Start from Real User Jobs: Design Around Core Human Moments

Effective product design begins by observing how people naturally use the platform and defining core "jobs to be done" around human behaviors rather than technical features. Users arrive with specific, often unspoken goals—to "join" a community (for a new hobby), to "ask" a question (to get instructional advice), or simply to "watch" content (for entertainment or information). Strategic focus is placed on users who traditional systems often underserve, such as Cold Start (low signals) users and Marginal users (low engagement frequency). For many marginal users, community products such as Groups and Reddit are a major—sometimes the only—use case; the objective is to identify their specific needs (commerce, niche interests, lifestyle) and show tangible value that helps them graduate to higher-loyalty users. Solutions are engineered to address these moments directly. For instance, Group recommendations are further tuned beyond power-user optimization to surface active local-utility communities and connection-driven groups for marginal cohorts, so the system serves real human needs.

Treat Ranking as a Product, Not Infrastructure

The ranking system is the primary product interface. Understand what content delivers the highest value for users and the business, then build objectives and feedback loops around that. Users frequently experience low diversity and perceive some recommendations as too broad; quality is characterized by relevance, diversity, and usefulness. To quantify subjective quality, we deploy foundational LLM-powered signals—for example, classifying content as utilitarian, and analyzing comments for relevance and focus. These signals improve retrieval for Search and Feed before ranking begins. Objective functions then optimize weighted outcomes (a weighted sum of user and community actions, with weights set by association with future sessions), and success is confirmed by long-term metrics—engagement that lifts graduation from low to high activity, not just one-off CTR.

Balance Friends, Creators, and AI

The key is to safeguard Friends and Family while integrating creator and generative content in proportions that enhance discovery but don't swamp or steal from it.

Most social platforms are evolving into discovery engines for unconnected content that grow interest-based exploration while still exploiting known interests. Cannibalization is managed explicitly, and signals are bridged across surfaces in three steps: close signal gaps, apply signals in retrieval and ranking, and improve responsiveness so new interests appear quickly. During the Reels scale-up, portfolio targets preserved Friends and Family visibility while unconnected recommendations expanded. Signal bridging allowed a group to join to gently steer Reels toward adjacent interests without flooding the feed, preserving core social value while broadening discovery.

Design for Uneven Ecosystems: Speed and Trust

AI/ML systems must deliver fast, trustworthy experiences across diverse conditions—network quality, device constraints, and digital literacy. A persistent challenge is the lag between a new signal (for example, joining a group) and the system surfacing relevant content; slow adaptation erodes trust. We prioritize real-time signals from current/recent sessions and use decay counters so fresh interests appear in hours, not days. For search, we use hybrid retrieval that combines keyword matching with embeddings and advanced language models, enabling relevance even for vague queries (for example, mapping "small individual cakes with frosting" to "cupcakes"). The result is responsiveness that users can feel and a system they can rely on.

Make Experimentation Safe at Scale

With billions of users, a 0.1% negative change can meaningfully damage trust and business. We measure rigorously across interconnected surfaces and focus on the ROI of cross-surface signal bridging—not just time-spent spikes, but graduation to higher-engagement states. Operational guardrails include forensic analysis to deepen the understanding of content quality gaps, routine content reviews, automated quality monitors, and kill switches / auto-rollback. For evaluation at scale, LLM-as-a-Judge (increasingly multimodal) provides stable, reliable quality signals where manual labeling is variable and expensive. The art is empathy for high-leverage human moments; the science is generative AI plus scalable ranking systems. Together, they turn a massive social network from a static directory into a responsive discovery engine.