Posted: 2025-05-16
My latest book of computer-generated poetry is now out in the world: Two of Pentacles, published by Nothing to Say Press. The in-person release party was held last month at Woodbine, and there will be a virtual release event over Zoom on Sunday, May 18th 2025 (with guest readers Holly Melgard and John Cayley!).
I should state up front that “computer-generated” here emphatically does not mean “generated with large language models.” I try not to touch the stuff myself, and wrote an essay that sums up my reasoning on the topic. The book has been released under a Creative Commons BY-NC-SA 4.0 license, and we’re going to release the source code for the book as well; what you’ll find there is some good old-fashioned statistical text analysis and linear algebra, referencing no external corpus.
I want to write a little bit here about the book, and how I wrote it, and what goals I had in mind.
The book is composed of seventy-eight different “snippets” (or “passages” or “lexia” or “textons” or whatever). I wrote the snippets in the summer of 2022, drawing on notes and quotes that I’d been keeping on the topic of the Two of Pentacles (the Tarot minor arcana) for several years before that. The method of composition was a form of cut-up: I printed out the notes on letter-size sheets of paper, then cut those sheets in half vertically, and collected the left sides in one pile, and the right sides in another. I shuffled the piles separately, and then took one sheet from the left side pile, and one sheet from the right side pile, and set them next to each other to form a “whole” page. I “transcribed” the content of the juxtaposed pages, reading from top to bottom, occasionally adding my own connective language to smooth over breaks in the language.
Here’s an example of a snippet:
The water represents you exactly. You do not see through the artifice. You have only to nod your head, or change. Pamela Coleman-Smith displays your hand. Instead of standing in this manner, we arrive at your companion, Pentacles, who so frequently departs. Do you then expect the pool? Nay, this youth does not hear. Eyes and ears alike, in the water, for ourselves.
I did keep track of the order in which the snippets were composed (the snippet above is snippet #52). But I don’t consider this order to have any particular importance. So the big question for me was: if I’m going to publish these snippets, what order should they occur in? For a while, I entertained the idea of publishing the work as a deck of cards, with one snippet on each card, but that seemed a little bit too on the nose. The version that I submitted to the Nightboat Poetry Prize was ordered completely at random.
But I found uniform randomness to be unsatisfactory—I didn’t just want to play 78-card pickup with the snippets. So for the version of the book I developed for Nothing to Say, I devised a process to order the snippets that would hopefully draw out their harmonies, rather than leaving the potential sparks of juxtapositional energy entirely to chance. The process I eventually decided on goes like this:
The linear sum assignment algorithm is necessary because I wanted all of the snippet pairs to be at least somewhat similar, even if that meant the most similar pairs didn’t always end up next to each other. It occurred to me that the pairing process is, then, a variant of the assignment problem, for which the linear sum assignment algorithm can produce good solutions. An alternative would have been to, say, pick a snippet at random, and pair it with its most similar passage, and keep going until all snippets had been chosen; but then the pairings that happen near the end of the process might not be very similar at all (since many of its ideal pairings will likely have already been removed from the pool).
The process I’ve described so far is entirely deterministic. But of course I wanted to make sure that each output was unique. To accomplish this, I “cut the deck”: at the beginning of the process, my program splits list of snippets randomly into several smaller groups, and the process described above is applied separately to each group. The resulting pairs from each group are then drawn back together into one pile and shuffled. The goal here was to add some variation in ordering while still maintaining good snippet pairings.
The resulting pairings are sometimes obviously similar: there will be one word, or segment of a word, that is noticeably frequent in both snippets. But often the effect is subtle. Because the analysis is based on n-grams, rather than whole words, sometimes the parts of the snippets that occur with particular frequency are located inside of words, or across word boundaries. As I reviewed outputs of the pairing algorithm, I found myself puzzling frequently enough over finding common n-grams in snippets that I made a little tool to visualize the similarities. Some screenshots of this tool are found below. In the screenshots, blue highlights show shared n-grams between the two passages, and the saturation of the highlight indicates how infrequently an n-gram occurs in the entire corpus. For example, here are two passages that are very similar to each other, according to my algorithm:
You can see that n-grams like those in the words “talisman” and “script,” and the word “under,” are particular to this particular pair of snippets. For contrast, here are two passages that are very dissimilar (i.e., they don’t share many rare n-grams). The visualization shows that there are barely any shared rare n-grams at all:
(Note that very common n-grams like “are” and “is” etc. are not highlighted in these examples, since they are shared by almost all of the passages.)
In Two of Pentacles I attempted to do something that I don’t usually do in poetry: write about something. The book really is meant as an exegesis of the divinatory interpretations of the two of pentacles, and for this reason I think of the book as being just as much an essay as it is a poem (or a collection of essays? a collection of poems? a “poem essay”?). The structure of the book, and the way we published it, are both important elements of how the book goes about presenting evidence and making arguments (fractal parataxis on the level of phrase, page, and volume). But my formal goal from the beginning was to try to make a text that clearly exhibited “aboutness” while still remaining as open as possible. The two of pentacles became the target of that “aboutness” because the card has always suggested deep meanings to me that seem only faintly hinted at in the text of others’ interpretations.
As for the form: I think many poets finish reading Maggie Nelson’s Bluets and immediately set about writing their own anthologies of “propositions.” Two of Pentacles is mine. I also had Anne Carson’s Short Talks on my mind during the composition process. I didn’t read them until well after I’d done the bulk of my writing, but Marosa di Giorgio’s books of poems collected in I Remember Nightfall were constantly on my mind while I was figuring out how to order and organize Two of Pentacles.
It was Andrew Yoon at Nothing to Say who initially suggested the idea of including illustrations in the book, and I think it was a good idea. I’m grateful to Taeyoon Choi for letting us use paintings from his “Uncomputable” series for this purpose. The paintings hover perfectly between abstraction and surrealism and really complement the text. Our cover artist, Ian O’Hara, was a dream to work with, and produced a fantastic cover that draws inspiration from non-figurative, pre-Golden Dawn designs of the two of coins.
That’s all for now! Please keep your eye out for the next book in the Two of Pentacles franchise: Two of Pentacles 2: Rise of Narcissus.