LinkedIn cuts AI job-search model training from days to hours
LinkedIn re‑engineered its job‑search ranking pipeline, moving from two‑day single‑node training to a distributed multi‑teacher distillation system that finishes in under five hours—a roughly eight‑fold speedup. By caching teacher outputs per shard and coordinating relevance, engagement, and embedding teachers across heterogeneous GPU clusters, they now iterate faster and serve a smaller student model at low latency.