Helena Bibowa, the oldest and only surviving native speaker of a tiny language called Wymysiöeryś is turning 102 today.
In honor of that, let me take you along on the story of Wymysiöeryś.
It's of death and resurrection, and a hefty deal of random, but I promise it's a happy one.
This is such a nice feature of some Sinitic languages (Mandarin and Canto both do it): the possibility of splitting up bysillabic loanwords to form A-not-AB questions, so "is it ok?" becomes "o唔ok?" or "o不ok?"😀
Making these visuals using @manim_community has been a satisfying part of this work. We found it fruitful to dedicate a good amount of effort to visual communication.
So how do get to the latent space? It's a straightforward adaptation of LLM models:
1. Convert events into tokens.
2. Include longer-term trajectories in the loss function.
@ajwchin did some amazing work not only getting this working but with these visualizations.