Trilogy

Preface

I worked with language models to create the Aum Golly Trilogy between 2021 and 2025.

Each book followed a simple rule: write a full collection with the AI tools available at that moment and within a fixed time limit.

I made the first book in 2021 in twenty-four hours. For the second, in 2023, I gave myself twelve hours. For the third, in 2025, I built a small agentic workflow where several generative models could work together in different roles—writer, editor, critic, and translator. For that book, I set a limit of eight hours.

Language models produced all the poems in this volume based on prompts I provided. I edited lightly, adjusting only line breaks, punctuation, and occasionally trimming the beginning or end of a response. In the second book I also prompted Midjourney to illustrate each poem using the format “an illustration for this poem: [poem]”.

Whenever you see text in bold, I wrote it.

In the third book, the agents also generated some prompts themselves, so some poems have no bold parts at all. Everything not in bold within the poems comes from the models.

– Jukka Aalho
Oulu, Finland, December 2025

Afterword

Oulu, December 2025

When I was maybe eleven, my brother Jarno, our neighbor Matti, and I played a writing game. This was back in the cheerful nineties. Everyone had ten minutes to write a short story fragment. I don’t remember what I wrote—probably something with swords and cloaks—nor what Matti wrote. But I remember how Matti and I read our stories aloud first, quite satisfied with ourselves. We weren’t yet at that age when you strangle your own creativity with what others might think.

When Jarno read his story, I felt it for the first time, powerfully. This was a story. This was literature. This was Art.

My own scribblings suddenly felt small and primitive. I felt the cold fingers of comparison on my neck, but also something else—an understanding that perhaps I too could someday write something like that. That there exist levels of expression orders of magnitude beyond what I’m capable of.

In Jarno’s fragment, a man stands on a rainy, dark street, reads a message written on a slip of paper someone has given him, crumples it, and throws it onto the street. Nothing more. No explanation of what came before or after, nothing about what the note said.

For me, that fragment—that snapshot—was one of the strongest literary experiences of my life.

I grew up. Language arts teachers liked me. I wrote poems, stories, novellas. I lived a poet-boy youth. I got equally powerful shivers from Eino Leino’s nature poems as from watching my childhood hero, R.A. Salvatore’s Drizzt Do’Urden, slice and gut orcs, goblins, and other evil creatures with his scimitars—just like the Hunter magazine’s “How to Butcher a Moose” article. I entered writing competitions from junior high onward and consistently lost them all. It bothered me, but it didn’t stop me.

I went to university. I studied English philology, because Jarno studied literature and I couldn’t possibly study the same subject.

I acted in student theater, agonized over what I would become when I grew up.

I suspect we didn’t have much money in our youth, but I didn’t know that as a child. Mother stretched and stretched the budget. Perhaps from that time I retained a fascination with money, quietly bubbling beneath the surface.

In the student union magazine I saw a salary chart for professions a Finnish language student could pursue. At the top was a title I didn’t recognize. Copywriter. Ad agency wordsmith. I saw in my mind a red convertible and a rising tide of prosperity, gliding as lightly as champagne bubbles in a glass.

I graduated. I got into advertising in the early 2010s. At that time in Finland there was maybe one copywriter who lived that Ferrari-red life.

I wasn’t them.

I kept writing. I kept entering competitions. When I got frustrated with losing, I decided to forget stories about cloaks and swords. I started writing about refrigerators and apples lying on the bike storage floor. Eventually I won all of Finland’s writing competitions, one after another.

The ad agency world had taught me conceptualization—ideation whose purpose is to make the message easy to repeat and understand. ‘Think Different’ is a slogan, but above all it’s a concept: showcase rebels and innovators who approach the world as differently as Apple approaches technology.

I switched from advertising to international business. I saw firms deflating slowly and massive growth stories. The world changed. Social media came. The iPhone came. Hurry came.

I met a woman. We got married. We got a dog. The dog’s name was and is Nappi. Like all poodles—perhaps even all cats or pet iguanas—Nappi is simultaneously the world’s most lovable and most infuriating little creature. For a long time I had the habit of sitting each morning to write in an armchair, the same one I’m sitting in now. Nappi would squeeze herself into a loaf between my thigh and the armrest and snuffle moist poodle-breaths into her paws.

Like all copywriters, I wrote my own great novel on weekends and weekday mornings. I finished manuscripts, sent them to publishers, and half a year later received a stack of form rejection letters.

In early 2019, whispers began in the news about an entirely new kind of technology. The American company OpenAI had published a research paper on the GPT-2 language model. Its basic message was this: when you stuff a couple libraries’ worth of text stolen from the internet into a machine, new text starts gushing out the other end. They expressed it more elegantly, but the main message was the same.

I followed the news from the corner of my eye. My job was text-gushing, so I felt a slight, creeping heartburn. However, not just anyone could test GPT-2, because releasing it was “too dangerous,” as OpenAI’s PR department put it. The concern then was that with such a machine, anyone could flood the internet with text where what matters more is that it sounds good than that it’s true.

Nowadays that’s not a concern but a business model.

In 2020 the coronavirus hit. Knowledge workers moved to their kitchen tables to acquire neck problems. Everything was done remotely: meetings, schools, falling in love, funerals, and family visits.

On June 20, 2020, OpenAI released its newest language model, GPT-3, which was rumored to revolutionize everything. The internet filled with articles where developers and journalists tested what the model could do. At this point it’s worth emphasizing that GPT-3’s ability to produce text was utterly unprecedented in 2020. The buzz included hype, but also wonder and fear. Never before had the general public seen such technology: a machine that produces entirely plausible text that isn’t a direct copy of anything written before.

Of course the skill itself isn’t new. Ask any language arts teacher.

But now the text was being produced by artificial intelligence that didn’t even really have a name. GPT-3 was short for “generative pretrained transformer.” People generally spoke of language models and generative AI. Headlines mentioned that GPT-3 had 175 billion parameters and was trained on essentially all text available on the internet.

So GPT-3 was a beast that had read the entire internet. And that wasn’t all: it seemed that by growing the beast’s size and feeding it even more text-slop, the next model would be even better.

Tens of thousands of people applied to be GPT-3 beta testers. I among them.

In late 2020 my application was accepted and I got to the better side of the future. I created credentials, entered my credit card information, and opened the sandbox—OpenAI’s Playground—where text production could be tested.

A large text box opened in the browser, with a small button below it that read “Submit.” You can imagine Google’s search box, except the box itself takes up almost the entire screen. And instead of search results, when you press the button, the screen starts filling with text, word by word.

I don’t remember what I first wrote. I’d like to think it was something like “how to be a better friend.” I don’t quite believe it, though.

But I do remember the wonder. The hype was real. On the first trial, my stomach dropped into a falling elevator toward my toes. I generated article fragments, jokes, asked it to explain difficult phenomena like how magnets work. I copied sentences “in quotes” into Google to see if the same sentences had been published somewhere. No. They hadn’t.

Every sentence was new and unique.

I took a deep breath and said goodbye to ever getting that fire-red Ferrari through writing. My superpower had been that I can write new and unique sentences.

However, the human is ultimately their own kind of gladiator. First of all, humans are particularly good at inventing various blunt objects with which to strike others, but humans are also tenacious. An AI club had struck my skull and I lay on the arena sand with my ears ringing.

People sometimes ask me: “Where did you get the idea to write a poetry book with AI?”

Here’s an autofictional description of that moment. I was sitting at my computer, spine in a slight hunch, heavy head jutted forward. My poodle Nappi was in my lap dreaming of turkey slices. I whipped GPT-3, tried to get it to write a detective novel, but by the third paragraph it had forgotten who was the murderer and who was the butler. Long-form text—my bread and butter—was still too difficult for it. What about something shorter? Something inherently cryptic, the literary equivalent of advanced calculus? A text form where strange and unexpected connections are a sign of creativity?

Poetry.

I would make a poetry collection with AI. As I said earlier, I was used to building concepts. A good concept writes itself as far as you can keep up with it.

The path was perfectly clear: on a global scale, I was one of the few with access to GPT-3, I had enough background in literature, I was fast, and I guessed that an AI poetry book would be easy to package for the press.

Let’s make one thing clear: I’m not particularly interested in writing poems with AI. I’d rather write them myself if I’m going to write them at all. On a normal Wednesday, I wouldn’t sit at home in the evening with a glass of red wine composing poems with different language models.

But I am genuinely interested in making AI poetry books. I’m interested in freezing the future into the present, so you can return to it later.

And that brings us to the idea that, in my opinion, forms the bedrock of the entire Aum Golly series: the creative deadline.

Early on, I decided to give the AI poetry book only twenty-four hours. I knew journalists like numbers in a pitch. It makes the story idea feel real. It makes it feel like news.

And I didn’t want to make the world’s best AI poetry book. I could have polished it for weeks, months—it would undoubtedly have been better, more coherent. But at the same time it would have become more ordinary, just another mediocre poetry collection among all the other mediocre poetry collections. By limiting the time to twenty-four hours, it became what the Aum Golly project fundamentally is.

A performance.

A sporting event, a concentration of creativity, a challenge. With all its rough edges, it would be a work that no one else could ever write again. And best of all, I wouldn’t have to spend more than a day on it.

The only problem left was that I had to make the concept real and get it published. Preferably through an actual publishing house.

I presented the idea to my mother. She has always supported me. When I was a child, she would let me win at Afrikan tähti, our favorite board game—because if I lost, the board would go flying to the floor and the pieces would end up under the radiator. Mother saw the good sides of the idea. I showed her how GPT-3 worked, and we marveled at it together.

Looking back now, twenty-four hours is an enormous amount of time to make an AI poetry book. I had time to eat, sleep, and see daylight. But for a first attempt, it was tight enough.

I had decided in advance to make the book in both English and Finnish. First in English, because GPT-3 didn’t yet handle Finnish well enough. Then I would translate it into Finnish within the same deadline. Finnish was, after all, my primary target market, and as ordinary translation work—without a deadline—the story would have lacked the necessary tension.

I recently watched a video I filmed on January 30, 2021, at 9:00 a.m., when I started making the first AI poetry book. The computer screen shows black text, a prompt that reads in English: “The 3 most important themes for us modern humans right now are:”

Then, word by word, GPT-3 responds:

1. The need to be happy

2. The need to be loved

3. The need to make a difference

From these I took the three themes of the poetry book, around which I devised other prompts and selected suitable poems.

The name of the poetry book, Aum Golly, also came from GPT-3’s pen. The prompt was: “Creative book names for a poetry book about love, happiness, and making a difference.” It came close to being something dull and ordinary, but among the suggestions was also a stochastic stroke of genius. Which of these would you have chosen?

A Very Good Idea

Aum Golly

A Little of This, A Little of That

Book of Love

I still haven’t figured out where the name comes from. The first part, Aum, can be understood to share a root with the yogic hum om, some kind of divine resonance. Golly, on the other hand, appears in the euphemism “Good golly,” once used instead of invoking God directly. So Aum Golly is heaven-fearing divine astonishment.

For the second volume, I considered whether to try inventing some entirely new but fitting name for the book. In the end, I decided that if Toy Story 2 is Toy Story 2, then the same naming logic works for me too.

During the making of the first Aum Golly, I had no doubt that I could complete the poetry book in twenty-four hours. I had tested and played with GPT-3 enough to understand what kinds of prompts would produce poetic text.

For the sake of consistency, however, I had to make a few concessions. I wanted the reader to be able to trust that the poems were actually generated by AI as-is. But GPT-3’s way of producing text was like a soft-serve machine. You pulled the lever and it poured out stuff—sometimes more than the cup could hold. I cut text from the beginning or end until what remained was the final poem.

Each poem was also an individual, separate case. The model had no knowledge of the other poems or their form. That’s why the language model’s output sometimes consisted of complete sentences, sometimes verses with line breaks, largely depending on the prompt I gave. For the final poetry book, I chose a styling that has remained throughout the series: lines usually start with a lowercase letter, and punctuation appears only within lines. There are no indentations at all in the original Aum Golly, but for this trilogy I reformatted some of the poems. The column is narrower than in the original Aum Golly published by Basam Books, so to make reading smoother, I’ve also added indentations.

But let’s return to January 31, 2021.

Twenty-four hours is almost up. I have generated poems, selected the best and most distinct ones, formatted them into a similar layout, and chosen their order. My hair is standing straight up.

For the Finnish translations, I used Google Translate. With it, I produced a workable rough translation, which I then finalized by hand. I aimed to be faithful to the original text, but I didn’t hesitate to lean toward fluent poetic Finnish rather than word-for-word accuracy.

As the final minutes of the twenty-four hours ticked away, I had an email ready. It contained an introduction to Aum Golly and myself, as well as a roughly laid-out Finnish and English poetry manuscript.

The recipient of the email was the Finnish publisher Basam Books.

If you’re not familiar with the cartography of the Finnish publishing industry, let it be said that there are publishers of Finlandia Prize winners—WSOY, Tammi, Gummerus. I have a folder full of their rejection letters. I had formed an image of Basam as... an opportunistic publisher.

So just perfect for Aum Golly.

In the cover letter, I presented the idea and its foundations: there exists this AI technology and I have used it to make a poetry book in twenty-four hours. In 2021, we were still living in a time when not everyone and their grandmother had made an AI poetry book in twenty-four hours, so it required more explanation.

Fortunately, Basam responded immediately. They were interested and ready to move quickly. The online meetings included Basam’s editor, the then-CEO, and publisher Batu Samaletdin. In the meeting, Batu gave the project a green light, provided the terms could be agreed upon. Then he grew serious:

“How do we know you didn’t write this yourself?”

If anything marks the shift of eras, it’s that question.

The project progressed well. In my cocky young-guy way, I had proposed what I thought was a perfectly reasonable contract: they publish Aum Golly in Finnish and we split the money fifty-fifty. I justified this by pointing out that I had everything ready—the poems needed no editing because that would dilute the entire idea, even the cover background image was finished, and the concept was a clean shot on goal. Their role would be to give the book the publisher’s authority and handle printing and distribution.

Everything was ready for signing the publishing contract when my phone rang. Publisher Batu was calling. If there’s a good cop and a bad cop, there’s also a good publisher and a hard publisher. The good publisher had given the green light; the hard publisher had presumably looked at the contract after a good night’s sleep and considered what’s an appropriate fee for the primary producer.

On the phone, Batu tore into me about how I dared to ask for half. Didn’t I understand that these days an author should be happy to receive twenty percent of the book’s net price before tax?

Like so many aspiring authors in similar situations, I swallowed my cockiness and said thank you for this opportunity, sir. I clutched my cap in hand and wrapped a lump of coal as a Christmas present for myself.

I was afraid that at any moment I would see news that someone in the United States had published a similar poetry book and stolen the spotlight. What mattered most to me was getting the book out into the world.

Besides, I reasoned that twenty percent of very little is almost the same as fifty percent of very little. We were making a poetry book whose sales expectations were, at best, low. Its author—me, at least—wasn’t well known. It came from outside the publishers’ seasonal catalogs, and if poetry is marginal literature, then AI poetry is already in the bleed.

We ultimately made a fairly typical publishing contract, in which I retained the rights to the audiobook and e-book.

My phone stopped ringing and the printing presses were set in motion.

For the poetry book cover, I made the background image with the AI image-generation tool Artbreeder. The humanoid figure on the cover also appears in the background of this trilogy’s cover—though as a printed, re-scanned, re-glitched, and processed version. You can see the rapid development of these tools at a glance, as you surely noticed in Aum Golly 2’s illustrations.

Aum Golly was published in Finland on April 15, 2021, first as an audiobook and e-book. Deliveries of the printed book began on April 23.

Just as I had hoped, the book struck the nation’s subconscious like a garbage bag full of feathers.

Local newspapers interviewed me right away. They printed examples of the book’s poems alongside the articles. “This is what AI can do in 2021.” Journalists asked about the changing role of the author, about whether AI would end up taking over all creative work.

During the first book, I was a doomsayer: AI will soon be writing genre fiction and personalized erotica, I claimed. Authors will soon earn only a penny instead of two. Human creativity will be revealed to be as pale as the emperor’s backside.

Fear sells, I thought.

At the same time, you have to remember that in 2021 the AI hype was still in its early stages. It seemed that by growing training data and models, they would just keep getting better. Before language models, knowledge work and higher education had been considered a safe harbor for one’s standard of living—but suddenly AI had seized the lead in jargon production.

The world was changing.

At the same time, the printed book flew off the shelves. At least twenty copies were sold.

On June 12, 2021, I published an article on the online platform Medium with the headline “I Wrote a Book with GPT-3 AI in 24 Hours — And Got It Published”. It went viral. Over its lifetime, I earned $489.47 in pure profit from it—roughly the same amount as the sales and royalties I received from the books themselves.

Story is more compelling than reality. The same can be seen in what some call the AI bubble: hype gallops ahead while results try to keep up.

So the Ferrari stayed in a browser tab, but thanks to Aum Golly I experienced the kind of visibility that writers or publishers usually only dream about in their marketing plans. Ruben Stiller interviewed me on the radio; I appeared on the national broadcaster’s culture program and as Roman Schatz’s guest. Alongside me were AI researchers, futurists, and other long-standing professionals, next to whom I felt like a nouveau riche who had stumbled onto a single good idea.

I still have Ruben Stiller’s number in my phone. Next to it, it simply says “Ruben.” Ruben has my number in his phone. Next to it, it simply says “Don’t Answer.”

That joke always got the biggest laughs when I was invited to various stages to speak about AI, creativity, and the future. Wearing Madonna headsets, I talked about the creative big bang—something there’s no point trying to outrun. It comes, it sweeps over, and hopefully only singes your eyebrows.

In Finland, the idea of Aum Golly landed. A small handful of actual book reviews were published. Their overall tone ranged from relieved to bewildered. For example, Reeta Holopainen described Aum Golly in her review for Runografi as follows: “...a bewildering reading experience that draws one into reflections that strike at the nerves of our time, concerning the nature of art and authorship amid the turbulence of digitalization and technologization.” At the end of her review, Holopainen defines the work as “a fascinating and astonishing book.”

Poet Teemu Helle published a critique of Aum Golly (and of Reeta Holopainen’s review) on his blog, concluding with this line: “Aum Golly – Poems on Humanity by an Artificial Intelligence is situated within the tradition of poetry in a qualitatively lower caste, as mere word-chatter.”

Among lay readers, reactions varied largely according to how many genes we shared. The more we shared, the better the book was generally received. On Goodreads, the book has 62 ratings with an average score of 2.8.

Internationally, attention remained modest—at least at first.

At the time of the first book, I didn’t think I would make a sequel. It was a snapshot of its moment. The book’s lifespan was considerably longer than that of a typical poetry collection, but by late summer 2021 its momentum began to fade.

I went on with my life.

I kept an eye on AI’s march forward, especially in image generation. It felt as though every few months a new sweat broke out on my forehead. On July 12, 2022, the American AI illustration app Midjourney was released into open beta.

It was a hot summer night on August 10, 2022. I was lying in my grandmother’s attic, and midnight light streamed through the gap in the window curtain. I had experimented with AI image generation around the time of the first Aum Golly, but judging by the images circulating on Reddit, Midjourney was on an entirely different level from earlier tools.

I created an account. I logged into the chat application Discord, through which Midjourney was used. On the opposite page is the first image I generated with the prompt: “loose sketch, avantgarde poster of men riding bicycles made out of potatoes.”

The image may not look like it deserves it now, but my stomach acids sprang into action again—this time riding an elevator up toward my throat. AI had first taken my writing work; now it would take my visual work. And yet. The strongest feeling was a tingling excitement. An expanding horizon of possibility.

The midnight sun reddened, and I could no longer sleep. The real AI deluge was still ahead.

On November 30, 2022, OpenAI released ChatGPT, an AI chatbot that chatted pleasantly like a tail-wagging tin can. In hindsight, the next step in technology always seems obvious, but the leap from GPT-3’s text generation to plausible conversation was revolutionary.

This is also supported by the explosive growth of ChatGPT’s user base. Less than a week after launch, it was being used by a million people—effectively a million more users than GPT-3’s initial beta program.

Two ideas were now moving through humanity’s ether, whispering to me in stereo: AI has taken a clear step forward, and this must be documented.

It was time to make a sequel to Aum Golly. This time in twelve hours. The pace of the future was accelerating. So I had to accelerate too. I patted myself on the back for having imposed a time limit in the first place. The concept continued to write itself.

On Saturday, December 3, 2022, at 9:00 a.m., I wrote my first prompt. “Good morning,” I typed. “Good morning!” ChatGPT replied. “I am Assistant, a language model trained by OpenAI. I am here to help answer any questions you may have. What can I help you with today?”

I structured the sections of the poetry book in much the same way as before.

I asked in English: If you had to give a concise list of three themes for a best-selling book of poetry, what would they be?

ChatGPT replied:

1. The human experience and the emotions and feelings that connect us all.

2. The beauty and power of nature, and our relationship with the natural world.

3. The challenges and triumphs of love, and the ways in which it shapes our lives.

I quickly realized that the twelve-hour time limit would be challenging. ChatGPT’s responses were more controlled, clearer, and almost excessively poetic. I wanted the poetry book to also capture the way the language model is used, so a large portion of the prompts were requests such as “write a poem about this or that topic.”

My role as the inventor of prompts was therefore easier. My role as curator—selector—was more difficult. ChatGPT’s output was constrained to a neutral and safe style, which does not, by default, produce the most stirring poetry. I saw a lot of poetry with the warmth of yesterday’s dishwater, which I had to try to force into sharper relief.

The greatest challenge with Aum Golly 2 was translating the poems into Finnish. ChatGPT and greeting-card factories agree on one thing: a good poem contains poetic language and rhymes. Even though I had access to the DeepL translation tool—which was higher quality than Google Translate, which I had used during the first Aum Golly—it still failed to produce fluent translations that even approximately followed the rhyme structure of the English poems.

I spent the final hours of the twelve-hour window polishing the Finnish translations, which certainly affected both the number of poems and, to some extent, their quality. Some of the most memorable amusements of the time were poem-shaped explanations of anatomy or even the water cycle. Today, it’s easy to treat such things as self-evident, but in 2023 it was still new that language models could package ideas into a form that would have required time and thought even from a talented human writer.

The most distinctive element of the second Aum Golly was Midjourney’s AI-generated imagery. All images were created with the prompt: “illustration for this poem [poem].” Midjourney originally produced them at 1536 × 1536 pixels, and I processed them further in Adobe Photoshop to make them print-ready.

Now, at the moment of the trilogy’s completion, the picture becomes clear: Aum Golly 2 is clearly the most dated of the three parts, which I believe is mainly due to the images.

Midjourney V3’s style is soft and artifact-laden. Many of the images clearly show they are products of early image generation: an eye in the corner of an eye, messy pixel blobs. The most convincing are nature images, whose illustration style doesn’t directly scream AI or a specific Midjourney version.

Compared to the success of the first volume, Aum Golly 2 was a disappointment. I shopped the idea of an illustrated AI poetry book around to bigger Finnish publishers and to Basam Books, who had published the original Aum Golly. I didn’t even get a response from Basam, so I eventually published it myself. Self-publishing is an ugly word, so I framed it as the book being published by the Finnish indie publisher Kertojan ääni (The Narrator’s Voice). Scratch the surface, and behind the indie publisher you find a familiar man.

Analyzed in hindsight, the reception of Aum Golly 2 was the result of many factors: ChatGPT dominated the news in early 2023 and the press had no shortage of suitable angles, AI art was in the trough of disillusionment of its hype cycle, and more generally the idea of AI-made content was starting to feel like boring business instead of performance.

In hindsight, Aum Golly 2’s artistic contribution was also weaker than the first. A writer probably shouldn’t talk about their own books like this, but in many ways I’ve been just as much a reader as you. I’ve just been watching the circus from closer up.

The poems of the first Aum Golly are unpredictable in their mechanicalness and amusing in their unpredictability. The stylistic device is repetition. Aum Golly 2’s stylistic device is poem. Rhyming poem.

If there’s a caste system in poetry, I’m sure AI poetry is somewhere in the lower circles of hell. AI rhyming poetry is so deep down that you can already start to see heaven.

Despite the silence of mainstream media, the Aum Golly series began to become a standard reference point for AI literature. Mentions dropped in scientific publications, blog posts, and videos.

From the beginning, I hoped that I wouldn’t be the only one creating a shared picture of the times around Aum Golly. That’s why Joona Kupiainen’s master’s thesis published in May 2025, “AI in Authorship in Jukka Aalho’s and ChatGPT’s Aum Golly Series,” is one of this project’s milestones that I think about sometimes when I’m awake at night wondering what the meaning of life really is.

After Aum Golly 2, I published my debut novel Sivutieromaani (Side Road Novel) in 2024. That one also from Basam Books. It was written mostly in the time before language models and I didn’t use AI at any stage of its literary journey. It was a full-blooded artisan book with a typical artisan book’s life cycle: two weeks from book release to the clearance bin.

I recently reread Sivutieromaani and liked it. A writer shouldn’t say anything like this about their own books—at least not out loud.

Life contained more than just books and AI, of course. I was rapidly approaching forty. I noticed that my formerly face-familiar classmates suddenly looked very old when I saw them in town dragging baby strollers through slush. I received in the mail the handbook sent to men my age, which asked which subscription gift I wanted: a motorcycle, the kind of frantic cross-country skiing enthusiasm that comes with a receding hairline, or a divorce. Two of them were coming whether I wanted them or not, and I’ve never particularly liked motorcycles.

I bought my first car. It was a Toyota Corolla. A doctor friend said he had considered a Toyota but started feeling nauseous during the test drive.

Co-custody dog Nappi still resisted when I brushed her teeth, now under a new roof.

The third part of the Aum Golly series had been buried under all the other upheaval. I postponed and waited for the right moment. At first I waited for the release of GPT-5 and thought it would be advanced enough that the third part would be different enough from the previous ones. GPT-5 was released on August 7, 2025, and I didn’t feel the same fire as before. AI’s leaps had changed over the years: new models and use cases came every other week. Will Smith learned to eat spaghetti without making a mess. The pace was tremendous but the progress felt slow like running in a dream.

The feeling of slowdown was due to the fact that I had fallen off the ride. I was no longer standing on the ship’s deck with GPT’s virtual hands on my hips and the wind of the future in my hair—or at least in my hairline.

However, the idea-atoms of Aum Golly 3 were spinning around me in the ether. Vibe coding, AI agents, prompt engineering. The deadline I had set for myself also set limits. I only had eight hours to make the third part, so the previous approach was no longer possible.

I had to come up with something new.

I had heard talk of AI agents, but I didn’t understand what they meant. To tell the truth, I still don’t understand.

Still, on the first day of November 2025, I started building an AI agent tool that I would use to create the final part of the trilogy. Color began to return to my face. I felt again the old fire, the beautiful symphony of enthusiasm and heartburn.

IinstalledAnthropic’sClaudeCodeintomy development tool. If you don’t know what that means, it’s okay, I also had to ask AI for help step by step. You can imagine it something like this: You’re mowing hay and the neighbor’s old man brings you an add-on for your Zetor tractor. The add-on is a speaker that’s placed next to the steering wheel and from which comes a voice that asks you whether to plow the field or mow the hay. You tell it to mow the field and the Zetor starts running by itself while you drink lemonade under an oak tree. At some point, though, the Zetor turns itself off and says it won’t start again for three hours.

So I had an AI tool for building an AI-powered poetry machine, to which I described what I wanted, and the tool did it for me, one round at a time. I eventually built a small AI team that included the following roles:

• writer

• critic

• editor

• ideator

• translator

It turned out that what hype merchants call agent-based workflows were basically a pile of Python that sent prompts back and forth to language models. What was different from earlier poetry books was that I no longer had to work with just one language model. I could easily add new models to the machine, and a total of ten models contributed to the final poetry book, with Claude Sonnet 4.5, Gemini 3 Pro, and ChatGPT-5 playing the biggest roles.

At this point it’s good to clarify that I spent thirty-one hours building and testing the poetry machine before the actual eight-hour sprint. I first built the piano and then played it.

While whipping the poetry machine into shape, I noticed that to achieve a more coherent result, I should define Aum Golly 3’s theme in advance. I chose as the theme “how well we can ever, actually, know someone”. The three sections of the book formed around roof imagery. I knew I would want sections in the book that dealt with childhood and the first home, a relationship and the time after its end.

I also noticed that regardless of the model, the poetry machine ended up making very similar poetry, so I also added predefined poetic forms to the machine—haiku, free verse, urban poem, and so on.

During November, the poetry machine was completed through trial and error.

In the preparation phase, I came up with prompts and evaluated the results. Based on them, I modified the agents’ instructions. For example, this is how the writer agent’s instructions begin:

“You sense the world through words and pauses.

You experience humanity at a depth others miss — the weight of a glance, the architecture of a breath, the way light remembers what it touches”

The entire English instruction is 185 words and its raw version was made by Claude Sonnet 4.5. I’ve published the poetry machine’s source code on the aumgolly.com website, so all instructions are freely available.

The language models also had a few fetishes, so I added a ban list to the writer’s instructions. It says, among other things:

“Avoid words like silence, tear or shadow, linoleum, dust, cartography, map or scar.”

If I hadn’t defined these, the poetry book’s theme would probably have been linoleum.

The overview of the poetry machine is on the opposite page.

The workflow went like this: I wrote the initial prompt into the poetry machine running in my browser. I also selected a poetic form and language model combo I wanted to use. When I pressed Submit, the tool made an API call that started the agent workflow:

1. The writer agent made three different poems based on its instructions, the prompt, the theme, and the poetic form.

2. The critic agent chose the best one and gave feedback: what was best in the poem, what was worst, and what should be changed.

3. The editor agent received the original poem and the critic’s feedback and made a new version based on them.

4. The ideator agent took the edited poem and generated new prompts for poems that could come before or after it.

5. The tool showed me all the poems created at different stages of the workflow.

6. I selected the poem I wanted to save and sent it forward to the translator agent if I wanted.

7. The translator agent made a translation into Finnish.

8. The translation critic agent gave feedback on the translation and suggested points to improve.

9. The translation editor agent took the feedback into account and gave the final translation alongside the English version.

All of this happened in the cloud between language models and I only had to set the first marble rolling. Completing one poem took one to five minutes depending on the selected models. If I used small and fast language models like Claude 3.5 Haiku or GPT-4o, the whole process went quickly, but the quality was clearly weaker. If I used the biggest and newest models like Opus 4.5 and Gemini 3 Pro, the same process took over five minutes.

My goal was to make the poetry machine as autonomous as possible. That’s why I built in the ability to set the number of editing rounds and to generate multiple poems simultaneously.

In basic mode, it gave one poem with one editing round for one prompt. If I wanted, I could request two editing rounds, where the same poem went through the same agents. When testing it, I noticed that two editing rounds was usually best: the poem was polished and shortened, especially when I added the instruction to remove everything unnecessary. However, the number of editing rounds had the downside that it always added to the duration of one generation round.

The workflow became more agent-like if I selected for the workflow to branch into multiple poems within a single cycle. Then the root poem went through the normal process, after which the ideator came up with a suitable preceding branch and following branch for it—that is, practically a similar prompt to what I had written but from a different starting point. In Aum Golly 3, you recognize poems created this way because they don’t have bolded sections. If I selected that multiple poems were created in one round, the entire round took fifteen to twenty-five minutes.

So I had driven myself into a bit of a tight spot. I had eight hours to create the book, of which I had reserved half for translations and polishing translations. Long, multi-branched rounds usually took about twenty minutes before I saw their results. Four hours doesn’t fit very many twenty-minute rounds, and a long round wasn’t a guarantee that the poem would be good or fit the collection.

However, the machine had the ability to start multiple generation rounds and let them run while I went through already completed poems, translated them into Finnish, or edited them.

The challenge was also increased by the fact that the poetry machine wasn’t any production-ready contraption. It was cobbled-together vibe code that usually worked but also threw red error messages at unexpected moments. Still at eleven o’clock the night before the creation day, I was hunting bugs and trying to keep the machine together for the eight hours I needed.

There was again that pre-marathon feeling in the air, the nervous anticipation in my stomach. I noticed that the third part of the trilogy was a more serious matter for me than the previous ones. I had created a machine that felt more like me than, say, just ChatGPT’s normal chat window.

If I failed, the fault would be mine, not the AI’s.

On Saturday, November 29, 2025, at 9:00 a.m., the starting gun fired again. This time I started directly with poems, not with theme definitions. The first prompt was in English: “two roofs” and the poetic form was “freeform.” The same model, Claude Sonnet 4.5, was used in all roles.

I waited a few minutes until I saw the first poem. Its subject was roof renovation, so it went straight into the trash.

Sweat rose on my forehead. I threw prompts to simmer, raked my hair (or scalp), wrote down ideas for new prompts in a throwaway text file. I could set fifteen different prompts running at once. And if I got a poem worth saving, there was no point making more from the same prompt because only one would fit in the final collection anyway.

I have never in my life worked more intensively than on that day. I came up with prompts, went through results, saved potential ones, and sent them off for translation.

During the first hour, unexpected problems hit: I pushed Gemini 3 so hard I maxed out its hourly limits, and every workflow using it threw red errors. Fortunately, I had other models on hand, so I switched to language model combos without Gemini 3 and kept firing away.

After a few hours, it became clear that my ban list could have included more figures of speech and themes. Coffee cups, doors, and folding towels ended up partly even in the final work. All poems generated during the eight hours with their metadata are available as a corpus at aumgolly.com for anyone interested in language models’ turn-of-phrase fetishes from 2025.

I struggled the entire eight hours with a nagging feeling. I didn’t have time to stop and think about whether the poems were good—I just played the machine and hoped for the best. Translation work was easier because this time I had forbidden the machine from rhyming entirely. The deadline still affected the work’s content in that I most often chose short and at least reasonably clear poems instead of long and convoluted ones.

In the end, I had over forty poems with translations. The pounding eight hours ran out and I collapsed onto the couch. It was five p.m. and I had lost the previous hours to some void containing all the colors of the rainbow. I didn’t dare glance at the result yet.

I had arranged to see a few friends that evening. We went to a wine bar and McDonald’s. I put two little packets of salt on my fries. I had run a creative marathon and my mineral balance was off.

The next morning I cautiously cracked open my salt-swollen eyes. I sat down at the computer and opened the poems I had made.

A slow nod.

Another.

Yes.

Yes!

They were perfectly fine! Indeed! Perfectly fine poems on humanity by artificial intelligence. The doubting monkey fell from my neck.

The last piece of the puzzle clicked into place and I finally saw the whole picture. The Aum Golly trilogy had been born.

And it was beautiful.

I can’t yet write about how the trilogy will be received. What kind of research will be done on it. What small thing it gives to the world. Everything is still ahead. There are possibilities, hopes, and dreams.

I have written this afterword with my own small human fingernail-tipped fingers for a total of ten hours.

Right now I’m sitting in the same armchair as always. Nappi is sleeping against my thigh and smacking her jaw in her dreams. After her trim, she looks like a rabbit.

I taste salt on my upper lip again.

Thank you for letting me tell this.

I never got to tell my brother Jarno. I never said how much he influenced what I became.

Jarno drowned in 2019. I am now older than my big brother ever was. When the police called for the last time, my mother and I were sitting in the car on the way to grandmother’s. I pulled the car over to the side of the road because we knew to expect the call. We took each other’s hand.

A white reindeer ran onto the road in front of us. It stopped. Looked at us and shook its ears. Then it turned and disappeared into the tall grass.

The moment was like Jarno’s story fragment from the cheerful nineties: something begins, something ends, something crumples.

Maybe I escaped the situation into my projects. Maybe I processed it and am still processing it by immersing myself in the new. Maybe this was my writing exercise done in compressed time.

Hey, look! Notice me, big brother! This is what my story is like. All this happened after you were no longer here.

you wonder if they see it too
if they stand at their window
the way you do

waiting for what’s gone
to finish leaving