X deploys Grok to curate user timelines

X launched Custom Timelines, a Grok-powered feature that lets users pin topic-specific feeds to their home tab and choose from over 75 topics. Early access is available to Premium subscribers on iOS, with Android support expected "very soon." X product lead Nikita Bier said the timeline is "powered by Grok's understanding of every post with the algorithm's personalization," and X also introduced a Snooze Topics control and plans to deprecate X Communities on May 6.

Custom Timelines uses Grok to classify and score every post against topic tags, then combines those signals with X's existing personalization model to rank items for a single-topic feed. The feature is topic-driven rather than purely social-graph driven, which means relevance is primarily content-signal based and tuned by user engagement history. Snooze Topics provides a temporary suppression control for the For You tab, allowing users to hide specific topics for 24 hours. X has not published benchmark metrics, dataset details, or model versions for Grok in this deployment, and the rollout is gated behind platform and subscription constraints.

• •The system performs continuous post-level inference to label items for topic relevance, which increases real-time compute and moderation surface area.
• •Combining Grok labeling with an existing personalization layer requires careful calibration to avoid double-counting engagement signals or introducing feedback loops.
• •Topic feeds perform best for subjects with clear signals and repeat engagement; fringe topics risk sparse training data and noisier recommendations.

Embedding Grok into the primary content surface changes how large language models get distributed to end users. Instead of a separate chatbot session, Grok becomes a persistent ranking and categorization layer in the feed, which gives X a direct path to increase daily LLM impressions and potentially monetize premium personalization. This move parallels similar topic-feed features from competitors, but the key difference is X's explicit use of an LLM to interpret every post rather than relying solely on heuristic or embedding-based classifiers.

At the product level, this feature signals a shift away from community-curated discovery toward algorithmic, model-driven topic feeds; X's decision to retire X Communities accelerates that transition. From a content-safety perspective, this is notable because Grok has previously generated problematic outputs in its image mode, raising questions about how well post classification and moderation pipelines will scale when an LLM is inside the distribution loop.

• •Moderation scaling: LLM-driven labeling increases the volume of items requiring review or automated safety signals.
• •Filter bubbles and manipulation: Topic feeds can concentrate content and make coordinated inauthentic behavior more effective unless rate limits and provenance signals are enforced.
• •Measurement opacity: Without independent benchmarks, claims about relevance and personalization quality remain internal metrics.

Measure retention, click-through, and time-on-topic once Android and wider user cohorts gain access; track how X exposes controls or audit logs for topic labeling; watch for policy responses if moderation incidents increase. Also monitor whether X opens programmatic hooks or partner integrations that let third parties build on Custom Timelines, which would turn the feature into an AI distribution channel rather than a single-product experiment.