Varun Srinivasan
@v
Here's an idea for a Farcaster startup: - Sell labels for Farcaster accounts - Accounts can have multiple labels (e.g. airdrop farmer, onchain transctor) - Labels can be either binary or a confidence score (0 to 1) Warpcast would pay to use this service if you did a good job.
33 replies
32 recasts
458 reactions
Varun Srinivasan
@v
By labels I don't mean generic stuff like "proof of human" which is largely uninteresting. Labels should be valuable to a specific company building on FC. For example: - Onchain spender (useful to anyone launching a mint) - Geopolitics caster (useful for Warpcast feed algorithms) - Automated account etc...
8 replies
24 recasts
213 reactions
Varun Srinivasan
@v
Other mistake people make is trying to giga-brain it by automating fully or applying LLMs. This is really, really hard and unlikely to work. Start by manually labelling with heuristics and sell the data. If its working, try to develop a basic ML-model using some common features.
4 replies
0 recast
21 reactions
kbc
@kbc
Isn’t this still supervised machine learning? You pick 100 to 500 people at random, label them, and the label the rest based on similar behavior
1 reply
0 recast
1 reaction
Varun Srinivasan
@v
If only it was that easy in practice
1 reply
0 recast
0 reaction
kbc
@kbc
Do you mean that there is more value in manually labelling top 100 to 500 users and sell that(and then iterate), instead of going for a larger labels dataset with potentially less accurate labels?
1 reply
0 recast
0 reaction
Varun Srinivasan
@v
I mean that the extrapolation never works that well in practice. Lots of false positives and negatives.
0 reply
0 recast
1 reaction
