I skimmed your post and noticed "Harness Layer vs Core Intelligence?" This is an interesting topic! I've been wondering conceptually about how AI works vs a human brain, and also in applications that deal with lots of business data. Seems like retraining as it is today's tech is an very expensive task. So some people opt to use lora to add another trained matrice (?), or add RAG to allow for giant business datasets to be searchable.
I think retraining and this kind of additional weights is a key part of how humans actually learn. Context compaction can only go so far and "memory" (filesystem, notion, etc) doesn't even have compaction or affect the core intelligence/behaviour. Sort of like "instructions" vs "personality"
I also wonder if any sort of AI self-led training/retraining/etc will diminish the quality of the initial training.
If you do find an interesting group/slack/etc, please share it! I'd love to learn more about the harness side of things, and maybe long term, conceptually on the ML/AI side of things too. My background isn't really in ML so I doubt I'll be able to chat too much in that sense.
