Varun Srinivasan pfp
Varun Srinivasan
@v
I call this idea the "feed marketplace" as opposed to "feed-as-a-service" which is what @openrank and others are doing. The challenge building a "choose your algorithm" is that the most interesting feeds are no longer just "algorithms" or fancy SQL queries.
13 replies
22 recasts
173 reactions

Varun Srinivasan pfp
Varun Srinivasan
@v
Feeds are performance sensitive, and you cant just run a SQL query when someone is scrolling. Ideally, you have a very long cached feed for your users that can be quickly delivered on demand. This is computed and stored in a very fast in-memory datastore like Redis.
1 reply
0 recast
18 reactions

Varun Srinivasan pfp
Varun Srinivasan
@v
Finding interesting content also relies on a lot of dervied data. Warpcast will take the number of times you've interacted with someone and divide it by the frequency with which they cast. This is "affinity" and it helps prioritize casts in your feed. Such metrics are difficult to calculate in real time.
1 reply
0 recast
19 reactions

Varun Srinivasan pfp
Varun Srinivasan
@v
It would be hard for Warpcast to fetch and service a feed over an API. A more interesting approach is to get the algorithm and run it inside Warpcast's infrastructure. We tell you what data is available, you write a binary that can ingest data and produce a feed for a given user.
3 replies
0 recast
17 reactions

Varun Srinivasan pfp
Varun Srinivasan
@v
This is similar to what Numerai does for their hedge fund model. The only problem is that this stuff takes a long time to build and get right, so unfortunately not on the near term roadmap đŸĨ˛
2 replies
0 recast
13 reactions

Yassine Landa pfp
Yassine Landa
@yassinelanda.eth
Performant AI based personalization solutions are sub 50-100ms even using state of the art ML feeds (things that trains in realtime on huge derived data). Integrated them at scale as an APIs in many Fortune 500 companies.
0 reply
0 recast
1 reaction

Samuel pfp
Samuel
@samuellhuber
assuming a feed as a service customer runs their own redis one could stream it to their and push feeds as a service. Though precomputing only for paid users of that client and not for all farcaster users for cost reasons
0 reply
0 recast
0 reaction