How to grow on X as a founder in 2026: optimize for the five actions the algorithm actually pays for. On August 13, 2026, X open-sourced the code behind the For You feed, including Phoenix, the transformer model that scores every post, and the real weights file. A copy-link share is worth 20 points. A reply, a quote, or a DM share is worth 5. A new follow is worth 4. A like is worth 0.5. A mute costs you 58.8 and a report costs 234. So the playbook is: post things people send to a specific friend, start conversations instead of collecting likes, turn big accounts in your niche into mutuals (mutual-follow replies get a 15x class boost), and never post anything that makes a stranger reach for mute.
Every algorithm guide written before mid-2026 is describing a dead system. The old heuristics, likes as the goal, links as poison, video worship, were guesses about a black box. Now the box is open at github.com/xai-org/x-algorithm and some of the folklore turns out to be wrong: the published weight on video quality views is literally 0.0, and opening a link carries a small positive weight, not a penalty.
This guide translates the code into a founder's playbook: how Phoenix works in plain English, how the new-author boost helps you reach 1,000 followers, what happened to the reply-guy strategy, whether to grow a founder or brand account, and the posting cadence and formats the weights actually favor. Written for a founder doing their own marketing, probably at 11pm.
How does the new X algorithm actually work in 2026? (Phoenix in plain English)
The repository, published under xAI's GitHub org on August 13 and 14, 2026, shows a pipeline with three candidate sources running in parallel: Thunder pulls recent posts from accounts you follow, Phoenix retrieval embeds you and every post as vectors and returns the nearest matches, and SimClusters finds posts from engagement-pattern communities you belong to. A separate component called Grox runs LLM classifiers on every post at publish time, tagging spam, adult content and violent media. (Per TechCrunch, the systems that use Grok itself to predict rule violations were held back from the release.)
Then Phoenix ranks. For each candidate post, the transformer predicts the probability that you, specifically, will take each possible action: favorite, reply, repost, quote, share, DM it to someone, copy the link, click your profile, open the post, expand a photo, open a video, dwell on it, follow the author, or hit not interested, mute, block, report. The final score is the sum of each probability times its weight. The repo is explicit that weights scale predicted probabilities, not raw engagement counts. Nobody's like is worth more than anyone else's; what varies is how likely each viewer is to act.
Three adjustments follow: an author diversity decay (your second post in someone's feed scores less than your first, down to a floor), an out-of-network discount (posts from accounts the viewer does not follow are multiplied by a factor below 1), and a new-author boost (authors below an impressions threshold get lifted toward a target position).
The whole thing means content quality is now measurable in advance. Phoenix is asking one question about your post: what will this specific viewer probably do with it?
Do this now: skim the README for 15 minutes. It is the rare primary source in a discipline built on secondhand guesses.
How do you get your first 1,000 followers on X?
Two lines of code matter most when you are small.
First, the new-author boost: posts from authors whose impressions are below a threshold get lifted toward a target position in the feed. The system deliberately spends impressions on unproven accounts to find out what they are. Your early posts get distribution you have not earned yet. Most new founders waste that free sample on "excited to announce" posts that teach Phoenix nobody engages with them.
Second, FollowAuthorWeight: 4.0. Phoenix scores each post partly on the probability that the viewer follows you after seeing it, eight times the weight of a like. The posts that convert strangers into followers are consistently the same kinds: a specific opinion about your niche that some people will disagree with, real numbers from your own business ("we spent 4,000 dollars on ads last month, here is exactly what happened"), and useful artifacts (a teardown, a checklist, a before-and-after). Generic motivation converts nobody because following you promises nothing.
So the first-1,000 playbook: pick one lane and stay in it for 90 days, because Phoenix retrieval embeds your content and shows it to people near it in vector space, and an account that posts about everything embeds as nothing. Write a bio that says who the account is for, since follow decisions happen on your profile, not on the post. Publish one genuinely specific post every day while the new-author boost is spotting you impressions. And spend more time in other people's replies than in your own composer, for the reasons in the next section.
Do this now: write down the one topic your account is about, then draft five posts that only someone running your kind of company could write. Post one a day this week.
Does the reply-guy strategy still work?
Yes, but the mechanism changed, and knowing the real mechanism makes you better at it.
The old theory was that a banger reply rides the big account's post into thousands of For You feeds. The code says mostly no: an OONRetweetReplyFilter removes replies from accounts the viewer does not follow. Strangers scrolling For You generally do not see your reply as a feed item. Your reply's real audience is the people who open that thread, plus everyone who taps through to your profile.
What the code giveth instead: BidirectionalFollowReplyWeightBoost: 15.0. Predicted replies between accounts that follow each other get a boost thirty times the weight of a like. Mutuals are the most valuable relationship in the entire weights file.
So the reply-guy strategy in 2026 is a mutuals acquisition strategy. Pick 20 or 30 accounts in your niche, from slightly-bigger peers to the big names, and reply with actual substance: add a data point they did not have, a counterexample, a sharper way to say what they said. The goal is not going viral in their thread; it is that they notice you, check the profile you tightened up in the last section, and follow back. Every mutual you convert becomes a standing distribution deal: your posts get a heavy boost in their feed, and their engagement (replies weigh 5.0) lifts your posts for everyone else.
The bar for reply quality is also economically enforced. A mute costs -58.8 and hits every future post's probability for that viewer. Paste-grade AI replies are not a growth hack, they are compounding self-harm.
Do this now: list 20 target accounts in your niche. Leave one substantive reply on each per week. Track who follows back; those mutuals are the asset.