Meta Adaptive Ranking Model: Bending the Inference Scaling Curve to Serve LLM-Scale Models for Ads

Meta developed Bayesian Optimization for Concrete (BOxCrete), an AI model to design high-quality and sustainably produced concrete mixes in the US. BOxCrete allows rapid exploration and validation of new concrete formulations, reducing the need for months-long lab testing and empowering US-based concrete producers to incorporate more locally sourced materials. Collaborations with Amrize and the University of Illinois at Urbana-Champaign have showcased the model's effectiveness in creating stronger and faster-curing concrete mixes for large-scale data centers and infrastructure projects.

Here's a 3-sentence summary of the engineering blog post: Meta's Friend Bubbles feature on Facebook Reels enhances social discovery by highlighting Reels liked or reacted to by friends, making it easier to connect over shared interests. To achieve this, the system combines social and interest signals to rank friend-interacted content and estimate relationship strength using machine learning models, emphasizing connection quality over quantity. The friend bubbles recommendation system includes components like Viewer-Friend Closeness and Video Relevance, which work together to surface relevant, high-quality content and improve user engagement.