Meta engineering blog

Meta explains how Private Processing extends confidential computing to AI glasses, using trusted execution environments, remote attestation, anonymous routing, and encrypted in-boundary storage. The design addresses stateful AI, non-targetability, operational observability, and independent verification while keeping personal context inaccessible to Meta.

Meta open-sources Rebalancer, a high-performance library for modeling and solving large-scale assignment problems. The post explains its specification language, expression-graph architecture, MIP and local-search solvers, debugging tools, and production performance across millions of daily infrastructure optimization workloads.

Meta details Petal, a planned 7,000-kilometer transatlantic subsea cable designed to deliver 1 Pbps using 2-core fiber at scale. The post explains the fiber, repeater, Fan-In/Fan-Out, power, and manufacturing innovations that double capacity without proportionally increasing infrastructure or energy requirements.

Meta explains why it placed ZGateway in front of ZippyDB to replace a fragile client-to-database connection mesh with a bounded, centrally managed proxy tier. The post details connection fan-in reduction, cross-client batching and coalescing, admission control, caching, load balancing, and cross-region resilience.

Meta describes an AI “second brain” that combines structured, auditable knowledge files with composable reasoning recipes and a human-guided self-improvement loop. Expert feedback becomes minimal, regression-tested edits without model retraining, preserving institutional expertise while reducing assessment time and preventing regressions.

Meta explains how MTIA 300 co-designs compute and communication for recommendation-model training. Its integrated NIC chiplets, dedicated message engines, near-memory reduction hardware, and compiled HCCL collectives deliver up to 940 GB/s of in-rack communication while reducing concurrent GEMM degradation to under 0.5%.

Meta introduces MetaRoCE, an open RDMA transport designed for AI-scale Ethernet. The protocol moves congestion and ordering intelligence to NIC endpoints, enabling out-of-order delivery, native multipathing, loss tolerance, and graceful recovery, with benchmark results showing improved throughput and resilience over RoCEv2.

WhatsApp is evolving its anti-scam protections to help keep people safe while preserving the privacy of end-to-end encrypted messages. The post says Meta is sharing an early look at Scam Alert and how it aims to provide safety improvements with verifiability guarantees.

Every day, Meta’s recommendation platforms handle billions of user interactions, generating rich temporal signals that capture individual preferences and intent across products, ads, and content. The post discusses how Meta uses sequence learning and a multi-stage architecture to improve ads ranking at scale.

Meta’s Generative Ads Recommendation Model (GEM), the foundation model behind ads recommendations across Instagram and Facebook, now trains at LLM scale on several thousand of the latest-generation GPUs. The post explains how Meta doubled end-to-end training efficiency to 20–25% Model FLOPs Utilization while scaling training FLOPs 4x.

Hierarchical Interest Representation is a research area for Meta Ads focused on an upstream representation layer across users, advertisers, products, and services. It aims to learn unified embeddings that connect inferred user interests with what advertisers offer in deep funnel ads.

At Meta’s scale, even small latency regressions can hurt ads performance. The team used sched_ext, an upstream BPF-based extensible scheduling framework, to create a scheduling policy tailored to Ads delivery after a Linux kernel upgrade introduced latency risk.

Over the past several years, model capabilities and training dataset sizes have grown rapidly, and Meta frames storage as a critical part of keeping AI innovation fast and cost-effective. The post appears to discuss the storage architecture needed to support frontier AI workloads at scale.

Meta highlights its 10th consecutive year sponsoring the Python Software Foundation and its long-term support for the Python ecosystem. The post emphasizes Python’s importance across Meta’s engineering stack and its role in the company’s open-source efforts.

Privacy controls such as retention, access, purpose limitation, downstream sharing, and anonymization depend on correctly understanding what data they are acting on. This post uses an asset classification case study to show how Meta approaches that problem in AI-native infrastructure.

Smart glasses like Ray-Ban Meta and Oakley Meta Vanguards need batteries that can power cameras, speakers, AI workloads, and even a display while fitting into the temple arms. This post explores how Meta engineered ultra-narrow batteries to meet those constraints.

Meta shares the multi-year effort behind adopting AV1 for real-time communication, including codec selection, device eligibility, rate control, and error resilience. The post covers the technical and operational challenges of deploying AV1 at scale and the improvements made to call quality.

We’re introducing Instantaneous PowerLoss Storm, a new testing paradigm within Meta’s infrastructure for handling and mitigating instant or zero-notice power loss in our data centers. The post explains how Meta built readiness to tolerate instant failures using defense-in-depth strategies, the tradeoffs involved, and how the team validated that readiness.

We’re introducing SilverTorch, a reimagining of recommendation systems that unifies all retrieval components for user generated content under a unified architecture. It reports up to 23.7x higher throughput and 20.9x better compute cost efficiency than a CPU-based solution while also improving accuracy.

On its face the new Friend Bubbles feature looks simple enough: it highlights Reels your friends have watched and reacted to. The article frames the feature as deceptively straightforward and emphasizes the deep engineering work required to make social discovery scale.