Hello! Project's super rookie group Rosy Chronicle is about to make their official debut! 🌟✨ Come check it out and feel their charm! 💖🎉 #RosyChronicle# #HelloProject# #JPOP# #IdolDebut# #JapaneseIdols# #MusicRecommendation# @Rosy_Chronicle

Modal is the AI-native cloud provider that just crossed a $1B revenue run rate and raised $750M at $15.75B. Back in 2024, CEO @bernhardsson revealed how they threw out traditional infrastructure and built @modal from scratch for AI: “I spent seven years at Spotify building the music recommendation system. And I realized 90% of the work is building infrastructure, not the application. So what’s broken? First, traditional infrastructure isn’t built for fast feedback loops. When you’re building AI applications, it’s annoying to test locally and then deploy to production. But you can split it into two environments. A local environment you test in. And a cloud environment you deploy into. But then you replace it with a different problem. Going from one environment to the other. This isn’t as much of a problem in traditional application development. A lot of infrastructure is pretty good at mocking the cloud when you develop locally. But it gets really hard when you have GPUs and very large models. The initialization step is very complex. You need to load all these models into GPUs. Originally, I thought about it as developer experience and fast feedback loops. I wanted to write code, hit enter, and run it in the cloud immediately. It turns out that’s also really good for running applications in the cloud. It lets you scale up and down very fast. Then third, there’s something you have in AI that you don’t have in backend land. GPU capacity problems. Now you have to run multi-region. You need to aggregate capacity from all these different places. And no region and no cloud provider has enough capacity to handle this dynamic, flexible scaling. And most of our workloads are inference. So a lot of things break with traditional backend infrastructure. That’s why we decided to start over and build our own.” Full episode here:

What if you could leave your recommendation retriever completely untouched and just bolt on one lightweight "verifier" to significantly boost recall? A team at Meta just showed exactly that. Title: Recommendation Retrievers Need Verifiers: Universal Generative Reranking for Sequential Recommendations URL: The idea is a "drafter-verifier" pipeline: freeze the existing retrieval model (the drafter) entirely, and add a lightweight verifier that generatively scores candidate items via their identifier tokens. 🌐 Highlight 1: It scales with catalog size On the YaMBDa music recommendation dataset, gains grew larger as the catalog got bigger — at the largest scale (4.65B interactions), Recall@10 improved by up to 29.8%. 🧩 Highlight 2: Works across wildly different drafters The same training recipe worked on 4 architecturally distinct drafters: attention-based SASRec, recurrent GRU4Rec, convolutional NextItNet, and even the LLM-based MiniOneRec. 🔬 Highlight 3: The gain really comes from generative verification Injecting content embeddings directly into the drafter actually hurt standalone performance, but adding the verifier on top recovered it — proving the improvement comes from the output-side generative verification process itself, not simple feature injection. The fact that it bolts on without touching the retrieval index feels like a big deal for adopting this in large-scale production. #RecSys# #MachineLearning#