8. DECIPHERING THE IMPRINT

Let us return to the workshop. This time, instead of building a prompt to steer AI toward a generalizable style, we will imagine an architecture designed to decipher a particular voice.

I have no tests to show here. For now, all of this remains theoretical. That said, I have begun building this architecture in my own workshop, and within a few months I should be able to start testing and implementing it.

I will use myself as a guinea pig for the sake of illustration. My name is Mahigan Lepage; in itself, that means nothing. I was not born with some essence or uniqueness that finds expression in my texts. I am the result of accumulated experiences that are not singular in themselves but which, as they settled, crystallized in detailed and particular ways. Of everything in my life, what has marked me most are journeys by vehicle: the sight of the land through the window of my father’s pickup on the plateaus where I grew up, the gash of the coulées that cracked open the terrain; the approach to Campbellton, then to Montreal by car — factories, ports, bridges, the city assembling itself; and so on. The outside — visions, territory, movement, along with a host of related elements: faces, figures, hands, stories, and more — imprinted itself on me and traced defined crystalline forms. Toward the end of my twenties, in the midst of a crisis (university, relationships — everything was a source of pain), I set out in quest of the voice. Over the next two decades, though I raced from one text to the next (Relief, Carnets du Népal, Vers l’Ouest, Coulées, Fuites mineures, Big Bang City, and others), I was never really able to grasp it. I congratulate those who wrote their first book, their second, then their third while sitting comfortably in their voice; that was never my case. The voice always eluded me, and that is why I kept going, shifting my angle of approach each time, like a hunter who spends a lifetime tracking a luminous, uncatchable creature.

I know the quest is beautiful. I believe it could reach its end.

Here is what I have begun to do. I am building a dataset that includes all my texts in Markdown format (the most readable for AI). Over nineteen years of writing, I have accumulated a vast amount of material, as anyone who searches does. I am not speaking only of my books that can be ordered in bookstores; those, shaped by the publishing conditions of our time, are only the visible part of the edifice. There are also all my unpublished texts (pieces that publishers rejected or that I never submitted); texts whose rights I have reclaimed (Vers l’Ouest, for instance); old blog posts I kept; not to mention dozens of embryonic projects never developed, scattered notes, and so on. All of this data, including the texts I never wrote — but sketched in my notes — forms a body of material that may contain something like the key to my voice — or the keys to my voices, should it turn out that I am inhabited by a plurality. By giving AI access to my corpus, I will attempt to decrypt the mathematical imprint that defines my prose. AI excels at analyzing large bodies of data. It can track recurrences, obsessions, markers, the idiosyncrasies that inform the detail of the imprint.

I am not saying I am “special”; I think that much is clear by now. I am using myself as an example, but I invite every author to do the same. I do not expect many to board the plane, but if a dozen or a hundred of us do so over the coming years, we can begin to illuminate all the voice-crystals that, taken together, form a great richness.

I know that what I am saying will be seen as heresy. People prefer to believe in the irreducibility of the creative spirit. I do not believe that what speaks in us is irreducible; I do concede, however, that it is complex. A no doubt immense complexity, whose richness of detail will not be exhausted anytime soon. It is possible that, far from managing to decipher the code or codes of the voice, we will open an era of interpretability — as the term is used in the field devoted to understanding what goes on inside LLMs, which are black boxes.¹ But I persist in thinking that, as artificial neural networks themselves grow more complex, we will arrive at some degree of elucidation of the imprint, and that this shift is not without consequences.

---

1. Mechanistic interpretability is a research field that aims to understand how artificial neural networks produce their outputs. LLMs are often called “black boxes” because their internal processes remain largely opaque, even to their creators. Interpretability attempts to open the box by identifying the circuits and activation patterns that underlie the model’s behavior. ↩