X Algorithm Code: Copy-Link Shares Weigh 30x a Like

This timeline only includes posts from the personal accounts of Elon Musk and X head of product Nikita Bier. Quoted text is verbatim from the original posts; everything else is paraphrased. These are plans and statements: the code is the arbiter of what shipped. Dates are UTC+8.

The current README states that For You “ranks posts using a transformer model” (README L3).

Per the README, the model predicts how likely the viewer is to take each action on a post, multiplies each prediction by a weight, and sums the results (README L352–L356). The default weights live in home-mixer/params/param.rs. The table compares them to the 2023 Twitter algorithm weights in twitter/the-algorithm-ml; “—” means that version has no equivalent entry.

The three biggest shifts: replies dropped from 13.5 to 5.0; author-engaged replies went from 75.0 to no separate weight; and share signals were added, with copy-link sharing (20.0) now the highest positive weight in the file.

Caveat 1: weights multiply predicted probabilities, not counts. The comment at param.rs L291–L311 says each weight multiplies the “predicted probability” of an action (or a continuous value such as watch time) and that “the weights do not multiply raw engagement counts.” It explicitly rejects the reading that “one report cancels 468 likes,” noting that a report's baseline probability is more than 1000x lower than a like's, so it needs a larger weight to affect ranking at all (param.rs L291–L311; also README L354). So “copy link weighs 30x a like” (20.0 / 0.66 ≈ 30.3) describes how strongly each prediction is scaled, not “one copy-link share equals 30 likes.”

Caveat 2: param.rs holds defaults that experiments can override. The README's “Experiments and Configuration” section says X runs experiments on a small share of traffic, reads many tunable values from a configuration system, and uses cron scripts to set the repo defaults to the “primary production values” (README L414–L418). Treat the table as a snapshot of the main production config; a given user may be in an experiment bucket with different values.

A fair reading: the scoring step really is model-driven, but filtering, dedup, discounts and cold start around it are still hand-written rules. vm-ranker's value_model.rs reads the author-diversity switch from params instead (vm-ranker/scoring/value_model.rs L104–L106), so which path production uses, and whether diversity is on, cannot be determined from the repo. Unverified.

Each point cites its code. Items marked “inference” are my reading; the code does not prove them directly.

1. 48 hours is a hard cutoff. AgeFilter drops posts older than 48 hours (README L374).
2. The first two hours matter for small accounts. Cold-start parameters: max post age 7,200 s, follower cap 50,000, impression threshold 200 (param.rs L489–L518). Inference: new posts from smaller accounts may get extra placement in that window; the mechanism was not traced line by line.
3. Content people pass along is weighted most. Copy-link share 20.0, DM share 5.0 (L357–L368). Inference: reference material people save or send to colleagues beats chasing likes.
4. Replies still matter, but there is no 75x bonus anymore. focal_author_reply remains a model feature (reply.rs L62–L69). Inference: author replies still influence ranking, by an amount the code does not reveal.
5. Out-of-network reach is discounted. See README L360 and L376 (OONRetweetReplyFilter).
6. Negative feedback is still heavily weighted. Report −234, mute −58.8 (L426–L432). Per caveat 1, this acts on the predicted chance that a viewer reports you, not a tally of reports.

Self-check: the README describes Under the Hood, a pilot tool that shows visibility-affecting labels on your account and posts (README L452–L456; entry point https://x.com/i/jf/under_the_hood ).

I ran into a similar gap between official guidance and actual mechanics when building my own site; see bilingual site SEO.

• Checked on 2026-10-15 (UTC+8).
• Code: xai-org/x-algorithm commit 17c96bf346df3d62b90f1bb01830d432dba74485 (2026-10-14 03:41 UTC); every line number verified locally with rg/sed.
• 2023 baseline: twitter/the-algorithm-ml commit b85210863f7a94efded0ef5c5ccf4ff42767876c, recap README L31–L40, each weight linked to its own line.
• Statements: original posts by Elon Musk and Nikita Bier, linked in the timeline.
• Left out as unverified: any Premium ranking boost, the live out-of-network discount and author-diversity values, the repo's star count.

1. Copy-link share 20.0 vs like 0.66 (L320 / L363–L368).
2. Author-engaged reply: 75.0 in 2023 (L36) → no separate weight in 2026, still a model feature (reply.rs L62–L69).
3. “deletion of all heuristics within 4 to 6 weeks” (Musk, 2025-10-18) vs the 48-hour AgeFilter (README L374) and enable_author_diversity: false (value_model.rs L45).