I started a new For You experiment today that gives less priority to popular posts. 50% of the users on a given day will see popular posts demoted: score = score / power(popularity, 0.2) 50% will see no change (control arm). "popularity" is the number of likes and reposts a post has.

The algorithm behind For You is very simple so it's easy to see why any post shows up. Curious? Then: 1. Go to linklonk.com/bluesky 2. Enter your user name 3. Enter the post url to explain 4. Hit Submit

Have been running this For You feed experiment for a week now when we demote posts that were liked only by a single curator (ie, a user that shares a like with you). In the experiment the users are less likely to do all of these: - show less like this - show more like this - like a post

I may have found a bug with the Android app client in sendInteractions. It looks like Android app does not forward the "feedContext" from getFeedSkeleton response to sendInteractions request. It is empty. @hailey.at would you know who could help with that? Context below

Found and fixed a bug in For You. The intention is to count likes and reposts of the same post only once per user. From For You perspective we just care that user B endorsed item A. For example, if you like and repost something it should be treated the same if you just like or just repost.

A new For You experiment started. Now that we have "show more like this" and "show less like this" we can tell not only which For You posts people like but also what they don't like. This is helpful to identify some potential algorithm improvements.

I ran a little experiment on the For You feed for the past 6 days. The experimental change prioritizes posts that have more "paths" from you. A path is: 1. You liked post X 2. Someone else liked post Y 3. They also liked post Z today. More details in the quoted post. Results below 🧵

You can now use "show more like this" and "show less like this" in the For You feed. (this is being rolled out so you may need to update your Bluesky app) Here is how it works in For You 🧵

The term "collaborative filtering" sounds weird to me. Why is it "filtering"? Filtering is binary - something either passes the filter or does not. Filtering focuses on the negative part - getting rid of the bad stuff.

Created a "For You" feed that only shows video posts. The algorithm and the codebase is exactly the same as the regular For You. It simply filters out posts without videos when reading the Jetstream firehose. Could be useful for something like @skylight.social

Custom feeds stopped loading for me, including For You and SkyFeeds Discover I'm using web client.

Welcome to your ✨For You✨ feed This feed is powered by your likes! The more you like, the better your recommendations become. To get started, tap the ❤️ on a few posts in your Following or Discover feeds and come back here for more great content.

For You sometimes shows posts that were liked by the a single user that you happen to have to have a few likes in common. If that user happens to like a mix of different topics then you get a bunch of random stuff in For You. I have an idea of how to solve it. (1/?)

The For You feed will be down for 10-20 minutes. I need to install an additional SSD which will hold the Bluesky data. Wish me luck!

For You feed is a simple algorithm that has no AI or Machine Learning. It does have a few knobs, parameters, rules that affect what you see in For You. One rule prevents multiple posts in a row to be coming from likes of the same user. I would push down consecutive posts like to get...

Good to see progress on a Bluesky change to support "Show less like this" for custom feeds like For You: github.com/bluesky-soci... Hopefully it lands soon!

For You feed is #2 most shared feed according to blueskyfeeds.com/feeds/top-sh... This is impressive given the #1 feed is filling up the gap for a missing feature - bookmarks. Thanks everyone for helping spread the word!

Got a backup power source for the For You feed server - EcoFlow DELTA 3 Portable Power Station. Now if the power goes down at home the For You feed should continue to function.

Do not worry about sustainability of the For You feed. It is a hobby project that has barely any cost and I'm planning to run it long-term. It runs on my home PC, fronted by a tiny VPS ($10/month). With the current setup it should be able to run 30x the peak load (1 qps -> 30 qps).

It's great to see so many people trying out the For You feed in the last two days! Also I'm happy to see people recommending it to others.

Common feedback about the For You feed - make content fresher. The staleness could be solved if we record what posts the feed return in the past for the user and then exclude those posts in the future.

If you see content in For You that you would like to see less of, feel free to DM me a link to that post and I will add it manually to your "disliked" posts. "dislike" makes you see less content in For You from people who liked that post. This is a manual way to do "show less like this"

Welcome to the ✨For You✨ feed! It finds people who liked the same posts as you, and shows you what else they've liked recently. 📌 Pin to add it to your top bar ❤️ Like the feed and repost to spread the goodness

The interview had a discussion of points I haven't considered like the fact that publishers will force people to sign in before accessing content so they can avoid the cost of repeated verifications youtu.be/9A-dTz5FsXA

Hey 📚book lovers! Ever find a great post and wonder who else liked it, and what else they liked? Wonder no more! There is a feed for that now: bsky.app/profile/spac... 📌 Pin it to discover your next great read! #promoted

The For You feed finds people who liked the same posts as you and then shows you what else these "curators" liked. To keep the feed fast I limited the number of curators to top 20000. A week ago I made performance improvements which allows us to raise the limit to 30000.

I want to see Bluesky as a sustainable platform. This means they need to monetize it to be viable. I’m curious to see what ideas the team has in store. I also want to suggest an idea that they probably have not considered. 🧵

Do you want to discover great posts? Try this feed bsky.app/profile/spac... It finds people who liked the same posts you liked and shows you what else they liked recently. #promoted

To spread the goodness of the For You feed I will try out the @graze.social ad system. In a blog post they highlighted one ad campaign that brought in 50 signups for a free app for $100 spent: www.graze.social/blog/how-sil... Not sure what to make of it or how it would translate to feed pins.

The non-official Discover feed bsky.app/profile/skyf... built by @redsolver.dev has many similarities to the For You bsky.app/profile/spac... feed I’m building. Both feeds find people who liked the same posts as you (curators) and then show you what else they liked. Let's see how they differ 🧵

The For You feed was down... due to thunderstorms. That's one downside of running the feed generator from a home computer. Time to get a UPS for it.

Great point! We often forget that the like button is used based on the affordances it provides. It notifies the author - so people use it as a read receipt. And it increments a counter visible to others - so people use it to express support for opinions. Which means political opinions attract likes

Enjoying the For You feed? Give it a like ♡ to help more people discover it: bsky.app/profile/did:... The more people use it -> the more feedback we get -> the better we can make it for you.

The For You feed has felt a bit too static because older posts (posted 12-24 hours ago) tend to sit at the top as they have more time to accumulate more likes. To make it surface fresher content more often I applied an exponential time decay with a half-life of 6 hours.

It's interesting that other retailers match Amazon's sale prices for the same products. I wonder how they coordinate this. popular.info/p/prime-day-...

The 96GB of RAM went in use right away as filesystem cache: MiB Mem : 95663.4 total, 4044.2 free, 12029.5 used, 80722.6 buff/cache The SQLite database that stores the index of the last 60 days of likes is 200GB. Half of it now fits in cache.

Brought down the "For You" feed generator for a "quick" ram upgrade on my desktop. Getting 96GB of RAM installed (instead of 32) and the first boot is taking forever! Supposedly the computer is training the correct configuration to use with the new memory chips. So I hope it works in the end.

I accidentally broke the For You feed - no new items were being imported from the Bluesky firehose in the last 24 hours and as a result the For You feed was empty. The import is fixed now and new items are being recommended. Though the recommendations will be sparse for now and a bit random.

The "For You" feed generator is not available for maintenance.

As part of making the "For You" feed I needed a way to test it outside of the Bluesky app and I built a simple web UI - it allows you to plug any Bluesky user id and see the recommendations it would generate for that user. Example for @zachweinersmith.bsky.social : linklonk.com/bluesky?user...

Still so strange that no rec system ever asks whether you like the person it’s trying to shape you into

Made a ton of optimizations to the "For You" feed (bsky.app/profile/did:...) generation today. It should be on average ~50% faster than before. Let me know if you have issues with responsiveness so I could investigate your case.

Created "For You" feed which works like this: - we find all posts you liked - then we find other people who liked the same posts - then we find what else they liked to create the "for you" feed The aim of this algorithm is to be simple and transparent. bsky.app/profile/spac... Feedback welcome.

I want to create an "Others also liked" feed. I'm reading the data from the Bluesky Jetstream to record who liked what. 1 day of data is ~6GB in SQLite table - just user and post ids. It would require a beefy server to operate on a non-trivial history of posts.

Hope you don’t miss the weather forecasts

Will simply scaling up LLMs get us to AGI? My feature for Nature, w/ inputs from @fchollet.bsky.social @rao2z.bsky.social @yoshuabengio.bsky.social @melaniemitchell.bsky.social @dileeplearning.bsky.social @andrewgwils.bsky.social, Raia Hadsell, Keyon Vafa, Karl Friston www.nature.com/articles/d41...

It’s just ~content~ all the way down

Made the "Hot Links" feed much faster now (<0.6sec) by precalculating the list of the hot links. Also added a time based decay so older links get gradually lower in the feed. This is similar to the time based decay used on Reddit's Hot and on HackerNews.

Created a feed that shows posts with links. For now sorted by the like count. Will add time-based decay similar to Reddit/HackerNews. There are a bunch of duplicated links which differ in some url parameters. Will need to add some logic for deduplication. bsky.app/profile/did:...