I kick-started my AI deep dive into the world of biblical genealogies, a sea of names, connections, and implied stories. As I navigated this intricate network, using OpenAI's embedding model (text-embedding-ada-002), I observed intriguing patterns in the AI responses and grappled with the mysterious algorithms behind them. This unexpected journey reminded me of the art of interpretation that biblical scholars partake in, a constant dance between the words on the page and the world they attempt to illuminate.
Deep Dives and Divine Dialogues
The world of biblical genealogies is an intricate web of connections. I asked our AI chat agent to identify important elements within this complex tapestry, and I noticed a recurring theme - the names Jesus, David, and Abraham. These characters seem to hold a prime position in the AI's responses, leading me to question how the model weights each element in its vector database.
Could the order of data uploaded be influencing the output? The AI's responses might be sequenced by the chronological order of the texts, with Jesus from Matthew (Latest), David from 1 and 2 Samuel (Middle), and Abraham (First). Or is the AI picking up on some inherent significance to these figures? After all, we've woven in GPT-3.5-Turbo's parameters and all of the secondary text around the Bible.
As I reflected on this, I formed a working theory around semantic and thematic relevance. Could the model be clustering together similar items by their meaning? If that were the case, the RLHF (Reinforced Learning and Human Feedback) that guides the AI would be organizing the info around central themes and their meanings.
This led me to a critical question, one that I believe needs to be asked before fine-tuning can begin: How does the AI weigh its responses? Knowing this could drastically impact our results.
The Research Workflow
I'm using a basic implementation of LangChain's Conversational Retrieval Chain. Essentially, it's taking the selected books of the bible and creating a vector database with Chroma.
When we ask a question, it searches Chroma for relevant information and retrieves the items related to our questions as context.
If you're interested, I wrote an article explaining my AI Research Workflow.
AI Conversational Retrieval
A workflow based on LangChain's conversational retrieval agents that I used in the AI bible research project to upload documents and ask them questions.November 15, 2023
Differences from ChatGPT
For reference, here is ChatGPT-3.5-Turbo's response without the verses on genealogies as context.
It's not bad as is, but if you notice, the response is more about the information instead of my question.
Ideally, I want the response to be more like a thematic narrative. A story based on our questions and the relevant parts of the Bible. Something that enhances our reading of the Bible, rather than information retrieval.
AI Perspectives and People of Faith
During my interaction with the AI chat agent, I noticed that changing the system prompt influenced the results in unexpected ways, presenting a challenge for introspection.
These observations led me to an intriguing thought experiment:
How would a person who is an expert in biblical genealogies leading a bible study respond if they were an AI chat agent?
And how would putting ourselves into the imaginary shoes of the persona we create, shape the responses of the chat agent that can't understand the human mind or soul?
Here's a collection of responses that I found interesting.
A Journey of Discovery
Embarking on this project has been both challenging and thrilling. As I weave through the world of biblical genealogies and AI's mysteries, I am learning, unlearning, and relearning.
I am filled with a sense of curiosity, fascination, and at times, slight apprehension about the complex relationship between theology and technology. But as I delve deeper, I realize that this intersection of AI and the Bible is a goldmine of discovery, full of exciting patterns and intriguing puzzles.