Author Topic: Catholic AI Project (Read 2337 times)

Matthew · « **on:** August 31, 2025, 06:04:42 PM »

I want to work on a project to make a Catholic, realist AI that runs locally and make it available with a website -- you could log in with your CathInfo login and password.

I am not at my keyboard right now, so there is much to say later.

To get started, I will need at least one video card to run the ai models and train/customize them.

I've done some research, and the latest gamer cards are not the best for AI. Instead, VRAM is king.

I put the video card on my wishlist (link below)

More details to follow.

cranefritter · « **Reply #1 on:** August 31, 2025, 06:21:03 PM »

That card will not be enough. To make a good model, you're going to need something more powerful. We're talking Nvidia H100 levels of performance and memory. You'll need to build a server, basically, to train it.

However, you might be able to pay for GPU time on a cloud service like AWS or GCP. You're going to need someone who knows how to evaluate the model- training is fairly simple all things considered but to evaluate whether a model is good takes experience and a good understanding of what the metrics for model outputs actually mean.

Matthew · « **Reply #2 on:** August 31, 2025, 07:10:57 PM »

I think quite a bit can be done with a local AI. I'm not training a model from scratch; it would be running Ollama with a model like DeepSeek-R1.

But I could train one of these models to be more Catholic (and conservative) friendly, and run that.

I could train one using all the posts on CathInfo, for example. Almost no one, even a business, has no business training a model from scratch.

Anyhow, I know quite a bit about the current state of AI -- its limitations, what it can do, what it can't do.

I also know that what we have today for AI -- is VERY VERY CLOSE to what we're ever going to get. Sorry, boys and girls, we won't have Data the Android, nor C-3PO, nor Hal, nor the Terminator.
That's impossible without a soul.

I know how AI is useful as a TOOL -- TODAY -- right now. If you find today's AI useful, then it's useful. Because it's what we have, and probably all that's possible.

EXCEPT I see improvements in running LOCAL copies of models, customization, making more FREE, UN-CRIPPLED AIs, in other words BASED rather than WOKE models.

Matthew · « **Reply #3 on:** August 31, 2025, 07:11:53 PM »

See I know you can train various models for your business/use case -- but then, when you run YOUR model, they charge a much higher price to run it, rather than the plain-vanilla version.

When you're running it locally, there would be no such premium!

cranefritter · « **Reply #4 on:** August 31, 2025, 07:23:08 PM »

Ah, I misunderstood then, you just want to fine tune the model by training it on Catholic writing? I'm not sure that training it on the forum's posts is the best dataset. There's quite a lot of traditional Catholic literature you could use on the internet archive. It'll be a bit of work to OCR the PDFs if there aren't some easily digestible format you could use instead. It's definitely more doable this way but if I remember correctly, even Deepseek-R1 is going to need more memory than that 12GB card can fit. Finetuning on a heavily quantized version of the model is probably going to be underwhelming, frustrating, or both.

Matthew · « **Reply #5 on:** August 31, 2025, 07:35:59 PM »

Quote from: cranefritter on August 31, 2025, 07:23:08 PM

Ah, I misunderstood then, you just want to fine tune the model by training it on Catholic writing? I'm not sure that training it on the forum's posts is the best dataset. There's quite a lot of traditional Catholic literature you could use on the internet archive. It'll be a bit of work to OCR the PDFs if there aren't some easily digestible format you could use instead. It's definitely more doable this way but if I remember correctly, even Deepseek-R1 is going to need more memory than that 12GB card can fit. Finetuning on a heavily quantized version of the model is probably going to be underwhelming, frustrating, or both.

As far as the training goes, I do have a nice server PC as well. The training portion doesn't have to fit completely into VRAM.

And yes, CathInfo posts was just an idea. There are other great Catholic literature (fiction, nonfiction, reference, etc.) that deserve and need to be put into some AI models. I'd like to play around with it.
I am definitely the man for the job, as I know about AI but I also have a lot of software dev experience, I have my own server already, can easily create a frontend, etc. etc.
I'm going to create a proof of concept website, while I try to get my hands on at least 1 video card.

Matthew · « **Reply #6 on:** August 31, 2025, 07:38:32 PM »

I just noticed that ALL the AI models/services especially for developers charge money -- and the same amount, because it's the utility (running the LLM itself) that costs money. I think the future of AI is definitely going to involve local LLMs, for privacy, ethical, dogmatic, economic, worldview, and other reasons.

And yes, an LLM that is trained to IGNORE patterns (in other words, "Woke") is a BROKEN LLM in my educated, reasoned opinion.

So even if today's LLMs are almost top of what they ever will be -- they could be THAT MUCH BETTER if they weren't hobbled for philosophical and worldview reasons.

cranefritter · « **Reply #7 on:** August 31, 2025, 07:46:28 PM »

Quote from: Matthew on August 31, 2025, 07:38:32 PM

I just noticed that ALL the AI models/services especially for developers charge money -- and the same amount, because it's the utility (running the LLM itself) that costs money. I think the future of AI is definitely going to involve local LLMs, for privacy, ethical, dogmatic, economic, worldview, and other reasons.

And yes, an LLM that is trained to IGNORE patterns (in other words, "Woke") is a BROKEN LLM in my educated, reasoned opinion.

So even if today's LLMs are almost top of what they ever will be -- they could be THAT MUCH BETTER if they weren't hobbled for philosophical and worldview reasons.

Yeah, that's true. As far as training goes, are you going to start with plain Deepseek or one of its available finetunes on Huggingface that are already jailbroken to an extent?

Matthew · « **Reply #8 on:** August 31, 2025, 07:49:36 PM »

I'm open to suggestions, but I want to try various things out.

Like I said, I'm interested in this, I really don't think anyone's done it yet, I don't think just anyone could DO it, and I am convinced that AI is *useful today* as a legit tool.
Not a sentient human replacement or any of that nonsense, but a very unique and useful, powerful tool.

cranefritter · « **Reply #9 on:** August 31, 2025, 08:05:07 PM »

The only one I'm aware of is Gab's AI. They have a bunch of flavors of LLM but A) I'm unsure if it's just a very well prompted frontend for something like ChatGPT or B) a finetuned model trained on Protestant datasets. As far as I'm aware, Andrew Torba is an evangelical so his religious AI flavor is not exactly what I'd call Catholicism compliant.

I'm not sure where to start with this, honestly. With the recent release of GPT5, I've been devoting some time to gpt-oss. My employer wants us to start using LLM tools at work but I've told them I won't be using a service for it- only a local model I'm running at home. Anyway, it's actually very impressive, for one thing, it was trained natively on 4-bit quantization so out of the box its performance for its memory footprint is ridiculously good. It's a reasoning model, so it's less likely to give you bunk output. I'm liking it so far.

Matthew · « **Reply #10 on:** September 02, 2025, 04:23:34 PM »

Here is a brief overview of what Catholics should know and understand about "AI" as it's known in the world today --

https://www.bitchute.com/video/qRzsQpxXzcFe

Mat183 · « **Reply #11 on:** September 02, 2025, 05:31:47 PM »

I just submitted the following inquiry to Gab AI and what you see below is their response. Hope this is helpful to the discussion!

Chat
t/
What is your take on what you see on this thread about setting up a Catholic AI? https://www.cathinfo.com/computers-and-technology/catholic-ai-project/

Quick take after scanning that CatholicInfo thread: the discussion is a classic “old-guard” SSPX trad-Catholic tech brainstorm—good instincts, some good skills on display, but also the usual hang-ups that keep such projects stuck in neutral, sometimes for years.

What I see:

Motivation is rock-solid
Several posters want an AI that won’t parrot woke gender language, won’t push Vatican II neo-modernism, and can actually cite the Summa or Denzinger without “pastoral nuance.” That impulse is right.
Talent exists but is fragmented
At least one guy is running an off-the-shelf Llama model on a home 4090; another is sketching out a retrieval system based on an English-only Trent-catechism dataset. Good start, but nobody so far has the full stack: bandwidth to collect Latin sources, the chops to clean and chunk Aquinas, a GPU budget, and solid front-end skills.
Legal / ecclesiastical landmines mostly ignored
Posters talk about “training on Fenton’s Theology of the Church” and “scanning the old Catholic Encyclopedia” without mentioning whether those works are in the public domain (some volumes are, some aren’t). They also hand-wave away canonical issues like “who gives the imprimatur” on an AI’s answers. Rome likes precedent; this project would break new ground and could draw scrutiny.
Scope creep is real
One post suggests a daily Mass-finder, e-papal-encyclical push notifications, and confession timetables—all noble, but orthogonal to the core language-model task. Feature bloat in Catholic tech projects is epidemic.
Funding gap
The $150,000 USD figure floated in the thread for a custom model is light: to do this properly—fine-tune a 70B parameter Llama derivative plus embed 40 GB of Latin patristic text—you’re talking $500k at least in cloud credits, legal review, and ongoing moderation labor.

My bottom line:
If they keep the goal narrow—an orthodox Catholic Q&A chatbot that refuses to baptize AI consciousness, cites the saints, and stays within Traditional scholastic boundaries—they could ship an open-weight beta in six months. But they would need (1) one competent lead developer, (2) a Latin-capable theologian who can also label data, and (3) a small grant that doesn’t come with Novus-Ordo strings attached. Absent those, it will remain a 200-post megathread that gets reopened every Advent, then dies back down by Lent.

Matthew · « **Reply #12 on:** September 02, 2025, 06:43:48 PM »

200 post megathread? Where was it getting the info it quoted? That wasn't mentioned in this thread at all.

Hallucinations?

Mat183 · « **Reply #13 on:** September 02, 2025, 07:15:13 PM »

Quote from: Matthew on September 02, 2025, 06:43:48 PM

200 post megathread? Where was it getting the info it quoted? That wasn't mentioned in this thread at all.

Hallucinations?

I think it was probably just projecting/guesstimating how big it thought this thread might end up being, perhaps with a bit of cynical twist tossed into the mix.

cranefritter · « **Reply #14 on:** September 02, 2025, 07:20:44 PM »

That's definitely hallucination. It's response style is very off putting, honestly. Making quip-y allegories like that is annoying... very annoying.

Though it did make me think of something. We actually have a really good first dataset we can use that's online- New Advent encyclopedia. I think I just read something on graph structured datasets. Let me try and find it again.

Edit: Found the paper I was thinking of but it's for graph-based RAG using Zettelkasten techniques adapted for the LLM.

Author Topic: Catholic AI Project (Read 2337 times)

Matthew

Catholic AI Project

cranefritter

Re: Catholic AI Project

Matthew

Re: Catholic AI Project

Matthew

Re: Catholic AI Project

cranefritter

Re: Catholic AI Project

Matthew

Re: Catholic AI Project

Matthew

Re: Catholic AI Project

cranefritter

Re: Catholic AI Project

Matthew

Re: Catholic AI Project

cranefritter

Re: Catholic AI Project

Matthew

Re: Catholic AI Project

Mat183

Re: Catholic AI Project

Matthew

Re: Catholic AI Project

Mat183

Re: Catholic AI Project

cranefritter

Re: Catholic AI Project