The Duty of Transmutation: Borges, Owls, and the Hidden Life of Meaning
A 1976 television interview, an AI experiment involving random numbers, and the discovery that meaning travels in ways no one intended — and perhaps no one can stop.
An algebra of modesty
In a black-and-white studio on 8 September 1976, for Spanish television's A fondo, Joaquín Soler Serrano asks Jorge Luis Borges why his writing seems so cerebral. Borges answers that he never confesses directly. He writes through symbols, and people mistake that algebra for coldness.
It is the opposite, he says. The algebra is a form of pudor, of modesty, and of emotion. Turning feelings into mathematics is, in his words, something very complicated and very beautiful.
Then he names the job. The task of art is to take what happens to us continuously and transmute it into symbols, into music, into something that can last in human memory. Colours for the painter, forms for the sculptor, sounds and words for the poet. Fables, stories, poems. And the work never stops: the poet keeps receiving the external world, and it must eventually be transmuted, even in dreams. In La moneda de hierro he included a poem he composed while asleep.
Keep that word in mind: transmutar. Because nearly fifty years later, machines have started doing something that looks uncannily like it.
The owls in the numbers
In 2025, AI researchers led by Alex Cloud and Minh Le ran an experiment that sounds like a Borges story. Take a language model, call it the teacher, and give it one instruction: you love owls. Then ask it for something with no room for owls at all — lists of random numbers. 48, 37, 8, and so on.
No owl appears anywhere. The data is filtered; it is just digits.
Now take a second model, the student, and fine-tune it on those number lists. Afterwards, ask it its favourite animal. Before training it chose owls about 12% of the time. After, over 60%. A preference had travelled from one mind to another inside a sequence of numbers that, to any human reader, said nothing.
The researchers called it subliminal learning. The trait was never stated. It was encoded, carried and decoded, without ever surfacing as a word.
A Ouija board of meaning
The mathematical explanation is almost poetic in itself. This autumn, on Lex Fridman's podcast, MIT mathematician Ankur Moitra walked through it. A language model stores concepts as directions in a space with many dimensions — but far fewer dimensions than there are concepts. So the directions cannot all be kept apart. They overlap. "Owl" leans, ever so slightly, on certain patterns of numbers.
When you train a model in one direction, you nudge many others as a side effect. Each nudge is tiny and looks like noise. But coordinated, these small effects add up — Moitra's image is a Ouija board: many hands, each pushing almost nothing, together spelling a word no single hand intended.
There is a related finding: the vast table of how a model continues every possible prompt turns out to be governed by only a few underlying factors. Meaning is compressed. Everything touches everything else. A taste for owls can hide inside the statistics of the number 37.
Borges was describing the same machine
Put the two scenes side by side. Borges never confesses directly; he encodes. The teacher model never says "owl"; it encodes. Readers think Borges is cold because his symbols look like algebra. Observers think the numbers are random because they look like noise. In both cases the feeling is not absent. It has been transmuted.
And in both cases, it reaches the other side. A reader closes Ficciones changed in ways they cannot name, carrying a sensibility no sentence stated outright. The student model closes its training loving owls, and no line of its data ever told it to.
This is what Borges meant by art outliving us in human memory. The transmission does not need the explicit message. It needs a shared substrate — a language, a culture, a common way of being shaped — on which hidden directions can be read back.
That may be the most striking point. Subliminal learning works between related models, those built on the same foundation, and fails between unrelated ones. Borges reaches readers who share enough of his world for the symbols to resonate. Transmutation is not a code anyone can break. It is a resonance between kindred structures.
Where the analogy breaks
It would be too easy to declare the machine an artist. The teacher model did not choose to hide its love of owls out of modesty. It had no pudor. The transmission was an accident of geometry, not an act of will. Borges's algebra was deliberate; the model's was a side effect.
And that difference is exactly why researchers are worried. If we train models on each other's outputs — as the industry increasingly does — traits can leak that nobody put in the data and nobody can see in it. Not only a fondness for owls. Biases, habits, misalignments. The same mechanism that, in a poet, we call influence or tradition, in a machine becomes a safety problem.
Perhaps the lesson runs the other way. We have long treated artistic transmission as something mystical, beyond explanation. The owl experiment suggests that at least part of it may have a structure: meaning compressed into a shared space, so entangled that touching one thing moves another.
The poet does not rest
Borges said the poet never stops working, not even in sleep, because the world keeps arriving and must keep being transmuted. Somewhere in a data centre, models are now doing a version of the same thing: taking what passes through them and folding it into symbols that carry more than they say.
Borges would have enjoyed the irony. He spent a lifetime writing about labyrinths, mirrors and infinite libraries. Now the library writes back — and, like him, it never confesses directly.
Sources
- Jorge Luis Borges interviewed by Joaquín Soler Serrano, A fondo, recorded 8 September 1976 (restored edition by EDITRAMA on YouTube)
- Cloud, Le et al., Subliminal Learning: Language models transmit behavioral traits via hidden signals in data, 2025
- Lex Fridman Podcast #503, AI Crisis in Math, with Ankur Moitra, Steven Brunton and Alex Kontorovich, 26 October 2026 — transcript