When Elon Musk was at odds with Donald Trump, famously mocking him over the Epstein files before deleting the tweets, he briefly consulted Curtis Yarvin, a former computer engineer turned political blogger who advocates replacing American democracy with a CEO-style monarchy or an American Caesar.
Yarvin, though not widely known outside MAGA circles, is often cited by members of Team Trump and has influenced figures like Musk and JD Vance. Known for speaking his mind, he recently criticized Musk, comparing the $47 billion Twitter acquisition to a Yanomamo Indian trading gold dust for an AK-47, noting Musk's lack of understanding of the platform's mechanics.
The criticism stemmed from a post by BLΛC, an artist who claimed his reach on X dropped by over 40% in three months despite consistent posting. He discovered that artists he followed were still active but had become invisible to him, suggesting a systemic shift.
BLΛC investigated X's publicly available code using an AI agent swarm and found that the platform no longer waits to see how a post performs; it predicts performance using a system called Phoenix. This system evaluates posts with multiple prediction heads, including favorite_score, reply_score, dwell_score, dwell_time, and photo_expand_score, along with negative signals like not_interested and report.
These predictions are combined into a single score that determines initial reach, creating a prediction trap: low predicted engagement leads to fewer views, confirming the prediction, while high predicted engagement boosts visibility. Follower count, retrieved via Gizmoduck, is used for display but not directly for ranking.
The process begins with candidate generation from followed accounts, similar users, and behavioral patterns. Each candidate is described by features like recency, format, and past engagement. The Phoenix model then predicts outcomes, which are combined into a ranking score. Filters like AuthorDiversityScorer limit multiple posts from the same author in a session, reducing the impact of frequent posting. Reposting others' content can reduce visibility, while self-reposts are subject to deduplication using Bloom filters and RetweetDeduplicationFilter. Posts are stored in Thunder, an in-memory store, for about 48 hours before being removed from the active pool.
This marks a significant shift from older Twitter, which was built around the social graph—users followed accounts and saw their posts with ranking layered on top. Now, visibility depends on predicted user behavior rather than follower count or direct interactions. Content lifespan is compressed, and the connection between effort and reach becomes less intuitive.
Yarvin's analogy captures this gap: the platform retains its power, but its operation is no longer transparent to users. The machine works, but not in a way users recognize.
