Yoyoo great post and talk - @mantequilla-soft just closed that gap on https://3speak.tv 👩🚒
The new algo is live! It's based on viewer retention + the interests users pick themselves. We reused the LLM we already run for captions: it takes the transcription + the post data (body, community, tags, etc.) and works out what the video is actually about aka finds a general "topic".
Also 3Speak struggles with viewer numbers right now as every Hive frontend but we've got one big advantage: everyone else uses our player. So on the retention side, we are covered :P
The algo runs on the feeds, plus the recommended list on the watch page and the shorts feed, of course. We'll be tuning it over the next days/weeks, but I think this is a huge milestone towards actually feeling like an app in 2026 :D
Thats great to hear. Would be great to see an algo-race on Hive with seeing who can better use the onchain data. The actual onchain interactions could help create social graphs that could inform the algos at a higher level.
Yeah, it would be good, but on 3Speak we don't only use pure on-chain data. The most important metric is viewer retention, which is not stored on Hive. Currently we're experimenting with social graphs but first we need to get some watch time to actually play around with good weights on all parts of the algo :P
Support engagement with a delegation to topcomment:
50HP - 100HP - 250HP - 500HP - 750HP - 1000HP - 1500HP