The system uses a Resource Core Model with 6B sparse parameters to predict marginal utility, directing compute to high-impact requests across a heterogeneous fleet.
The system uses LLMs to validate user-stated interests and apply them responsively to ranking and delivery, balancing new content visibility with existing engagement.
The Triton Low-Level Extensions (TLX) framework enables explicit warp-group orchestration, delivering up to 2x forward speedup in GDPA kernels and reducing RecSys inference latency by up to two-thirds.
The system preserves F3's shared computation efficiency while allowing independent feature validation and rollback.
The system achieved a 60% true-positive rate and engaged 589 engineers through a human-on-the-loop workflow.
The new architecture uses a Mixture-of-Experts model and fused GPU kernels to deliver near-ranking-level performance at 300x fewer FLOPs than the 2024 system.
The library separates system integration from privacy modules, reducing new requirement implementation from months to days while supporting AI agent controls.
The announcement covers new hardware for the Prometheus cluster, including a petabit-scale BAG super-spine and the Tahan Ethernet scale-up switch for MTIA 300.
The system uses stable semantic tokens to improve cold-start generalization and cross-campaign transfer for advertisers.
The 1,300-line router delegates to 14 specialized agents, reducing manual configuration prep from days to under an hour.
The New Albany facility supports Muse model training with over 90% RL goodput and aims for a 100k GPU-hour mean time between interruptions.
Meta engineers achieved core feature parity with Facebook Groups in eight weeks by treating AI agents as autonomous participants in the development cycle.
The new model improves conversion rates by 18.9% and click-through rates by 12% across Meta's advertising surfaces.
Meta introduced the IcePack liquid-cooled rack with 102.4T IceCube switches and open-sourced its MetaRoCE transport protocol.
Meta aggregates fragmented idle inference capacity into a shared pool to enhance ad relevance without requiring new hardware.
Meta has developed a structured knowledge-based AI agent that incorporates expert feedback without model retraining.
Meta developed a sequence learning platform that separates offline user modeling from online ranking.
Meta doubled the training efficiency of its ads recommendation foundation model GEM while scaling training compute 4x.
Meta leveraged the sched_ext framework to build a kernel scheduler optimized for ads workloads, reducing latency by 28% and improving power efficiency.
Meta is marking its 10th year of sponsoring the Python Software Foundation (PSF), continuing its technical investment to ensure the sustainability of the Python ecosystem.
Meta operates a hybrid asset classification system combining LLMs and deterministic rules to address the complex data privacy challenges of AI-native environments.
SilverTorch is a new paradigm that unifies microservice-based recommendation systems into a single neural network, maximizing retrieval efficiency and accuracy.
Meta released a Reels podcast covering the development process of Friend Bubbles.
Facebook Groups search quality was improved through hybrid retrieval and automated evaluation.
AI Briefing