How it's going.
@karpathy posted about LLM knowledge-bases. Very similar to the LLM Obsidian Brain I'd been working on for a few weeks. Had already built or designed most of the components, and I'm ahead in a number of ways.
My approach is built around rich (but flexible) self-evolving data types (called "artefacts"), either living or temporal. Things like "Designs" and "Reports" that learn from, and document, the things that you do, and make them reusable. This is very powerful.
These artefacts had BEEN the graph. His approach made me realise I needed to rethink my approach to ingestion & graph compilation.
1. An ingestion/graph layer beneath artefacts:
raw input --> ingestion --> granular semantic graph --> THEN produce artefacts.
2. Process input in real time, turn by turn, comment by comment, not by batch. Data (universal, append only) Context & Meaning (subjective, maintained).
3. Artefacts become the layer for humans (and agents) to collaborate. Stuff like, research -> a report -> changes to a design -> an implementation plan.
Attachment:
plan | raw input | processing report | ingested output