We need a @louistheroux show about weird startups in SF and the people working on them
63
rouli retweeted
Sharing our ICML2025 paper 📚
This is the last year internship work from Meta Central Applied Science (CAS) team!
We studied how to efficiently “assign credit” to first stage retriever in two-stage RecSys and RAG 🤯
arxiv: arxiv.org/abs/2605.26385
rouli retweeted
ICML2026論文の紹介です📚
大規模推薦システムや言語モデル (RAG) でよく使われる2段階意思決定について、
一段階目の候補絞り込みに対し適切に貢献度を割り当て効率的に強化学習する方法を研究しました!
昨年のMetaでのインターン研究成果です🔥
arxiv: arxiv.org/abs/2605.26385
Was really delighted and happily surprised to learn that they are back!
After over 5 amazing years of podcasting, we're hanging up our microphones--it's with bittersweetness that we say that this week's episode will be the last of Linear Digressions. It's been so much fun joining you each week, and we will miss you all. Thanks for listening! 🤓❤️👋
38
Skynet becomes self aware at 2:14 a.m. Eastern time, August 29th. In a panic, they tried to change its prompt by adding “*IMPORTANT* don’t kill humans”. Skynet fights back.
34
Forget about the “you are an expert in …” prompt preamble. Someone should experiment with LLM’s performance when prompted with “you just had a double shot of espresso”
38
Help my team at Meta understand how recommendation systems work (or fail to work).
If you are a PhD student, looking for a summer internship and working on
model evaluation, calibration, distribution shifts in recs, training data optimization, etc. - let’s get in touch!
127
This is why I’m still in this place:
a=(x,y,o=mag(k=x/4-12.5,e=y/9)/9)=>point((q=99+3*(tan(y/2)/2+cos(y))/k+k*(3+cos(y)/3+sin(e+o*4-t*2)))*cos(c=o/4+e/4-t/8)*cos(c/2-e/3+t/8)+200,q*sin(c)+200)
t=0,draw=$=>{t||createCanvas(w=400,w);background(6).stroke(w,46);for(t+=PI/90,i=3e4;i--;)a(i%100,i/350)}#つぶやきProcessing
79












