Varun Srinivasan pfp
Varun Srinivasan
@v
Warpcast now uses an ML model to order casts. If you have many unread casts, the model decides which ones go at the very top of your feed. It's main purpose is to learn your preferences and show you the most interesting stuff first. I'll go into some more details on how it works in the thread below.
52 replies
20 recasts
179 reactions
Varun Srinivasan pfp
Varun Srinivasan
@v
Today's model uses 4 main features to "rank" a piece of content: 1. How often you like casts from the author 2. How often the author casts 3. How many interactions the cast has 4. How old the cast is
6 replies
4 recasts
35 reactions
Varun Srinivasan pfp
Varun Srinivasan
@v
If you like someones casts a lot, their stuff will move higher in your feed. This is the primary factor used to rank your feed. The model's predictive power (y-axis) increases as the number of likes increases (x-axis).
2 replies
1 recast
23 reactions
Varun Srinivasan pfp
Varun Srinivasan
@v
If you cast a lot, the predictive power begins to diminish. A high prediction score means that the model will bump the cast up in the feeds of people who like it and down in the feeds of people who wont.
1 reply
0 recast
8 reactions
Varun Srinivasan pfp
Varun Srinivasan
@v
Importantly, this doesn't mean you should cast less! Your casts still show up in feeds, the model is just less confident about changing its position in your feed. But if you are casting a lot of stuff that doesn't get engagement, trimming some of it might help the model get your content to the right users.
2 replies
0 recast
13 reactions
Varun Srinivasan pfp
Varun Srinivasan
@v
The third factor is engagement - the more likes a cast has, the more the predictive power is. Interestingly, the relationship isn't linear. It increases from 1 to 10, then decreases till 50 and then increases again. I'm not entirely sure why this happens yet, warrants more investigation.
3 replies
0 recast
11 reactions
Varun Srinivasan pfp
Varun Srinivasan
@v
The fourth important factor is time. This acts as a balance to the other factors - even if something is from someone you like, if its many hours or days old it's unlikely to show up at the top of your feed.
4 replies
0 recast
9 reactions
Varun Srinivasan pfp
Varun Srinivasan
@v
And that's a quick overview of how our model decides what casts to show you. We expect to be adding in more features over time like language, channel data etc which will make the model even better at finding stuff that is important to you.
4 replies
0 recast
8 reactions
🪱 pfp
🪱
@worm.eth
Will this be open source? Feels like it would help people building alternative clients
1 reply
0 recast
0 reaction
Varun Srinivasan pfp
Varun Srinivasan
@v
@openrank is building this as a service for developers, which is the right way to make it easy for clients. open sourcing our implementation is not that useful because it only works with our specific toolchain and infra. if you use a slightly different tech stack youd have to rewrite a lot of it.
1 reply
0 recast
2 reactions
🪱 pfp
🪱
@worm.eth
thanks for the response!!
0 reply
0 recast
0 reaction