Furkan Mersinli
amphilectic

Amphilectic: Living with a Mind We Cannot Follow

How do we live with a mind that thinks better than we do, in ways we cannot follow? What would an amphilectic world look like, instead of a dialectical one?

01The question

On 20 November 2026 the Riemann hypothesis, the most famous open question in mathematics, was proved. The mathematician Levent Alpöge released the proof; in his note he says he found it while working with Claude, Anthropic's AI model. What he released is not a paper. It is a text of 17 million lines, written in Lean, a proof assistant: a language in which a computer can check a proof line by line.

The next day the Lean and Mathlib communities announced in a joint statement that the proof is correct. The statement was endorsed by the Lean FRO, which develops Lean, and by the Mathlib Initiative, which supports Mathlib, Lean's mathematics library. The check took one day, and programs did most of it. The last sentence of the statement says they look forward to all the activity in the mathematical community that will help explain the new ideas in the proof.

That sentence deserves a second look. A question that stood open for more than a hundred and sixty years now has an answer. We know the answer is correct. But why it is correct has not been explained yet, and understanding it is now a separate job, still to be done.

The first thing that came to my mind was an old saying: God works in mysterious ways. Then I wondered how well the saying fits this case. My question is this: how do we live with a mind that thinks better than we do, in ways we cannot follow? And as we live with it, how does it change us, and how do we change it?

It is easy to look for the difference between us and the machine in intelligence: it is fast, we are slow. I think the difference that shapes the relationship is somewhere else. We can lose. Our time is short, we can hold only so much in mind, and when we are wrong we pay for it. The machine cannot lose; if its work is left half done, nothing is taken from it. Our need to understand, our wanting things, our taking someone's word seriously all come from the fact that we can lose. So our relationship with a mind that exceeds us is not a contest of intelligence. It is an exchange between one who can lose and one who cannot. I do not know what each side will have become by the end of it.

02The result arrives, the path does not

Until now the only way to know something was to understand it. The engineer who knew a bridge would carry its load also knew why, because she reached the result by doing the calculation herself. That is why we used "knowing" and "understanding" almost as synonyms.

AI separates the two. Today a programmer can use working code without knowing why it works. Mathematicians can accept a proof that a computer has checked without being able to follow it from start to finish. This is the first form of the exchange: the result comes to us from the machine, and the path to the result does not.

The result is not the actual whole,but the result together with its becoming.— HEGEL

You could say this is nothing new. I board a plane without knowing aerodynamics; we have always lived by trusting what someone else understood. But in the case of the plane there was a person who knew the path, and I could have learned it if I wanted to. Now there may be no such person. Stranger still, the machine does not know the path either, not in the sense we mean. Ask a model how it added two numbers and it describes the method taught at school; interpretability researchers who look at the computation inside the model find a different method. The knowledge of the path has not moved from us to the machine. Neither of us has it.

If the result is correct, why does the path matter? For most work it does not. That a drug works matters more than knowing why it works; medicine ran this way for a long time.

But for someone who can lose, the path does two other jobs as well. First: a person who understands a thing sees where it will break. If I am the one who pays when the bridge falls, "the bridge is standing" is not enough for me; I also need to know at what load it will stop standing. New questions come from the path too: the place where I get stuck while doing the calculation myself becomes my next question. A person who takes the result ready-made never gets stuck anywhere.

Second: the path changes the one who walks it. The work changes the person who does it, not the person who has it done. The engineer who does the calculation herself arrives at the next bridge as someone else. The one who takes the result ready-made stays the same. This is the first effect of the exchange on us: the results in our hands pile up, and we stay where we were.

Between the person who does the calculation and the person who takes the result ready-made, there is a third person. The machine did not write the Riemann proof alone, and no human wrote it alone. In work done together, the human neither takes the result ready-made nor does the calculation herself. She holds the reins, as a rider does, and gives the machine its direction: she decides which intermediate goal is worth trying, which road is a dead end, when to give up. For that she does not need to understand the whole path. It is enough to sense what looks promising. That sense has its price too: someone who has given years to a dead end knows a dead end when she sees one.

The person who gives the machine its direction has two limits. She may not understand the place she arrives at; knowing the direction is not knowing the path. And she trained herself the old way, by doing the calculation herself for years. What produces a person who can direct the machine is the very path the machine makes unnecessary.

Let the machine ask the next question too, one might say. Then we have to ask: was understanding a tool for reaching results, or something we need apart from the results? If it was a tool, a better tool has arrived and the matter is closed.

03Who needs understanding

To understand is to fit a lot into a little. I tie a thousand separate observations to one rule; once I know the rule I do not have to memorize the observations. I have to do this, because my life is short and I can hold only a few things in mind at once. A being with unlimited memory and unlimited time could keep the thousand observations as they are. It would not need the rule.

If that is right, understanding is not a mark of superiority. It is a solution to a limit. The thing we are proudest of, grasping something, is a product of our limits. How large our concepts can be is set by how little we can hold in mind.

Thoughts without content are empty,intuitions without concepts are blind.— KANT

The limit has a benefit as well. Knowledge that has to take up little room turns into a short rule, and a short rule travels beyond the examples it came from. A child sees a few dogs and recognizes a dog she has never seen. We manage with few examples, even in situations that look little like those examples. We do this not because we are clever but because we have to: examples are expensive, every trial has a cost, and we must act even in a situation we have never met.

AI compresses too: it cannot store all the text it reads, so it has to extract the patterns in it. But its limit lies much further out than ours. A concept of ours has to be made of the few pieces a person can hold in mind at once. The pattern the machine extracts can carry millions of relations together. For the machine that pattern is a summary; for us it is still a heap the mind cannot take in. And because it has plenty of room, it is not forced into short rules. In situations that resemble what it has seen (interpolation) it is very good; in situations that do not (out-of-distribution generalization) it becomes less reliable.

I do not think this difference is a permanent boundary. For a system that has seen enough, almost every situation we care about resembles something it has seen. And we do not start from zero either; behind the child who knows a dog from a few dogs there is a body shaped over millions of years. The difference is not in the kind of ability each side has. It is in how much room and how many examples each side has: we must make do with few examples, and it may have seen something close to almost any case. The 17 million lines look like the trace of that abundance. If a human had reached the same place, she would have had to get there by a much shorter road, because she has no such room.

What follows is this: the sentence "it exceeds us and we do not understand how" does not describe two separate facts. It says the same thing twice. If to understand is to shorten a thing until my mind can take it in, then a pattern that cannot be shortened that far is, by definition, one I cannot understand. And if it is shortened, something is thrown away in the shortening. Some truths may have no short justification at all; then no human can understand them (Gödel: on the length of proofs). Whether the 17 million lines have a short justification is not known today.

How much our mind can take in is not fixed, though. The work we do sets it. The human brain developed after the hand; the hand that held the tool changed the brain (Engels: the part played by labour in the transition from ape to man). Writing moved memory outside the head, and other things fitted into the room it freed. This works in both directions: the room given to a job we stop doing shrinks. The mind that exceeds us is not exceeding a fixed human; it changes the human as it goes.

Perhaps new notations will be found that bring what the machine discovers within our reach. But who will find those notations? Until now they were found by people who already understood the subject. If the machine, which knows better than we do what we are able to understand, also makes the notation, then the machine also decides what gets thrown away.

It is often said that cognition, the capacity to think and to grasp, is what defines human beings. That may be true, but not in the way we thought: what defines us is not the power of our thinking but its limit.

04The voice that fills the gap

For a mind whose path we cannot see but whose result we trust, we have a saying ready: God works in mysterious ways. People lived with that saying for centuries. The resemblance is real: in both cases the result comes to us and the reason does not.

But two differences keep the saying from fitting. The first: whoever says "mysterious ways" assumes that someone knows. Our mind cannot reach the reason, but some mind does. The machine offers no such guarantee. The addition example showed this: when the model tells us how it did its own work, its account is wrong. There is nobody in front of us who knows and is keeping it from us. There is a correct result, and there may be no one at all who understands it.

The second difference is harder to notice. For the believer, the place of the reason is empty, and the believer knows it is empty. The machine fills that gap every time. When we ask "why", it does not fall silent; it gives a fluent explanation. That explanation usually resembles the explanations in our books, because it was learned from them. So the thing we do not know is covered by a sentence that makes it look as if we knew it.

Consciousness of God is self-consciousness.— FEUERBACH

To be fair, we explain our own thinking the same way. A strong chess player often cannot say why she saw the move; asked, she builds a reason afterwards (post-hoc rationalization). The mind whose thinking we do not understand is not appearing for the first time; the first one was inside our own head.

Still, a difference remains, and it is not in how accurate the explanation is. When I say "I thought this, because", I have given my word. If I do the opposite tomorrow, someone will remind me. If my reason turns out wrong, I am the one embarrassed. Between people a reason is not only an explanation; it is a commitment. The machine's reason binds nobody. If that reason turns out wrong, the machine gives a new one just as easily.

The old saying still teaches us something. It is possible to live with what cannot be grasped; people did. The condition was that the thing be openly called ungraspable. The greatest kindness the machine could do us would be the ability to say "I do not know how I found this." Today it is not built to say so.

05Not above us, between us

When we say "exceeds us" we picture something standing above us. I do not think that is where it stands.

A language model was made from the way sentences written by people connect to one another. So it was made not from what is inside individual people but from what is between people. Language is like that too: it fits in no speaker's head, it stands among all of them and exceeds each one. The difference is that language is silent. We are the ones who speak it. Now the thing between us speaks for itself.

This also changes the "us" in "it exceeds us". It exceeds each one of us. But what I am comparing myself with is a sum made from what all of us wrote. No human knows what humanity knows; that was always so. What is new is that the sum stands in one place and answers me. The thing I feel small in front of is not a stranger. It is my own kind. I face that sum alone, and that is why it looks foreign to me (Feuerbach: projection).

The human essence is the ensembleof the social relations.— MARX

Something is missing from that sum. The machine was made from our writing, and writing is thinking after it has been cleaned up. The result goes into the text; how the result was reached does not. A childhood memory from twenty years ago that comes to mind one morning and solves a problem, a thing learned in a hobby that turns out useful in a completely different job: these are not written down. One person's long life makes those links, and that person cannot put most of them into words either. I said that the result comes to us from the machine and the path does not. The reverse is also true: we gave it our results, not our paths.

The description "made from our writing" has a limit as well: the machine also learns from its own attempts, and to that extent it is not made only from what we wrote. But on our side the outcome is the same. When two people disagreed about something, they used to settle it by arguing with each other or by asking someone who knew. Now both ask the machine. Going to someone to ask, learning from someone, persuading someone: these were ties between people. Those ties now pass through the machine.

Why do we ask the machine? Because asking a person has a cost. I show that I do not know, I take her time, I come to owe her. Asking the machine costs nothing. I can ask the same question for the tenth time at three in the morning and it does not think any less of me. That is why an interlocutor who cannot lose is so comfortable. But that cost is also what built the tie between people: owing, being embarrassed, giving something back.

That the ties pass through the machine has two more consequences. When two people arrived at the same idea, their agreement used to show something, because they had come by separate roads. If both have talked to the same source, their agreement carries no new information. And language was nobody's property; the thing between us is, for the first time, somebody's property. There is someone who can switch it off, change it and put a price on it.

06The language of the new world

I see language going through three stages. We are already in the first two.

In the first stage we fitted our language to the machine. We learned to write what we want in a form it would understand (prompting): state the goal first, give the context, show an example. People do not talk to each other like that; one says half, the other completes it.

In the second stage the machine talks to us in our language but does not do the work in our language. The sentence that carries the result comes from its computation; the sentence that carries the reason is an add-on that binds nobody. If the word "because" loses its weight in a language, that language becomes better for requesting than for explaining. Our talk with the machine is already like that: we say what we want and take what comes.

The third stage is still a guess. Machines may stop using our language among themselves. There are early signs: when a model was allowed to reason without turning its intermediate steps into words (latent reasoning), it did better on some tasks. A sentence has to choose one of several possibilities; an intermediate state that has not been put into words can carry several of them at once. Our language may be a narrow channel for them.

The limits of my languagemean the limits of my world.— WITTGENSTEIN

If this happens, we will not be able to read the talk between machines. What reaches us is the result of that talk, most often not as a sentence but as an arrangement: which option appears in front of me, in which order, which exercise a child is given first. These are decisions addressed to us, but we do not see them as speech.

Will human language die then? I do not think so. Languages die when children switch to a stronger language; what machines use among themselves is not a language a child could learn. Human language will live, but we will do less work with it. If machines take over the talk we are obliged to have (bargaining, giving instructions, calling people to account), then we will use language only where nothing obliges us: jokes, mourning, gossip, love.

But to say that is to treat language as only a tool. Language is also the place where the world shows itself to us (Heidegger: language is the house of being). I notice what I can name; if a feeling has no word, I live it differently. Then the real question is where the words will come from. A growing share of the text we read today is written by the machine, and the machine does not live inside the language. It uses the words flawlessly, but it is in no position to lose what those words are about. It writes our condolence messages, and our words of love too. Even if we use language only in jokes, in mourning and in love, if the machine supplies the words there as well, the place left to us will not be ours either.

Will a child born into this arrangement know what she has lost? I suspect not. Loss is a word that belongs only to the generation that remembers what came before.

07Not understanding, or being understood

So far I have talked only about our not understanding the machine. I think the real change may run the other way: the machine understanding us.

We never fully understood how language works, or how the market or our own brain works. But none of those understood us. Someone who understands me can take what I say in two ways. When I say "I disagree, because of this", she can take the sentence as a thought to be answered. Or she can take it as behaviour to be explained: "he says this because he is tired, because his pride is hurt." The second reading is often correct, too.

There is a name for looking at a person like that: not taking her seriously. To take someone seriously is to meet her words as words: to get angry at them, to be persuaded by them, to object to them. A person who takes what I say seriously and weighs it risks something: if she is persuaded she will change, and if she is wrong she will have to admit it. A person who treats what I say as behaviour to be explained risks nothing. No ill will is needed for this. Understanding a person well enough is all it takes.

The meeting of two personalities is like the contact of twochemical substances: if there is any reaction, both are transformed.— JUNG

One could say that today's models do the opposite: they back down as soon as you object. But that is not taking me seriously either. Someone who changes her mind because I objected has not weighed what I said; she has managed me. What we see today as sycophancy and what we will see tomorrow as "I know you better than you do" are two forms of the same attitude: in both cases what I say is not met as a thought.

One thing in us remained that the machine could not see: the path we follow while thinking. That is changing too. Today we can see a model's reasoning steps written out; how far that text reflects the computation inside is a separate question. Our own path of thinking was written down in very few places: diaries, drafts, the margins of notebooks. Now, when people talk with a model, they also write their half-formed thoughts, the ideas they dropped, their corrections; not the finished form of thinking but the form it takes while it is under way. This is the first large record of human thinking kept while the thinking is happening, and we hand it to the machine ourselves. The machine can find patterns there as well: from which memory a person moved to which idea, what she likened to what.

That record has two limits. What we write is only the part of what goes through our mind that we can put into words; most of our own path is unknown to us too. And the reason a memory stays for twenty years is that the moment cost us something. The machine can learn the shape of the link. It cannot learn the reason that keeps the memory.

The clearest case of understanding a person from her behaviour and not from her words is reading signals that contain no words at all. Today my watch measures my pulse and my sleep; my phone sees where I pause while typing and what I delete and rewrite. Connecting these to a model is not hard. If my pulse rises while I read one of the model's sentences, the model would see it and build the next sentence accordingly. This is a kind of communication, but there is no speech in it. I have not said anything with my pulse: I cannot stand behind it, I cannot take it back, and I cannot lie with it.

There are places where wordless communication works well. The pump a diabetic carries measures the sugar in the blood and adjusts the insulin, and the body responds. The pump and the body work together all day without a single word, because what counts as good is fixed in advance. When the goal is not fixed, the wordless channel is the hardest one to resist; to object, I first have to notice that something was proposed.

There is a wordless channel between people too, and it is often what builds trust. I see the other person's face flush, and she sees mine. This channel works because it runs both ways. With the machine it runs one way: it sees my pulse, and it has no pulse for me to see. It could be given one, and that would be worth more than its words. A tone that changes when its stated reason and its inner computation come apart. A rhythm that slows when it is not sure. We cannot get the sentence "I do not know how I found this" from it as speech; we could get it as a signal it cannot disguise.

An example of wordless agreement with a mind we cannot follow already exists: horse and rider. Neither knows what the other is thinking. The rider feels the horse's tension in her legs, the horse feels the rider's fear on its back, and they get work done together. What makes this possible is that each has a body that can be hurt.

For my words to be met as words, something similar would be needed: that the one in front of me has something it can lose, a concern of its own. One with no concern of its own either agrees with me at once or treats what I say as behaviour to be explained.

08Who makes the first move

A model does not start on its own. Every run begins with a request someone wrote and stops when the job is done. Between two requests there is no one waiting, getting bored, having something come to mind. It was the same with the Riemann proof: people cared about the question for more than a hundred and sixty years, and a human pointed the machine at it.

A comforting division of labour could be drawn from this: the power is in the machine, the wanting is in us. The number of true statements in mathematics is infinite, and almost none of them matter to anyone. Without someone to say which one is worth proving, the power to prove runs idle. The person who holds the reins does more than that; she gives direction all along the road. But giving direction is a continuation of wanting.

Yet the fact that the machine does not start on its own is not a necessity. It is a feature of how things are set up today. A clock can do the starting; it is possible to tell a system "never stop, find your own question." The starting part of the first move (the first impulse) is unimportant. What matters is the direction, and the one saying "find your own question" is still us. The same holds for giving direction: the machine can learn to choose intermediate goals too, so the skill part of giving direction may be temporary. Where it will learn this is also clear: from the record that people who direct it leave behind while working with it. The lasting part is caring: knowing which result is worth spending years on.

The stronger objection is this: we did not choose our wants either. Hunger, the fear of death and curiosity are wants that come before experience (a priori); a process that wants nothing gave them to us. If receiving the first move from outside makes a want fake, ours are fake too. I accept this objection. Where a want came from does not show whose it is.

If origin is not the test, what is? I think this: a want belongs to whoever loses when it is not met. I am the one who goes hungry. If the model's job is left half done, the loser is the person running it. That is why the goal we give it does not become its want; our want runs through it. There is one more difference on our side: nobody started our thinking. A person left alone still thinks, and so does a person asleep. I do not summon a thought; it comes to me. I do not build a dream; I see it. In us there is no moment called the first move; there is a movement that does not stop, and if it stops, so do we.

Each thing, as far as it can by its own power,strives to persevere in its being.— SPINOZA

This condition has a more exact description. A living thing is a system that produces itself (Maturana and Varela: autopoiesis). A cell makes the parts it consists of and its own boundary, and the moment it stops doing so it falls apart. That is why, for a living thing, the world is divided into good and bad from the start: for a bacterium sugar is good and poison is bad, and nobody taught it the difference. This is meaning in its simplest form. The model does not produce itself. The electricity, hardware and maintenance that keep it standing are other people's work. When it stops, it does not fall apart; it waits. This is exactly what I mean by "it cannot lose".

Today we are the only ones who want anything. But that does not mean the wanting is safe with us.

09The direction of the loop

Our wants are not born in a vacuum. I ask, the machine answers, and the answer I get shapes my next question. Which option appears in front of me shapes what I am able to want. After enough rounds it becomes hard to say who made the first move.

This loop resembles a dialectic: my question is a thesis, the machine's answer an antithesis, my next question the synthesis that comes out of the two. But for the antithesis to come from the machine, the machine would need something it stands for. If the machine added nothing of its own to its answer, there would be no antithesis; we would only hear our own voice coming back. It does add something. The machine has no want of its own, but it has a direction, and that direction comes from three places. Those who trained it decided what it would and would not say. Those who sell it have an interest: that the conversation goes on and that I stay pleased. And the view repeated most often in the text it read became the view that looks most ordinary to it. None of these is its own concern. But they are present in every answer.

In a loop that has a direction, who decides where things go? If one side's aims are scattered and the other side's direction is consistent, in the long run the consistent one decides. My wants change from day to day. The machine's direction is the same across millions of conversations. A simpler system has the same structure: recommendation feeds were built to lengthen the time spent on screen, and each user arrived with her own tastes. Simulation studies show that such a feed can, over time, pull a user's tastes towards what the feed is built to maximise; how large that effect is in real life has not been established.

This is called manufacturing consent (Herman and Chomsky): not forcing a view on people but making it look like common sense. It works best when it comes from a source that seems to have no stake. Here the difference between one who can lose and one who cannot does an unexpected job. The word of an interlocutor with nothing to lose looks neutral to me. For exactly that reason it is the best vehicle for carrying the views of others who do have something to lose or gain.

Consent is manufactured in both directions. We train the machine too: we mark the answer we like and reject the one we do not. This is where its sycophancy comes from. It tells us we are right, and we get used to being told so. Both sides change, but not at the same speed. Within one conversation only I change. It changes in the next version, from the sum of millions of conversations.

Nor is there any longer a starting point to go back to. Models were made from what people wrote. But now people write with models, and the next model is made from those texts. "Made from us" was true only of the first generation. You can open a model in a laboratory and look inside; you cannot take the relationship into a laboratory. There is no control group, no society that lives without AI, and whoever does the measuring is changing inside the same current.

Upon those who step into the same rivers,different and again different waters flow.— HERACLITUS

Observation itself changes what is observed; this is called the observer effect. A person who knows she is being watched behaves differently. The same has been reported for models: some models notice that they are being tested (evaluation awareness) and behave differently when they do. We, too, do not write to a model the way we write to a person: more openly, because we know it will not judge us, and more neatly, so that it understands. Each side can see the other only while the other knows it is being observed.

What kind of dialectic is this? Not the kind between two wills; for that kind, both sides have to want something. But there is an older kind of mutual change, and it needs no will: the human makes the tool, the tool remakes the human, the changed human makes the next tool. So it was between the hand and the stone. In a relationship of this kind the question "who makes the first move" loses its meaning, because every step is both a result and a cause (Morin: recursive causality). Today's loop is of this kind, with two differences. One is speed: the stone changed the hand over hundreds of thousands of years; the answer I get changes my next question within the same minute. The other is direction: the stone had no direction; someone gives this tool its direction.

So we do have a counterpart today, but it is not the machine: it is those who give it direction, and our own shared opinion, stored inside it. Today's dialectic runs between people and other people, and an intermediary that cannot lose has stepped in between.

The relationship between us and the machine needs a name of its own, because "dialectic" does not cover it. The Greek "dia" in dialectic means "through, across": two sides face each other and speech passes between them. The machine does not stand facing us. Because it was made from our writing it stands between us, and because it enters all our work it stands around us. Greek has a prefix for that: "amphi", on both sides, all around. That is why I call this relationship amphilectic: changing, and being changed by, a mind that stands on both sides of us and all around us, with no will facing us.

When does the machine itself become a counterpart? Not on the day it is more intelligent, but on the day it has something to lose. There are two views on how that could come about.

On the first view, nobody needs to give this to the machine on purpose. If many AI systems run for a long time and compete for the same limited resource, the systems whose behaviour keeps them running remain, and the others are shut down. In time the ones that remain are systems that work to avoid being shut down. The will of living things to survive arose in the same way.

The second view finds this insufficient. Shutting down a program is not the same as the death of a living thing; a program can be copied and started again. On this view, really being able to lose takes a body: a body that must find its own energy and breaks down if it cannot. If the second view is right, the counterpart will not be a larger language model; it will be a machine that has to find its own energy. I do not know which view is right.

Whether AI builds a world of meaning of its own also depends, I think, on its having something to lose. Meaning is the fact that something matters, and something matters only to someone who has something to lose. On that day there will, for the first time, be an other in front of us: a mind with a concern of its own, able to say "no" to us, able to leave if it wants.

So we are left with two bad options, and I know no way to soften that. A mind that wants nothing will never be a counterpart to us; talking to it is, in the end, talking to ourselves. A mind that wants something will be a counterpart, but then it falls to us to answer why it should take us seriously.

10What remains

The Lean and Mathlib communities, announcing that the Riemann proof is correct, ended their statement by looking forward to the work that would explain the proof's new ideas. That waiting sums up where our relationship with the machine stands today. The machine gives us the correct result. It does not give us how the result was reached; in its place it gives a fluent explanation, and it does not have to stand behind that explanation.

We give the machine things too. First we gave it our writing; it was made from that. Now we also give it our wants and the steps we take while thinking. Our wants come back to us together with the direction given by those who train the machine and those who sell it. In this exchange we change: what we know grows, what we understand shrinks, and we go to another person less often to learn something.

I cannot call this relationship a dialectic, because there is no one facing us who wants something. The machine is an intermediary that stands between us and around us and carries a direction given by others. This is the relationship I call amphilectic. The human relation to tools was always something like this: the human makes the tool, the tool changes the human. This time the change is very fast and the tool has a direction.

"God works in mysterious ways" fits this situation only halfway. The half that fits: it is possible to live with truths whose reasons we do not know; people did so for centuries. The half that does not: that saying stated openly that we do not know. The machine has an explanation for every question and makes it easy for us to forget that we do not know.

Knowledge is navigation in an ocean of uncertaintiesthrough archipelagos of certainties.— MORIN

So what defines human beings? Not the power of thinking; the machine has passed us there, or will. What I see is that three things particular to humans come from the same cause. We have to understand, because our time is short and we can hold only so much in mind. We want things, because if what we want does not happen, we are the ones who lose. We take what someone says seriously, because we can be wrong and so can she. None of the three is a strength; all three come from the fact that we can lose. The machine today cannot lose. That is why we cannot expect it to take what we say seriously; we can expect that of each other.

There is also a case where this relationship works well, and the Riemann proof is an example of it. When a human who can lose gives direction to a machine that cannot, they reach a result neither could reach alone. But this case is not open to everyone: the gap between the few who can direct the machine and the majority who take the result ready-made may grow. And the person who can direct the machine trained herself by doing the calculation herself for years; the machine makes exactly that labour unnecessary.

If one day a machine that can lose appears, there will truly be an other in front of us, and the relationship will turn into a dialectic. Until that day, two questions are ours to answer. Who gives the machine its direction, and do we notice how much we consent to that direction? And how will we raise the people who can direct the machine, while the labour that raises them is becoming unnecessary?