Kasra Rahjerdi
@jc4p
I got nerd sniped by the new Trending Topics and wanted to see if I could do it in a better way. Checkout the approaches and results and let me know your fav: https://kasra.blog/2025/04/16/farcaster-trending-topics-analysis/?1
8 replies
15 recasts
62 reactions
shoni.eth
@alexpaden
i think the only “serious” improvement is efficiency around personalization nonetheless nice job / cool share
1 reply
0 recast
2 reactions
Kasra Rahjerdi
@jc4p
i think the RHLF approach would be a ton better cause you can make it personalized out of the box by weighing the users own interactions, but who wants to fine tune an LLM and host it on a GPU!
2 replies
0 recast
2 reactions
shoni.eth
@alexpaden
hmm maybe. where my mind goes is more nested classification sports > football > super bowl that’s the LLM output (just a classifier) and the personalization and efficiency is in determining if i at all care about sports so maybe that would be the best place to apply rhlf
1 reply
0 recast
1 reaction
Kasra Rahjerdi
@jc4p
two other ideas in this vein: 1. have the LLM classify into top level buckets, cluster those and compare engagement, cluster the top ones to find specific topics 2. have the LLM classify like 20 topics, filter to the ones that have the highest positive sentiment (if you want to encourage happiness and joy)
1 reply
0 recast
0 reaction
shoni.eth
@alexpaden
yeah those seem reasonable my only gripe with engagement vs (i.e these comments) is that i engage so much stuff i don’t at all care about i keep coming back to thinking of it as subpar to some degree. usually in a convo like this i don’t like your comment just respond, but i find it deeply interesting. guess that heavy lifting shouldn’t be on the app builder / if i don’t die of liver failure this week i hope the new huggingface releases help 🤣
0 reply
0 recast
1 reaction