# What X's open-source algorithm code actually rewards

In 2026 xAI open-sourced the X (Twitter) For You algorithm again. Most write-ups quote weights from the old 2023 code, but the new release deliberately hides the numbers. What it does show is better: the exact structure of the ranker, and the precise list of things the model predicts about you.

## The one formula everything hangs on

Both in-network posts (from people you follow, source: Thunder) and out-of-network posts (discovered across the platform, source: Phoenix retrieval) get scored by the same Grok-based transformer. The final score is a weighted sum:

    score = sum( weight * P(action) ) over 19 actions

The release states they "eliminated every single hand-engineered feature". There is no keyword score, no follower-count bonus, no hashtag trick. The model learns what you engage with and finds more of it. The weights that combine the 19 predictions are the one thing kept private, so nobody outside X can quote them honestly. The 19 predictions themselves are the answer.

## The 19 things it predicts (this is the reward function)

- 14 positive signals: like, reply, repost, quote (and click a quoted post), three separate kinds of share (share, share via DM, share via copy-link), profile click, follow the author, link click, photo expand, video quality view, and dwell.
- 4 negative signals: not interested, mute the author, block the author, and report. These are separate heads with negative weight, so a post likely to earn a mute or report is pushed down.

Two positive heads matter enormously for growth: the model literally predicts P(profile click) and P(follow the author). Your visit-to-follow funnel is something the ranking directly optimizes. And the "pass it on" side is carried by three share signals including DM shares: content that travels is content people forward, not content that collects quiet likes.

## The weights everyone quotes (2023, directional)

Numbers like "a reply is worth 13.5x a like" come from the 2023 release (twitter/the-algorithm), not the 2026 Grok rewrite, which does not publish its weights. Directional, but the shape matches the 19 heads.

| Action (2023) | Weight | vs a like |
| --- | --- | --- |
| Reply the author engages back on | +75 | ~150x |
| Reply | +13.5 | ~27x |
| Quote tweet | ~13.5 | ~27x |
| Profile click that leads to engagement | +12 | ~24x |
| Bookmark / dwell 2 min+ | +10 | ~20x |
| Retweet (plain) | +1 | ~2x |
| Like | +0.5 | baseline |
| Video watched 50% | ~0.005 | minimal |
| Mute / block | -74 | catastrophic |
| Report | -369 | nuclear |

A reply the author answers is roughly 150x a like; a report is a four-figure negative. Likes and follower counts sit near the floor.

## Answers to the questions people actually ask

- Quote tweet or repost? Quote. A plain repost is ~1x; a quote is a high-weight action because it adds content and starts a conversation.
- How fast to reply to comments? Within 30 to 60 minutes. Early velocity is the biggest factor, and a reply the author answers is the top signal.
- Morning or evening? There's no time-of-day weight; post when your audience is online, because that drives the early velocity the ranking extrapolates from.
- Does posting more dilute reach? Past a few posts a day, yes. Author-diversity decay (about 0.5 per successive post, 0.25 floor in 2023) caps how many of your posts one person sees. Three to four quality posts beat ten.
- Do threads count as multiple posts? No, a thread is one post; its job is dwell time.
- Text or video, and do links kill reach? Native content out-distributes link posts; an external link in the body gets little reach, so put the link in the first reply.
- TweepCred? The 2023 PageRank reputation score was a hand-engineered feature; the 2026 rewrite eliminated all hand-engineered features, so there's no separate reputation number to chase.

## How to test any of this

Change one variable at a time, give it two weeks, and measure reply rate and depth, bookmark and share rate, and profile-visit-to-follow rate, not likes and not raw impressions. If a change moves replies, saves, and follows, it works even when likes don't budge.

## Out-of-network is down-weighted, on purpose

Posts from accounts you don't follow are retrieved by a two-tower similarity model, then multiplied by a factor below one before ranking. In-network posts keep full score. Discovery is real but starts at a disadvantage, and the only way past it is to be relevant enough that the retrieval model confidently matches you to an audience. So: stay on a consistent topic, and reply to accounts whose audience is already your people. A reply borrows their distribution and skips the out-of-network penalty.

## Author-diversity decay

After scoring, each additional post from the same author is attenuated by exponential decay. Hammering one big account's replies has diminishing returns; spreading the same effort across many relevant accounts nets far more total reach.

## What the code tells you not to do

The four negative heads predict a mute, a block, a not-interested, and a report, and subtract them. A salesy reply, a dunk, an over-promising thread can rack up impressions while generating the exact signals that suppress you and cost the follow. Reach without the right reaction is not a win. A Grox content-understanding layer runs spam and policy classification, and hard filters drop posts that are too old, from muted or blocked accounts, or matching muted keywords. None of it is gameable with keywords.

## The playbook the code implies

- Optimize for high-effort actions (replies, shares, profile clicks, follows), not likes.
- Reply to relevant mid-size accounts early to clear the out-of-network penalty.
- Spread across accounts; diversity decay punishes concentration.
- Stay on topic so the retrieval model can match you to an audience.
- Stay constructive; the negative heads are real.

Argus reads the real outputs of these prediction heads from your own X analytics (profile visits, follows, saves, shares) and ranks the exact posts to reply to next, then drafts each reply in your voice. You always press send. Run the free X growth check at argushq.cc/share, or read how the X algorithm works in 2026 at argushq.cc/guides/x-algorithm-2026.