upvote
As much as I love all these technologies, they don't have any direction that approaches the efficacy of a transformer
reply
You can have a transformer based index built on top of a graph database. Much of the explainability tech (people who prefer "alignment" also prefer "mechanistic interpretability") is reverse engineering the real world graph that's hiding in the weights.

https://research.google/pubs/the-case-for-learned-index-stru...

reply