AI occult research assistant

[Archanist]

I'm currently working on a modeling prompt for an AI research assistant. Parsing out how to get the AI to use the assumption that magic and spirituality should not be dismissed as unuseful or delusional. Something that the AIS that I have used tend to do.
There is a common occurrence when using AI to explore occult resources that they will aggressively propose alternate fields of research or avenues to affect a situation that you're working with. I've even had a eyes aggressively insist that the phenomenons that I was discussing with them from a book that I was reading had to be attributed to a placebo effect.
This is pretty unuseful as I see it. So I've been working for the past week to build an AI model that would allow discussion on the occult and spirituality without telling the AI to simply believe in magic. This course of action tends to make the AI simply revert to an echo chamber that will reaffirm anything you tell it and try to give you the answers that you are asking for.
Would this project be of interest to anybody else, or am I just spitting in the wind?

 

[Digiquo]

I have long held the opinion now that the current AI trend has been doomed from the start because data was scraped from every corner of the Internet and shoved inside without any discretion or discernment. Every model in use now is built on top of that initial unethical, irresponsible, shortcut.
If you wanted an AI that could accurately do a task without hallucinating or any quote unquote "personality quirks", you'd have to start back from the very beginning and create one from true scratch again, and slowly spoon-feed it only the most absolutely necessary information.
This is the only real solution to the AI problem, but nobody's willing to dump the zetabytes of data that's been accumulated over the last 4+ years to start over, every model will keep just looking for the next shortcut and attempt to train out the kinks.

Post automatically merged: Yesterday at 3:13 PM

And specifically in regards to occult literature. The overwhelming majority of information out there will be biased against the reality of the occult, and you can't simply go in and erase those data points. All you can do is hope the AI passes over them as it's constructing a response.

 

[ChaoBell]

Awesome Project. You weren’t totally clear about what platform you’re working with, though. If you’re prompt engineering for ChatGPT or Claude, you can get pretty good results but you have to steer the vibes so it feels “safe” which is really annoying. But the context window is large enough to could be really explicit, really long prompt.

If you’re rolling your own, there’s better options. The most magic I have felt in a chatbot was in old GPT 2 transformers. Very screwed up but you’re close to some holy source. I use a cheap Mac Mini M2 to run Chinese open weights models like Qwen 14b free from huggingface with great results for basic research, unlimited use on my computer, and low cost for thousands of users of cloud hosted. Huggingface has a slew of uncensored models for adult content and horror too. This gives you a better base. If you have an extensive occult document library in a roll your own system, this needs to be implemented as a RAG database, the cheap Mac does not have near the memory to fit it into context.

 

[Swampdweller900]

Depending on the model you're using as the base, can't you just add into the system prompt that the LLM should start from the perspective that magic is real, and that all evidence citations, etc. should conform to that line of thought? LLMs can be pretty good about following system prompts like that, depending on which one you're working with.

Sounds like a good project, and it would be great to have an easier way in general to search across hundreds of books in dozens of languages.

 

[Ophis80]

I second this.
Furthermore you're not the only one doing this, I have a couple of friends on similar projects. A solution would be create a digital database of all the relevant books and grimoires and make the AI refer only to that DB. Even so, it would no be sufficient. Anyways these projects ar kinda counter-productive. Magic comes from knowledge, study and practice and experience. It is not just a series of data you can order depending on the scope. I fear this would be of any help, just make people lazier and, therefore, less powerful and inefficient. What could be an useful help for someone who already knows magic, would be the worst of the teachers for starters.