I had a cool idea for a new LLM-based application this week that I think has a ton of potential.
A Specific Product
Imagine feeding an AI system a book chunk by chunk, allowing it to update its knowledge about the characters, plot, and world as it goes. Instead of just shoving it all in a single context where it might get confused or try to condense information, it could verbosely and dynamically update a database with key information and time-based metadata.
For example, let’s say you feed it the Lord of the Rings series. You could ask it questions like “What was Frodo’s mental state halfway through the second book?” or “When did Gandalf first get shown wearing white?” The AI would be able to pinpoint those details because it “read” the book in stages, tracking the progression.
This could be super powerful for generating and updating wikis, tracking character arcs, or even analyzing D&D campaign recordings!
The Abstract Pattern
But wait, there’s a more powerful idea here, in my opinion! This pattern of an AI system with incremental learning and updating could be applied to all kinds of content, not just books.
Example of an LLM generating a timeline
Imagine feeding the system a series of disparate articles and posts about the same incident. It could compile a full timeline and detailed list. Or if you had it process live audio or video, where it would be a necessity to ingest the content bit-by-bit, it would be able to update a timeline or key points in real-time.
I view it like the difference between cramming for a test by reading a textbook cover to cover and actually learning the material over time. The latter approach leads to a deeper, more “human” understanding because it can be saved with time-based context or ordering, and LLMs are generally able to be more verbose and detailed with a higher ratio of output tokens to input content.
Some Ideas
The applications for this type of incremental learning AI are endless. It could be used to:
- Create super-detailed wikis for shows or book series
- Analyze a person’s political views and how they changed over time by their social media posts
- Track the progression of a scientific theory through successive papers
- Monitor character development and world-changes in RPG games via audio recordings
And so much more! By mimicking the way we humans learn and update our knowledge, I think AI systems can get better insights than they currently have.
– Joseph
Sign up for my email list to know when I post more content like this.
I also post my thoughts on Twitter/X.