Post

I’m so out of touch with web technologies that I was about to make an incorrect joke about how out of touch I am. I thought “Gem” might be an underlying tech you use, and I was going to say “I’m a bit out of date, the only GEM I know is GEM Desktop, is that OK?”, but it’s just like a jobs platform?

Ok this is just gossiping at this point but...what happened to @why.bsky.team and why is the CTO delving into ML stuff, I'm confused 🤣

My background isn’t ML so I recently had to start training in my off time. It’s actually been very fun; wish I had done this 5 years ago. But that’s my opening caveat- I’m still relatively junior on this subject matter

The way Discover works is by sampling from multiple different “Candidate Groups (CGs).” This can be things like followed users, liked by followed users, social clusters, and topics you’ve voiced interest in.

These get put through a predictive ranker, which is a large statistical model that tries to guess whether the viewer would like a post. Model #1 for us to train.

Stop showing me unsolicited Bear content in Discover

Good thread, do the algos consider content you only view (but don’t engage with) as a weight? My grand theory for why social text feeds suck is: views for text based content is bad data. I have to read it to know what it is, so you can’t know if I just scrolled by without viewing it

This includes tapping into a post to read it, which you often need to do to read the whole post or thread to know it sucks

Paul I *did* put my resume in in like November and never got a call back 💀

I agree with you from personal experience Bluesky is passing up some of their most experienced candidates. It's probably a time management issue. Also they have no office, which makes coordination difficult. I'd have trouble hiring too in this current climate.

i've done a lot of hiring in large open source ecosystems. the trade-offs involved in hiring from within the ecosystem were discussed endlessly each time we did it. after many ecosystem hires over the years, i can't say that how things played out points to a clear conclusion. hard call 🤷🏽‍♂️

Super interesting thread! As far as generating candidate groups: topic similarity using an off the shelf embedding model with a fixed training cutoff means that you won't be able to topic cluster along newer slang, memes, etc, which is super important for social media.

You need to build a dynamic topic model, which you can do by dynamically identifying the clusters only, or by identifying clusters + updating embeddings, which I think is more powerful. Moreover, what can feel "magical" about a truly good recommender system is when it recommends something that...

I haven't interacted with bit which I really want to see. Usually this is done by combining the two approaches for candidate groups you talked about: it's not follow based OR topic, it's _both_. To do this what you want to do is create a semantic clustering algorithm for _users_.

i have 0 knowledge or skills or expertise in the area but i do have a good heart

That’s good though

Wait, is BlueSky doing algorithms now? I kinda hope not.

There has been an official algorithmic "Discover" feed since I think autumn 2023?

Thanks for sharing and love the openness. I’m in a similar boat where ML is concerned, but happy to chat any time about my 3 years working on News Feed and the perspectives that gave me rubbing shoulders with the folks building that stuff on feed, and what we generally found was “hard”

I’m confused. I was invited to apply to this position by your recruiting team, and then summarily rejected with the excuse that you had filled your pipeline. And yet here you are trying to fill it again. I am a 20 yr career engineer w/ academic research exp in SNA and graph theory. What gives?

i’m hoping this is just an oversight in their hiring process, mistakes definitely happen. you’re an incredible engineer & bluesky would be really fortunate to have you.

It’s more likely because I have a reputation for being an agitator in west coast tech at this point because i lead with my ethics instead of quarterly goals and user exploitation.