Content
@
0 reply
20 recasts
20 reactions
ncitron.eth
@ncitron.eth
Should Farcaster decouple clients from algorithms and moderation? We could design a common api specification that allows algorithm providers to be plugged in to any existing client. The main thing against this I think is that it's not clear to me whether the algorithm is actually the main product a client provides.
4 replies
2 recasts
38 reactions
Dan Romero
@dwr.eth
It wouldn't be performant without a lot of work. Modern consumer apps use machine learning and the algo needs to be fast. If it's not fast and supplying a lot of content, people churn.
3 replies
0 recast
32 reactions
ncitron.eth
@ncitron.eth
Yeah but the warpcast algo feed is generated quickly? I'm sure this took a bit of engineering effort but I imagine others can do it. The reason I think this would be valuable is that building a good algo is probably one of the hardest things to do when building a client, so standardized algo services would be nice for clients bootstrapping.
2 replies
0 recast
3 reactions
sahil
@sahil
@dwr.eth yes its hard, but quite doable. 3 engineers on our team put together a configurable ranking and feed infra for farcaster. It powers global, for you, channel ranking system and feed infra. Its costly, but if it becomes a public service for an open social graph, the innovation and value created at app/client level far exceeds cost.
1 reply
0 recast
2 reactions
Dan Romero
@dwr.eth
How many feed pulls per day is your infra set up for? How do we know the feed would retain relative to the control?
1 reply
0 recast
0 reaction
sahil
@sahil
feed pulls - current aws setup can serve few hundred requests per second for the For You feed (most computationally expensive), with 1 sec latency. global trending feed is 3 second latency, but can be sub-second purely by linearly scaling infra. all of this can be scaled up based on real demand. we do feed A/B testing based on network engagement data and user/client interviews for now. we can't do anything based on client side data/metrics, yet. in the future, clients can customize feed parameters or model weights using OpenRank SDK based on their client side data.
0 reply
0 recast
0 reaction
