Intelligent Delivery: Making Every Unit of Compute Serve Better Ads · Engineering at Meta
Science, Technology & Innovation · Nov 11, 2026
Meta reports that ML-driven, request-level compute allocation improved ads-ranking performance by about 10% beyond model and serving-system advances from 2017 through the first half of 2026 by optimizing marginal utility rather than average efficiency.
Intelligent Delivery: Making Every Unit of Compute Serve Better Ads · Engineering at Meta
Science, Technology & Innovation · Nov 11, 2026
Meta’s shared deep-learning Resource Core Model improved offline identification of high-impact ad requests by about 20% by predicting which requests merit extra compute and how resource settings affect outcomes, latency, and serving cost.
Intelligent Delivery: Making Every Unit of Compute Serve Better Ads · Engineering at Meta
Science, Technology & Innovation · Nov 11, 2026
Meta reports using AI agents to analyze serving systems, propose and test targeted optimizations, and reduce optimization cycles from months to weeks while retaining engineer approval before deployment.