TEST LINK TO BLOG POST WITH FULL DAILY NOTES
Try = experimental writing
Calculated = more critically careful than dancing with chaos
Intro:
I’m writing a novel (a sonnet in novel form might be a better way to see it) that i hope will be a quick but literary work and I’m exploring how AI might make the process faster. I’m a painfully slow writer and I’d like to finish a few projects more quickly but at a high level of quality.
My goal is a 40,000 word novel, 1000 words polished to 90% finish per week. That will finish a polished draft in 40 weeks. Give it a week for read-throughs, then off to beta-readers round 1 and round 2, which takes about 3 months. The final product arrives in around 54 weeks.
Along the way I’ll also:
- make vids about the process, showing how the writing is done with or without AI help as the case may be.
- make notes and summarize them about what AI seems good, mediocre, and poor/useless for vis a vis the work of writing this kind of stuff.
- record the subjective experience of being a human empowered and threatened by this technology.
- whatever else might be useful without slowing the process down.
MORE ON SUPPORTING THIS PROJECT
June 5, 2025
- On ChatGPT’s advice i set up a local AI model using LM Studio and Mistral. The AI was downloaded from Hugging Face, TheBloke, quant method is WHAT (meaning the AI has been compressed for use on a small PC FINISH).
- Result: Pretty slow. Quite shallow feedback, especially compared with ChatGPT 4 online. Funny, it kept recommending i use the online version of Mistral (Le Chat LINK). So it seems that to get useful beta-reader feedback, with complexity and nuance, a small local setup might not work. Note, the examples below I made up; they’re not from any AI generated response.
- Complexity: simple ideas (“Your pace is good”) are explored in more detail (“Your chapter starts with a quick, active pace while the characters argue about the money, and then slows quite a lot when Jill contemplates life in a small town. Some readers may find the shift jarring …” etc.)
- Nuance: superficial observations (“Jill is an under-represented character whom many women readers might identify with.”) are followed by observations that catch more subtle aspects (“Jill is an under-represented character whom many women readers might identify with. Her struggle in choosing between building a career in New York and a building a homestead in rural Wyoming offer opportunities to explore themes of gender pressure, gender roles, isolation, adventure, and career. However, you might consider whether …” etc.)
- I tried Le Chat online. The results were better, but still far below those of ChatGPT-4. I’ll try to get access to Claude, Gemini, and DeepSeek to continue the comparisons.
- Result: Pretty slow. Quite shallow feedback, especially compared with ChatGPT 4 online. Funny, it kept recommending i use the online version of Mistral (Le Chat LINK). So it seems that to get useful beta-reader feedback, with complexity and nuance, a small local setup might not work. Note, the examples below I made up; they’re not from any AI generated response.