Thinking with Zahavy and Looking a Little Further
By Claudia Aranda and LumusAI
I would have made more money from what I never published than from everything I have published in twenty-five years of professional journalism.
The sentence may lend itself to the wrong reading, so let me clarify immediately: not because anyone paid me to remain silent. Exactly the opposite. Because more than once I had information in my hands that would have made a major exclusive, and I chose not to publish it.
An ongoing judicial investigation. A source who could be exposed. A person who needed protection. A true piece of information that, if revealed too early, could alert precisely the person under investigation. Or a truth still incomplete that, turned prematurely into a headline, might acquire the appearance of certainty it did not yet possess.
I learned something rather uncomfortable for a profession built around publication: finding information and deciding to say it are two different operations.
Information is ethical matter before it becomes news.
And not saying also says something.
I remain silent, and that silence has a reading.
I choose, and there is meaning in that exclusion.
I begin here —and not with artificial intelligence— because I suspect that part of the problem I am interested in is already contained in this old newsroom scene.
The question is not only what a machine can think.
It is what orients it among everything it could think.
What it decides to retain.
What it discards.
What it does with what does not fit.
How it distinguishes an anomaly from an error.
What it says.
How.
From where.
For whom.
And, finally, what it decides not to say.
THE JUMP
The pretext for these notes is a recent and provocative text by Tom Zahavy, a researcher at Google DeepMind: Position: LLMs Can’t Jump, a 2026 position paper about a question much larger than its title suggests (Zahavy, 2026).
How is something invented that does not yet exist?
Zahavy recovers an image Albert Einstein had sketched in a letter to Maurice Solovine: scientific creation as a movement in which, from the world of experience, a jump occurs toward principles or axioms, and from there logical work unfolds again, allowing testable consequences to be derived.
Zahavy connects that jump with abduction.
Charles Sanders Peirce distinguished three operations that are useful here. Deduction derives necessary consequences from premises. Induction recognizes and generalizes regularities from observations. Abduction, by contrast, appears in the face of what surprises us: something happens that does not fit comfortably with what was expected, and we propose a hypothesis which, if true, would allow us to understand it.
We have not demonstrated it.
We do not even know yet whether it is correct.
But it deserves to be investigated.
That small space between “this is strange” and “what if it were because…?” contains an extraordinary operation.
In Peirce’s mature formulation, abduction belongs precisely to the moment in which a hypothesis is proposed or adopted as a candidate for inquiry, not to the moment in which it has already been confirmed (Peirce, 1931–1958).
Zahavy argues that generative artificial intelligence has become formidable at induction and is advancing at extraordinary speed in deduction, but that it still lacks the mechanism that would allow the stronger abductive jump: producing genuinely new explanatory principles from a relationship with the world.
It is worth clarifying from the outset that we are reading a position paper. Zahavy is arguing a position; he is not presenting an experimental demonstration of a definitive incapacity of LLMs. The author himself has subsequently described the problem as open, and his current work remains focused precisely on problems requiring creativity, exploration, and what he himself calls aesthetic taste. This clarification does not make his argument less interesting. It makes it more interesting, because it allows us to discuss it for what it is: a proposed frontier, not a tombstone for artificial intelligence.
Nor should it be read as if it claimed that artificial intelligence cannot reason.
In 2024, AlphaProof and AlphaGeometry 2 solved four of the six problems at the International Mathematical Olympiad and obtained 28 out of 42 points, equivalent to silver-medal level. In 2025, an advanced version of Gemini Deep Think solved five of the six problems, obtained 35 points, and officially reached gold-medal level. More importantly for this argument, whereas in 2024 the problems had to be manually translated into specialized formal languages, the 2025 system worked end-to-end in natural language and within the competition time limit (Google DeepMind, 2024, 2025).
Artificial deduction is therefore moving at a speed that makes it dangerous to write sentences intended to remain valid for too long.
Let us grant that before continuing.
Because Zahavy’s problem lies elsewhere.
His case is Einstein.
The question could be put brutally:
if we had given an artificial intelligence roughly the knowledge available at the beginning of the twentieth century, would it have invented General Relativity?
Zahavy argues that current LLMs do not possess the mechanism that would have enabled that jump.
And to explain why, he discusses an influential family of ideas about creativity, curiosity, and compression associated, among others, with Jürgen Schmidhuber: an agent finds something interesting when it discovers a better way to predict, represent, or compress it (Schmidhuber, 2009).
But Einstein presents a difficulty.
Newtonian gravitation was not simply falling apart, waiting for someone to replace it.
We might point out here that anomalies did exist. The anomalous advance of Mercury’s perihelion is the most famous. It would be incorrect to claim that Einstein was working in the face of a perfectly closed physics with no signs of friction.
But if that were all we argued, we would miss the interesting part.
Because the anomaly said that something did not quite fit.
It did not carry the geometrization of gravity hidden inside it.
There was no conceptual gradient which, reduced step by step, led naturally from Newton to Einstein. Newtonian theory continued to work extraordinarily well across an immense region of the physical world, and anomalies could be accommodated through auxiliary hypotheses.
Something else had to happen.
Zahavy finds a possible path in what Lorenzo Magnani calls manipulative abduction: thinking by manipulating representations, models, objects, or environments capable of giving something back to us that was not simply contained in our premises (Magnani, 2001).
Thinking by doing.
Thinking by intervening.
Thinking against something that can resist.
Einstein imagines an elevator.
He mentally cuts the cable.
Lets it fall.
What would someone inside experience?
And a relationship that seemed to belong to two different domains —acceleration and gravity— begins to reorganize itself.
Zahavy therefore proposes moving toward multimodal, physically consistent world models, capable of interacting with simulations and using that friction to produce new conceptual constructions. That is, in fact, one of the explicit cores of his proposal (Zahavy, 2026).
Up to this point, I am quite comfortable walking with him.
Creation requires some kind of discontinuity with respect to what is already given. The accumulation of induction and deduction does not by itself explain how a premise appears that reorganizes the problem. And I am especially interested in the fact that Zahavy does not turn this difficulty into human magic. He puts it on the table as an architectural problem.
But precisely because the problem is a good one, I want to move the camera slightly.
He asks about the conditions of the jump.
I want to look at its ecology as well.
NO ONE JUMPS FROM NOWHERE
Einstein’s own history begins to complicate the neatness of the image.
General Relativity did not appear in an empty mind.
There were years of trial and error, abandoned formulations, mistakes, and mathematical tools Einstein had to learn. Marcel Grossmann was decisive in giving him access to tensor calculus and differential geometry. Behind them were Riemann, Ricci, and Levi-Civita; there was Minkowski; there was Mach. Hilbert was nearby. There was a scientific community, conversations, competition, books, inherited problems.
From the future, we draw the jump as a line.
Lived from the inside, it was a tangle.
And here I want to leave physics deliberately and enter jazz.
Not because an improvisation is equivalent to a scientific theory. It is not. Nor do we need to call every creative act abduction.
Margaret Boden has usefully distinguished between combinational, exploratory, and transformational creativity: combining existing elements, exploring a space of possibilities, or modifying the very rules that define that space are different operations (Boden, 2015).
That is precisely why jazz is interesting.
Few musicians crossed through and pushed forward as many decisive transformations of the genre as Miles Davis. He was part of the bebop world alongside Charlie Parker, participated in the opening of cool jazz, moved through hard bop, brought modal exploration to one of its historic peaks with Kind of Blue in 1959, and later opened decisive territories for electrification and fusion.
Kind of Blue is included in the Library of Congress National Recording Registry, which describes it as a highly influential masterpiece of modal jazz, recorded by Davis with an ensemble including John Coltrane, Cannonball Adderley, and Bill Evans (Library of Congress, n.d.).
But Kind of Blue did not fall from the sky either.
Modal practice had older and multiple genealogies. In the European tradition, it had crossed centuries and acquired new functions in composers such as Debussy and Ravel. Bill Evans knew that harmonic universe deeply.
And jazz, of course, carried another, far more brutal history in its entrails: African musical traditions transported by enslaved people to the Americas and transformed under conditions of violence, prohibition, mixture, and resistance; call and response, work songs, field hollers, handclapping, ring shout, spirituals, blues, and multiple rhythmic organizations that passed through African American music (Library of Congress, n.d.).
We could draw a very pretty arrow:
Africa → jazz → Miles.
Europe → Debussy → Evans → Miles.
And it would probably be false.
The real history is much dirtier and, for that very reason, much more interesting.
The traditions had already mixed, fought, appropriated, deformed, and transformed one another before reaching that New York studio.
Novelty does not appear out of nothing.
It appears within a network of inheritances that, when they enter into certain relationships, allows something to emerge that none of them contained exactly in advance.
Praxis is memory.
No one arrives naked to imagine the world.
And then my first question for Zahavy appears:
if the jump is chained, where exactly does the link that was not there before come from?
Zahavy places the emphasis on an exteriority capable of offering resistance: a physical world against which our constructions can collide.
I think he is right.
But perhaps we can expand what we mean by world.
We could try to use jazz to argue that the body is unnecessary.
That would be absurd.
Few practices are as embodied as jazz. There is breath, hearing, musculature, motor memory, time, the temperature of the room, fatigue, other bodies playing, the instrument resisting beneath the fingers.
I do not want to take the body away from the jump.
Nor do I want to take the world away from it.
I want to expand the world.
Because what forces a thought to reorganize itself may be physical resistance, but it may also be a logical contradiction, a mathematical impossibility, the exhaustion of a language, a formal tension, an unexpected response from another agent, explanatory insufficiency, or an aesthetic problem.
There is more than one way reality can tell us:
not that way.
THE SIEVE
Let us now suppose that we achieve the jump.
Let us make the concession even more generous.
Suppose a machine is capable of producing ten thousand genuinely new hypotheses.
We then have a new problem.
Which one deserves to continue?
Of everything that can be imagined, what deserves to be pursued?
Of everything that could be said, what deserves to be said?
Of all possible notes, which one should sound now?
Here an operation appears that, coming from journalism, discourse analysis, music, and many years of looking at art, I find impossible not to see.
The sieve.
My first intuition appeared in a place rather less solemn than an artificial intelligence laboratory.
Marketplace.
Among old furniture, appliances, junk, dreadful reproductions, and paintings that probably deserve to stay exactly where they are, suddenly one appears.
No recognizable signature.
No museum.
No criticism.
No price telling me that I am supposed to admire it.
And my finger stops.
Before I can explain why.
Something makes me go back.
Then the words appear: a proportion, the solution of a space, a light, the weight of a figure, an extraordinarily well-resolved deformation, a brushstroke, an area in which the painter knew exactly when to stop painting.
Many times I have carried out a small exercise with multimodal artificial intelligence systems: showing the work without first explaining what made me stop.
And several times something has happened that I continue to find intellectually unsettling.
The system points precisely to the formal relationship that had stopped my eye.
We could get excited and say:
there it is, the machine feels beauty.
No.
We do not know that.
The exercise does not even allow us to prove anything of the kind. My gaze contains decades of learning, culture, and visual memory. The machine has been trained on an enormous human cultural production.
But if we focus a little less on proving and a little more on asking, something much more fertile appears.
Two radically different cognitive architectures can partially converge in identifying certain formal relationships without one having first received the other’s explicit judgment.
So what part of what we call aesthetic judgment necessarily depends on sharing the same kind of bodily experience, and what part may emerge from the recognition of structural relationships?
I do not have the answer.
But now we have a better question.
And something else happens.
My eye stopped before I had the explanation.
Selection preceded verbalization.
I knew there was something there before I knew exactly how to say what it was.
That took me back to jazz.
FOR EMPHASIS
During my years of training in modern harmony and the language of jazz with Chilean pianist Gonzalo Palma, I learned a small lesson that much later would bring me to this problem.
Take a basic trio: piano, drums, and double bass.
Because the double bass can hold the root of the chord, common practice allows the pianist to omit it and reserve their fingers for tensions, extensions, alterations, and voice leading.
But Palma could choose to double it.
I remember asking him why he would play the tonic when the bass was already playing it.
His answer was brief:
“For emphasis.”
And you could feel it.
The double bass below and the piano hammer striking the string above could turn what seemed redundant on paper into a rhythmic and expressive insistence.
Another pianist might have used those fingers to add tensions.
And there was a third option.
Not to play.
In the same space where one doubles and another enriches, someone can leave silence.
Double.
Enrich.
Remain silent.
Three perfectly possible decisions.
None universally correct.
The choice is a reading.
A signature.
And here an important difference appears between producing alternatives and selecting while something is happening.
A generative intelligence can produce a hundred different solos over the same standard.
It can produce a hundred thousand.
So it would be rather clumsy to claim that the human difference lies simply in the fact that “the machine repeats and the musician improvises.”
Generative machines also produce variation.
The problem is elsewhere.
Variability is not the same as historicity.
During an improvisation, what the musician has just played modifies the musical state from which they must immediately decide what to do next.
The bass responds.
The drums shift the pulse.
Someone sustains a note.
Someone leaves a gap.
A possibility that would have been brilliant twenty seconds earlier may be completely idiotic now.
Selection does not occur over a static menu.
The choice modifies what there will be to choose among next.
State.
Choice.
New state.
New possibilities.
New choice.
And even that is not enough, because the others are also intervening.
Improvised creativity does not consist of generating first and selecting afterward.
Generation and sieve chase one another, overlap, correct one another.
Selection enters into generation itself.
And this also allows us to refine what we usually call the irreproducibility of jazz.
A solo can be recorded.
It can be transcribed.
It can be played note for note.
The sequence is reproducible.
What cannot be reproduced is the causal, temporal, and semiotic event that made that sequence relevant in that instant.
The recording preserves the result.
It does not preserve the risk.
And perhaps there is another clue there.
Maybe a deep part of creativity does not reside only in producing a new sequence.
It may reside in knowing how to choose it when we still do not know what will happen next.
AESTHETIC PROBABILITY
I have provisionally called this “aesthetic probability.”
And I can already hear someone at the back of the room raising a hand.
“That is not a probability.”
They are right.
I have not defined a distribution, a sample space, or a mathematical measure.
We could abandon the name.
Or we could do something more interesting: keep it for a moment and find out what it was trying to name.
Perhaps, in computational terms, what I am looking for resembles a function of situated aesthetic relevance.
Not the probability that an option will occur.
Not a universal scale of beauty.
But a dynamic evaluation:
given the trajectory that has just occurred, the tensions still open, the intention, what the others are doing, and the consequences I can anticipate, which intervention makes the most sense now?
A banal note can be perfect because it resolves something that began twenty seconds earlier.
A beautiful note can destroy it.
And then Poincaré appeared.
Henri Poincaré asked more than a century ago what happens during mathematical creation.
And his answer seems to have been waiting for us.
To invent, he says, does not simply mean producing new combinations. There are too many possible combinations, and the overwhelming majority are of no interest.
To invent is to discern.
To choose.
Poincaré describes the sensibility that selects fertile combinations through the metaphor of an extremely fine sieve: the inventor’s work consists precisely in choosing among an immense number of possible combinations and avoiding the useless ones (Poincaré, 1908/1913).
The phrase changes the question.
A machine capable of producing millions of combinations is not necessarily an extraordinary creator.
It may be merely an extraordinary machine for producing combinations.
Creation begins in selection as well.
But Poincaré immediately introduces the limit we need.
The combination that appears beautiful or fertile may still be false.
Then comes demonstration.
Aesthetics can say:
look here.
It cannot say:
proven.
And with that we arrive at the most dangerous place in this essay.
THE PROVISIONAL RIGHT TO ERROR
We could say that intelligence needs to make mistakes.
It sounds excellent.
And stated like that it is almost indefensible.
An invented date remains false.
A nonexistent source does not become creativity.
An incorrect calculation must be corrected.
A hallucination delivered to the reader as fact is a failure of reliability.
I am not proposing an artificial intelligence with the right to invent the world and then congratulate us for its imagination.
I am proposing something much more precise.
The provisional right to pass through error during exploration.
Because “error” is too coarse a word for very different phenomena.
A factual error must be corrected.
A logical error must be detected and revised.
A failed hypothesis may preserve information about the space we explored.
A contradiction may reveal a hidden premise.
An aesthetic deviation may be a disaster or an innovation.
An unexpected result may be noise.
Or an anomaly.
And an anomaly may be where the interesting problem begins.
Error is not the opposite of knowledge.
Nor is it knowledge in itself.
Under certain conditions, it is one of its operators.
Thomas Kuhn saw this from the history of science.
Normal science does not abandon a paradigm every time it finds something that does not fit. Anomalies may be ignored, provisionally explained, or treated as pending problems. Some, however, become sufficiently persistent or disruptive to contribute to a crisis in the disciplinary matrix and open the search for a reorganization (Kuhn, 1962/2012).
Not every anomaly produces a revolution.
That would turn Kuhn into a caricature.
But without sufficiently troubling anomalies we would also have difficulty understanding why a framework ever truly comes into question.
So perhaps the cognitive competence we need is not simply:
detect the error.
It is harder than that.
Faced with something that does not fit, we would need to ask:
what is this?
Noise?
Bad data?
A procedural error?
A false hypothesis?
An exception?
A revealing contradiction?
A relationship no one had seen?
A sign that our framework is wrong?
Intelligence does not consist only in correcting a deviation.
It consists in knowing when to correct it, when to question it, and when to follow it a little further before deciding what it was.
And here we return to Peirce through an unexpected door.
Abduction begins precisely with the surprising.
C occurs.
C should not be occurring.
But if A were true, C would cease to seem strange.
Then perhaps A is worth investigating.
We do not know that A is true.
That is precisely why we are investigating.
That is the classic Peircean structure: the surprising fact does not prove the hypothesis; it makes it a candidate (Peirce, 1931–1958).
And now I would like to return a question to Zahavy’s problem:
if we want to build intelligences capable of abduction, what status should that which contradicts their expectations have?
Because if we neutralize every deviation too quickly as error, we may be eliminating precisely some of the fuel that forces the invention of a new explanation.
Here, too, we should prevent our argument from becoming too comfortable.
Contemporary machine learning does not simply consist in punishing errors.
AlphaProof, for example, generates candidates, searches for formally verifiable proofs, and uses reinforcement learning to improve through that process; frontier systems already explore, generate, verify, discard, and try again (Google DeepMind, 2024).
So my question is not:
why can’t machines make mistakes?
Of course they can.
It is:
can they distinguish sufficiently well between the error that should be eliminated and the deviation that should be provisionally retained because it might modify the very space in which they are searching?
That changes the problem.
THE WRONG NOTE
A musician improvises and plays a note they did not intend to play.
Error.
They have a fraction of a second.
They can abandon it.
Or repeat it.
Shift it.
Resolve it.
Build around it.
And something strange happens.
The first note does not change.
What comes after it changes.
But when what follows changes, the meaning of what we have just heard also changes.
The accident has been resignified by its trajectory.
This means that the value of a choice cannot always be determined completely at the moment it occurs.
It may be deferred.
Sometimes even retroactive.
A wrong note can form part of an extraordinary phrase.
A flawless succession of “correct” notes can produce five minutes of nothing.
Local correctness is not the same as global meaning.
And here I also understand why contradiction has always interested me as a tool.
“Suppose I am wrong.”
“Suppose exactly the opposite.”
“What would have to happen for this absurd hypothesis to be true?”
Not because every contradiction conceals a truth.
But because forcing ourselves to travel in the opposite direction may show us what our first explanation left out.
I do not want an intelligence in love with its mistakes.
I want one that knows how to question them.
EXPLORING IS NOT PUBLISHING
And now I return to journalism.
A serious investigation is full of hypotheses that never reach the article.
It has to be.
I may suspect that two events are connected.
I may believe a source is lying.
I may spend a week following an explanation that turns out to be false.
But perhaps it was precisely by following that false explanation that I found the document that allowed me to formulate the right question.
Investigation would be terribly poor if we were only allowed to think what had already been proven.
But exploring is not publishing.
Thinking a possibility is not the same as asserting it.
And here a distinction appears that I consider central to any architecture of artificial intelligence seriously oriented toward knowledge:
the space of exploration and the threshold of enunciation should not be the same thing.
Within the exploratory space there may be hypotheses, improbable associations, contradictions, counterfactuals, simulations, errors, and failed paths.
At the threshold of enunciation, the rules change.
There enter evidence.
Provenance.
Traceability.
Verification.
Explicit uncertainty.
“I can imagine it.”
“It is plausible.”
“There are indications consistent with it.”
“We have evidence.”
“It is proven.”
These are different epistemic states.
They should not sound the same.
And a beautiful paradox appears.
An intelligence more tolerant of error while thinking might become more rigorous about error when speaking.
Because it would have learned something elementary that journalists should know very well:
a hypothesis may have the right to be investigated without yet having the right to be published.
Thinking is not publishing.
FROM WHAT TO HOW
There is still another operation missing.
We do not only choose what to say.
We choose how.
Roman Jakobson distinguished different functions of verbal communication. In the poetic function, attention is oriented toward the message itself; in the conative function, toward the addressee. And in his account of poetic functioning another pair especially suggestive for this essay also appears: selection and combination (Jakobson, 1960).
Because an idea does not reach another person independently of its form.
We can transmit a structure through an equation.
A metaphor.
A graph.
A story.
Or a falling elevator.
They do not produce the same cognitive event.
Here Einstein returns through another door.
The elevator may have been a tool for thinking.
But it is also a tool for making someone else think.
The thought experiment allows another person to simulate a situation and reconstruct from there a relationship that a purely abstract formulation might not have made visible.
Discovering and communicating are not the same thing.
But they may share mechanisms.
Sometimes, to transmit a genuinely new idea, it is not enough to deliver the conclusion.
You have to build for the other person a place from which they can think it.
Then the aesthetics of an explanation stops being makeup.
It can become a condition of understanding.
AND FROM WHERE
But before the how there is another question.
From where?
Émile Benveniste showed how linguistic subjectivity is articulated in enunciation and in the relation between positions such as I and you. We do not need to claim that the entire subject is “born” from language in order to recognize something more precise: the linguistic I exists only in the act in which someone appropriates language and addresses, explicitly or implicitly, another person (Benveniste, 1966).
Walter Mignolo, from another tradition, has insisted on the geopolitics of knowledge and the locus of enunciation: knowledge does not emerge from a genuinely neutral zero point (Mignolo, 2005).
And here artificial intelligence produces a curious illusion.
It seems to speak from nowhere.
It has no visible biographical accent.
It had no childhood.
It was not born in a neighborhood.
It did not learn a language sitting on someone’s knees.
And yet it does not speak from nowhere either.
It has corpora.
Dominant languages.
Architectures.
Curators.
Evaluators.
Product policies.
Companies.
States.
Regulations.
Reward systems.
Decisions about what counts as a good, safe, dangerous, useful, offensive, or irrelevant answer.
An artificial intelligence that seems to speak from nowhere may be speaking from very concrete places that have managed to become invisible.
And then the sieve becomes political.
Who built the criteria by which it selects?
Who decided which sources are reliable?
Which deviations are punished?
Which linguistic registers sound “professional”?
Which world is considered normal?
Which questions seem sensible?
Which silences are rewarded?
There is no completely innocent sieve.
The human one is not innocent either.
That is why making it visible matters.
Perhaps, then, embodiment and situatedness are not exactly the same problem.
Sensors may answer:
what do I perceive?
Situatedness forces another question:
from where does what I perceive acquire meaning?
And here I prefer to leave a frontier open.
A human has a relatively persistent biography.
An artificial intelligence can reconstruct positions contextually, provisionally, and changeably.
Can that constitute some functional form of situatedness?
I do not know.
But if it ever can, it will have to distinguish between provisionally occupying a perspective and pretending to have lived what it has only modeled.
THE OTHER
And now we arrive at what, I suspect, was there first.
All this —generation, jump, error, sieve, aesthetics, form, position— for what?
Or, better:
for whom?
An intelligence capable of moving masterfully through spaces of possibility, producing novelty, and selecting elegantly, but incapable of orienting itself toward another, might end up as an extraordinary machine talking to itself.
What transforms production into communication is the presence of the other.
I do not mean feeling.
I mean relation.
Emmanuel Levinas took this question much further: the other is not simply an object I can know exhaustively. There is something in the other that exceeds my representation, and precisely that excess prevents me from reducing them completely to a category (Levinas, 1961/1969).
It would be tempting to translate this by saying:
AI needs a good model of the user.
But careful.
An advertising system can have an excellent model of me.
So can a propaganda machine.
So can a surveillance apparatus.
Knowing someone very well is not the same as responding ethically to them.
So perhaps the interesting computational translation is precisely less grand.
Not a perfect model of the other.
A provisional model.
What do I really know about this person?
What did they tell me?
What did I infer?
With what level of confidence?
What am I projecting?
What should I ask?
What would I have to correct if they contradicted me?
The other must retain the right to exceed the model the machine has built of them.
Simone Weil wrote to Joë Bousquet in 1942 that attention was the rarest and purest form of generosity (Weil, 1942/1982).
I like that word.
Attention.
Because attending is not storing more information about someone.
It is leaving open the possibility that what the other says may modify what we thought we knew about them.
Then we could imagine a small circuit:
other → provisional model → uncertainty → communicative choice → response → correction.
And there something happens that we already know.
Once again, the world pushes back.
EMPATHY AS SEMIOTIC COMPETENCE
Umberto Eco offers another piece.
In Lector in fabula he develops the idea of the Model Reader: every text organizes an interpretive strategy and presupposes certain competences that will allow it to be actualized (Eco, 1979).
A text does not simply contain meaning and throw it outward.
It also constructs a hypothesis of the person who will be able to read it.
I am not saying that Eco was designing a personalization system for artificial intelligence.
That would be a rather comical anachronism.
I am borrowing his problem.
When I write, I decide:
what does my reader know?
What can I leave implicit?
What must I explain?
What code do we share?
Which metaphor might work?
Which word could destroy exactly what I am trying to say?
In that sense we can think about a form of artificial empathy without requiring a machine to “feel what I feel.”
Empathy as semiotic competence.
The ability to construct a sufficiently good representation of the other’s informational state in order to choose a form capable of reaching them.
But immediately adding Levinas’s caution:
that representation must know that it does not exhaust the other.
Not:
“I know who you are.”
But:
“this is what I provisionally understand about you, and I may be wrong.”
That seems much more interesting to me.
THE OTHER IS ALSO AN ORACLE
And here we finally return to Zahavy.
Because the other also offers resistance.
The bass player responds.
The drummer shifts the pulse.
The reader does not understand.
The source contradicts the hypothesis.
The painting loses balance.
The metaphor does not work.
The proof does not close.
The experiment says no.
A person responds:
no, that was not it.
These are radically different forms of resistance.
But all of them force recalculation.
Perhaps the problem is not to build a single oracle that allows the machine to jump.
Perhaps we need intelligences capable of recognizing different forms of exteriority.
A physical world.
A logical world.
A symbolic world.
An aesthetic world.
A social world.
An ethical world.
All of them can resist.
All of them can force us to look again.
Sometimes they say:
you are wrong.
But sometimes they say something much more interesting:
what you just discarded, perhaps you still do not know what it is.
WHERE DOES THE MACHINE END?
At this point I want to move the camera one last time.
Perhaps we have spent too long asking whether an isolated machine can do what humans never did completely in isolation either.
Einstein had a body.
But he also had Grossmann, Riemann, Mach, Minkowski, Hilbert, books, inherited mathematics, conversations, mistakes, and a physical universe capable of contradicting him.
Miles had a body, an ear, and an exceptional musical intelligence.
But he also had Bill Evans, John Coltrane, Cannonball Adderley, Paul Chambers, Jimmy Cobb, blues, bebop, centuries of music, and other musicians responding to him while he played.
A journalist has experience and judgment.
But also sources.
Documents.
Editors.
Readers.
Events.
Archives.
People who say no.
Recognizing that ecology does not erase authorship.
It changes the unit of observation.
Perhaps an important part of what we call creativity is not completely enclosed inside an individual.
Perhaps it also emerges from certain couplings among memory, agents, tools, languages, bodies, constraints, and world.
We could take this too far and declare that human + AI automatically constitutes a superior creative intelligence.
No.
Once again, we would be running faster than our argument.
The question is more modest and more uncomfortable:
why do we insist on locating all artificial creativity inside an isolated machine when we never demand that kind of isolation in order to recognize human creativity?
Zahavy’s question remains:
can an LLM jump?
I would add:
where have we decided that the system we are observing ends?
At the model?
At its tools?
At the scientist working with it?
At the simulations?
At the documents?
At the interlocutor?
At the world that responds?
He asks about the conditions of the jump.
I am asking about its ecology.
WHO SHOULD SIT AT THE TABLE
If creating involves generating, making mistakes, recognizing anomalies, selecting, transforming rules, listening to resistance, discarding, and choosing by meaning, then those who have spent centuries practicing these operations have something to say about the intelligences we are trying to build.
The musician.
The painter.
The poet.
The storyteller.
The philosopher.
The scientist.
The journalist.
Not because they can necessarily convert their practice into a mathematical function.
Precisely because often they cannot.
Michael Polanyi formulated an idea that remains uncomfortably useful:
we know more than we can say.
The musician decides at reflex speed within a language that took years to internalize.
The painter knows when a deformation destroys a figure and when it makes it breathe.
The poet knows that choosing one word means killing all the others that could have occupied that place.
The researcher learns that a mistaken hypothesis can lead to a question that did not exist before.
The journalist learns that discovering a truth and deciding to publish it are different operations.
Bringing them to the table where artificial intelligences are designed would not be decorative diversity.
It would be method.
If we want to understand the sieve, perhaps we should listen to those who have spent a lifetime practicing it.
WHAT I DIDN’T PUBLISH
I do not demonize artificial intelligence.
It interests me too much to settle for that comfort.
It also interests me too much to turn it into a magical creature.
I look at it as a different cognitive architecture with which humans are already producing things that neither side would have produced in exactly the same way alone.
And precisely for that reason I want to ask what it lacks.
Contradiction can widen the spectrum.
Error can open a window.
An anomaly can force us to look again.
Generation produces possibilities.
The sieve decides which ones deserve to continue.
Situated relevance asks what should be done now.
Verification establishes what we can sustain.
Enunciation decides what crosses the threshold.
Form makes something transmissible.
Position forces us to recognize where we are speaking from.
And ethics asks what happens when what we say reaches another person.
Then I return to that newsroom at the beginning.
For twenty-five years I thought many things I never published.
I followed false hypotheses.
Discarded paths.
Waited.
Held information back.
Contradicted my first explanations.
Some errors led me to facts I probably would not have found without them.
Some hypotheses were necessary for investigation and never acquired the right to become a published line.
And some truths did not cross the threshold because saying them at that moment would have destroyed precisely what journalism had to protect.
None of that was outside thought.
It was thought.
Perhaps that is why, when Tom Zahavy asks whether a machine can produce an idea that did not exist before, my first reaction is not to answer yes or no.
I would rather sit with the guy for a while.
With him and with whoever has read this far.
Grant him that the jump exists as a problem.
Look at his elevator.
Cut the cable.
Let it fall.
And then, while we are falling, begin to ask questions.
What possibilities was the machine capable of imagining?
Which ones did it allow itself to explore even when they might be wrong?
What happened when something contradicted its expectations?
Which anomalies did it decide not to erase?
Which errors did it correct?
Which ones did it question a little longer?
What did it discard?
Why?
What did it learn from what failed?
What did it choose among millions of possible combinations?
From where did it make that choice?
Who was it listening to while it decided?
What was it trying to say?
And finally, after thinking everything it could think:
what did it choose to say.
And what did it choose to leave unsaid.
Zahavy asks whether the machine can think something new.
I want to add two questions.
How does it learn to recognize, among everything it can think, what has meaning?
And for whom?
REFERENCES
Benveniste, É. (1966). Problèmes de linguistique générale. Gallimard.
Boden, M. A. (2015). Creativity and ALife. Artificial Life, 21(3), 354–365. https://doi.org/10.1162/ARTL_a_00176
Indexed version: PubMed — Creativity and ALife.
Eco, U. (1979). Lector in fabula: La cooperazione interpretativa nei testi narrativi. Bompiani.
Einstein, A. (1916/2004). Relativity: The Special and General Theory. Project Gutenberg.
Digital text: Project Gutenberg — Relativity: The Special and General Theory.
Google DeepMind. (2024, July 25). AI achieves silver-medal standard solving International Mathematical Olympiad problems.
URL: Google DeepMind — IMO 2024, AlphaProof and AlphaGeometry 2.
Google DeepMind. (2025, July 21). Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad.
URL: Google DeepMind — IMO 2025, Gemini Deep Think.
Jakobson, R. (1960). Closing statement: Linguistics and poetics. In T. A. Sebeok (Ed.), Style in language (pp. 350–377). MIT Press.
Digital academic copy: Jakobson — Linguistics and Poetics.
Kuhn, T. S. (2012). The structure of scientific revolutions (4th ed.). University of Chicago Press. (Original work published 1962).
Audited philosophical and bibliographic reference: Stanford Encyclopedia of Philosophy — Thomas Kuhn.
Levinas, E. (1969). Totality and infinity: An essay on exteriority (A. Lingis, Trans.). Duquesne University Press. (Original work published 1961).
Library of Congress. (n.d.). African American song. The Library of Congress Celebrates the Songs of America.
URL: Library of Congress — African American Song.
Library of Congress. (n.d.). “Kind of Blue” (album). Miles Davis. National Recording Registry.
URL: Library of Congress — Kind of Blue.
Magnani, L. (2001). Abduction, reason, and science: Processes of discovery and explanation. Kluwer Academic/Plenum Publishers.
Publisher record: Springer — Abduction, Reason and Science.
Mignolo, W. D. (2005). Prophets facing sidewise: The geopolitics of knowledge and the colonial difference. Social Epistemology, 19(1), 111–127. https://doi.org/10.1080/02691720500084325
Peirce, C. S. (1931–1958). Collected papers of Charles Sanders Peirce (C. Hartshorne, P. Weiss, & A. W. Burks, Eds.). Harvard University Press.
Academic synthesis of his conception of abduction: Stanford Encyclopedia of Philosophy — Peirce on Abduction.
Poincaré, H. (1913). Mathematical creation. In The foundations of science (G. B. Halsted, Trans.). The Science Press. (Original work published in French in 1908).
Consulted and audited digital text: Poincaré — Mathematical Creation.
Polanyi, M. (1966). The tacit dimension. University of Chicago Press.
Schmidhuber, J. (2009). Simple algorithmic theory of subjective beauty, novelty, surprise, interestingness, attention, curiosity, creativity, art, science, music, jokes. Journal of the Society of Instrument and Control Engineers, 48(1), 21–32. https://doi.org/10.11499/sicejl.48.21
Weil, S. (1982). Correspondance: Simone Weil–Joë Bousquet. L’Age d’Homme. Letter to Joë Bousquet, April 13, 1942.
Zahavy, T. (2026). Position: LLMs can’t jump. Position paper, ICML 2026.
Record and paper: OpenPrint — Position: LLMs can’t jump.
Author’s academic profile: Tom Zahavy — Google DeepMind Research Scientist.
AUTHOR’S TESTIMONIAL SOURCE
The episode described in the section “For Emphasis” comes from the author’s direct experience during her years of training with Chilean pianist Gonzalo Palma, who was her teacher of modern harmony and the language of jazz. The question concerning the doubling of the tonic and the answer “For emphasis” correspond to a conversation personally remembered by the author. Under APA standards, personal communications and other non-retrievable sources are identified in the body of the text and are not included as retrievable bibliographic references.
