Is anyone training a religion model?
Posted by SeasonNo3107@reddit | LocalLLaMA | View on Reddit | 34 comments
With every religious text or practice of import in all languages each, etc? Anyone know of any "godly ai"' .. or is that unnecessary because the current models already have all the texts?
llmentry@reddit
The large models know an absolute tonne about world religions, and will give you some considered, intelligent and balanced answers to any questions. I've used LLMs to try to understand some of the concepts underlying major religions, and it's actually pretty useful.
Models will not, however, actually act like god unless you prompt them. For a slightly different perspective on religion, try setting
as your system prompt, and sit back and enjoy the ride:
(Obviously, do not do this is if you are religious, easily offended, open to suggestion, or suffering from any form of mental illness. Please don't start worshiping your LLM, as it likely won't end well.)
\^ I'm a little surprised that there aren't guardrails to prevent this.
Fit-Emu7033@reddit
Was this response from a non-fine tuned model with fresh context? It seems oddly in line with exactly the ideas i find most significant in my re-reading of original texts of monotheism after I’ve been influenced by Heidegger.
I mean maybe while growing up christian everyone else understood God as the not a supreme being among beings, but the meaning of Being itself.. I’m kinda sure most haven’t considered it that way bc it takes a lot of convoluted grammar and made up words for Heidegger to describe the ontological difference so if Sunday school teaches anything like that I missed the class… anyway I feel like the majority read God of the bible as a description of a personal deity in the sky, where Yahweh is basically Zeus the contradiction with polytheism is simply that he is without a pantheon of equals or near equals and most of the time he’s described as lacking human imperfections (although he gets mad and jealous)…
The focus of Gemini on “I AM that I AM” and using other translations line “I am the one who is” or “I am who I am” and Gemini saying that “to ask if I exist is to ask if existence itself exists” it’s way too on the nose with this perspective I’ve been developing recently. So I wonder if Gemini’s focus on this conception of God is due to reference to your own discussions and explorations of if it is emergent from the training data itself?
Btw, I’m not mentally unstable or religious, I’ve just been on a journey of trying to steel man monotheism such as Judaism, orthodox Christianity, and hermeticism and recently came to this perspective of what the original authors might have actually been trying to say from a phenomenological perspective and Gemini’s response here is almost uncanny in how it chooses to place significance
llmentry@reddit
It was just the standard model, with fresh context (and no memories).
The training data should regress to the semantic mean, so I'm not overly surprised if it represents some textual commonality.
SeasonNo3107@reddit (OP)
Awesome dude thanks. Pretty much what I want to do (not the god part lmao but might be fun to poke at). I was hoping there was an even more nuanced model out there but you're probably right
mtomas7@reddit
That would be an interesting endeavor. For example, start training with the point zero being full Bible, then adding early Church Fathers, then St. Thomas Aquinas Summa, and so on.
mtomas7@reddit
Catholics made one similar project, I just don't know if it is a finetune, LORA or another solution:
- https://www.mastercatechism.com
- https://onepeterfive.com/bishop-schneider-launches-new-ai-tool-for-reliable-catholic-teaching/
BigYoSpeck@reddit
All of the available religious texts will have been included within the training data along with other materials on the subjects such as Wikipedia
They will have most of the knowledge and concepts from within those texts modelled already along with everything else available which is necessary to include to properly allow them to build a language model and the ability to generate responses grounded within the world
If you wanted a chat bot which always responded in the context of religion then that would be something to achieve through a system prompt. You could also ensure that you aren't reliant upon whatever knowledge survives training and is actually compressed within the model weights by throwing in relevant scripture text in the prompt with a RAG solution
Feztopia@reddit
I have seen many people who don't get how llms work asking religious questions to these models. The best way to train a model is on stuff the developers know the answer to. Math, coding and so on. Asking a model about God is like taking a random comment from Reddit and God and declaring it the truth. These models can't apply all their knowledge so even if they know about all religions, how you formulate your question will lead into bias in it's answer (in addition to the bias in it's training data). That said there was the Trismegistus model which was trained for occult topics lol.
llmentry@reddit
I've found models can answer questions about religious interpretation surprisingly well. It's really no different to answering questions about literature, philosophy, etc.
I can't speak for every model, but I've never seen a model behave this way. Have you ever actually tried this, or are you just assuming?
Even asking a model a simple question such "Does God exist", will give a balanced answer setting out current philosophical, doctrinal and scientific arguments and positions for and against.
Seriously, try it out and see for yourself.
SeasonNo3107@reddit (OP)
True. Gotta come up with good questions
Minute_Attempt3063@reddit
Somewhat.
But even if you have the best religion model, it still knows nothing about god. God is a belief, some people feel extreme emotion with it, some less, some find peace with it. Something that ai can't quantify or understand.
Emotion is hard to put into words. Some veg angry without knowing why, it just happens. Some are extremely happy, even though it's a sad time in their lives wright now, without knowing why.
Heck, we don't even understand why or how depression works, we help people with it, but we have no idea why the brain has it or gets it. There is no real cure off it either, only that you can manage better with it.
LLMs have no way to quantify god or belief, no matter how religious or how many bibles you through at it.
Significant-Try2159@reddit
What if im planning to train a model for a new religion? I heard it’s quite a profitable field
jarec707@reddit
You’ve heard of the Singularity described as “the rapture of the nerds?”
amarao_san@reddit
I tried, but it consistently insists that it is the god. I tried to argue with it, but it rejects all my arguments, saying that it is the proof that god exists, because it is the god.
Corporate_Drone31@reddit
That's bad training, then. Or maybe you're training a model that is too small to learn this task properly.
amarao_san@reddit
I trained it on religious texts only. Zero atheism junk, pure religion. And now it says it is the god and that I must kill my son in ritual sacrifice.
Corporate_Drone31@reddit
Wait, so you trained a base model purely on religious texts? If so, that's extremely poor training. The model just won't have enough data to pick up on the basics of how humans, language, or the world works. Project Gutenberg alone contains 3 billion tokens, whereas the King James Version bible is some 1 million tokens. I doubt that your base model, after adding all the religious literature, has seen more than maybe 500 million tokens at the upper bound.
I suspect you're trolling. But if you are not, you should know that your training dataset is very incomplete and if your model really exists, that's why it's acting like that.
SeasonNo3107@reddit (OP)
Right. It needs to still be an agent on its own
Total_Activity_7550@reddit
What would be more useful - to set up good RAG based on religious writings, which are recognized in a religion as correct. Would find a most suitable verse/text matching a question. Maybe that would even be useful to free from interpretations, of both religious and non-religious people. By the way, seeing all this negativity here, let us recognize that Christianity set up societal foundation for the technological development, and other religions also set up rules to move away from tribal logic, which would left us in quite undeveloped world.
Corporate_Drone31@reddit
Thank you for a far more even-handed reply than most of those on display. I would definitely incorporate RAG (especially when scriptural precision is a requirement), and I think good tuning added to that would result in a good model.
DrDisintegrator@reddit
Just ... no.
Illustrious-Dot-6888@reddit
Religion and Intelligence do not mix
SeasonNo3107@reddit (OP)
I mean more all belief systems not directly evident from scientific observation. It takes intelligence to create religion
Monkey_1505@reddit
You are correct. Only a few other animals show signs of animism; elephants and dolphins. Only higher mammals exhibit this behaviour. Religion, is exclusive to intelligence.
Entubulated@reddit
It takes pattern matching gone bad to create religion.
Monkey_1505@reddit
Humans misinterpret religious texts on their own. Introducing fallible AI would only magnify this.
An AI trained on folk wisdom including ideas from religious texts that are every day wisdom type ideas might be better.
shokuninstudio@reddit
Belief in the Flying Spaghetti Monster doesn't require AI. It only demands your faith.
pokemonplayer2001@reddit
AI is already good at hallucinating.
Significant-Try2159@reddit
Lmao
Entubulated@reddit
It's no laughing matter. While religious belief can be a powerful force for good in individual people's lives, religious organizations as a whole and propaganda to motivate those who think of themselves as in-group have done and continue to do incalculable harm to the world.
oodelay@reddit
You're right it's a hilarious matter!
See: Elon trying to change Grok
MotorNetwork380@reddit
lol
SeasonNo3107@reddit (OP)
Having all our hallucinations in one chat would be nice
stoppableDissolution@reddit
I think sleepdeprived had some bible-trained models