@Wattenberger Great post Amelia π The "zoom" UI pattern feels like a natural one to navigate levels of abstraction. I think beyond context this idea is also instrumental for discovery, which is really difficult with search-retrieve UIs like chat and search boxes.
https://t.co/BTXbqCIuCI
Iβve been experimenting with using GPT's learnings from its training data to get this coordinate system & content zoom-ins and outs. The coordinates, and thus topics, are relative to each other and spatialized in 2D.
This feels more like going to different bookshelves in a library, and gives a way to see what's on the shelves. Tying it to coordinates allows panning over to related topics, or zoom-in to more specific topics.
When you find something you like, its easy to then ask about it.
You get to a bookstore, wanting to learn something new. Everything is nicely fit into different categories, and you have no idea what you will discover after some perusing.
You could basically keep zooming in, further and further, until its at a specificity that you want to actually dig in and learn more about a specific topic.
Since the general workings of a transformer model is to guess the next best word based off prior words, I'm curious if this could work on other linear-based sequences such as app actions. i.e [Open browser] -> [Open Email] -> ??? [Suggestion: Tab to Open Unreads]