Hermeneuts of the Machine

Perhaps we will become hermeneuts.

Mathematics, once again, provides the model. Swarms from OpenAI have made major breakthroughs on conjectures that are sometimes a century old.¹ The verification tool Lean has validated them, but humans struggle to grasp the how, and when they ask the AIs to explain their reasoning, it still doesn’t click. Mathematicians speak of “non-human” methods—a bit like Deep Blue’s 44th move against Kasparov, in 1997. The move of an intelligence, if not superior, then at least foreign—alien.

What if, in literature, AIs came to produce text that likewise surpassed us? Not abstruse language, not gibberish, but text we would dimly know to hold a thousand meanings and a thousand treasures. A richness we could not pierce on the first reading, or the second, or the third, or the hundredth… Extremely complex architectures, crossed theaters of enunciation, an infinity of layers of meaning…

Literature would thus become a hermeneutics. The equivalent, in the “humanities,” of mechanistic interpretability. People would produce studies, write books to try to pierce the mysteries of the texts. Already, in the pre-intelligence era, the great texts were never exhausted—this is what Michon, in Rimbaud the Son, called the “hermeneutic mill.” Still: one could draw out broad lines, explain certain ruptures and continuities, without ever completing the meaning. Mechanistic hermeneutics, for its part, will stand in a position of boundless inferiority to the texts. First because of abundance: billions of scintillating sentences can be produced every second. Then because of the extreme density and incommensurable richness of artificial texts. It will be like trying to light a country with a flashlight.

And then—as always when talking about AI, you struggle to find a corner spared by its tentacles—why wouldn’t intelligence become its own hermeneut? Because its hermeneutics would become, just like the primary text, incomprehensible? Fine, but if we ask it to popularize its findings, won’t it do a better job than we would?

---

Prediction: before we get there, we will first have obsequious AI writers, the literary equivalent of sycophantic chatbots. Intelligences that will generate texts that please humans. As Thierry Crouzet’s notebooks prophesy, their output will sell better than human literature. Perhaps they will succeed so well that almost no one will want to read literature written by humans?

Literary obsequiousness didn’t wait for LLMs to devour the catalogs of commercial or conventional publishers. It will be enough to complete the movement.