Meta Feeds Public Instagram Posts to LLMs to Improve Recommendation Algorithms
Meta CEO Mark Zuckerberg revealed that Instagram feeds all public posts and Reels into large language models (LLMs) to analyze their themes and tones. The resulting insights are then used to train and refine the platform's recommendation systems, contributing to a 10% year-over-year increase in user time spent on the app. This represents a major real-world application of generative AI in social media, demonstrating how LLMs can be utilized to enhance user engagement through highly personalized content recommendations. However, it also highlights the ongoing tension between algorithm optimization and regulatory scrutiny regarding social media addiction among teens. The system upgrade, which includes faster AI inference and a new architecture for Reels, led to a 15 basis point increase in Instagram session volume. Meanwhile, Meta is facing lawsuits from over 20 states regarding addictive product designs and has set aside $2.4 billion for legal expenses.
## BACKGROUND
Traditional recommendation systems rely on user history and collaborative filtering to suggest content. Integrating Large Language Models (LLMs) allows systems to better understand the semantic meaning, context, and sentiment of content, leading to more accurate and personalized recommendations.