Devin Conley
@dcon.eth
Anybody in the farcaster community thinking about ML recommendation systems for a decentralized social network? Lots of interesting challenges across data collection, training, model deployment/inference, asset management, etc
5 replies
0 recast
2 reactions
Dean Pierce 👨💻🌎🌍
@deanpierce.eth
I had an idea a few years back to do this sort of thing on the frontend with tensorflowjs. It was my vision for bbly.io, a firehose of news and comments (similar to Reddit) delivered over IPFS pubsub, curated clientside by models trained by users as they vote/label content. Pure static HTML social news aggregation 👍
2 replies
0 recast
0 reaction
Devin Conley
@dcon.eth
Cool idea. Were you planning to use individually fine tuned models or some kind of federated learning approach?
1 reply
0 recast
0 reaction
Dean Pierce 👨💻🌎🌍
@deanpierce.eth
Definitely individually trained, in the browser. Every post is labeled by the community with any number of seven labels (hilarious, enlightening, etc). Users start with a handful of default scoring/viewing profiles, letting them prioritize funny, or inspirational posts etc. They can then make their own. Then ML boosts:
1 reply
0 recast
0 reaction
Dean Pierce 👨💻🌎🌍
@deanpierce.eth
After there's enough labeled data in the user's local storage (indicating their labeling preferences on things they've voted on), they can train custom ML boosts to augment the scoring of each post. Not huge, maybe 5-10 mins of training, and nearly instant inferencing. These designs were from before LLMs were a thing.
2 replies
0 recast
0 reaction
