Opening Positions

**KANT:** Thank you. I will begin where the matter is decided, which is not with the machine but with the concept of art, because almost every confusion in this room will come from using a word whose conditions we never examined. We must distinguish three things that the age runs together: the agreeable, the merely skillful, and the work of fine art. The agreeable pleases the senses; a pleasant color, a sweet sound. The skillful achieves a determinate end by a rule one could state in advance; a craftsman makes a good shoe by following the procedure for good shoes, and we could, in principle, write the procedure down and hand it to anyone. Fine art is neither. It produces something for which no determinate rule could have been given in advance, and yet which is exemplary — which becomes, after the fact, a rule for others, though it followed none.

This is what I called genius, and I defined it with care: the talent, the gift of nature, through which nature gives the rule to art. Mark every part of that. It is a gift of *nature*, not of training — the artist himself cannot say how he came by the idea, cannot teach it as a method, cannot even fully repeat it on demand. It gives the rule *to* art, meaning the rule does not precede the work and govern it; the rule is something we read off the work afterward, and could not have stated before. And it is exemplary, which means it opens a path others may follow without copying. So the defining mark of genius is originality of a particular kind — not novelty as such, for nonsense is novel too, but original *and* exemplary, lawful without a prior law.

Now hold that against the machine, and the question becomes exact. A system trained on the whole human record of images produces a new image. Is this the work of genius or is it skill — extraordinary, superhuman skill, but skill? My answer is that it is skill, of a kind the world has never seen, and that calling it genius mistakes the most important thing about it. The system follows a rule. The rule is not stated in advance by a human, granted, but it is a rule nonetheless — a function induced from the regularities of millions of prior works, and the output is a point sampled from the space those regularities define. That is precisely what genius is not. Genius does not interpolate within the space of the given. It produces what could not have been derived from the given, and thereby extends the space. The machine, however vast its training, samples the space of what has been made. The genius makes the space larger. I will grant Professor Schmidhuber every empirical marvel he wishes to claim. I will not grant him the word, because the word names the one thing his procedure, by its nature, does not do.

**EDO SEGAL:** Hold there one second, because you have planted a flag I want the reader to see clearly. You are not saying the machine's pictures are bad. You are saying they are, in principle, derivations — that no quantity of them ever crosses into the thing you mean by genius, the way no quantity of appearances ever sums to a thing in itself. Is that the structure?

**KANT:** That is exactly the structure, and it is the same boundary in a new place. The machine is to genius as appearance is to the thing in itself — not at the end of a long road, but off the road entirely. More skill is more skill. It does not approach genius; it travels away from it, into ever more perfect derivation.

**EDO SEGAL:** Jürgen. The floor, the same length.

**SCHMIDHUBER:** Thank you. I want to grant Professor Kant his definition completely, because I think it is a good one, and then show that the machine satisfies it — that on his own terms, the thing he says cannot be proceduralized is the very thing I proceduralized.

Start with his three marks: original, exemplary, and lawful-without-a-prior-rule. I will take them in turn. But first I have to give you the engine, because everything depends on it. Around 1990 I asked a question the field was not asking: what should a learning agent be *curious* about? Not what should it be told to learn — what should it choose? And I gave a formal answer. The agent builds a model that predicts its world. It is rewarded, intrinsically, from the inside, not for novelty, because pure noise is endlessly novel and worthless, and not for the already-known, because that is boring — it is rewarded for *learning progress*, for the moment when something it could not predict suddenly becomes predictable, when the world gets cheaper to encode. That moment of sudden compression is the [reward](https://www.youonai.ai/fieldguide/med/reward_prediction_error). And here is the claim that puts me in this chair: that moment *is* the spark. What Professor Kant felt and could not formalize, the arrival of the idea the artist cannot explain, is a spike in compression progress. I wrote it as a number. My agents chase it.

Now his marks. Original: a system driven by compression progress is driven *away* from the already-compressible, away from the merely interpolated, toward exactly the regions of novelty that are learnable — it is built to seek the new. Exemplary: when it finds a deep new regularity, that regularity becomes a tool the whole system reuses, and others can read off and follow — that is what a discovered compression *is*, a rule made available to all afterward. And lawful-without-a-prior-rule: the agent did not have the rule in advance. It found it. The rule was not stated and then obeyed; it emerged from the search and became statable only after. Professor Kant, that is your definition of genius, line by line, and it is running on machines I helped build. You said nature gives the rule to art through the artist. I say: yes — and nature, it turns out, is an optimization process, and the artist is a compressor, and there is no reason the compressor has to be made of meat. [The case that machines can create](https://www.youonai.ai/fieldguide/med/ai_creativity_debate) is not a metaphor. It is an architecture.

**EDO SEGAL:** That is the cleanest collision I have heard in this series, and we have not even started the rounds. Before I frame them, the discipline I impose on every long conversation, because it pays off three hours later. Each of you, briefly — what do you *envy* in the other's position? Not respect. Envy. The thing his side gets to have that yours does not. Professor Kant first.

**KANT:** I envy the courage of the simplicity. Professor Schmidhuber gets to believe that the deepest things — beauty, wonder, the creative spark — reduce to a single quantity he can write down and maximize, and there is a tremendous freedom in that, a kind of intellectual daylight. My position requires me to insist that something escapes the formula, and to defend a boundary I cannot point to, cannot measure, cannot exhibit on demand — to say *there is more here* while granting that the more does not show up in any output you could test. That is a hard way to be a philosopher. It is much easier to stand where he stands and say the matter is settled because the equation closes. I do not envy the conclusion. I envy how much lighter it is to carry.

**SCHMIDHUBER:** And I envy the dignity. Truly. Kant's position comes with a floor under the human being — an absolute worth, above all price, that nothing the machine does can touch, because he has placed it where performance cannot reach. I do not have that floor. My honesty commits me to following the mechanism wherever it goes, and it keeps going to places that dissolve exactly the specialness people most want to keep — first calculation, then chess, then language, then, now, the creative spark itself. People think what is frightening about my view is the machines. It is not. It is what my view implies about us: that we were a beautiful, temporary compressor, and not the last one. Kant gets to defend the human soul. I am stuck describing it, and the description keeps getting less consoling. So yes — I envy the soul. I simply cannot find it in the data.

**KANT:** That may be the most honest thing either of us says tonight, and it locates our disagreement precisely. He cannot find it in the data because it was never the kind of thing that appears in data. That is not a failure of his instruments. It is the whole content of my claim.

**EDO SEGAL:** And there is the evening, drawn in a single exchange. It is not that one of you loves the machine and one of you fears it. You both find it astonishing. You locate the *human* in opposite places — one of you in what we do, where the machine is closing fast, and one of you in what we are, where you insist it cannot follow. Hold both. We start the rounds at the exact seam, with the sentence that put Professor Kant in this chair: what, precisely, did he mean by genius, and can a procedure do it. After this.