gods.txt / AI Hexcrawl

AI-generated Scripture and Fantasy Locations

About

@gods_txt is a Twitter bot that publishes text generated by a natural language processing model trained on on a growing corpus of world religion texts built on top of GPT-2. This custom model produces outputs that employ the same style, themes, and vocabulary as the source material and combinations thereof with additional vocabulary and context afforded by the base model.The resulting passages are full of interesting cosmogony, heretical sermons, and silly syncretism.

Frequently Asked Questions

Q: WHY DON'T YOU USE GPT-3?
A: Bigger is better, right? Well, no, not in this case. GPT-3 cannot be trained beyond a one shot due to its immensity. At most the AI riffs off of prompts in a realistic fashion, but it doesn't synthesize the material in the manner demonstrated by my narrower, custom models. It's an important but underappreciated distinction. It's not as sexy, but there it is.For GPT-3 content, check out @ai_hexcrawl!

Q: HOW DOES THE BOT WORK?
A: I use Max Woolf's gpt-2-simple aitextgen colab notebook and sometimes Adam King's (creator of Talk to Transformer) Inferkit to train and generate text. Instructions for how to do this are publicly available. Tweet-sized passages generated by the trained model are entered into a queue which pushes them to Twitter on recurring intervals.

Q: WHY DO I SEE WORDS AND CONCEPTS NOT TYPICALLY ASSOCIATION WITH RELIGION OR MYTHOLOGY?
A: As noted above, the model is not comprised exclusively of the training material, but is instead built on top of a general purpose language model. This allows for greater flexibility and coherence.

Q: HOW MUCH LIBERTY IS TAKEN WITH THE TEXT?
A: The text generated by the model is itself not modified or edited. However, in the interest of transparency, I believe it's worthwhile to identify all the steps where human intervention does occur and potentially affects the outputs.The process of training and generating text isn't without its biases. First, there is an inherent bias in the pre-trained model. The base model was trained trained on "websites and documents," and I've heard Reddit was in the mix. I'll let you draw your own conclusions.Second, the training data was selected by me, and its content is governed by its language (English), its availability in plain text, and its trainability. Things like long lists of genealogies for example have a tendency to get the model stuck in loops, so those are excised. The model also doesn't do well with unfamiliar proper nouns, as it will place them at a whim in an inappropriate context. I actually would love to include more Chinese folk religion, but haven't had the time to solve this problem.Third, I curate the generated outputs, so more selection bias. I choose what seems cool to me, and since I've had some time to gauge reactions, that feeds back into my selection criteria. The good news is, I'm fairly permissive in what makes the cut. I do not permit things which are overtly sexist or ableist, or that bring up sexual violence. I do permit antiquated references such as "man" and "mankind" when referring to humanity because that is a consequence of the source material, but is not in any way meant to exclude or otherwise harm people of any gender identity or expression.

Q: WHY DO YOU DO THIS?
A: I wanted to add verisimilitude to my audio fiction project and homebrew RPG campaign setting. I thought text synthesized from in-world religions could be used, or in the very least, serve as inspiration. I was also curious what an AI would make of the world's religions, and perhaps perform some sort of analysis. Anyways, I thought it was worth sharing, hence the Twitter bot.I don't do this to lampoon or denigrate anyone's beliefs or faith traditions. Regardless of how you or I may view religion, it has been with us in one form or another for millennia. From early shamanism to megachurches, faith and society have impacted one another and evolved together. Maybe we can reveal more about humanity in the latent space of what has been held as divine.

Q: WILL YOU ADD AI ART (E.G. LATENTVISIONS) TO EVERY POST?
A:
I have experimented with advadnoun's Colab workbooks, and while it is extremely cool, I will not be adding this into the process. @bokar_n has popularized this practice, and I love sharing, but I would prefer the interpretation be left to the reader as far as the original output. If you would like to generate art based on the outputs, quote retweet the passage with your creation, and I'll retweet it.Once again, for this feature, check out @ai_hexcrawl, where each post (as of December 2021) is accompanied by pixel art generated by Pixray.Update — 4/23/2022: I will begin including images generated by vqgan + CLIP. The prompt is the GPT output.

Q: WHY DO YOU HAVE A KO-FI?
A: The services I use to access the necessary processing power aren't exactly free. Now I've sunk enough of my funds into this project, and I'd happy to keep doing so in order to keep it going, but financial support will ease the burden, not line my wallet.

Q: WHAT OTHER PROJECTS ARE YOU WORKING ON?
A: The following are in various states of completion:

Contact

I'm always happy to respond to questions and lend advice if you want to attempt your own custom model, but I do have boundaries. Please do not ask me to give a detailed tutorial, as these have already been made available elsewhere.

Testimonials

I had to unfollow @gods_txt because I was starting to worship it.
— @The_AlexWallerFwiw, @gods_txt is gimmick account but it's also a great statistical representation of every human transcendental experience. As an AI guy, if I had to pick a religion, this would probably be it.
— @StatesWarringHoly shit this goes HARD.
— @fearjunkie