It’s not Friday…

This is huge. Amazing work by @mmccue.bsky.social and the Surf team, and also to Vox for making some really bold moves embracing the open social Internet. And, look how good it looks!

teaser image for Dune 3, except it's called Doug, and Paul and Chani are replaced with Doug and Patti

Wikipedia screenshot of https://en.wikipedia.org/wiki/Dust_II

Happy 25 years of dust2

Cyberspaces you could navigate as well as any place you’ve ever lived.

Personalization at Bluesky | Gaurav Chakravorty If you’re interested in building personalization systems where architecture still matters, this is worth a read. Ian Wesley-Smith — someone I hired at Discord and worked closely with — just published ...

Posted here as well. www.linkedin.com/posts/gaurav...
I hope this gets the visibility to the the right audience

Also thanks to @recsysml.bsky.social for proofreading and providing me a platform to post. Follow them for more great recsys content.

A picture of the bluesky discover feed from the Apple App Store, showing a post by Jerry, and more importantly, a very cute corgi.

you and my dog are equally famous

Bluesky Jobs Bluesky Jobs

If this post interests you come work with me!
jobs.gem.com/bluesky/am9i...

Personalization at Bluesky The past, present, and future of personalization of the Discover feed

I made a post about personalization at Bluesky.

open.substack.com/pub/recsysml...

Bluesky Jobs Bluesky Jobs

jobs.gem.com/bluesky/am9i...

It’s impossible to watch any political thriller now because whatever incident is generating dramatic tension pales in comparison to reality.

Sign in PRO for presumed liability. app.leg.wa.gov/csi/House?se...

its on the roadmap!

A picture of house minority leader Hakeem Jeffries.

Every time I get in a Lyft they cal me Lan. Serifs could fix this very specific me problem.

Op-Ed: It’s Time to Imagine a Safer, More Connected Rainier Avenue » The Urbanist # Rainier Avenue has too many cars traveling at dangerous speeds, but conventional planning practices make it hard to change that and design a future where Rainier Avenue thrives. Let's stop making ha...

If we want to create a more affordable, livable Seattle, we need to design for *reducing* car volumes. This is the only way to meet our climate, public health, vision zero, accessibility and housing goals. Our fight to design highway on-ramps with LESS capacity epitomizes why it's still hard.

Ranking Models explained: Deep Dive into RecSys Architecture (Features, Embeddings, & Attention) Watch now (18 mins) | What happens post retrieval in a recommender system

In this three part series I compare #LLM and #RecommenderSystems to show the gaps and opportunities. They are surprisingly fewer than one would think.

Part 1 open.substack.com/pub/recsysml...

GitHub - vec2text/vec2text: utilities for decoding deep representations (like sentence embeddings) back to text utilities for decoding deep representations (like sentence embeddings) back to text - vec2text/vec2text

This project is absolutely nuts. You can take the vector embeddings for text and convert them back to the text that generated them.

github.com/vec2text/vec...

hah maybe. I feel like I would start with TransAct, and maybe PinSage to cover the world of recs at Pintrest

also a bunch of practical stuff on how they serve these long user histories. Clearly a lot of work went into this, and a lot of goodies in this paper.

- ranking model that tried to predict the next user action in a sequence. Use a cross-entropy loss to predict labels/actions on the sample and a next action loss to predict the next action the user will take.
- impression based negative sampling doubled some offline metrics

TransAct V2: Lifelong User Action Sequence Modeling on Pinterest Recommendation Modeling user action sequences has become a popular focus in industrial recommendation system research, particularly for Click-Through Rate (CTR) prediction tasks. However, industry-scale CTR models o...

Read Xia et al., “TransAct V2.” earlier today, some notes:
- long and short term user histories (long term [10k items], short term [100], impressions [100])
- sampled soft max loss function seems useful in environments with very weak negatives

#recsys #ml
arxiv.org/abs/2506.02267

A corgi lying on the ground facing the camera. A red ball is sitting next to his snoot.

Atticus wants to play

but some more than others

you most of all

watching Alien: Earth and I think I'm rooting for the alien

todays discover outage was brought to you by the letter <SPACE>

anyone got opinions on data orchestration tools? Specifically Dagster vs Temporal?

This person has clearly never used protobufs, especially in python.