Yeah. I thought it was just long chat history, so I partly solved that with my projectContext.md file over at this post Workaround for lack of memory between chats
That has helped a lot with starting new chats more readily. But recently seems increasingly sub-par from the get-go. Hard to tell whether its model and / or auxiliary functionality thats the cause, probably a little of both.
I wonder if large models like 3.5 Sonnet experience degradation the longer they are out in the wild getting RLHF’d to oblivion, can string some parallels with the long context conundrum.
Of course nothing empirical to back any of that up, just some rants. I’ll probably focus on keeping my own engineering sharp for a bit, while we wait for whatever’s next.