3 Promising Paths to improve Local Llamas

[nikshepsvn]

Full credit goes to @alain40 who posted this here: turboderp/exllama#92 (comment)
I wanted to raise the points he made there here as this is a larger forum to have this discussion and get more input, I've pasted a screenshot of his message for convinience:

Personally for me the third point is most interesting, if the base model is able to pick-up and use LoRa's in a tool-like fashion (like how we do for plugins in langchain now) we could have a base-model that can adapt responses based on various datasets in almost realtime, I'm going to dive deeper into this as well