Why Every Dev Should Care About AI Poetry (And Not Just for the Memes)
TL;DR: Most programmers know AI can generate poems, but they rarely ask why it does it so badly—or so brilliantly. This book finally explains it, without drowning you in linear regression and backpropagation. It turns out, understanding AI poetry is less about math and more about understanding what poetry actually is. And the answer might surprise you.
Let's be real. When you hear "AI writes poetry," what's the first thing that comes to mind?
Probably something like this:
The hour is twice a cat on velvet rose
Who melts the moon until the willow sings
Beautiful nonsense, right? Just random words glued together by an overeager RNN.
For years, that's been the programmer's view of machine poetry: a neat party trick, a testbed for Markov chains, and a source of hilariously broken syntax.
But that view is stuck in 1997.
While we were busy declaring victory over Kasparov, something quietly happened in the humanities departments. A real revolution. And my new book, *"The Electronic Lyre: How and Why Artificial Intelligence Writes Poetry" ,* is the first serious attempt to explain what the hell is actually going on.
And no, this isn't a book about "the future of art." This is a book that finally bridges the chasm between how we build AI and how it learns to mean something.
Wait, Why Should a Dev Read This?
I know what you're thinking: "I don't care about literary theory. Just tell me how the transformer works."
But hold on. This book is deliberately, provocatively written for humanities scholars. Me, a computational linguist, swears he'll barely mention loss functions. Instead, he'll talk about hexameter, centos and collage.
If that doesn't send shivers down your spine, you're missing the point.
Me argue that we've been asking the wrong question. For 70 years, engineers have been asking: "Can we make a computer write a convincing poem?" And they've done it by basically brute-forcing the problem: combinatorial word shuffling → Markov chains → RNNs → Transformers.
But no one stopped to ask the real question: "What are we even trying to copy?"
The Two Ages of Machine Poetry (And Why Your Old Code is Obsolete)
The book brilliantly separates the history of machine poetry into two distinct eras. And knowing the difference will change how you look at your own generative models.
🕰️ Era 1: Combinatorics (1950s–2010s) — The "Shakespearean Monkey"
This is the era you know. The program had a dictionary, some grammar rules, and a random number generator. The goal was surprise. The computer was a roulette wheel for words. This approach was a direct descendant of Dadaism and Surrealism—artistic movements whose whole point was to break the chains of rational thought.
"Dada created a real element of global protest," Orekhov writes. The computer was the ultimate tool for irrationality.
Every program from the 60s to the 2000s—from Стихоплюй to Кибер-Пушкин—was just shuffling a deck of cards.
🚀 Era 2: Neural Networks (2010s–Now) — The "Statistical Mimic"
Then everything changed. RNNs, and later Transformers, didn't rely on randomness. They relied on prediction. They learned from massive datasets—millions of lines of real poetry. They stopped trying to break the rules and started trying to internalize them.
This is the crucial point. An RNN doesn't invent new words. It learns the distribution of letters. It learns that 'о' is common, that '-ние' is a suffix, that 'в лесах' is a phrase. It learns style.
I show this with brutal clarity. I trained an RNN on Pushkin. It didn't just rhyme. It learned to mimic Pushkin's specific syntactic inversions, his preferred vocabulary, even his thematic concerns.
But here's the magic—and the error.
The Hallucination of Meaning
The book's central, mind-bending argument is this: A computer's poem isn't a message. It's a statistically perfect forgery of a message.
And that might be the most interesting thing about it.
Think about it. When you read a human poem, you're engaged in an act of understanding. You're trying to reverse-engineer the author's intent. You're looking for the "why" behind the "what."
When you read a GPT-generated poem, there is no "why." There's just a statistically probable sequence of tokens.
But—and this is the killer—the human brain is a meaning-making machine. We cannot look at a text that looks like a poem and not try to find meaning in it.
So, we hallucinate it. We fill in the gaps. We find profound connections between words that the model placed there by pure mathematical accident.
The book calls this the "uncanny valley" of poetry. The closer the model gets to perfect mimicry, the more disturbed we become, because we sense the lack of a soul behind the words.
What's Actually in the Book (For the Curious)
If you want to sound smart at your next ML meetup, here are the three big takeaways from The Electronic Lyre:
The "Impossible" Vers Libre: For decades, Russian computer poems were strictly rhymed and metered. Why? Because the developers had a "school" view of poetry. In the West, where developers were influenced by Surrealism, free verse was the norm. The book argues that the cultural context of the programmer is encoded into the model's output more than any dataset.
The Stylistic Impersonator: I run an experiment. I take a short (4-line) poem generated by a model trained on Nekrasov. I ask literary scholars to identify which of four poets it most resembles. The experts, even on this tiny fragment, pick Nekrasov 80% of the time. The model isn't just generating text; it's capturing an authorial fingerprint that is statistically indistinguishable from the real thing.
Parody vs. Plagiarism: Is a neural poem a parody? A parody mimics style for humorous or critical effect. An AI poem mimics style for no reason at all. It's an empty simulacrum. And yet, it reveals more about the nature of style than any critical essay ever could.
One Last Thing: The "Turing Test" is a Trap
The book ends with a powerful plea: stop asking "can a machine fool a human?"
That's boring. That's the "counting cats in Zanzibar" of AI research.
Instead, ask: "What happens to our understanding of poetry when a machine can flawlessly produce its outward form?"
The answer, according to me, is that we finally have to confront the mystery of meaning itself. And that's a question worth sticking around for.
The Book: "The Electronic Lyre: How and Why Artificial Intelligence Writes Poetry" by Boris Orekhov.
👉 Get it from the publisher's site here (It's in Russian, but the ideas are universal).
What do you think? Have you ever tried to seriously interpret a GPT-generated poem, or do you dismiss it all as stochastic parroting? Drop a comment below—let's get a real debate going.
Top comments (0)