Regarding tuning, RAG and Pre-Trained models

[MerovingianXOR]

I want to use ChatGPT for 2 purposes:
1# to find out things
2# to write articles

With Writing I have an account on chatgpt the problem there right now is that:

1 GPT forgets things between revisions.

2 GPT wants to start over from beginning for each revision - I would prefer to draft only what I want revised and then put the pieces together,

3 I want each revision to have a number I can refer too, so I can tell GPT that revision 1 was good for this, now go back to 5 and mix in this from 5.

4 also i want numbered revisions because in the middle of the way gpt misunderstand the purpose and then i want to rewind and say
ok we start again from revision 3.

5 I want what i tune in logic regarding writing to be implemented in all further tasks/projects, not to tune it again and again for each task
mostly it is the usage of adjectives that is the problem

6 and where a statement does not belong in relation the subcontext we are writing (but do in a generall main context which in the situation is beside the point)

7 Teach it a certain logic for a certain kind of article that is a category of my writing.

8 feed the GPT reasearch papers and books so it can have more precise data to work on an output

So I understood sofar that i need to get a pre-trained model, but where are those and how do I start work on it? and then how to tune it to my specification.

Thankful for any direction!

[mat.eo]

Great thoughts. What you want isn’t AI, but a system built for AI.

I wouldn’t be surprised if a model couldn’t whip up a revision state management server in Python using Flask or something.

For “permanent” tuning this could be as simple as having an evolving system instructions. If it gets too large then fine-tuning can be an option.

Then, for each of your “refinement” tasks you can dedicate a single “AI Agent” to.

This isn’t necessarily the best way, but it all points towards a “system”, and not a single model.

[MerovingianXOR]

Hello Mat and thanks for reaching out!

@mat.eo

but it all points towards a “system”, and not a single model.

Could you please tell me more about this?

and

For “permanent” tuning this could be as simple as having an evolving system instructions. If it gets too large then fine-tuning can be an option.

=)

[mat.eo]

@MerovingianXOR

Could you please tell me more about this?

I’d love to, but there is a lot to unpack and I’d probably be asking a hundred questions. I created a ChatGPT instance with the context, it may be of interest.

The edit button lets you talk with the context-aware model.

Ai System Analysis

I hope it helps. In a sense it’s just basically an application. I’m not sure how technical you are, but some tinkering with GPTs w/ Actions and then setting up an API might be a good start.

[MerovingianXOR]

Thank you for this, I will check it out. I am not technical in the sence that i cannot code.

[curt.kennedy]

It might be a bit of a stretch, but you can use AI based coding tools to write stories that have revisions, and you can accept edits from the AI, line by line if you want. Even choose whatever AI model.

There is probably a writer centric version of this, but I generated this with Cursor using o3-mini as the model.

No coding required.

Here you can see I had the AI create a “story” from scratch. Prompt to o3-mini in upper right.

After accepting the story. I had it modify with this prompt “modify this to include an evil villan”

After applying the changes, I get a diff where I can accept line by line or all changes:

There is a plugin called “Local History” where all file changes are saved in a history directory locally, so you can always pull any revision forward and start from there.

This particular tool is called Cursor, built on VS Code … but as you can see, it can also write stories, and do revision management.
For context, instead of RAG, you could attach multiple files to your query, and it will draw from all these files in the response. This is a “manual RAG” of sorts. But there are many more out there to choose from that have AI and revision management / “manual RAG” built in.

Might be what you need as a writer.

[MerovingianXOR]

@curt.kennedy

r applying the changes, I get a diff where I can accept line by line or all changes:

Hello,

I will read this in detail during the day. I want to edit and revise a new draft =) : that all though I can not code - I am not afraid of code. I have seen some code during my years and I would be comfortable on copy pasting code and I am open to gradually learn more about it.

I am eager to try your examples during the day.

[MerovingianXOR]

@mat.eo

a sense it’s just basically an application. I’m not sure how technical you are, but some tinkering with GPTs w/ Actions and then setting up an API might

So by system you mean a combination of different solutions?

“The edit button lets you talk with the context-aware model.” In what capacity is it contect aware? How did it become context aware?

I read the article there but i did not interact with it yet.
The problem with my question is that I listed all my problems at once, mainly because I was eager to find out if there are solutions. Therefore it has been difficult to digest the answers. So I will start over now with one step at a time…

[MerovingianXOR]

@curt.kennedy

There is a plugin called “Local History” where all file changes are saved in a history directory locally, so you can always pull any revision forward and start from there.

What is a file in this context?

[MerovingianXOR]

I tried to paste a link to my project here but could not.

Here is a project; So I am promting the GPT to give me individual senteces that will be building blocks for reports I use in my work. So each individual sentence can be copied and pasted after the other. My reports can cover up to 8 different subjects and so far i have finished material for 3. There is about 3-4 drafts that are in the feed that are incorrect and i would like to start over on a fresh workspace without these incorrect drafts, but still i dont want to loose the correct feedback that i want to bring with me forward. How can I do that?

[MerovingianXOR]

How can I share a link to the project?

[mat.eo]

Very creative. It’s neat to consider that writing code and story-telling share a lot of the same tools.

Cursor would be a great start to test things out.

@MerovingianXOR There’s good and bad news.

The bad news is that you’re going to have to get your hands somewhat dirty to make this happen. The good news is that ChatGPT can easily accommodate your request and make this happen, you just need to know how to piece it together.

Otherwise, it may make more sense to look for a service that is already offering this. As mentioned by curt, you could start off using Cursor, which is typically for writing code, but can be bent into story-telling (Similar to how ChatGPT Canvas can be used for code or story-telling).

I have my own personal service that I use for researching and posting. I’m actually tying it into an AI Agent, and will be releasing it as a GPT once it’s ready. If you’re interested I’d be happy for you to try it out.

@MerovingianXOR

“The edit button lets you talk with the context-aware model.” In what capacity is it contect aware? How did it become context aware?

What I meant to say is that I provided the model the necessary information so that you can pick up the conversation and it will understand the situation.

[MerovingianXOR]

OK, so the text I see when I press the link, that is what you have written and what will make GPT context aware;

[MerovingianXOR]

@mat.eo

Otherwise, it may make more sense to look for a service that is already offering this

I checked out cursors website. What services are already offering “this” =). And how can I try yours?

[MerovingianXOR]

Sometimes when you want to make a small improvement you have to do big changes, and sometimes when you want to make big improvements it will be enough with small changes.

Which makes my thoughts wonder to:

Is it possible to feed excisting project in chatgpt with my previous reports to make it more context aware for the drafts that are to come now?

I would really like to shre the link if possible …

[mat.eo]

Although Cursor doesn’t explicitly support what you’re looking for, the features it offers is quite similar to what you want.

@MerovingianXOR

Is it possible to feed excisting project in chatgpt with my previous reports to make it more context aware for the drafts that are to come now?

Yes, this can be accomplished using Actions. The idea is that you have a state management server that your GPT can interact with

@MerovingianXOR

And how can I try yours?

When it’s ready for public use I’d love to send you a DM.

[curt.kennedy]

Sharing links to a project in this context is normally the job for GitHub … ChatGPT can help if you get stuck.

Here you can even check in your revisions into GitHub, and have a solid trail of stories to revert back to if needed.

I’m glad to you tried this approach. It’s not a perfect fit for writers, but it satisfies all your basic requirements … gives you revisions, RAG context, line by line edits, similar to what I imagine a serious story editor would need.

Cool thing is, you can even hook these files into \LaTeX and have professional rendered and typeset PDF’s. There’s another rabbit hole for ya.

[MerovingianXOR]

EDIT: My account was blocked for 4 hours because I am new here.

@curt.kennedy

his is a “manual RAG” of sorts. But there are many more out there to choose from that have AI and revision management / “manual RAG” built in.

Sorry, I take my last entry back. I will open an account in Cursor and into the cave of Mephistopheles. Curts post was a little scary to look at because the screenshots are “coding black” but now I looked at it and know I have to enter the dungeon there …

[MerovingianXOR]

@curt.kennedy

ol thing is, you can even hook these files into \LaTeX and have professional rendered and typeset PDF’s. There’s another rabbit hole for ya.

I will investigate more into latex.

[curt.kennedy]

Don’t let my “Coding black” screenshots intimidate you … didn’t even know that was a thing

Oh and \LaTeX is a powerful typesetting engine. I learned it 30 years ago, and it is still going strong. Any respectable AI should be able to help format your output in \LaTeX or other \TeX variants (this forum uses one of the variants, hence the funky typesetting here, since I know the commands).