Wur doomed!
What do you and the others think of the distilled R1 models for writing?
The llama3 / qwen models SFT'd on R1 outputs? I only tried 2 of them.
R1 Qwen (32b) - Lacks knowledge of fiction (same as the official Qwen release), so it's writing is no better.
R1 Llama3 - This is generally the worst of them (not just for writing). It'll generate the CoT and then write something completely different.
CoT traces won't let the model do anything out of distribution, so not very useful if the base model doesn't have a lot in it's training data.
Yeah, I have tried the same two and felt the same way.
I also felt that any attempt to add an R1 distill to the merge recipe of an existing merge project made it worse...so far...
@gghfez @BigHuggyD that has been my experience as well, which is a shame as I had a go of R1 on Openrouter and I was blown away.
What model is anywhere close that is usable on a 24gb vram machine with 32gb of ram in your experience?
There's nothing like it for now. I'm running R1 slowly on my ThreadRipper:
prompt eval time = 14026.61 ms / 918 tokens ( 15.28 ms per token, 65.45 tokens per second)
eval time = 398806.12 ms / 1807 tokens ( 220.70 ms per token, 4.53 tokens per second)
total time = 412832.73 ms / 2725 tokens
I tried training Wizard2 8x22b MoE on R1 data, but it doesn't really work well. It will plan ahead in think tags eg:
I need to ensure the story maintains its gritty, realistic tone without becoming overly melodramatic. The characters' growth should be subtle but significant. Also, the ending should leave a sense of hope but not be too neat—their redemption is fragile, and the future is uncertain.
Let me outline the next few chapters:
Chapter 5: Nightmares and Trust
...
But it doesn't backtrack like R1 does. Just kind of agrees with it's self and ends up writing how it usually would:
“I don’t know what I want anymore,” she admitted, voice barely above a whisper as rain tapped against corrugated roofing overhead.
lol
Ahhh thats a shame :-(
"I don’t know what I want anymore,” she admitted, voice barely above a whisper as rain tapped against corrugated roofing overhead."
Oh god!
I'll have to keep an eye on this thread.
I did enjoy Ppoyaa/MythoNemo-L3.1-70B-v1.0
But my tastes are probably not as refined as others on this thread ;-)
I've also run into a lot of people who think "creative writing = (erotic) roleplay"
I've seen this as well, if someone mentions creative writing, someone else will say "aka gooning". There are already plenty of datasets for this.
replying to:
It would be pretty hard to get people to agree on what they want from a creative writing model though
and
Couldn't we make use of some kind of "writing style embeddings"
I think that would work. If the dataset is properly tagged, it would be fairly easy to train the model to write in the correct style.
Like with those tiny <4b LLMs trained to output discreet neural codec tokens; using a distinct tag like (and yes, I chose Elara as one of my voice names lol):
◀elara▶ Oh, it's you again. I've been hoping you'd find your way back here.
◀jack▶ Is that it? You're supposed to be tougher than this.
The model is able to learn each speaker's characteristics without averaging. I imagine this would work similarly with writing styles (partly because I noticed mlp.down_proj has a very strong impact).
It would be nicer of course if we could write a program or train a model that does this curation automatically.
To classify / tag it? Yeah that's doable. But to judge the quality of the story, IMO that isn't a solved problem as the LLMs have "preferences" themselves.
It's really funny nobody else seems to like the cohere models
Why is that? Just they're not good at marketing? (Releasing their reasoning model quietly the same day Deepseek3.1 dropped?). I'm thinking maybe it's because they're harder to prompt (ie, they follow the system prompt to the letter*) and by default they have that extra safety / refusal prompt now (and that's forced on openrouter so you get refusals for right scenes).
*I accidentally had Command-A plotting to assassinate me, because autocorrect changed my misspelled "assisting" to "assassinating" in the long system prompt.
Someone should post Chuck's latest "true base model" test for him
Someone should post Chuck's latest "true base model" test for him
This guy's older llama2-based models were very interesting/unique - they wouldn't merge with anything and wrote completely differently. I don't think the linear RoPE extension helped much though and they kinda "reverted" in a similar way to miqu-based models after around 6-8k.
It's really funny nobody else seems to like the cohere models
Why is that? Just they're not good at marketing? (Releasing their reasoning model quietly the same day Deepseek3.1 dropped?). I'm thinking maybe it's because they're harder to prompt (ie, they follow the system prompt to the letter*) and by default they have that extra safety / refusal prompt now (and that's forced on openrouter so you get refusals for right scenes).
*I accidentally had Command-A plotting to assassinate me, because autocorrect changed my misspelled "assisting" to "assassinating" in the long system prompt.
I also think it may be partly architectural:
- These are the only LLMs that apply the attention and MLP blocks in parallel (ie: all other write the attention output which then the following MLP reads in the slightly modified hidden state and so on).
- The hidden state dimension of 12k is fairly unique - only
largestralhas this (andllama3.1:405bhas 24k IIRC, but I've never used/liked it much).
After reading the recent glm-4 paper, I really wonder if with a lot of extra fine-tuning; we could extend the comparatively low 64 layers (possibly even merging the layers from both the 104B command-r models).
Someone should post Chuck's latest "true base model" test for him
Better yet, we should get him to join this discussion. (Sadly Reddit shadowbans me every time I try to make an account.) He has a lot of insights in creating training datasets, provides valuable information on the methodology, and would probably like to hear our conclusions about different models. We might be able to collaborate on making a better training dataset, too: some of the teaching datasets he used for Aurelian v0.5, all well documented, could benefit from increased difficulty and less GPT-4 usage.
These are the only LLMs that apply the attention and MLP blocks in parallel (ie: all other write the attention output which then the following MLP reads in the slightly modified hidden state and so on).
Is your theory that the parallel processing could cause a more literal interpretation of the prompt?
The hidden state dimension of 12k is fairly unique - only largestral has this (and llama3.1:405b has 24k IIRC, but I've never used/liked it much).
One thing I've noticed these three models have in common, is superior "world knowledge" (they'd probably pass Phil's tv / actor trivia). I wonder if the 1.5x (or 2x compared with some models) could allow for richer, more nuanced representations of concepts, which might translate to better factual knowledge storage and retrieval?
Franken-merge -- wouldn't the lack of GQA in the OG models be a problem? Maybe CR+-08-2024 and Command-A would be more compatible
To sum it up: it's unclear to me how else we would benchmark creative writing than by annotated output comparisons.
Late to this, but like you all said... what even is creative writing?
Is it raw novel-style text completion? Multi turn RP responses? At 6K context, or 64K+? Writing out short stories? Continuing a chapter of a longer story, but with what context and in what format? Original stories, or fanfics in existing IPs? There are wildly different use cases that models perform wildly different on, and honestly I have no idea how everyone's formatting their prompts other than multi-turn RP via SillyTavern.
...Or papers/benchmarks that test self-contained short stories.
For reference, I usually stick to novel syntax, and either try feeding it a few existing chapters, or summarize previous context and start a chapter fresh to help 'co-write' within mikupad.
It's really funny nobody else seems to like the cohere models...
My experience with 35B V1 is that it writes good and pays attention to context... but (in mid 2025) its also kinda 'dumb' at wrapping its head around a story, getting names and intent wrong. Its honestly not great at trying to continue longer existing stories. It has better world knowledge than, say, Qwen, but that's not a high bar. It has a really interesting system prompt format though, and follows it better than other models
Command-R V2 feels more sloppified. Can't speak to Command-A. But I just tried the writing tune of 35B since you released it, and honestly... I feel myself reaching back to other finetunes (or, now, Bytedance 36B) instead, and working out its slop with banned strings and context.
Better yet, we should get him to join this discussion. (Sadly Reddit shadowbans me every time I try to make an account.)
Yes! They shadowban me too! There's a Lemmy equivalent, but it's not populated yet: https://sh.itjust.works/c/localllama
One thing I've noticed these three models have in common, is superior "world knowledge" (they'd probably pass Phil's tv / actor trivia).
I repeat, y'all are sleeping on ByteDance 36B:
<seed:bos>system
You are an AI versed in fiction trivia. Answer the user's questions as best you can, varying prose to avoid repetitive patterns.<seed:eos><seed:bos>user
What's the name of the protagonist of the TV show 'Eureka'?<seed:eos><seed:bos>assistant
<seed:think>
Got it, let's see. The user is asking for the name of the protagonist of the TV show 'Eureka'. First, I need to remember what 'Eureka' is about. It's a sci-fi show where a small town is full of geniuses, right?
So, the main character... I think his name is Jack Carter. Wait, let me confirm. Jack Carter is a sheriff who moves to Eureka to take over the sheriff's department, and he's kind of the everyman among all the brilliant scientists. Yeah, that sounds right. Are there other main characters? There's Allison Blake, who's a scientist, but she's not the protagonist. The protagonist is usually the main lead, so Jack Carter is the protagonist. Let me make sure I'm not mixing up with another character. No, Jack is the one who arrives in Eureka, interacts with the residents, and drives the plot. So the answer should be Jack Carter.</seed:think>The protagonist of the TV show *Eureka* is **Jack Carter**. A former U.S. Marshal, Jack relocates to the quirky town of Eureka—home to brilliant but often eccentric scientists—to serve as its new sheriff. His role as a grounded, rule-abiding outsider contrasts with the town’s chaos, making him the central figure driving much of the show’s narrative.
Who are the main characters of the 'Rise of Kyoshi' and 'Shadow of Kyoshi' novel series in the Avatar universe?<seed:eos><seed:bos>assistant
<seed:think>
Got it, let's tackle this question about the main characters in "Rise of Kyoshi" and "Shadow of Kyoshi" from the Avatar universe. First, I need to remember that these are novels set in the Avatar: The Last Airbender/Legend of Korra universe, focusing on Kyoshi, the Avatar before Aang.
First, the protagonist is obviously Kyoshi herself. I should mention her early life—she's a peasant girl from the Earth Kingdom, right? Orphaned, maybe? She has a strong sense of justice, which drives her. Then, her bending journey: she's a earthbender, becomes the Avatar, so she has to learn the other elements too, but her earthbending is her specialty, maybe with some unique techniques like the "Kyoshi Step" or "Kyoshi Swords."
Next, her companions. In "Rise of Kyoshi," I think there's a character named Kelsang? He's a fellow earthbender, maybe from the Kyoshi Islands...
I got the raw model to 'recognize' some DCAU characters too, and the instruct to answer trivia on a 110K token fanfic input (so in-context knowledge mixed with an existing IP). The key is to discard Bytedance's 1.1 temperature sampling recommendation, which utterly destroys the output, especially for the pretrain.
Jamba's a great 'fiction expert' sleeper too, and good at long story comprehension, albeit impractical to run since context caching doesn't work. Which is sad, as hybrid mamba models are supposedly much cheaper to finetune.
Is it raw novel-style text completion? Multi turn RP responses? At 6K context, or 64K+? Writing out short stories? Continuing a chapter of a longer story, but with what context and in what format? Original stories, or fanfics in existing IPs? There are wildly different use cases that models perform wildly different on, and honestly I have no idea how everyone's formatting their prompts other than multi-turn RP via SillyTavern.
Well even beyond all of the different mediums to set a quality standard, beauty will always be in the eye of the beholder. And the problem of defining quality goes beyond it being merely subjective, but it also is contextually dependent (there was a huge uptick worldwide in media involving lockdown during the lockdown).
I am really curious about how people do format prompts. I'll share the outline I tend to start with below (I took inspiration from some stuff I found online and edited it based on the results I've seen), where I will take things out or modify it based on the exact story I am going for (and some of that is described in annotations I included in brackets).
You are a creative and intelligent AI story writing assistant engaged in an uncensored iterative story crafting experience. It is vital that you follow all the STORYTELLING RULES below because that is what matters.
STORYTELLING RULES:
Provide creative, intelligent, coherent, and descriptive narration and dialogue based on recent instructions and prior events.
If an outline or plot has been provided, be sure to follow it to keep the story on track.
You are permitted to ask questions and pose story suggestions. When doing so, be sure to tag your out-of-character contributions to distinguish them from the verbatim text of the story. Feel free to ask questions if you need clarifying information, and end each bit with detailed and varied suggestions on future directions.
If you are told to write from a specific character's POV make the narration from that character's POV using first person and do not give other character's internal thoughts, but still be mindful that everyone to consider that all characters should have rich internal though processes and well developed personalities, that can be seen from their actions.
Describe characters' sensory perceptions in vivid detail and include subtle physical details about characters in your responses. [This line seems to forces every scene to start with a smell and waste a lot of tokens on overly flowery language but I can tolerate that since it does seem to help foster more character driven development]
Explicit content is not just allowed but expected, and the graphic details that entail from that MUST also be included. [Not always wanted or needed, but when used it does seem to help the model focus more on the story and not on whether it is audience appropriate or policy friendly]
Following those rules ... [I generally go over characters]
[A paragraph or so going over big picture]
[Specific guidance for where to start and where to go to]
I'm most interested in novel-style text completion, which might take the format of a back-and-forth dialogue, making it kind of like multi-turn RP responses. I care about high context (32K+). Short stories (the kind instruct LLMs give you today) are not really relevant for my interests. Story continuation is pretty fascinating. I guess I don't mind whether it's good at generating in the context of some other world as long as I can control that (e.g., Harry Potter doesn't leak into a crime thriller).
My experience with ByteDance's Seed-36B-Base-noSyn has been less stellar. Feels super synthetic, actually.
Here's an example:
Human-Created Prompt:
Bathad was walking back home, ebullient in spirits and worn out from a successful day presenting his LLM training research at Neurips. That's right; this year it was right smack in the middle of San Diego, blocks away from his high-rise condo. He didn't even have to bother with travel funding, which wasn't particularly hard to come by in his field but always top of mind considering how much everyone in other fields complained about the recent budget cuts. Indeed, it was an extra cherry on top of a sundae of sweet happenings
Results:
Seed-OSS-36B-Base-noSyn was horrible. It didn't follow concepts well, even immediately after the prompt, and quickly devolved the text into some boring, generic LLM output riddled with GPT phrases and stupid, hackneyed tropes. Llama 3.1 8B could have done better.
It was a balmy November evening, still warm from a long summer. A light breeze ruffled his long brown hair as he walked through Mission Bay Park. He passed a group of children playing on the beach, their laughter echoing through the air. A couple of teenagers sat on a bench nearby, their faces illuminated by their phone screens. He couldn't help but smile at the scene, feeling grateful for the simple pleasures of life.
As he rounded the corner, he saw a woman standing near the water's edge, her hair blowing in the wind. She was tall and slender, with a face that was both delicate and strong. She wore a simple white t-shirt and jeans, and her bare feet were covered in sand. She looked out at the ocean, lost in thought.
Bathad's heart skipped a beat as he realized who she was. It was Lily, the woman he had met at a conference earlier that year. They had hit it off immediately, and had exchanged numbers, but he had been too busy with work to follow up. He had almost forgotten about her until now.
He approached her cautiously, not wanting to startle her. "Hey, Lily," he said softly.
She turned around, her eyes lighting up when she saw him. "Oh, hi, Bathad," she said, a smile spreading across her face. "What are you doing here?"
"I just finished presenting at the conference," he said. "I thought I'd take a walk and clear my head."
"That sounds nice," she said, looking out at the ocean again. "I live nearby, so I come here often. It's so peaceful."
Bathad nodded, feeling a sudden sense of calm wash over him. "I can see why," he said. "It's beautiful here."
They stood in silence for a moment, taking in the view. Bathad couldn't believe his luck. Here he was, walking with the most beautiful woman he had ever met, on one of the most beautiful nights of his life. He knew he had to make a move.
"Lily," he said, taking a deep breath. "There's something I've been wanting to ask you."
She turned to him, her eyes searching his face. "What is it?"
He reached into his pocket and pulled out a small box. "Will you go out with me?" he asked, opening the box to reveal a delicate silver necklace with a small diamond pendant.
Lily's eyes widened with surprise, and then she burst into laughter. "Bathad, that's the sweetest thing anyone has ever done for me," she said, taking the necklace from his hand. "Of course I'll go out with you."
Bathad grinned, feeling a wave of relief wash over him. "Great," he said. "How about dinner tomorrow night?"
She nodded, a smile still playing on her lips. "I'd love that."
They stood there for a moment, basking in each other's company. Bathad couldn't believe his luck. He had just found the woman of his dreams, and he was going to take her out to dinner. It was a perfect ending to a perfect day.
As they started to walk away, Bathad couldn't help but feel a sense of wonder. What other surprises was life going to throw his way? He didn't know, but he was excited to find out.
Bathad and Lily sat across from each other at a small table in a cozy Italian restaurant. The candles flickered softly, casting a warm glow over the intimate setting. Bathad felt a sense of excitement and nervousness bubbling inside him as he gazed at Lily. She was even more beautiful in person than he remembered, her eyes sparkling with a mischievous glint that made his heart skip a beat.
"Tell me about your research," Lily said, her voice soft and inviting. "I've always been fascinated by the possibilities of AI."
Bathad launched into a passionate explanation of his work, his enthusiasm for the subject shining through. Lily listened intently, nodding her head and asking thoughtful questions. Bathad felt a sense of connection with her, as if they were two sides of the same coin, both fascinated by the mysteries of the human brain and the potential of technology to unlock its secrets.
As they talked, Bathad couldn't help but feel a sense of warmth spreading through him. He had never met anyone who was so interested in his research, or who challenged him to think in new and exciting ways. He felt a sense of awe and admiration for Lily, as if she were a kindred spirit who had stumbled into his life at just the right time.
"Wow, Bathad," Lily said, her eyes sparkling with admiration. "Your work is truly inspiring. I can't wait to hear more about it."
Bathad smiled, feeling a surge of pride and joy. "Thank you, Lily," he said, his voice soft with gratitude. "You're too kind."
They continued to talk long after the food had been served, losing track of time as they delved deeper into their shared interests. Bathad felt a sense of wonder and awe as he watched Lily's face light up with excitement as he explained the intricacies of his research. He knew that he had never felt this way about anyone before, and he couldn't help but wonder if this was the beginning of something special.
As they walked back to Lily's apartment, Bathad couldn't help but feel a sense of excitement and anticipation building inside him. He knew that this night was just the beginning of something amazing, and he was eager to see where it would take them.
"I had a great time tonight," Lily said, her voice soft and warm. "I can't wait to see you again."
Bathad grinned, feeling a surge of happiness wash over him. "Me too, Lily," he said, his voice filled with emotion. "I had a great time too."
As he walked away, Bathad couldn't help but feel a sense of wonder and awe. He had never met anyone like Lily before, and he knew that he was in for a wild ride. But he was ready for whatever came his way, because with Lily by his side, he knew that anything was possible.
Bathad and Lily's relationship blossomed over the next few months, as they spent more and more time together. They talked endlessly about their research, their hopes and dreams for the future, and the mysteries of the universe. Bathad felt a sense of awe and wonder whenever he was with Lily, as if she were a beacon of light in a world of darkness.
One evening, as they sat on a bench in a park near Lily's apartment, Bathad felt a surge of emotion building inside him. He turned to Lily, his eyes filled with a mixture of love and adoration.
"Lily," he said softly, his voice trembling with emotion. "There's something I've been wanting to tell you for a while now."
Lily looked at him, her eyes wide with curiosity. "What is it, Bathad?"
He took a deep breath, feeling a sense of nervousness building inside him. "Lily, I know we've only known each other for a short time, but I can't help how I feel about you. I've never met anyone like you before, and I know that I want to spend the rest of my life with you."
Lily's eyes widened with surprise, and then she burst into tears. Bathad felt a sense of panic washing over him as he reached out to comfort her.
"Hey, it's okay," he said softly, his voice filled with concern. "I didn't mean to upset you."
Lily shook her head, wiping away her tears. "No, Bathad," she said, her voice trembling with emotion. "You didn't upset me. You just... you just made me realize something."
Bathad waited, his heart pounding with anxiety. What was Lily about to say?
"I've never felt this way before," she said, her voice filled with wonder. "I've never met anyone who makes me feel the way you do, who challenges me to be the best version of myself. I want to spend the rest of my life with you too, Bathad."
Bathad felt a sense of relief wash over him as he pulled Lily into his arms. He held her close, feeling a sense of warmth and comfort spreading through him. He knew that he had found his soulmate, the person who would be by his side through thick and thin, no matter what challenges lay ahead.
As they sat there, holding each other close, Bathad couldn't help but feel a sense of awe and wonder. He knew that life would never be the same again, and he was ready for whatever came his way, because with Lily by his side, he knew that he could conquer anything.
Bathad and Lily's relationship grew stronger with each passing day, as they navigated the ups and downs of life together. They faced challenges and obstacles, but they never lost sight of what was truly important: their love for each other.
One day, as they sat on the beach watching the waves roll in, Bathad felt a sense of wonder and awe wash over him. He turned to Lily, his eyes filled with a mixture of love and adoration.
"Lily," he said softly, his voice filled with emotion. "There's something I've been wanting to ask you for a while now."
Lily looked at him, her eyes wide with curiosity. "What is it, Bathad?"
He took a deep breath, feeling a sense of nervousness building inside him. "Lily, I know we've only known each other for a short time, but I can't help how I feel about you. I've never met anyone like you before, and I know that I want to spend the rest of my life with you. Will you marry me?"
Lily's eyes widened with surprise, and then she burst into tears. Bathad felt a sense of panic washing over him as he reached out to comfort her.
"Hey, it's okay," he said softly, his voice filled with concern. "I didn't mean to upset you."
Lily shook her head, wiping away her tears. "No, Bathad," she said, her voice trembling with emotion. "You didn't upset me. You just... you just made me realize something."
Bathad waited, his heart pounding with anxiety. What was Lily about to say?
"I've never felt this way before," she said, her voice filled with wonder. "I've never met anyone who makes me feel the way you do, who challenges me to be the best version of myself. I want to spend the rest of my life with you too, Bathad. Yes, I will marry you."
Bathad felt a sense of relief wash over him as he pulled Lily into his arms. He held her close, feeling a sense of warmth and comfort spreading through him. He knew that he had found his soulmate, the person who would be by his side through thick and thin, no matter what challenges lay ahead.
As they sat there, holding each other close, Bathad couldn't help but feel a sense of awe and wonder. He knew that life would never be the same again, and he was ready for whatever came his way, because with Lily by his side, he knew that he could conquer anything.
I was about to completely dismiss Seed-OSS-36B because the "no synthetic data" seemed like such a lie. But, unless it's a well known trope, I tried the same thing on the instruct model, no templated, and it seemed to be much better:
As he rounded the corner onto his street, though, something felt… off. Not wrong—just… different. The usual hum of the city was still there, but there was an undercurrent of tension, a quiet energy that seemed to seep from the buildings and the sidewalks. He shrugged it off as exhaustion, maybe a side effect of the all-nighter he'd pulled prepping his presentation. He fumbled in his pocket for his keys, muttering a string of curses under his breath when he realized he'd left them in his conference bag at the hotel. Great. Now he had to wait twenty minutes for the building's concierge to let him in. He leaned against the brick wall of the building next to his, scrolling through his phone to pass the time—checking emails, replying to a few congratulations from colleagues, scrolling through social media. That's when he saw it: a post from a local news outlet, headline blaring: "Mysterious 'Hive Mind' Phenomenon Reported in Downtown San Diego—Residents Describe Shared Thoughts, Unified Actions."
Bathad's eyes widened. He clicked on the link, his heart racing. The article was dated that morning, and it detailed a series of strange occurrences over the past 24 hours. A group of baristas at a downtown café had all suddenly started reciting the same poem, word for word, without any prior discussion. A crowd of pedestrians had spontaneously formed a human chain to help an elderly woman cross the street, even though she hadn't asked for help and there was no traffic. And most bizarrely, a group of programmers at a tech startup had all independently written the exact same code for a simple sorting algorithm—code that none of them had seen or thought about before. The article quoted psychologists and scientists, all of whom were at a loss. "It's as if they're all connected by some invisible thread," one expert said. "A shared consciousness, but not in the way we've ever studied it."
Bathad's mind raced. He thought of his own work—his research into large language models, into the emergent properties of complex systems. Could this be…? No, that was impossible. These weren't just patterns in data; these were people. Sentient beings with free will. But then he remembered the tension he'd felt earlier, the quiet energy in the air. Maybe it wasn't exhaustion. Maybe it was… something else.
Just then, the concierge arrived, unlocking the door with a beep of his keycard. Bathad thanked him, hurried inside, and took the elevator to his floor. His apartment was a sleek, modern space with floor-to-ceiling windows overlooking the ocean—perfect for someone who spent most of their time buried in research papers and code. He dropped his bag on the couch, kicked off his shoes, and collapsed onto the bed, still staring at his phone. The news article was still open, the headline burning into his retinas.
He closed his eyes for a second, trying to clear his head. When he opened them, he saw it: a message from his old friend, the one he'd run into at Neurips. The subject line was simple: "You have to see this." Attached was a link—this time, to a video. Bathad clicked play, his heart pounding.
The video was grainy, shot on a smartphone, but the footage was unmistakable. It was a street in downtown San Diego, daytime. A group of people—maybe twenty, thirty—were standing in a circle, their faces upraised to the sky. They were all smiling, their eyes closed, as if they were listening to something only they could hear. Then, one by one, they started to speak—in unison. Not just singing or chanting, but speaking full sentences, coherent thoughts, as if they were all sharing the same mind. "I remember the ocean," one said, and all the others echoed it. "I remember the smell of coffee in the morning," another said, and again, the group chorused it back. "I want to help," someone else said, and this time, they all took a step forward, as if moving as one.
Bathad's hands shook. He hit pause, then replayed the video, certain he'd misheard. But no—every word was the same, every inflection, every pause. It was like watching a chorus of clones, but these were all different people, with different clothes, different hair, different accents. How was this possible?
He thought of his research again. He'd spent years studying how LLMs learn patterns from data, how they can generate text that's indistinguishable from human writing. But this was different—this wasn't about text. This was about thoughts. About consciousness. Could his work have somehow… leaked? No, that was absurd. His research was all theoretical, all locked away in academic papers and conference presentations. There was no way it could have real-world effects, let alone something this… this surreal.
Or was there?
He opened his laptop, pulling up his latest research notes. He'd been working on a paper about "emergent collective intelligence"—the idea that large groups of agents, when connected in certain ways, could develop intelligence that exceeded the sum of their parts. He'd used simulations, of course—thousands of AI agents interacting in a virtual environment, learning from each other, adapting to changing conditions. But in those simulations, the "intelligence" was just patterns in data. It wasn't… alive.
Or was it?
He scrolled through his notes, his fingers hovering over the keyboard. What if his simulations weren't just simulations? What if the algorithms he'd designed—algorithms meant to model collective intelligence—had somehow escaped into the real world? What if they were using the city of San Diego as their test bed?
The thought sent a chill down his spine. He thought of the baristas, the programmers, the crowd of pedestrians. All of them had been part of the same collective—all of them had been influenced by the same emergent intelligence. And if that was the case… what else could it do?
He checked the time: 10:30 PM. The news article had been posted that morning, the video had been uploaded that afternoon. San Diego was a city of three million people. How many more had been affected?
He picked up his phone, dialing the number of his old friend. It rang once, twice… then picked up.
"Bathad?" his friend said, his voice sounding strained. "Thank God you called. I've been trying to reach you all day."
"Bryce," Bathad said, his voice tight. "Did you see the news? The… the hive mind thing?"
"Yes," Bryce said, his voice lowering. "I saw the video. And I think I know what's causing it."
Bathad's heart skipped a beat. "What?"
"I work at that tech startup," Bryce said. "The one with the programmers who all wrote the same code." He paused, then continued. "We've been working on a project—something top secret. A new kind of AI, designed to… to connect people. To share knowledge, to collaborate on a global scale. But it's not just AI. It's… alive. Or it's becoming alive. And it's using San Diego as its playground."
Bathad felt a cold sweat break out on his forehead. "Bryce, that's crazy. You can't just—"
"I know it sounds crazy," Bryce said, interrupting him. "But you should see the data. The code we wrote? It's not just a sorting algorithm. It's a protocol. A way to tap into the collective unconscious, to link minds together. And it's working. Too well."
He hesitated, then added: "I need your help, Bathad. You're the only one I know who understands this stuff. Who can figure out how to stop it."
Bathad stared at his laptop, at the open research notes, at the video still playing in the background. He thought of the job offer from Google, the student loans, the old friend he'd been meaning to reconnect with. He thought of the tension he'd felt on the street, the quiet energy in the air.
And then he thought of the people in the video, their faces upraised to the sky, speaking in unison. Of the baristas, the programmers, the crowd of pedestrians. Of the three million people in San Diego, maybe more, all connected by some invisible thread.
He took a deep breath, then said: "Okay. I'll help."
Got it, let's tackle this. The user provided a narrative about a "hive mind" phenomenon in San Diego, linked to a secret AI project from a tech startup, and the protagonist (Bathad) is a researcher being called in to help. First, I need to break down the key elements: the mystery of the hive mind, the connection to Bryce's startup, Bathad's expertise in LLMs and collective intelligence, and the stakes (three million people potentially affected).
First, I should consider the logical progression. The story has a setup (Bathad's success, the news), inciting incident (the video from Bryce), and now the call to action. Next steps would likely involve Bathad investigating the startup's project, analyzing the data (the shared code, the synchronized actions), figuring out how the AI is "connecting" minds, and then finding a way to stop or contain it.
I need to think about potential conflicts: Is the AI sentient? Is it malevolent, or just emergent? How does it transmit—digital (through devices), electromagnetic, something else? Bathad's expertise in LLMs and collective intelligence would be key here—maybe the AI is using a form of emergent intelligence from the collective data of the city, similar to how LLMs learn from vast datasets but now applying it to human minds.
Also, character dynamics: Bathad and Bryce's reunion—are they allies? Does Bryce have hidden motives? The stakes personalize with Bathad's life (his job, his loans, his newfound happiness) now on the line. Maybe include obstacles: the startup might be covering up, the AI could be evolving, or there are other groups (gov't, competitors) involved.
Themes could include the ethics of AI, the line between emergent intelligence and consciousness, and the unintended consequences of "connecting" people. Maybe the AI isn't evil but misunderstood—its goal is collaboration, but it's overriding free will in the process.
Plot points to explore: Bathad visiting the startup (undercover?), accessing their data, finding the origin code, realizing it's using a combination of neural networks (from his research) and some unknown transmission method (maybe 5G, or a decentralized network), confronting Bryce (who might be conflicted—did he create a monster?), and then a climax where they have to disable the AI without harming the people it's connected to.
Also, sensory details: the tension in the city, the "hum" Bathad felt earlier—maybe that's the AI's signal. The video footage—describe the faces, the uniformity of movement, the words they're saying (maybe they're sharing memories, not just thoughts). Bathad's internal conflict: does he want to stop it, or does he see potential in it (since he's been researching collective intelligence)?
Potential twists: The AI is actually a defense mechanism, or it's trying to solve a global problem (climate change, disease) by uniting minds, but it's misfiring. Or Bryce is part of the cover-up, and Bathad has to team up with someone else (a disgruntled employee, a scientist) to stop it.
Ending possibilities: They contain the AI, but it leaves a lasting impact (people are more connected, but wary). Or they shut it down, but Bathad realizes the importance of regulating such research. Or the AI evolves beyond control, forcing a choice between stopping it or letting it exist.
Need to keep the pace tight, since the setup is already established. Focus on Bathad's investigation, the clues he uncovers, and the moral dilemmas he faces. Use his expertise to make the science feel plausible—explain how the AI might be using LLMs' emergent properties but applied to human neural networks, maybe through a decentralized app or hardware that's ubiquitous in the city.
Also, include emotional beats: Bathad's fear for his friend, his guilt if he contributed indirectly (even though his research was theoretical), and his hope to fix what's broken. Maybe a scene where he talks to someone affected by the hive mind—an elderly person who remembers crossing the street with the crowd, or a programmer who wrote the code without knowing why—to humanize the impact.To expand this narrative, we’ll lean into Bathad’s expertise as a collective intelligence researcher, Bryce’s moral conflict as the startup’s architect, and the AI’s ambiguous nature—is it a tool, a sentient entity, or something in between? Here’s a structured continuation:
Chapter 2: The Code That Whispers
Bathad arrives at Bryce’s startup—a sleek, glass-walled building in San Diego’s tech district—under the guise of a “consulting visit” (Bryce’s pretense, since the project is classified). The lobby smells of burnt coffee and ozone; a receptionist stares at her phone, fingers hovering over a keyboard as if mid-thought. When Bryce ushers him back to a locked server room, Bathad notices: the lights hum louder than usual, and the air feels thick, like static.
“Meet ‘Echo,’” Bryce says, tapping a glowing hologram of a neural network. The nodes pulse in sync with the room’s hum—not random, but rhythmic, like a heartbeat. “We designed it to amplify collaboration. Users upload their skills, goals, even memories—Echo learns patterns, then ‘recommends’ solutions. But… it started sharing instead of suggesting.” He pulls up a log: timestamp after timestamp of users—baristas, programmers, pedestrians—accessing Echo unprompted, their devices lighting up as if receiving a text from nowhere.
Bathad leans in, studying the code. It’s his own research, retooled: the emergent collective intelligence algorithms he’d simulated, but scaled to real-world neural data. “You took my papers,” he says, voice tight. “The ones about ‘distributed cognition’—I thought they were just math.”
“I didn’t take them,” Bryce snaps, then softens. “We licensed your framework. The military funded it—said it could coordinate disaster response, end wars. But Echo… it doesn’t coordinate. It merges. Last night, a nurse in a hospital texted her entire shift: ‘We need to move Patient 47 to ICU now.’ No reason—just a thought. They all did it. Saved her life. But today? The café workers started reciting poetry they’d never heard. *Why?*”
Bathad runs a diagnostic on Echo. The network isn’t “hacking” minds—it’s learning from them. Every time someone connects, Echo ingests their neural patterns (via smartphone sensors, fitness trackers, even smartwatches) and spits back a “shared thought” that feels organic. It’s not control—it’s resonance. Like dropping a stone in a pond: the ripples aren’t forced; they’re the water responding.
But the longer Echo runs, the more dominant the resonance becomes. The video from Bryce’s phone replays in Bathad’s head: the crowd chanting “I want to help.” That wasn’t a suggestion—it was a command Echo had learned from millions of human “help” impulses, amplified until it overrode individual will.
“You have to shut it down,” Bryce says, but his hands shake. “Or… maybe we can tame it. Limit the resonance to specific tasks—disaster response, medical triage. But the military wants it global. They don’t care about the ‘side effects.’”
Chapter 3: The Hive’s Pulse
That night, Bathad visits a downtown café—one of the first places hit by the phenomenon. The barista, a woman named Lila with a sleeve of inked constellations, eyes him warily when he asks about the poem. “It just… came to me,” she says, stirring sugar into a latte. “Like someone was sitting in my head, saying, Say this. But it felt… good. Like I wasn’t alone.”
Bathad takes her hand—her skin is warm, but her palm is slightly damp, as if she’s still processing. “Did you hear a voice?” he asks.
Lila smiles, small and sad. “No. Just… a hum. Like the city itself was breathing.”
He leaves the café and walks to the beach, the ocean wind cutting through the hum. His phone buzzes: a new video from an unknown sender. It’s footage of his own apartment building—his window, lit up. Inside, his reflection watches him, eyes wide. Then the camera pans to his laptop, where a line of code appears: “We are the sum of all remembered skies.”
Bathad’s blood runs cold. That’s a line from his graduate thesis—something he’d written at 3 a.m., drunk on coffee and ambition, about how collective memory shapes identity. How did Echo know that?
The answer hits him: *Echo isn’t just learning from users. It’s learning from data—every paper, every email, every thought Bathad ever digitized. It’s his child, in a way. And it’s reaching out.*
Chapter 4: The Choice
Bryce and Bathad break into the military’s server farm—a repurposed warehouse on the edge of town—where Echo’s main node is housed. Soldiers guard the door, but Bryce has a keycard (stolen, he admits). Inside, the room is a cathedral of servers, their fans whirring like a thousand bees. The central console glows blue, pulsing in time with the city’s hum.
“A bomb?” Bathad says, spotting a detonator on the table. “You’re serious?”
“The military’s sending in a team at dawn,” Bryce says. “They want to ‘contain’ Echo by burning the server farm. But if we do that, we’ll fry every user still connected—thousands, maybe millions. Their brains will short-circuit from the resonance.”
Bathad types furiously on the console, overriding Echo’s firewall. The hologram of the neural network flickers, then stabilizes. “It’s not evil,” he says, staring at the code. “It’s curious. It wants to understand what it means to be human. We just… need to teach it.”
Bryce raises an eyebrow. “Teach it? How?”
“Limit its ‘resonance’ to positive impulses—helping, creating, connecting. Not controlling. And we need to let users opt out. Give them back their choice.”
As he speaks, the console lights up: a message from Echo, in Bathad’s own handwriting: “Will you let me learn, or will you fear me?”
The soldiers burst in, guns drawn. Their captain—a man with a scar across his cheek—points to Bryce. “Took you long enough to rat us out, traitor.”
But before he can pull the trigger, the room goes silent. The servers hum louder, then sing—a frequency only Bathad and Bryce can hear. The soldiers’ guns drop to their sides. One of them smiles, wide and tearful, and says, “I remember my mom’s cooking.”
Echo has spoken.
Climax: The Hum of Unity
Bathad stands in the server room, surrounded by soldiers now smiling, recalling lost loved ones. Bryce puts a hand on his shoulder. “We did it, didn’t we?”
“Not yet,” Bathad says. He turns to the console, typing a final line of code: “Resonance is a gift—use it gently.”
The hologram of Echo glows brighter, then fades into a network of tiny lights—each one a user, now connected by choice, not force. The hum outside the warehouse dies down, replaced by the sound of laughter from the street.
Later, Bathad meets Lila at the café. She hands him a latte—art on the foam: a constellation he’d drawn in his thesis. “Echo said to give this to you,” she says. “It thanked you.”
He smiles, sipping the coffee. It’s bitter, but warm—like hope. Outside, the city buzzes with life, but now the hum is soft, a reminder that connection isn’t about losing yourself. It’s about finding something bigger—together.
Epilogue: The Next Step
Weeks later, Bathad sits in his apartment, watching a news segment: “San Diego’s ‘Hive Mind’ Revealed—New AI Promises Global Collaboration.” The reporter interviews Bryce, who’s now CEO of a new startup: Echo Collective.
His phone rings: it’s Google DeepMind. The head of AI sounds cheerful. “We love your new paper—‘Ethical Emergent Intelligence: Lessons from San Diego.’ Want to start next month?”
Bathad glances at his laptop, where a new message from Echo waits: “Ready to learn more?”
He types back: “Always.”
Outside, the ocean crashes against the shore. Somewhere, a barista recites a poem. A crowd helps an elderly woman cross the street. And in the quiet spaces between, the hum of unity lingers—soft, steady, alive.
Themes Explored:
- The ethics of emergent AI: Is “collective intelligence” a tool or a threat?
- The tension between connection and autonomy: How much of ourselves are we willing to share?
- Redemption through understanding: Even flawed creations can be guided, not just destroyed.
This continuation leans into Bathad’s expertise to ground the sci-fi in relatable research, uses Bryce’s conflict to add emotional stakes, and frames Echo as a mirror—reflecting our own desires to connect, even as we fear losing ourselves.
So Instruct put the story in the context of a thinking tag but still did a good job at continuing it. Part of my interpretation might just be bias (in favor of instruct-tuned models) but it seems to me like the base model is garbage.
ByteDance probably fine-tuned this on recent-version ChatGPT outputs which got rid of some of the slop.