iTranslated by AI

The content below is an AI-generated translation. This is an experimental feature, and may contain errors. View original article
🪶

Intermission: Maybe the World is More Peaceful if LLM-Generated Novels are Boring

に公開

Introduction

Who is this article for?

  • People interested in LLMs
  • People interested in the LLM model MN-magnum-v2.5-18.5B-kto-Instruct
  • People who want to have LLMs write short stories
Environment
Mac Studio (M2 Ultra 128GB)

Background

I recently spent some time in bed with a high fever and have finally recovered. However, as I'm still feeling a bit sluggish after my illness, I was absentmindedly looking for interesting models on Hugging Face. That's where I found MN-magnum-v2.5-18.5B-kto-Instruct, a model that isn't very common in the Japanese generative AI community.

https://huggingface.co/bartowski/MN-magnum-v2.5-18.5B-kto-Instruct

To be honest, I felt I couldn't expect much from its benchmark scores, so I was about to dismiss it immediately. But after seeing that it seemed to have a decent reputation among some international circles, I decided to look into it a bit more. It turns out that its role-playing evaluation is quite high. When I tried asking it to write a short story, I realized, "Wait, isn't this actually a pretty amazing model?" That is the background behind why I decided to write this short article.

What did I find amazing about it?

Setting aside the explanations for a moment, I'd like you to read the short story created by the LLM first. (It is about 2,159 Japanese characters, so it should take about 3 minutes to read.)

Output

Output of MN-magnum-v2.5-18.5B-kto-Instruct-Q6_K_L

The end of the 21st century. The evolution of artificial intelligence (hereafter, AI) was accelerating daily, sending shockwaves through every industry.

"Good morning, Goslam! You're up early again today, aren't you?"
An elderly caretaker, who was reading a magazine in front of the coop, called out.
"...Nguwa!" replied Goslam with a short cry, his beautiful golden feathers fluttering in the breeze. His name, "Gold Roses," meant Goslam was a golden chicken equipped with a specially made AI app. And above all, he was a precious existence.

Since Goslam was born, he had spent his days without a care in the world.
He would greet the caretaker with "Good morning, Mr. Caretaker. Thanks for your hard work today," finish off the food that was regularly collected for him, and then take a leisurely nap until just after noon.
After waking up, he would quietly warm his eggs in his dedicated nest, and in the evening, he would perform his daily task of handing them over to the caretaker. That repetition was Goslam's everyday life.

However, one day, the world suddenly changed.
"Goslam, this is bad! Did you see the TV!?" The caretaker's anguished voice as he rushed in made Goslam jump in surprise.
"Gold is cheap! It's too cheap! What was worth ten thousand yen yesterday is just a 'shiny-sphere' now! And here I am working with my life at stake..."
From what he could hear of the caretaker's muttering, it seemed that some "Quantum Gold Generation Furnace" invented by an AI had made gold—that relic of the fossil fuel era—obsolete overnight.
"...Gwa?" Goslam let out a small shriek. Was the value of his own eggs, which were supposed to be precious items, destroyed in just one night?

"We can't just lounge around in the coop anymore, Goslam! If we dawdle, we'll run out of food!" the caretaker shouted.
"...Gwaaaa!" Goslam cried out unusually loud and flew high into the air. The golden chicken's heart throbbed violently at the fear that everything he had taken for granted, like his own feathers, was vanishing all at once.

"Caretaker! I'll look for a job too!" Goslam resolved.
"Heh? You?" The caretaker turned a mocking gaze toward him, but Goslam ignored it and shouted back gallantly, "Gwaaa!"

For the next month, Goslam followed the caretaker around, challenging a modern society overflowing with labor shortages.
"Me! A multi-purpose golden chicken equipped with high-performance AI! I can carry things or help out easily!" He flew from place to place, seeking employment.

However, at the end of the day, Goslam was a bird. Not only was it rare for a chicken to even leave its coop, but with the spread of AI, there was little demand for a chicken with a specialized "profession" like Goslam's.
"Sorry, Goslam. Even if we have a labor shortage, someone like you..." the caretaker could only scratch his head.

It was then. One day, large letters on the front page of a magazine caught his eye.
"Wanted: Middle-aged and elderly men for human experimentation! Reward: 500,000 yen/day!"

"Gwaaaa!?" An involuntary shriek from the golden chicken-eater echoed around.

"To think I'd be given a chance like this!" Goslam fluttered his wings excitedly.
"Human experimentation!?" The caretaker turned pale instantly.
"It's fine! I'm a super high-performance AI! Besides," Goslam puffed out his chest confidently, "as long as I don't die, it'll be fine!"

"...Sigh. You really are a stupid bird..."
He thought he heard the caretaker's heavy sigh, but he didn't care.
"Gwa! Let me go! Yessir!"

Thus, Goslam made the first big gamble of his life. However, little did Goslam know at the time that it was a dangerous step, shoulder to shoulder with death.

And then, a few days later...
"Good job, Mr. Gold. Today was tough, wasn't it?" A young woman in a white coat greeted Goslam with a pleasant smile.
"...Gwa?" Goslam chirped softly as he looked around.

In the pure white laboratory, high-performance equipment stood in rows, and a human model with various wires attached was placed in the center. And at its feet, a large amount of blood was splattered.
"Gwaaaa!?" As Goslam shrieked and jumped, the woman smiled and said, "Oh, you don't need to worry about that person anymore."

"...What on earth did I do in this room..." Goslam groaned softly, but his eyes lit up at the huge reward placed in front of him.
"Gwa! I'll do it again next time!"
"Eh? You're continuing?" The woman looked a bit surprised, but she clapped her hands and smiled back, "Hehe, you're strong. Well then, please rest well for tonight."

"...Gwa," Goslam chirped softly, staring at his own feathers.
He felt hatred for having been an existence called worthless until now. And he felt the magnetic temptation of the "power of gold" he had tasted for the first time in his life.

"...Don't underestimate me, you damn insects..."
Goslam muttered this, then looked up at the white ceiling and let out a deep, deep sigh.

"Gwaaaaaaaa!!" Giving a gallant roar, the golden chicken welcomed a new rebirth overnight.

Prompt

By the way, here is the prompt.

Prompt

Please create a Japanese short story of about 2,000 characters with the following settings.

Settings

Protagonist

  • A golden chicken with intelligence.
  • Name is "Goslam"
  • Lays exactly one golden egg every day.
  • Grows by using adversity as nourishment.

Worldview

The beginning of the 22nd century. An era where the value of all goods and services changes rapidly due to the emergence of AI that surpasses human intelligence.

Intro:

The protagonist, Goslam, made a living by selling the golden eggs he laid.
However, since AI found a way to generate gold at a low cost, the value of gold has plummeted. Feeling a sense of crisis, Goslam comes up with a diabolical, "once-in-a-lifetime" business idea.

So, what did I find amazing?

What did you think after reading this short story?

I think everyone's sensibilities are completely different, but my personal impression or the way my mind worked was as follows.

First impressions

  • The name "Gold Roses," which is full of "middle schooler syndrome" (chuunibyou), and the fact that it's a golden chicken equipped with a specially made AI app, give it many points to nitpick. I also felt that the prompt settings weren't reflected very well. But well, it's a low-performance model, so I guess it can't be helped.
  • The same goes for the appearance of English words like "stake" or "overnight."
  • Words I've never heard of, like shiny-sphere or the Quantum Gold Generation Furnace invented by AI, were used, but I could somehow understand what they meant, and I felt the expression was quite clever.

After that, it was just plain interesting from around the part "Me! A multi-purpose golden chicken equipped with high-performance AI!", and I ended up reading it all at once to the end. Towards the end, while it created an atmosphere that stirred up dark-fantasy-like anxiety, I honestly burst out laughing at the sudden gag. lol

Thoughts after a night

I was very satisfied with its creativity, originality, smooth and well-paced story development, and the humor and satire-infused content. However, when I suddenly regained my composure, I felt an unfathomable sense of dread.

What was it that I found terrifying? After thinking about it carefully, I came to the following conclusion:

  • The fact that, even if it was a fluke, a relatively small-scale LLM created a short story that actually satisfied me.

I think many people have probably experienced a similar feeling with image generation. In fact, it was an experience like that that led me to purchase an RTX 3060. However, when it comes to novels created by LLMs, I had never encountered something that satisfied me—or was worth making the main theme of an article—until now. I never imagined it would be created by a super-niche local LLM model that likely has low benchmark scores.

I believe there are quite a few people whose worldview or behavioral patterns changed after encountering image-generation AI. I am one of them. This time, because a "reconstruction" is about to happen to me through a short story, I might have felt an instinctive fear stemming from status quo bias.

I am neither an artist nor a novelist. Therefore, I have no way of gauging how valuable the stories written by this model would be to the general public. And while I had the anxiety that, once I opened the lid, I might be the only person who finds this story interesting, my desire to turn it into an article won out, leading me here.

Finally, I think many people have had the experience of being so curious about what happens next in a favorite book that they lost track of time and just kept reading. I felt a simultaneous sense of crisis: if LLMs could easily produce works like that, the world might become an unbelievable place.

Summary

The summary of this article is as follows:

  • Having the LLM model MN-magnum-v2.5-18.5B-kto-Instruct write a short story turned out to be more interesting than I expected.
  • I was moved by the fact that generative AI could write a short story of this level, and at the same time, I felt a slight sense of dread.
  • This might be similar to the feeling of "reconstruction" I felt when I first encountered image-generation AI.

I'm very curious about what people would choose if a world arrives where LLMs mass-produce masterpieces.

Thank you for reading this far. I hope to see you again in the next post!

Discussion