness. It looks like thinking at the very moment it has traded certainty for plausibility. And the plausible, by definition, resembles what a human would say, whether or not it is true. Hence the mirage: it sounds like intelligence to us, and it was trained to sound like us.

Fog and mind

Something did change in earnest, and it is the source of the unease. The calculator’s algorithm was legible: Pythagoras’s trace could be followed step by step, twenty-five centuries on. The one inside artificial intelligence is so vast and so tangled that not even those who built it can say, before a given answer, why it produced that one and no other. It remains an algorithm, without magic and without a ghost inside, but one that stopped being legible to its own authors: nobody’s insight, crystallized in a thicket even they cannot decipher.

It is tempting to read that opacity as a sign of mind: if its own creators cannot follow its decisions, might it not be growing more like us? It runs the other way. A human being is unpredictable too, for another reason: because they judge from a purpose of their own, because they want. Their freedom has a root. The machine is unpredictable in another way: so many parameters, so many statistical correlations, that the thread cannot be followed. In the human, the unpredictable grows from an intention. In the machine, from a thicket. Freedom on one side. Fog on the other. Confusing them means believing that because we fail to understand it, it understands.

Choosing the what-for

Beneath calculation and beneath prediction lies an operation the machine never performs: judging. Judging means more than finding the best way within a given goal; it means deciding what the goal should be. What is worth pursuing. What line to leave uncrossed. What counts and what does not.

Think about programming. The machine makes generating candidate code cheap and returns it tirelessly: whole functions, variants, corrections. But someone had to decide which problem was worth solving, and what better means here: the fastest code, the most readable, the cheapest to maintain, the easiest to extend when the system grows and someone else inherits it? These are goals that compete, and which one should prevail cannot be calculated. It is judged. The machine solves brilliantly for how might this be done. For this, and what for it has no way out: the question falls outside what it can compute. It can optimize within a goal. But that a goal can be optimized leaves unanswered why that goal is worth pursuing. Judgment begins in that difference.

Beneath judgment, wanting

And judgment rests on something the machine lacks. A what-for does not float in a vacuum: it exists only for someone with something at stake. Judging demands an interested point of view, not in the mean sense but in the literal one: someone to whom the work matters, who would rather it turn out one way than another, who answers for it.

The machine wants nothing: it has nothing to lose. Wanting grows from knowing oneself finite. Whoever has limited time must choose; whoever can lose, takes care; whoever can be wounded knows what a wound in another is. Intention is the child of vulnerability. Only a being whose things can go wrong, and run out, has reasons to prefer one outcome over another.

The machine, which never dies, never wants.

Power it has in excess; what it lacks is something at stake. Whether some other class of machine will one day be able to want is an open question. What exists today, however much it surprises us, generates and optimizes without anything of its own resting on the result.

The single root

These differences are not several. They are one, seen at different depths.

The machine predicts; it does not judge. Judging means choosing a what-for, and choosing a what-for belongs to whoever wants. Wanting has no part in what the machine does. It is not a deprivation it suffers: its task is another, to produce a plausible continuation. Pull the root and the whole tree comes down. Where there is no finitude there is no wanting, where there is no wanting there is no judgment, and without judgment prediction runs blind: enormous power fired at a goal someone else set.

That is why it pays to say carefully what artificial intelligence is. Living intelligence takes many forms, and none belongs to us alone. All of them share one condition: life is organized around something that can be lost, from the bacterium seeking food to the crow solving a problem. Persisting, feeding, reproducing, avoiding harm. Statistical prediction lacks that root. It resembles it from outside: it solves, it gets things right, it surprises. It lacks what founds it: nothing is at stake for it. It produces the plausible without being able to want the true. Our mortality is not the limit of the mind. It is its root.

What is at stake

None of this diminishes the machine: it places it. At generating the plausible it has no rival, and it would be foolish to do without it, as it would be foolish to divide by hand when you can avoid it. Using the tool is sensible. Believing it wants what it produces is not. The real danger, for now, is less that it replaces the human than that the human, dazzled by plausibility and cowed by opacity, abdicates what the machine lacks: the judgment that grows from wanting, and the wanting that grows from knowing oneself finite. Confusing the plausible with the true, or fog with mind, is the shape of that abdication.

The labor market has just produced a striking signal. Figures from the Federal Reserve Bank of New York record a reversal that would have sounded unthinkable a decade ago: in 2024, unemployment among recent philosophy graduates ran at 5.1%, against 7% for computer science. Researchers at that same bank estimate that remote work, rather than artificial intelligence, accounts for 64% of the general rise in unemployment among young graduates: training a beginner at a distance is harder. The reversal is real; its cause is contested. Few philosophers, many programmers: it could be plain supply and demand. From the trade, however, comes a converging testimony. Marco Argenti, the Goldman Sachs CIO, wrote in Harvard Business Review that an engineering degree alone no longer suffices, and advised his daughter to pair it with philosophy. His reason is the one every team runs into: AI can produce high-quality code, and yet “it can work well, but not do what you want it to do.” Building and tuning a model means converting objectives and criteria into functions, data and evaluations. Deciding what counts, what goal to pursue, what line to leave uncrossed, is no task of calculation. It is a judgment. The machine can optimize a goal we hand it. That alone supplies no reason to choose it.

The small prodigy

Artificial intelligence lives in data centers that cover acres, feed on vast quantities of text and consume electricity on the scale of nations. The figures give the measure of the appetite. The International Energy Agency estimates they drew some 415 terawatt-hours in 2024, and projects around 945 by 2030. To make the scale concrete: the 2024 figure runs close to what France consumes in a year, and if the Agency’s projection holds, by 2030 the infrastructure will draw more than Japan does today, or better than three times all of California.

It predicts, and it predicts wonderfully. Across from it sits an organ of three pounds that runs on some twenty watts, about the LED lamp in our living room, and fits inside the small vault of the skull. The human brain is barely two percent of our weight, and it takes close to twenty percent of the energy we spend. It competes in neither speed nor memory nor scale; that arm-wrestle it lost long ago. That organ does something the country of silicon never touches. It has something at stake in what it does: it wants, it judges, it answers, it understands what it says. It competes where it matters, and it wins. The highest thing we know fits in no continent of servers: it fits in a head, and lights up on what a bulb costs.

Using the machine hard does not end, contrary to the fear, in the humiliation of the human. It ends in wonder at what we are.

That the highest should fit in the smallest.

What the machine reveals

The machine shows its limit, and in the same gesture ours. It forces us to discover what we are at the root.

For centuries we believed language was the properly human thing. Speaking in words set us apart from the animal; the word was the mark of mind, the proof that we thought. And then came a machine that talks, writes, argues and answers, that handles many languages without any of it mattering to it, and that certainty collapsed. If something that commands language without inhabiting it can chain words and concepts better than we can, then language was never the summit.

It was the jungle: dense, beautiful, tangled, and now imitable as well.

We discovered, almost with wonder, that something sits above that jungle. We were never the language. We were what made language want to say something. We were the intention beneath the word, the judgment that chooses, the wanting that grows from knowing oneself finite. The machine took language from us as an emblem, and in taking it pushed us higher, into the light above the jungle, where what imitation alone never reaches lives.

The advance of knowledge was never the accumulation of answers that sound good; of those the machine produces without end. It was the pursuit of something true and wanted, sustained by someone who tries, risks an assertion and answers for it. As prediction grows abundant, intention matters more rather than less. It becomes the scarce thing, the decisive thing, the human thing.

The machine multiplies what we know how to do. It returns to us, without meaning to, the question of what we are. And the answer is no longer the tongue we share with it.

Man is not the language that speaks: he is the language that wants the true.

The closing line comes from Reason Under Siege, forthcoming, which puts a single question to twenty thinkers from Heraclitus to Arendt: what it takes to keep judging when everything pushes toward the verdict that came first.

A companion piece on the labor-market signal discussed above, written for an engineering audience, appeared in Techstrong IT: Philosophy Majors Beat CS Majors on One Labor-Market Metric.