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.

Technowatermelon. Elder Millenial. Building Farcaster. 

nf.td/varun

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).

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.

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.

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.

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.

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.

Huge shoutout to @akshaan and @cmlad.eth who have built out all of our ML and feed infrastructure!

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

ML, robotics, DeFi | building /garagesale /gysr /tournament ...

is the model trained to predict whether a user interacted with a cast, given those four input features?

i.e. basically scoring "likelihood of interacting"

IRL: Creative Director who makes movies and ads.

is this changing the way the notifications tab works too?

Unlearning. Built @nftsoffc. Full time product manager. Part time builder/tinkerer on farcaster. Working on something.

This model isn't (yet) focussed on finding new stuff. 

It's mostly reordering the stuff that is already in your feed to present it in the best possible order so that the most interesting stuff is at the top. 

That seems to make feeds a lot more useful for people, both qualitatively and quantitatively.

Relying purely on the metadata of a cast - likes, author, frequency - instead of the content of the cast somehow doesn't seem complete. 

Wouldn't that work against surfacing relevant content and new users casting relevant content?

Wow this thread is super interesting, thanks for taking the time to go into the details !

Is there a way to somehow opt out of this or control it? Because I don’t want my likes to dictate what’s showing in the home feed.
I just want things to be shown according to when they are posted and not how much I approve of them.

Is there anyway a user can reset the data the model has about them or the user would have to change his behavior for the current data to be overwritten by the new data with time?

Fascinating - thanks for sharing this and expounding on the topic in your subsequent casts, Varun. Looking forward to how my WC feed is affected.
1000 $DEGEN

This one was at the very top of my feed, so it's working 😝

great thread Varun! have question!

If i like many people who cast alot, is it possible for high ranking unfollowed people to be seen in my feed at all?

Can we please have a simple feed with casts of people we follow in reverse chronological order, just to be able to catch up with ALL our friends, without any algo that wrongly decide to hide them. Our friends are always equally important, whatever the likes, the frequency of cast or the number of replies 🙏

Thats why i saw now account that i did not follwed and account that have boosted likes .. i can say it boosted because its have 300+ likes even his/her cast is nonsense.

Once again, we are seeing the importance of likes. When balancing out reactions per unit?