X Algorithm Update: Next-Gen Phoenix Pipeline, GPU Scorer Runtime, and Abuse Enforcement Integration

The latest open-source release integrates the next-generation 'Phoenix' candidate pipeline and scorer with a specialized GPU scoring runtime (BDSM). Key updates also introduce post-engager feature hydration, a zero-favorite codivert filter, and an overhaul of confidence-based abuse enforcement controls.

Simple Explanation

X is moving towards a modern, GPU-accelerated scoring architecture dubbed 'Phoenix', allowing faster and more sophisticated neural ranking. The feed now hydrates candidate tweets with real-time data on who engaged with them and aggressively filters out zero-engagement content using a specialized 'Zero Fav' mechanism, while tightening automated anti-abuse controls.

Key Points

• 1Integration of next-generation 'Phoenix' candidate pipeline and GPU scoring runtime
• 2Real-time feature hydration capturing early post engagers to refine relevance
• 3Implementation of 'Zero Fav Codivert Filter' targeting unengaged candidate content
• 4Fine-grained action control and confidence-driven rollback in the abuse enforcement service

Technical Details

The commit introduces `phoenix_candidate_pipeline.rs` and `phoenix_scorer.rs` into Home Mixer, interfacing with the Python-based BDSM GPU runtime (`gpu_scorer.py`, `batch_prefetcher.py`) for low-latency batch model inference. Feature engineering is expanded with `post_engagers_hydrator.rs` to leverage graphs of recent engagers. Additionally, `zero_fav_codivert_filter.rs` implements traffic and routing control for posts lacking initial engagement, while `abuse-enforcement-service` gains granular action control levels, confidence bounds, and rollback pathways.

Affected Users

All X users browsing the 'For You' feed, content creators, and accounts utilizing automated or high-volume publishing

Before & After

Previously, candidate scoring relied on legacy ML pipelines with more static hydration; now, Home Mixer utilizes high-throughput GPU batch inference via Phoenix, paired with dynamic early-engager hydration and stricter automated abuse boundaries.

Action Items

• Optimize posting schedules to capture immediate early engagement from core audiences
• Focus on high-engagement hooks to prevent posts from triggering the 'Zero Fav Codivert' low-engagement pruning
• Avoid aggressive auto-posting and low-quality external links to steer clear of automated abuse confidence thresholds

Monetization Tips

• Because 'Post Engagers' are hydrated into candidate scoring, interactions from reputable, niche-relevant users heavily boost distribution.
• Prevent initial dead-zones: release high-value content when your most active subscribers are online to avoid being suppressed by zero-engagement filters.

FAQ

Tags

Changed Files