When Elon Musk was up in arms with Donald Trump, when he memorably mocked Trump for being on the Epstein files before deleting his tweets, he briefly consulted Curtis Yarvin, a former computer engineer turned political blogger who often argues for replacing American democracy with a CEO-style monarchy or an American Caesar. Yarvin might not be particularly well known outside MAGA circles, but his work is often cited by members of Team Trump and has influenced figures such as Musk and JD Vance.Yarvin has a reputation for saying exactly what he thinks, regardless of who it offends. Despite being close to Musk, he has no hesitation in criticising him, as he did recently when he mocked the version of Twitter that has emerged under Musk. Sharing a post about the platform’s engagement algorithm, he wrote: “Elon paying $47B for Twitter was like an Yanomamo Indian in the Amazon trading a sack of gold dust for an AK-47. He has a sense of its awesome power. He doesn’t know how to load it, how to fire it, or what end the bullet comes out. But he has the coolest war club in the tribe.”The post he was referring to was not a casual complaint about low likes, nor the usual algorithmic folklore that surfaces whenever someone’s reach dips. It came from an X user called BLΛC, an artist who said he had spent four years on the platform, built more than 60,000 followers, posted every day with discipline, and then watched his reach fall by more than 40% over three months without changing anything about his work.What unsettled him was not just the drop in numbers, but the shift in visibility. Artists he used to see constantly had stopped appearing in his feed. He assumed they had left. They had not. They were still posting, still creating, often multiple times a day. They had simply become invisible to him, and presumably he to them.This is something many X users have begun to notice. The “For You” section often feels filled with content that bears little relation to one’s interests. But instead of guessing, BLΛC decided to investigate the system itself.What he claims to have found is not a broken platform, but one that operates on a very different logic from the Twitter users once understood.
Goodbye Digital Townhall
The central claim is that X no longer waits to see how a post performs. It predicts how the post will perform and assigns reach based on that prediction.This prediction is generated through a system referred to as Phoenix, which evaluates each post using multiple “prediction heads.” These are separate outputs that estimate the likelihood of different user actions.Among the variables identified are favorite_score (likes), reply_score (replies), dwell_score (whether a user pauses), dwell_time (how long they pause), and photo_expand_score (whether an image is opened). Negative signals such as not_interested and report are also factored in.Each of these is a probability, and the system combines them into a single score that represents expected engagement. That score determines how widely the post is initially shown.This creates what BLΛC describes as the “prediction trap.” If the system predicts low engagement, the post is shown to fewer people, which reduces the chances of engagement and confirms the prediction. If it predicts high engagement, the post is shown more widely, increasing the likelihood that it performs well.The system is not just measuring performance. It is shaping it.The thread also highlights the role of follower count. According to BLΛC’s reading of the code, author_followers_count is retrieved through a service called Gizmoduck and used for display, but not fed into the ranking system. In practical terms, this suggests that follower count may no longer directly influence reach.
AI Agent Swarm
BLΛC says he used an AI agent swarm to analyse X’s publicly available codebase. Instead of reading it manually, multiple agents were tasked with examining different parts of the system, from input pipelines to ranking and filtering.The process begins with candidate generation, where the system gathers posts that could potentially be shown to a user. These are drawn from followed accounts, similar users, and behavioural patterns.Each candidate is then described using features, which include attributes such as recency, format, past engagement, and user behaviour.The Phoenix model then applies its prediction heads to estimate likely outcomes. These predictions are combined into a ranking score.After scoring, filters are applied.One of these is AuthorDiversityScorer, which reduces the visibility of multiple posts from the same author within a single session. This prevents feed domination but also means that frequent posting can reduce the reach of individual posts.Reposting is handled differently. According to the thread, reposting another user’s content can significantly reduce visibility for that specific post, while self-reposts are subject to deduplication rules based on whether the content has already been seen in a session. Systems such as Bloom filters and RetweetDeduplicationFilter are used to manage this.Finally, there is Thunder, an in-memory store that holds posts for roughly 48 hours before they fall out of the active pool. After that, they are no longer served by the algorithm.
How it’s different from older Twitter
Earlier versions of Twitter were built around the social graph. You followed people, they posted, and you saw it, with ranking layered on top.The system described here is built around prediction. Visibility is determined less by who follows you and more by how the system expects users to behave.This changes the role of familiar signals. Follower count becomes less relevant to reach. Posting frequency is constrained by diversity rules. Reposting can introduce penalties. Content lifespan is compressed into a short window.The platform moves from reflecting activity to anticipating it.The change is not experienced as a single break but as a gradual erosion of familiarity. Accounts that once appeared regularly become intermittent. Posts that would have travelled begin to stall. The connection between effort and reach becomes harder to understand.Earlier, users could follow how content moved. Now that movement is mediated by a system that operates on predictions rather than visible interactions.From the platform’s perspective, this prioritises engagement. From the user’s perspective, it reduces clarity.Yarvin’s analogy captures that gap. The platform retains its power, but the way it operates is no longer intuitive to those using it.The machine still works. It just no longer works in a way that users recognise.
Leave a Reply