I have been trying to get into writing short stories as a hobby. I have a couple of good ideas. But I tend to struggle when actually putting my thoughts to words.
Some issues I struggle with are as follows:
-
Inability to settle on the right words: I’ll write something and think that what I wrote could be written better or differently and then I keep on writing and deleting and rephrasing with different words. Thus making very slow progress.
-
Problems with continuity: I might think up a somewhat long plot line. But I have to write the whole thing in one go because if I don’t then my brain will splinter the story into multiple possible story branches when I stop and I am unable to choose the path to follow.
-
Lose interest in continuing if I take a break: If I stop writing mid way and take a break from writing for an extended period of time, I am unable to find the motivation to resume. Mostly because trying to catchup with the story up to that point feels hard. I have this same tendency with video games as well where I don’t feel like picking up a game after an extended period of absence.
So is anyone here who does writing as either a hobby or professionally? If so how do you cope with your condition?
What I’ve found that works for me is to just make up the story as I go without much planning. The issue with this is approach I’ve found is that it’s hard to find a conclusion to end the story.
I have the same problem.
You’re not going to want to hear this. But local LLMs help.
Hear me out.
Before they turn into the sycophantic chatbots you know, LLMs come in an initial “pre trained” state. They can’t answer questions in multi turn format or anything, all they do continue blocks of raw text. That’s it. Pretrains are like text improv machines, basically.
So what you do is use it as glorified autocomplete, not for whole paragraphs but a few words or a sentence.
Like, let’s say you’re stuck on how to word something mid paragraph. Mash the continue button, and it will stream a plausible continuation of that paragraph, which several thousand words as co text, until you tell it to stop. It’s not always usable, but it usually knocks your brain loose, though sometimes it comes up with something you wouldn’t even think of.
What’s more, if you use Mikupad as a UI, hover your mouse over text and it will show you the logprobs of text; basically glorified synonyms for every word in a sentence, as part of its “internal thinking”
https://github.com/lmg-anon/mikupad
I know, I know. AI.
But you can, and should, use more modestly/ethically trained use totally local models, that need zero internet. These are, in fact, basically the only ones you can use: ChatGPT literally does not work for this, for many reasons, which mostly boil down to “Sam Altman is a tech bro asshole”
It is a tool. It’s a wonderful autocomplete for ADD, to knock your brain loose. It’s helped me write hundreds of thousands of words, with zero slop and nothing leaving my PC. And most what is awful about AI doesn’t apply here.
I don’t personally use generative AI once I’m in progress, but it’s been an absolute godsend for me to overcome the “blank page” problem. I have it help me with an outline or to brainstorm a concept that needs fleshing out and write off of that.
The especially nice thing is I don’t really reference the outline a lot in the end. Something about the process is enough to rattle my brain into an ordered enough state to flow. Like I just needed to get concepts out of my brain and onto a page so I’m not thinking about it while trying to write.
It’s been a game changer.
I was dabbling in local llms recently using Ollama to generate stories from prompts. It’s fine but not something I’d consider something original like how I’d write.
But I guess it’s better than nothing if I get stuck.
Ollama is really terrible, especially since it uses chat mode by default and a 2048 context.
If you run a better pretrain with a “notepad” UI like Mikupad, it’s like night and day. It follows your writing style because that’s the context it has to go on.
If you’re interested, tell me your hardware config and I can recommend something specific, but generally you’re going to want to run ik_llama.cpp with a big MoE base model, like GLM Air base. Use something like Q6/Q5.1 cache quantization, enable the hadamard option, and then tune the GPU layer count until it fills your vram.
Well of course. If you give it a short prompt, and it generates a story, that isn’t going to be anything of value. But if you give it long prompts and have it give you ten different sets of three sentences that could follow, you have a goid shot that one of those either fits what you were thinking but couldn’t get in words, or will stimulate a better sentence in your mind. It can be a block breaker. Don’t ask it for whole stories, just sentences or paragraphs. Or even just to reword some thing you wrote if you don’t like how it flowed.