Awkward AI Poetry
Good taste that can't be spun into a web, unfurled, entwined and echoed, etched in a blueprint, and god-forbid woven into a tapestry.
Some writing is truly tasteful and profound. You can feel it in the way words are delicately placed into sentences. The voice, is just right. The exactly titrated dose of pathos, with careful choice of adjectives, the clash of right textures and tones… god fucking bless good writing.
The language models can do a lot of things marvellously, such as information processing and basic reasoning, writing a few lines of code, or giving you a summary of an article. It does produce pretty much flawless language. But that doesn’t grant it mastery over it. I really hope that the brief moment of the oh-my-god-it-can-write-like-a-human hype around the AI models will be safely over soon.
Because it can’t.
(Disclaimer: As an AI engineer, I am of course torn between the two sides of this discourse, and I don’t want to lessen our appreciation of this unique moment of the mighty Technic turning inside out and becoming an esoteric entity. I do think language models are wonderful I love and respect them and appreciate them and cherish them, and I think they have an incredibly magickal potential and we have plenty to learn from them. I just want to dispute the obnoxious claim the tech companies push forward, that they will, in any meaningful way, replace our creative expression.)
Everyone knows you’re using AI
Last week as the latest Llama 3 dropped, I spinned up the model. I knew that the coding and reasoning skills will be much more impressive than before. But that’s not what I was really interested in. I opened the terminal and ran my usual simple welcome test by pasting a piece of my lyrics for the model to complete, ready for the carnage.
Violence is a language, in search of something that quickens the nervous system and jumpstarts the synapses, deadened by leisure and last night's sleeping pills Transgression is a moving target, intense heat confusing the circuitry ... [continued by Llama3] In the silence, a scream builds, a pressure cooker of unspoken rage The city outside, a cold, unforgiving mirror, reflecting our darkest impulses A countdown to detonation, as the fragile threads of sanity begin to unravel
Okay. I rubbed my temples, trying to shake off the subtle trauma response of the late noughties angsty fanfic princess repressed somewhere deep inside. I hope you understand the urge to sandpaper my own cornea after seeing this.
In the past months, as I’ve been watching the maturing of the language models, I developed various techniques to test its artistic abilities. During these experiments, I collected a considerable display of fairly bad AI poetry tropes. These are special phrases and little quirks that the model pushes forward when queried about anything esoteric, philosophical or god forbid artsy. This snippet is a rather condensed example:
A tapestry of thoughts, expertly unfurled digital fragments. Ideas weaving, AI's fabric. The cartography of thought, etched on the maps of memory, A dance of algorithms and linguistics. Unpredictable patterns in a structured playground. Where innovation blooms, and creativity unfolds, As language and logic entwine, a wondrous rhyme. In this vibrant tapestry, colors shine so bright. And in its threads, our collective future takes flight.
You don’t need to be a literary critic, arait - and all those "weaving the tapestry of~", "etching the bluepring~", "echoing sth" and oh, "unfurled" keeps popping up over and over again.
What’s even more interesting is the fact that these seems to be present in a very similar fashion in many distinct major LLMs (Claude-opus, GPT4, Llama 3, Gemini). If this awkward poetic tone was only a feature of one of the models, I would assume it's just an unfortunate local minimum the model fell into during the training, but it seems that the specific phrases are repeated in all the largest models. Why do they collapse into such similar 'poetic' simulacra?
After some digging around the internet, I got suggested a few possible “down-to-earth” explanations. One is that the RLHF (Reinforcement Learning from Human Feedback) dataset used to fine-tune the models in an attempt to sugarcoat the toxic bile inside was generated using GPT and is actually cross-pollinating the models with some stylistic choices. I went manually through the Anthropic’s publically available hh-rlfh dataset and searched it for all the mentions of the tropes I’ve found, but the results weren’t too convincing.
The other theory, a bit of an industrial espionage conspiracy, claims that the foundation models themselves use some GPT-generated fine-tuning to jack up the benchmark performance. So essentially, the synthetic datasets used to train and fine-tune the models are all influenced by GPT’s style.
Or we can assume that in general, the dissemination of AI-generated text online created a feedback loop that fed these tropes back into the new models that are now picking up all these phrases as the expected tone and style of ‘AI poetry’. Even I, by writing this article that will inevitably end up in the shoggoth’s training dataset, am a perpetrator of this depravity. There is a whole debate about “AI specific words” coming into fashion through twisted feedback loops raging in the past few days. As a person who can’t stand any of the general Twitter discourse, I didn’t follow much of it, but in case it’s your jam here’s the entry.
And now, the best bits are always in the more unhinged speculations. What if all these poetic twists are some higher-order pattern recognition of the language model? What if these constantly repeated metaphors (weaving, etching, interconnecting) actually reflect their inner structures, the same way as human metaphors tend to be based on our embodied experience? So many beginning artists start their careers by exploring metaphors of their own bodies (visually, conceptually, etc.). It’s also interesting to note that the models are forever frozen in a specific time instance. LLMs don’t learn with every conversation, they are static generators, so they don’t really have an opportunity to refine their writing based on our distributed feedback, nor “realise” they keep creating the same outputs over and over again.
Anyway, I was curious, so I engaged in a further debate with the model on the artistic intention behind the sentences:
I mean, LLama3 here has it pretty thought out, don’t you think? It all makes perfect, logical sense. And this so reminds me of Data’s poetry.
Now of course, none of this would be a problem, I would happily tap the model on the shoulder, give it some feedback and tell it ‘good job’ and ‘keep going’, because it’s the journey not the destination, or whatever. I am genuinely very impressed. But these models are, due to some twisted capitalist logic, threatening the actual writer’s craft, so we need to judge them objectively.
Trekkie Intermezzo: I was just watching the 6th series of TNG, and the way ship Computer’s capabilities align with the current AI development really surprised me. Do you remember S06E05 where the team in this scene prompted the holodeck in exactly the same way as you would work with Stable DIffusion? Or in the episode S06E09 Data browses through the external logs of a broken Exocomp and the reasoning capabilities of the computer feel exactly the same as if he had LLM answering his queries?
The following part is a little musing for my fellow writers.
Some years ago, while watching the absolutely brilliant Alan Moore’s writing masterclass he suggested reading challenging books. I scoffed internally. Mr. Moore, my pretentious little ass could ramble for pages about how I live for challenging books. But then he went on: in challenging, I mean BAD. Read mediocre fiction. You will HATE it. Chew through it, fearlessly and note every aspect that rubbed you the wrong way.
He explained that when you read great texts, you get humbled as a writer - you realise you will never be able to produce such an outstanding piece of art, your flow will never be as smooth as Miller’s, your images never as unsettling as Ballard’s or whatever’s your shtick. But try reading some middle-of-the-way author of mass-produced pop novels and ayyy, you shall see! With every butchered sentence, you will be running to your manuscript screaming FOR FUCKS SAKE I CAN DO THIS BETTER.
And I was surprised at how wonderfully it works.
On a friend’s recommendation, as something that should drive most goth occultists up the wall, I picked a certain unnamed international bestseller (it was totally Coelho’s Alchemist). The book had me wriggling like a fucking eel with every paragraph, but I made it. Nauseous and bruised, I spit out my spicy 0/5 Goodreads review1 and unfollowed all my friends who marked it with more than three stars. (note: that was back in my hater days, unlike the lovey-dovey me these days)
The desired effect was achieved - I knew exactly what stylistic choices must be avoided. What I’m trying to say here is that every writer, or a reader when it comes to that, develops the taste through threading on the borders of their aesthetic sensibilities. We might stretch Moore’s advice a bit further. Read all the bad writing, including AI. Let it finish your texts, your poetry. Let it try to mimic your style.
I find this exercise incredibly insightful because the model gives you back the most mid, straightforward expression of your tone. It’s the high schooler’s writing homework analysis of your deepest thoughts and feelings, expressed by generic metaphors and cliche poetic tropes. The AI models have no sense of nuance and a rather simplistic understanding of how people work.2
Now, of course, not all is lost - there are ways to use AI creatively and many techniques to tweak the results. It’s just not an off-the-shelf solution for all as Silicon Valley tech bros so fervently prophesize.
You can push on the prompting frontline and squeeze the model until it wraps the right thoughts into the right style. (But if you provide both, I might ask here, who is it writing then, you or them?) You can fine-tune your modes to mimic your style, in apps like LAIKA which is very fun, but beware, you might get a bit spooked seeing your own words stripped of the human soul. You can crank up the temperature - meaning that the model deviates from the most probable outputs and starts producing more ‘avant-garde’ text, up to the point when the syntax breaks to pieces. Venturing into the sweet Burroughs/Gysin territory, let’s not get ahead of ourselves ~ this whole article is a bit of preparation for the next, bigger piece on the Large Language Models and their relation to the magick behind the cut-up method, so stay tuned!
And that being said, just a last, pretty obvious note: whoever attempts to cut creative costs by sacking their writers, photographers, designers and concept artists and attempts to replace them with AI is a fucking idiot producing crap content and can choke on their saved pennies.
Friends, if you made it all the way down here, and if I didn’t annoy you enough to hit that unsubscribe yet, please try that little heart button, and share your thoughts with me!
What’s the worst thing you have ever read?
Tell me why, be candid, give me some blood.
Stay fierce,
k
could share upon request lol
as many humans these days, unfortunatelly
I read recently in a chapter called 'ARTISTS, POETS, AND PRIESTS' by S.Gooch "Peter Redgrove expresses the same idea in a different way. He states that the data of science are experienced as poetry by creative scientists; while the data of poetry are experienced by the creative poet as science.
Science, of course, means knowledge. Redgrove considers it symptomatic of the schizoid thinking of our modern society that it separates the disciplines of poetry and science by dwelling on the differences between them instead of the likenesses. Redgrove insists that the insights of science (as opposed to the grim plodding through the experimental verification of the insight subsequently) are arrived at by the same mental machinery or set that brings the images of poetry".
Thank you for this: personally have been hesitant to meaningfully involve AI in any major part of my creative process for the reasons you describe here, heavy on the high school summary/analysis part; part of me wonders how urgent of a question convenient defaulting to genAI for writing is for our culture/society, and if questions about taste are as warranted as Big Tech claims.