recommender systems

#topic #lab #AI #🌱 #response

i've always wondered how there aren't self-hosted recommender systems trained on your RSS subscriptions, people you follow in social media, articles/videos you consume, etc.

i've been wanting to train models/host my own algorithms for many years, but still haven't made much progress on the development side, besides the designs i've been exploring with the omnichannel curation feed experiment.

it feels like such a massive strategic/leverage point, not just having the right recommendations being delivered to you (going beyond persona-based filtering, more towards matching content type/level with what you're looking for), but even the act of ever-more training and engaging with your recommender system to clarify and uncover what your knowledge gaps are and what you're really looking for.

a few emerging projects are obviously using LLMs to be start moving in this direction, but even in spite of the resource/energy constraints on that, i'm still very unsure about how much LLM-centric this architecture should be. end-user training on taste and knowledge design patterns seem much more promising directions. (see: sari azout/sublime's content for lots of examples/discussion on that)

for a few years i've been following the developments of a few projects in this space, to see how/when they can facilitate that:

trails.social (ex-tweetscape)
sublime (ex-startupy)
hive.one