1. Tag all content with a date (probably almost done). In long passages, estimate a date based on published + written rate. 2. Once every sentence or phrase has a date, create a few blockchains. One for short thoughts (128 tokens), and more chains for longer thoughts (3072+ tokens). 3. Run an LLM embeddings cosine similarity or some better metric for each tokenized idea with a threshold of say 75%, tokenized by natural punctuation if not found. 4. Only NEWish content gets stored in the block, chained if novel. Again, short and long thoughts recorded. Coordinates from the vectors. 5a. I remember reading once that working through a PhD was like pricking the inside of a balloon, I love that visualization. Every novel idea pricks the inside of a growing sphere of knowledge over time. 5b. Has anyone ever tried to map human knowledge in three dimensions? 5c. Alternatively, a tree or root system growing over time. Branching from the vectors.
MountainMan1312•3h ago
All this stuff about tokens and vectors is quite different from my approach. I'm always interested to see how other metamages approach this issue.
I think a hypergraph structure is the only thing capable of properly representing knowledge. Knowledge is not just a bunch of nodes with 1:1 connections; there's complex structure in the nodes.