Engineering at Meta
14 Intel
Is this you? Sign in with X to claim this profile.
Intel
Meta reports moving Catching JiTTests, LLM-generated tests produced at diff time to catch bugs before code lands, from research prototype into production infrastructure, with a human-on-the-loop workflow and a companion paper on the deployment.
Meta's first in-house training chip treats interconnect as a design primitive: built-in NIC chiplets plus HCCL, a communication library co-designed with the hardware. The target is embedding-dominated recommendation training, where collectives — not FLOPs — are the bottleneck.
Meta describes its thin orchestrator pattern for data-pipeline agents: a small router handles sequencing and reliability checks while 14 specialist agents do domain work, after a single all-knowing agent began hallucinating in production.
Meta details Prometheus, its gigawatt-scale Ohio supercluster used to train the Muse models. Covers phased buildout toward roughly one million GPUs, closed-loop liquid cooling, and reliability work that sustains over 85% training goodput.
Meta describes how sequence learning became a scalable ranking platform: a multi-stage design that moves heavy user modeling offline and keeps online ranking cheap, plus dense tokenization with target-aware attention. Reported lifts: 6% conversions on Instagram, 3% on Facebook.
Meta engineers explain how a small team built the Forum app with an agent running full plan-build-test-correct loops. Covers codebase and test adaptations, the human-on-the-loop model, and measuring verified, review-surviving changes.
AI compute performance roughly triples every two years; storage and interconnect do not. Meta walks through how its exabyte-scale BLOB storage architecture was reshaped around two goals — keeping GPUs fed and keeping geo-distributed researchers from waiting on data movement.
Meta's ads foundation model now trains on thousands of GPUs at LLM scale, and getting there required rebuilding the stack: recommendation-specific attention kernels, mixed ultra-low-precision MXFP8 training, and 5D parallelism. End-to-end efficiency doubled to 20-25% MFU while FLOPs grew 4x.
Meta details F3 Lightning, which lets ads ML features compile, deploy and roll back independently while preserving shared computation, reaching production about 3.6x faster with no added infrastructure compute cost.
Meta details a domain-expert agent built on two layers: an auditable knowledge store kept separate from the reasoning layer, and a feedback loop that turns subject-matter-expert corrections into verified, regression-tested knowledge updates with no retraining. Aimed at compliance-style domains gover
Meta describes an opportunistic inference system that scales ads ranking breadth and model strength with spare accelerator capacity, using isolation to keep baseline serving safe when capacity is reclaimed.
Meta explains Private Processing, its confidential-VM infrastructure that lets AI Glasses offload heavier AI models to the cloud while keeping requests cryptographically attested and hidden from Meta itself.
WhatsApp is beta-testing Scam Alert, an opt-in on-device model that flags likely scam messages without any content leaving the phone. Meta published the design and verifiability guarantees ahead of general rollout, explicitly soliciting scrutiny from bug-bounty and security researchers.
Meta open-sourced MetaRoCE through the Open Compute Project: a purpose-built RDMA transport for AI traffic on commodity Ethernet, shipped with a reference implementation and compliance tests. It targets the collective-communication stalls that strand compute at hundred-thousand-GPU scale.