THURSDAY ⚡️ Computational Poetry ft. Allison Parrish

Computer programming is not just for, say, B2B SaaS.

Aug 6, 2022

Our next event is something a little out of the ordinary for us: we’re inviting Allison Parrish, a poet and programmer who specializes in the theory and practice of computational literature. This will be our first time inviting a poet! Our first time also inviting a theorist to give a short talk before our customary Q&A! I couldn’t be more excited. I hope you’ll join us.

📖 Wendit Tnce Inf (& other work) with Allison Parrish

No, that’s not a typo; it’s the title of Allison Parrish’s newest work. Keep reading for our review and the explanation!

Join us next Thursday to experience some poetry, join the Q&A etc, you know the drill.

Free Registration

🤖 our take: meet us at the fault lines of language and computation

By Shira Abramovich

Allison Parrish’s latest work, Wendit Tnce Inf, is very much a physical object. It’s a hand-printed, hand-bound chapbook that looks the part: running a finger over each page, the light emboss of the letterpress is tangible; the thread binding of the book is tied neatly at the center page; the letters are shaped like the old-fashioned lead type that is still used on many presses today.

*Wendit Tnce Inf. Image via Aleator Press.*

You’re not wrong to expect a poetry chapbook from this physical presentation, but what you find inside doesn’t match many expectations of poetry: the words are blurred, as though printed sloppily; the text reads like gibberish. That’s because, like most of Parrish’s literary work, Wendit Tnce Inf is computer-generated. Parrish created it using a GAN, a Generative Adversarial Network. Instead of training the GAN on words themselves, Parrish trained it on a bitmap image dataset of English literature—hence the gibberish, and its resemblance to old-time English type.

Why go to all this trouble, when a GAN is perfectly capable of creating almost human-level text when trained on words and paragraphs? Because Parrish isn’t interested in making computers imitate the work of human authors: she uses computers to expose the ways in which we view and interact with language, and vice versa. Like other kinds of conceptual writing, the work’s interest extends beyond its physical, literary value (which, some might argue, is hard to find), and instead encompasses the techniques that go into creating it—which in turn might illuminate something about the shape of words, the relationship between them, and what we do or don’t expect language to do.

Speaking of technique: Parrish’s computer-generated work encompasses a broad variety of formats and techniques. She’s best known as a creator of Twitter bots, which is where she first found her fame; her bot “Everyword,” one of the earliest Twitter bots to get a large following, printed every word in the English language, one every 30 minutes, for 7 years. The project was later printed as a book, where each word listed a timestamp and the number of likes it received. Parrish has also created a website for the generation of nonsense, a board game for word creation, a bot that creates poetry for NASA’s space pictures (extra-relevant while we’re all fangirling over the Webb telescope pictures), a chapbook of beautiful ML-generated visual-sonic connections between words, and much more. She theorizes and teaches about all of this at NYU’s Interactive Telecommunications Program.

What unites all of these projects is Parrish’s attention towards the fault lines of both language and computation. Parrish takes the core desire of contemporary language model research—that is, to create a machine that will generate and understand language in a manner that is indistinguishable from a human’s—and turns it on its head. Instead of creating a machine that speaks like a human, why not make a machine that puts glitches into human language? Instead of creating an algorithm that obfuscates the connections it finds between words, why not write a bot to find connections between words through computational means?

In doing so, Parrish questions the place of authorship and readership both, as a computer-generated text is a collaboration between a human author-programmer and a program or machine, and the resulting texts may never be read by any human at all. Do texts require authors? Do texts require readers? What about poems? And if so, what kind? These are the questions that much of computer-generated literature leads us to ask, and they’re questions that I wrote about a bit more in a previous essay on AI-generated poetry.

As a technologist, what I often find especially interesting about Parrish’s creative work is that she directly exposes the issues that we, as folks interested in technology and ethics, often ask about the use of computer-generated language and human-computer interaction. As just one example, take the issue of overly-large, ambiguous corpuses as training for large language models. A tech ethics-driven research approach might be to expose the biases in the model, and to then attempt to solve for them. Parrish, on the other hand, chooses to expose the inherent bias of all models by training her models on well-defined, well-documented corpi, letting the models parrot back whatever oddities the training corpi may have.

Whether or not all this should be considered literature is a hairy question for another time (such as Allison’s Q&A!). Whether or not you like it is a question you can decide for yourself. I love Parrish’s work because it is full of play—Why not have some fun with language? Why not, after all, write language bots for fun? What those bots and generated texts do is make us pay attention to both very small and very large oddities in language that we otherwise might not encounter, and remind us that computer programming is not just for, say, B2B SaaS, which, as a working engineer, I sometimes forget.

Join the Q&A

Shira Abramovich is a poet, translator, and software engineer.

Reboot publishes free essays on tech, humanity, and power every week. If you liked this or want to keep up with the community, subscribe below ⚡️