I Made an AI Write an Entire Book | Using AI Agents and Local LLMs

[@carldennis9907]

Your KZbin channel is what I've been looking for in terms of practical applications for llm agents. Thank you for sharing and keep up the great work. Your channel is a gold mine.

[@GosuCoder]

Thank you that means the world to me!

[@zandanshah]

I am in the process of converting old books written 100 years back into current English and easy to read format. Project in progress.

[@alessandrosalvaggio6772]

great! last year i made a "book translation" tool that allowed also custom prompts, maybe that could help, with the right prompt, to automate the work

[@GosuCoder]

That sounds like an amazing project to work on? Are you running this all locally?

[@christenw.1726]

I hope you are having good progress ❤

[@marleymomo9582]

Woah, great video, this is why I come to youtube to see what other people are thinking. This is genius bro. Waiting for more videos on this, thanx I feel inspired to try one project like you.

[@451236mateus]

By some twist of fate I've landed here. This video, pure gold, thank you.

[@kawsarahmad]

Excellent video on your thought process! Looking forward to follow ups to this and whatever else you have planned.

[@DouglasClemens]

Great video. Thanks for sharing your detailed workflow.

[@ScottLahteine]

This is a great project to explore the narrative powers of LLMs and the challenges of building a system that does plotting and planning top-down. The results are likely to always have that AI aftertaste, but if you can get a coherent book generated and then use that as a starting point to make a decent edited novel, that’s not bad. I’ve been using LLMs to discuss and develop a story, and this process is giving me insights into how agents might be designed, but I think I need to go through the complete collaboration process first to get a sense of what agents will and won’t be able to do.

[@yanikjayaram]

Yo dude this is dope!

[@patruff]

I read the book, unimpressed. Very impressive how you made it and thanks for sharing.

[@GosuCoder]

Same definitely work to do!

[@SaveThatMoney411]

Awesome. Was just searching for this! I want to write 120,000-150,000 word romance novels.

[@JaysThoughts-q5e]

"Wow, I'm blown away by the complexity and depth of your project! The idea of using multiple AI agents to generate a full-length book is fascinating, and I love how you've experimented with different models and approaches. I'm definitely inspired to try something like this myself. One question I do have is, have you considered using a more structured approach to outlining and planning, to help keep the story on track and prevent some of the 'off-track' moments you mentioned?" - Llama 3.2

[@GosuCoder]

I definitely have considered that, my version 1 went more down that route! I need to do so much more experimenting

[@UpperCaseX]

lmao he had ai write this

[@JaysThoughts-q5e]

@@UpperCaseX everyone says like, comment, subscribe, at least this comment wasn't for boner pills.

[@JayantHudar]

Very Interesting. I see good ideas actually implemented. Once you read all, you may find many shortcomings. ( I found mine) The character development should be first part of it. And I have at least 3 more ideas, however will give in person. Based on what I have tried doing using prompts alone. And choice of words, tone, personality etc. Will contact you. Great work by you. Regards.

[@alessandrosalvaggio6772]

great work! i made something like that in python and laravel last year (jan 24), it works with a gpt4 api key and creates guides, books etc.. and you also have the option to keep consistency (it uses a lot more tokens).

[@GosuCoder]

That’s so cool! I have thought about hooking it up to Claude or gpt4 to see how the agents perform. It’s pretty easy to swap out the LLM

[@JaysThoughts-q5e]

9:37 My first test was basically a random character generator that could loop forever. I think of the LLM as an unlimited stack to pop items from. I started with the family ancestry of the character, then build them up with other details the bot chooses. Sub-functions to decide their occupation, hair color, etc. It is a lot of fun and opens a lot of opportunities in my opinion.

[@GosuCoder]

That’s amazing!

[@Koroistro]

What I can see this being very interesting for would be to build a "book tree". Inspired to the tree of thoughts paradigm. Every chapter could branch off in several different narrative directions, the story could basically explore different timelines in parallel. It'd require more book-keeping, maintaining some sort of multiversal consistency (involved characters, their personalities) would be challenging, but it'd be something very different from "normal" books. (Sure there are choose your own adventure kind of books, but this isn't quite the same).

[@GosuCoder]

This is brilliant I like that idea a lot!

[@psylsph]

I've forked you're excellent work with a few small updates, the first to use DeekSeek v3 and the second for the agent to "Each chapter should target to be at least 5000 words. If the content is shorter, return it to the writer for expansion. This is a hard requirement - do not approve chapters shorter than 3000 words. " this completed with about 12 hours at the cost of about 1 dollar. Oh and weirdly caching seems to work with DeepSeek for me at least but with local models it eventually get into an infinite loop and just stops using the AI all together. I've disabled caching on the main branch on my fork for now

[@GosuCoder]

Woah thank you for playing around with it! Thank you for the kind words.

[@psylsph]

@@GosuCoderno problem it’s turning into an interesting time sink 😂 still not managed to get a book to generate locally I’ve now seen the wondering off AI

[@buffaloraouf3411]

any fork for that im so interesting

[@1MinuteFlipDoc]

Nice work! i'm doing similar experimentations! hahahaha! Thanks for sharing!

[@burgerbee5169]

Try a 70B Q4 model, (Lllama 3.3 70B Q4 K_M is my recommendation) and you will be amazed what that does to instruction following. Its like night and day. Every model bellow that feels like a joke now. It does take 2x24GB VRAM, but I run it with eBay eWaste here, 2xP6000 does the job nicely with large context.

[@sitedev]

I suspect the issues you’re encountering with respect to LLM output length is due to your local computer resources. A locally hosted LLM on consumer hardware (even a 4090) will result in a small output length (and small context window). If you use something like Gemini Flash 2.0 you’ll have access to a massive context window (1 million tokens) and longer output length.

[@isekaidev7803]

But note that the prices get very high at high context. So, it's not worth it.

[@SilconOne]

It's true they can't do large text blocks they will repeat or Weaver form context they also try to get to the end really quick

[@benjaminbirdsey6874]

I believe you might need to set the max output length in the LLM parameters, rather than trying to specify it in a prompt.

[@isekaidev7803]

It'd be better to use TypeScript instead of Python. That way it can be made into a plugin for VSCode or Obsidian or LogSeq or even a website.

[@carloscamacho6133]

@gosucoder, can I use LM Studio for the local LLM? I have the server running but not successful generating all chapters

[@GosuCoder]

What model are you running? Could be worth posting an issue on GitHub to see if we can figure it out

[@RickPadilla-g9d]

Great stuff. How do I contact you to talk? Work in film. And publishing. High level and also have tech background. Thx

[@GosuCoder]

You can shoot me an email at [e-mail]!

[@spicer41282]

Curious? What were your reasons that you chose the 7900?

[@GosuCoder]

I upgraded from a 3070 8GB, primarily went with the 7900xtx for the VRAM. I really want to get a 2nd so I can run even bigger models.

[@gareththomas3234]

Amazing I do non-fiction with AI

[@mrd6869]

You know they have longform AI tools that do this now.

[@GosuCoder]

Do you have an example of one I could test out? I’d love to compare the results.

[@farouthere6994]

and is it any worth reading ?

[@GosuCoder]

Probably not lol!

[@SHEMBOBKIPRUTO]

How much did it cost?

[@GosuCoder]

I ran it all locally so it was free

[@Pyala-f7x]

​@GosuCoder can you share the computer config... ram, vram,cpu,gpu as such