Comment by leobg
3 years ago
Sounds very interesting. Do I’m having trouble picturing how it works. How do you get the links into the system? I’m assuming this won’t work for links that you find on your phone, i.e. when not on your on your computer? And how do these links get indexed? By the stuff that’s on the page? What’s an example use case of where you use that system to find a link?
It's pretty manual at this point. The indexing is done by hand. The idea is kind of crazy, but I think it can be made to work, in the same way that Wikipedia is maintained by hand.
https://digraph.app/
https://github.com/emwalker/digraph
If you can crowdsource the indexing, you get yourself a manually curated search engine with a nice topic graph that can be traversed. A piece of this puzzle that hasn't been tackled yet is a reputation system to keep the signal-to-noise ratio high and deal with spam.
> What’s an example use case of where you use that system to find a link?
An example use case is that I come across some interesting long-form article on a topic I'm following, e.g., Shackleton's expedition, that's published on a nice website and that I don't have time to read. I can just drop the link in the right topic and get back to it without too much difficulty. Or that's the hope, anyway. (Doesn't always work out that like that.)
Another thing I'm interested in is what the topic structure ends up looking like as it's more fully fleshed out. So sometimes I'll drop in random links even if they're not that interesting, just to build out the topics.
Thanks for the link. Quite a collection. Have you thought about auto suggesting categories when adding new items? Cheap and fast way: Use fastText or Fast Sentence Embeddings to calculate average embeddings for all items in each category. Then, when user enters a new link, create embeddings and rank categories based on the new embeddings’ similarity to those of the categories.
I think there's plenty of scope for improvements like this. I haven't had the energy to pursue further improvements along these lines, but I suspect they could make it a better app.