a thoughtful web.
Good ideas and conversation. No ads, no tracking.   Login or Take a Tour!
comment

    1) It is possible as an individual to write 5,000 words of mediocrity not just with your own elbow grease but now with "Siri, write me a story" as well.

The 5,000 words of elbow grease will represent an individually-tuned run through the "training data" that human has consumed. The direction it will take will be refined on that human's own reward system. Each training data and each reward system will be different. If you ask that human to give you 5,000 words of the same story four times over, the human will naturally refine the story whether they like it or not - the human will learn the story.

"Siri, write me a story" will be the stochastic middle of a uniform training set, trimmed and edited to keep the LLM's owner out of trouble. It will be the same training data and reward system regardless of who asks for the story and the direction it will take will be defined by the committee that chose the general shape of its outputs. If you ask that LLM to give you 5,000 words of the same story four times over, it will make four generic runs through the training data, each iteration completely unrelated to the iteration before.

    2) It is possible to improve written stories in a way you want as often as you want. "Siri, take that writing advice I heard Brandon Sanderson talk about and apply it to this paragraph." Aka the aforementioned vibewriting.

It is not possible, however, to maintain the corpus of the story. "Siri, take that writing advice I heard Brandon Sanderson talk about and apply it to this paragraph without fucking up every other paragraph" requires a context awareness that LLMs CANNOT have. It can know that humans have five fingers if it has a model for "human" and a model for "fingers" and a model for "five" and ridiculous amounts of effort have been expended to ensure that the models produce fewer six-fingered men. LLMs have a hard time maintaining consistency from frame to frame; there is no latent context awareness to LLMs and there cannot be. There can only be scaffolds to refine the runs through the model such that they stay out of the Uncanny Valley.

    3) With LLM memory becoming larger and larger it is becoming increasingly more viable to manage consistency across larger pieces of text. "Siri, foreshadow this event in the previous 200 pages in 5 different places."

See, with each step you stray further and further from "things LLMs do" to "things writers do." "this is foreshadowing" is symbolic thinking and LLMs don't do that. They classify. You can do a lot with a classifier! but the thing you can't do is create. You can ape. You can imitate. You can reproduce. If you show an LLM a duck and a snake you can ask it to draw you a ducksnake. If you ask an author to give you a story about a ducksnake they will immediately think about what the symbols "duck" and "snake" imply. If you ask an LLM:

    Once upon a time, in a distant swamp hidden deep within a lush forest, there was a creature unlike any other. It was known as the Ducksnake, a strange fusion of two animals that should never have crossed paths. The Ducksnake had the body of a long, slithery snake, covered in shimmering emerald scales, but with the head of a curious duck, complete with a bright orange beak and soft feathers that glowed like sunlight on water.

    This creature, named Quackers, was a solitary being who roamed the swamp, always on the lookout for adventure. Quackers had a unique gift: it could swim gracefully through the water like a snake, yet waddle across the land like a duck. The other animals in the swamp were both perplexed and amazed by Quackers. The frogs croaked in confusion, the turtles stared in wonder, and the birds above laughed in a melody that sounded like they were trying to mimic Quackers' peculiar waddle.

    Quackers was a gentle soul who loved to explore, but it also had a problem: it didn’t quite fit in anywhere. The snakes didn’t understand it, thinking it was too fluffy, and the ducks thought it was too slippery. One day, feeling particularly out of place, Quackers set off to find a place where it truly belonged.

    As it ventured deeper into the heart of the swamp, Quackers came across a group of animals trapped in a rising tide. The beavers were trying to shore up a dam, but the water was flooding faster than they could build. The ducks were flapping in a frenzy, and the snakes slithered frantically, unable to offer much help.

    Without thinking, Quackers jumped into action. It slithered through the water with ease, using its long, snake-like body to create a path for the beavers to work. Then, with its duck head, Quackers used its beak to peck at nearby branches and bring them to the beavers, offering its strange yet helpful assistance. Quackers waddled along the shore, urging the ducks to get in line, using its unique body to coordinate efforts between the two species.

    By the time the tide began to recede, the animals were able to finish their work and prevent the flood. As the swamp settled back into calm, everyone gathered to thank Quackers. The frogs croaked in approval, the ducks quacked in delight, and the turtles clapped their little paws in appreciation.

    From that day on, Quackers was no longer an outcast. The creatures of the swamp learned that being different didn’t mean you didn’t belong. In fact, it was Quackers' very uniqueness that had saved them all. And so, Quackers, the Ducksnake, found a new sense of purpose, becoming the swamp’s most beloved and revered hero—a creature that proved you don’t have to fit into one mold to make a difference.

    And every now and then, you could still hear the soft quack and hiss echoing across the swamp, as Quackers continued to roam, always ready for the next adventure.

There's no there there. It knows what a duck is, it knows what a snake is, and it gives you a tale of forest creatures and cooperation because anybody asking for a story about an (x) is likely to be doing it for their kids.

    4) By chopping up tasks to multiple layers of agents, it is becoming increasingly more viable to cover all bases of writing by delegating tasks to specific agentic tools, from the abstract ("Siri, to what degree does this text convey this idea I have?") to the specific ("Siri, go through every sentence of the entire manuscript and ensure apostrophes are set correctly.")

(waving hands continues) Sure. Which is why all projects-by-committee always kick the shit out of individual effort, right? You're at the infinite monkeys again, taking it from the other direction - "if we break it up into smaller and smaller parts surely a machine can do it."

    5) The above can lead to a text which is good enough that it can be hard to distinguish by readers as being AI written.

Presumes facts not in evidence. Look at it another way - you can write so badly that people think it's AI. Anyone can. It's become an epithet. But we've been saying "hey it could get there some day!" since Sunspring.

WATCH THE ABOVE. It's from an era where you had to build your own LLM. And they did. And it output hot nonsense. It's ridiculous. And it's charming.

Now watch the below.

You can't even make it through, can ya? Sure... it looks like a movie. AI did all of it! Except - and here's the funny thing - the script. Yeah, that's purely human dreck up there. Even the guys spending months tuning bullshit AI movies don't fucking trust them to come up with the story. They recognize that dreck as their shit is, the AI dreck is so much worse that it's not even worth using it for clout points.

    Hell, it could even be enjoyable to read!

And that, right there, is you going from "arguing" to "hoping."

    I mean, we've talked before about how LLMs are a specific tool which in the hands of creatives will lead to new and better art, even if using said tools feels heinous at first. Why would that be different for fiction writing?

BECAUSE THERE ARE NO TOOLS TO FICTION WRITING. Speak it out to your daughter while you're going blind, write it in your Moleskine, carve it on the cave wall, dictate it to Siri, type it out into Word... there are tools to get it from your head to the paper but storytelling and singing are the only arts we practice with no tools at all. Sure - AutoTune will get you closer to hitting the right key and autocorrect will make sure your grammar and spelling are perfect. But there's nothing else you can do where you can stand there naked in the dark and create art indistinguishable from what you'd make with every tool and every dollar available.

    But I do think that you can create 80% of the scaffolding of a book in an afternoon and work from there.

Who fucking cares? Nobody reads 80% books. Go look up "AI short film." They all look like cutscenes to a game no one wants to play. It's the last 20% that gets you there. in everything. And AI has consistently not even begun to cross that 20% in all the years we've been yammering about LLMs.

    I believe you can forego writing groups and sharpen your thinking by loading 40 of the best books on writing into 40 LLM agents and have them have a go at your manuscript. I'm noticing that the barrier for myself to write longer form has been significantly lowered, because I know I can use these tools to create a version of my writing that is much better than I'd be able to pull on my own.

Right - you've made this argument about code before. You get more coding done because you aren't a coder. The argument always boils down to "I don't know what I'm doing, and AI helps me fool myself into thinking I don't have to."

    But I do think it matters a great deal what we compare that mean to, I do think it matters in whose hands the tool is whether what it churns out ends up, in the final product, as anything good.

Okay, how's this for an expert opinion: I have been paid multiple thousands of dollars for my writing, and I have yet to see AI offer me anything other than crude translation.

There's this idea that if AI can help the 79% skillful make it to 81% competence, the 99th percentile shall be out of business. Now - I'm out of business for writing because it doesn't pay well enough for me to bother. LLMs sure as shit aren't going to fix that. Chuck Wendig gets about $5k/manuscript. Your AI gonna write better than Chuck Wendig? And let me be clear - Chuck Wendig is a mediocre writer. We're setting a low bar here. The prior argument is if we let the LLM run a hundred thousand times it might maybe somewhere in there crank out something as well as Chuck Wendig because infinite monkeys theorem.

Why not just fucking pay Chuck Wendig?

    But it's hard to argue against LLMs being able to produce writing that is useful to some people under some circumstances, no?

No, it is the easiest fucking thing in the world. Especially when where we started was "this is the death of writing" and you're now backed into the "useful to some people under some circumstances."

    Dr. Makena Okafor manipulated the holographic projection of Jupiter with her prosthetic left hand, the embedded sensors in her artificial fingers interfacing seamlessly with the lab's quantum imaging system. The gas giant's northern aurora bloomed before her eyes—a spectral dance of charged particles rendered in false color, swirling in patterns that had consumed her waking thoughts for the past three years.

AYFKM

    Renowned curator Jacques Saunière staggered through the vaulted archway of the

    museum's Grand Gallery. He lunged for the nearest painting he could see, a Caravaggio.

    Grabbing the gilded frame, the seventy-six-year-old man heaved the masterpiece toward

    himself until it tore from the wall and Saunière collapsed backward in a heap beneath the

    canvas.

That's the first paragraph of The Da Vinci Code, widely regarded as some of the most execrable writing of the past 100 years.

    "It was a dark and stormy night; the rain fell in torrents--except at occasional intervals, when it

    was checked by a violent gust of wind which swept up the streets (for it is in London that our

    scene lies), rattling along the housetops, and fiercely agitating the scanty flame of the lamps that

    struggled against the darkness."

And that's the first paragraph of Bulwer Lytton's Paul Clifford, widely regarded as the worst writing until Dan Brown.

BY INSPECTION the AI is worse than both. Both Dan and Bulwer were going for it and you can read it in every pustule of purple prose. But at least they don't sit there like limp pieces of shit.

yet for some reason errbody wants to pretend that limp pieces of shit aren't limp pieces of shit.

we have taken so. many. steps. backward.