34 sats \ 2 replies \ @elvismercury 28 Jul 2023 \ on: Daily discussion thread
@koob et al, here's a simple idea for a SN feature:
For every SN post (and possibly treating aggregated comments separately) compute its embedding. Then, at the end of the post, display the N closest embeddings as "similar posts". It would be cool to have some more continuity, and promote some ever-green-ness, esp given some of the topics that tend to garner attention.
You could even make it a paid feature, cost 10 sats to get 5 most similar comments to a target comment or whatever -- I think comment similarity would be especially interesting.
I've been doing something sort of like this w/ my personal notes / ideas journal, and it's been awesome. The connections it unearths are super interesting.
We have a related section under every post. Is it just not very good as is?
reply
Wow, I am an idiot -- I never clicked on that before. Jesus. Sorry.
That said: embeddings feel like magic when you work with them -- I've been amazed at the level of "insight" (for lack of a better word) that they find, e.g., entires that have a deep semantic similarity, but little surface similarity, are identified as being close together in embeddings space. Very much like real human insight feels in practice. Have a look sometime when you get a minute.
reply