No Boilerplate @noboilerplate@tech.lgbtWHY IS AUTOCORRECT SO SHIT #GPT IS LITERALLY AUTOCORRECT!
Oh wait.
You can't sell "better autocorrect" to investors 
(if you've not seen it, my thoughts on #AI here https://youtu.be/kul0z3OTmVM)
@noboilerplate literally what LLM powered code completion is about -
Also, mobile keyboards have been using smol LMs for a while now, trying to recommend the next most plausible word (based on one or two of the previous words).
Currently using FUTO keyboard, some FOSS board with similar functionalities and even voice recognition (which is a rarity among Foss keyboards)
Recently mister Matt Williams showed off a thing which essentially does the code-completion thingy, but for obsidian [www.youtube.com]
I use Logseq [logseq.com] cuz Foss, so I couldn't care less, but maybe you are interested!
No Boilerplate @noboilerplate@smorty exactly. Give me a better model on my phone!
But... writing is the part I want to do, it's the REST of the boilerplate I want to automate! 
(BTW I use and recommend obsidian over logseq despite it being closed source because obsidian uses markdown. And while logseq started off using markdown, recent developments have made it look like they're moving to a propriatory-schema db that then is presented as markdown.
I need my brain to be files on disk.
In a perfect world I want open source and open date. But I've realised that open data, in the case of my second brain, beats open source.)
@noboilerplate very much so! The writing is what makes your content so interesting.
Putting bigger models on mobile is hard though. I tried running Qwen2.5-0.5B (300 MB in size) on my Pixel 6 and woah, that's slowing things down dramatically. The device becomes unusable and it generates about three tokens (words) per second. But the output quality is pretty good.
Yeah, that whole weird binary conversion i too don't like (the Logseq thing). Apparently it allows for quicker search, which could be useful.
They want to achieve a hybrid system now, which combines the raw markdown files in the main directory and the binary conversions for fast lookup stored elsewhere.
People pushed-back against the "binary only" workflow, so now that is the goal. Maybe said, they want to keep it in seperate files for LLM integration (which is reasonable I think)
I totally understand why you chose obsidian. It just seems like the better tool on general.