TLDR Tech / Marketing
How Amazon Uses LLMs to Recommend Products
Amazon's COSMO system tackles a fundamental limitation in product search by teaching recommendation engines why people buy things, not just what they buy. The team fed millions of purchase behavior pairs into large language models, then built a multi-stage pipeline to filter out the roughly 65%-91% of outputs that were circular, generic, or trivially obvious. A smaller instruction-tuned model called COSMO-LM now generates fresh commonsense knowledge on the fly at a fraction of the original inference cost. A/B tests showed a 0.7% lift in product sales and 8% increase in navigation engagement on just 10% of US traffic. Amazon projects billions in potential revenue if extended across all traffic.
Read the full story