+100
04.01.2025 19:28
👍 0
🔁 0
💬 0
📌 0
+100
When I first learned about RAG, I couldn’t believe it was just “ “.join(docs)
It does feel like
RAG -> Variable Selection
CoT -> Boosting
sure there are others too…
It also - for me at least - takes the joy out of it. Science is a participatory discipline and “playing the game” makes “scoring the points” all the more enticing (especially for kids!)
What app are folks using for e-readers these days? I’m a physical book + digital notes guy but the confines of my Manhattan apartment is causing me to start buying digital - any recommendations?
I get mad about this once a week. Just terrible brand positioning
Wait, no he’s right
Oh no - didn’t mean to scoop any takes from the bad place 😅
The most impactful CRM personalization is
select user, song, count(*)
from plays
group by 1, 2
order by 3 desc
Always has been. Always will be.
I’ve been implementing some time series type estimates at work this week and man oh man the indexing / naming conventions have me 😵💫😵💫😵💫
Got an email, "Thank you for coverage․py." I immediately thought, "they're going to want something, or complain about something."
Nope! It was all positive. Ended with "Thanks again for this awesome Python contribution!"
If you like what someone has done, tell them!
Getting a new laptop Monday - so excited to set up everything “right” this time only for it to explode immediately.
Yeah something beautiful about a NYC bar on Saturday morning with first round coffees, second round are beers
It’s also a religious tradition at this point
I was agonizing over why our SPR performance took a nose dive last week. Spent times in the logs, staring at feature values, investigating the FPR.
The front end was using rankings[::-1] not rankings 🙃 Sometimes the 10x work is not trusting a thing that comes out of the API.
One of my big ah-ha moments was “wait… the hidden layers are just doing feature engineering”.
😅😬
I forgot how nice it is to open an app and be flooded with cool ML/Stats-y stuff. Bluesky has big happy hour at a conference vibes and it’s 👌
A meme from the Simpsons where the top text is "Don't make me tap the sign" and the bottom text is him tapping a sign that reads "Don't peak at XP results early"
I regularly send marketers this picture which I feel like is the right level of passive-aggressive grump for this group
🎯🎯🎯
I made one for stats papers
We really have known the ramifications of RecSys tech (especially in news) all along. It's just that consumers and companies wanted it more than a non-personalized alternative.
What we want vs what is good for us can oftentimes be two different things.
Quote was from 30 years ago...
With GroupLens, users may choose to read articles only from a small group with whom they share many common interests. Over time this could lead to a fracture of the global village into many small tribes, each forming a ... community ... isolated from each other.
Forget ggplot vs base R graphics. Bring back terminal renderings and clip art.
Picture of an article on an online bulletin in 1994. Can only be described as ~old~
Picture of a systems architecture from 1994. Can only be described as comically dated.
I’m revisiting some original RecSys papers in prep for a new team-member joining and the figures in the original GroupLens paper are just so 🔥
dl.acm.org/doi/10.1145/...
Network growth models that accommodate starter parks (e.g. vertex —> block edge expansions) is already so fascinating me.
Thanks for sharing!
The more I read post-PhD on the history of statistics (i.e. Stigler’s works) the more dark stuff I find. I thinks it’s really concerning that one can attain a formal education in statistics and just never discuss these issues. Context as important as content.
Hey folks👋
My name is Ben and I’m a Data Scientist at Uber working on RecSys and CausalML. I’m a proud statistician with interests in matrix analysis and graph inference.
I live in NYC where I enjoy running long distances and reading a bit of everything. Find me here dravesb.github.io!
