Varun Srinivasan
@v
A quick overview of how @akshaan built Trending Topics on Farcaster. 1. Casts are selected and preprocessed daily. 2. Model categorizes thousands of casts with a prompt. 3. Topics are post-processed, merged and winners are picked It costs about $1/day and runs in < 1 hour.
19 replies
45 recasts
353 reactions
Varun Srinivasan
@v
Each day, a few thousand of the best performing casts are selected and pre-processed. All of the text content is turned into structured JSON that can be fed into the model as a prompt. Right now, we're classifying 1-5% of all casts.
3 replies
5 recasts
112 reactions
Varun Srinivasan
@v
Casts are fed into a model with a prompt to provide topics. A condensed version of our prompt is below. We do about 2 million input tokens a day processing casts, and it usually takes ~30 minutes to fetch all the related topics for each cast. Models are very good at figuring out subtle context. They can tell that a post is about basketball even with an oblique reference to a team.
2 replies
0 recast
13 reactions
Varun Srinivasan
@v
Topics aren't very clean when they come out. For example, you'll get "Pacers NBA finals" and "Indiana Pacers" which should be the same topic. A post-processing step removes filler words, merges topics with stemming and prioritizes topics with high uniqueness (TF-IDF score). This produces a very small set of unique topics.
1 reply
0 recast
9 reactions
Varun Srinivasan
@v
Topics are manually reviewed each day, which takes a minute or two for a human to tap through. Reviews are needed because models can output banal topics on occasion like "user experience". As models get better we'll likely eliminate the human in the loop completely.
0 reply
0 recast
14 reactions