> where an LLM can be useful [...] RAG, summarizing, filling gaps
I'd also emphasize onboarding. To smooth "oh joy, yet another ux", and simplify/speed/clarify "yes, that'd be nice - but no, we don't support it yet". One might imagine a pkm ui where you just start telling what you want, and it executes keyboard-visualizer style. And clearly characterizes what it can't do. Rather than 'intro-doc:"to do X, fiddle Y" -> human:"ok... I need to fiddle... Y?" -> ui:fiddle-Y-event"'. Keybindings as efficiency "expert mode", rather than onboarding barrier.
> If you were to try a small fragment, what would you pick as the smallest useful scope[?]
Last month's items which each prompted a "I really wish I had better tooling for this" search for something to try, were: A largely self-contained todo space, with lots of fiddly bits (life todos for another person). An exploration of a project idea - chats, notes, links, papers, sketches (MetaPaint: given stickers for ui elements, a custom paint app UI is simply another kids painting). A nudge on a backburnered area of interest - chats, papers, jupyter (towards a didactic perceptual color space, where the features you notice are real features of human vision, not mere model artifacts).
I actually looked at Obsidian for the first one, but I really wanted lightweight hierarchical task decomposition with implicit but overridable dependencies. Currently doing a one-big-markdown with frequent manual-ish passes. Tempted to fork an existing "mutant-markdown as executable literate-Julia" kludge.
But "smallest useful scope"... hmm. Of those, "project" and ""log"" seem for me plausible - "team" perhaps more a follow-on. But... I've been wanting better so often for so long, I'm mostly gated on there being a either a clearly satisficing task fit up front, or some hope of a plateau escape. A slog of extending/kludging/adapting in hope of incremental gain has diminished appeal. Hmm, a thought: having just looked at several pkm docs, current onboarding seems more "start with the basics, and build out", rather than a clear "with a fully tricked out set of extensions, here is the envelope what you can/cannot/by-workaround manage with us today".
I'd really like a plateau escape. Outliners, Hypertext, tiny manual force-directed graphs... it's been a fun half century, but I'd like to move on now. I wonder what an old tidlywiki-style self-hosted pkm bootstrap might look like, in a time of git, whisper, discussion and coding LLMs, extensive community code repos, and graph libraries less hopelessly behind graph HCI research. 1080p HMDs, head-tracked 4K, and gestures. Hmm...