360Brew Decoder-Only Foundation Model for Personalized Ranking and Reco (web.archive.org)

The 360Brew model, developed by Foundation AI Technologies at LinkedIn, is a new decoder-only foundation model designed to enhance personalized ranking and recommendation systems across the platform. This model, featuring 150 billion parameters, consolidates numerous predictive tasks into a single architecture that can efficiently address over 30 tasks without the need for task-specific fine-tuning. Its strengths lie in leveraging natural language interfaces for task definitions and simplifying the complex feature engineering that typically burdens traditional recommendation systems. By facilitating rapid iteration on new objectives and data sources, 360Brew aims to revolutionize the AI development process at LinkedIn. This development is significant for the AI/ML community as it showcases the growing capabilities of large foundation models in domains beyond natural language processing, particularly in recommendation systems. The ability of 360Brew to generalize across different tasks and efficiently handle cold-start problems enhances its effectiveness, challenging the reliance on extensive handcrafted features that characterize conventional models. With its centralized prompt engineering approach, 360Brew not only reduces technical debt associated with multiple specialized models but also empowers product designers to innovate more freely. This transition towards a more holistic, scalable AI architecture at LinkedIn could inspire similar advancements across various online platforms.