xAI put the entire "For You" feed algorithm on GitHub this morning — xai-org/x-algorithm, Apache-licensed, the real thing. Within an hour there were 19,000 stars and a hundred threads breaking down the transformer architecture.
I don't care about the transformer architecture. Neither should you, unless you're building one. What I care about — what you should care about if you're trying to grow a brand, a product, or yourself on X — is one question: what do I do differently on Monday morning?
So I read it. Here's how I read it as a marketer.
The big one: there are no features left to game
The single most important sentence in the whole repo is this — X says it has "eliminated every single hand-engineered feature and most heuristics from the system."
Sit with that. For fifteen years, social algorithms ran on hand-built signals: this many hashtags, this much recency, this ratio of likes-to-impressions, this keyword in the first line. Growth "hacking" was, mostly, reverse-engineering those signals and feeding them.
That game is over. The ranking is now a single Grok-based transformer that reads your post the way a person reads it — semantically — and reads each viewer's genuine engagement history the same way. It is not counting your hashtags. It is asking one question: is this the kind of thing this specific human actually engages with?
You cannot trick a model that understands meaning. You can only be genuinely relevant to a real audience. That's not a motivational poster — it's now literally the architecture.
Your reach has two doors, and one of them is ML
Every timeline is built from two pools. "Thunder" serves in-network posts — people who follow you. "Phoenix Retrieval" serves out-of-network posts — content pulled from the entire global corpus by an ML similarity model, shown to people who've never heard of you.
Both pools get ranked together. Which means out-of-network reach is real, it's significant, and it is decided by a model matching your content's meaning to a stranger's interests. Follower count gets you the first door. It does nothing for the second. If you've ever wondered how an account with 800 followers lands a 2-million-view post — that's Phoenix. The door is open. It opens on relevance, not status.
The signal everyone will miss: negative weights
Here's the part the ML threads will skim past. The model doesn't predict one "good post" score. It predicts about fifteen separate probabilities for every post: the chance you'll like it, reply, repost, quote, click, expand the photo, watch the video, dwell on it, and — critically — the chance you'll follow the author.
Then it predicts the bad ones: the chance you'll hit "not interested," mute the author, block them, or report the post.
The final score is every one of those probabilities multiplied by a weight and added up. Positive actions have positive weights. Negative actions have negative weights. They don't just fail to help you — they subtract.
This is the quiet death of ragebait. If your post earns ten thousand angry quote-tweets but a meaningful share of viewers mute or block you, the algorithm is mathematically dragging that post down at the same time. Outrage that gets you muted is now self-defeating by design. The model is explicitly trained to find content people quietly dislike and bury it.
The flip side is the opportunity: P(follow_author) is a weighted signal. Posts that make a stranger think "I want more of this person" are directly rewarded. Write for the follow, not the like.
There is a literal "slop" score
Buried in the content-understanding service ("Grox") is a vision-language classifier that screens posts before they ever get ranked. It outputs a quality_score, runs spam and safety checks — and carries a field called slop_score.
X has, in code, a named penalty for low-effort content. The "banger" screen is real and so is its opposite. AI-generated filler, engagement-bait templates, recycled thread formats — there is now an explicit model whose job is to notice and discount them.
Posting 12 times a day is a worse strategy than it was yesterday
There's an "Author Diversity Scorer" that attenuates repeated scores from the same author within a feed, so no single account can carpet a timeline. Translation: your eighth post of the day is competing against a deliberately dampened version of your own first seven. Volume has diminishing returns baked into the math. One post that genuinely earns dwell-time and follows will out-perform ten that don't — and now you can see exactly why in the source.
What I'd actually do Monday morning
- Write for the reply and the dwell, not the like. Likes are cheap predictions; replies, dwell time and follow-conversion carry real weight. Make posts people stop on and respond to.
- Earn the follow with every post. Ask of each draft: would a stranger who saw only this want more of me? If not, it's a weak post regardless of likes.
- Kill the ragebait instinct. Mutes, blocks and "not interested" are negative weights. A post that splits a room can be net-negative even with big numbers.
- Drop the hashtag-and-format tricks. There are no hand-engineered features left to feed. Spend that energy on saying something genuinely useful to a specific audience.
- Go for out-of-network relevance. Phoenix matches meaning to strangers' interests. Pick a clear topic lane and be unmistakably about it, so the retrieval model knows exactly who to show you to.
- Cut the slop. There's a model scoring it down. Fewer, sharper posts beat a content treadmill.
None of this is new advice, exactly. "Be genuinely good and genuinely relevant" has always been the boring right answer. What changed today is that X published the proof — in code — that the boring right answer is now the only answer. The shortcuts didn't just stop working. They got deleted from the system.
That's good news if you were ever willing to do the real work. It's bad news for everyone who built a growth strategy on tricks.
— Tom