Not to mention, Joplin allows more than structured text (native embedded images/media, even more kinds of content with plugins) and the text-only joplin.ai.chat() will definitely "need" to be supplemented with multi-modal AI support.

As for new techniques, I imagine the declining AI subsidies will raise the popularity of context reuse (reducing token use by not having the LLM re-read the whole notebook; is that what it's called? I don't remember much from videos/articles about AI trends because I don't use it). This will require a huge framework for managing data structures distilling every cached context window... And I think non-LLM functions will be added too to save tokens, for example an LLM may want to know "is there an unfinished TODO?" without reading all of them.