What changed in X's For You ranking code—and what didn't
Quick answer: X’s public home-mixer defaults changed seven ranking coefficients in commit 021afa44. The like coefficient rose from 0.66 to 0.708, while video-open fell from 0.02 to zero. These weights multiply predictions—not counts of likes or views. The commit does not establish when every user received a production rollout.
The update is dated Nov 5 in UTC. For readers in Central Time, its timestamp is Nov 4, 2026, at 11:16 PM CT. GitHub lists both the author and committer as “CI agent.” That’s a repository timestamp, not a confirmed launch time for a change to your feed. (Commit record)
The useful story is narrower than “X has a new algorithm.” There are specific changes we can inspect—and limits to what those changes tell us.
What changed in the Nov 5 X algorithm commit?
Here are all seven changed scoring coefficients in home-mixer/params/param.rs. Each before-and-after pair comes from the commit diff; the links in the first column open the corresponding new defaults.
| Signal / parameter | Before | After |
|---|---|---|
| Like — FavoriteWeight | 0.66 | 0.708 |
| Video open — VideoOpenWeight | 0.02 | 0.0 |
| Click — ClickWeight | 0.3 | 0.274 |
| Share — ShareWeight | 2.0 | 2.118 |
| Dwell time — ContDwellTimeWeight | 0.006 | 0.0075 |
| Click dwell time — ContClickDwellTimeWeight | 0.4 | 0.385 |
| Not interested — NotInterestedWeight | −52.272 | −53.07 |
The like, share and dwell-time coefficients increased. Click and click-dwell coefficients decreased; the not-interested coefficient became more negative. That describes the settings, not a measured change in anyone’s reach. (Before-and-after source)
Zero video-open weight does not mean video is dead
One video-open term is now zero. That isn’t the same as removing every video-related signal from ranking. The same historical file retains a separate VideoCarouselVideoOpenWeight of 1.0. Treating one coefficient as a verdict on every video would ignore that distinction. (Carousel default)
Nor are these seven numbers the whole update. The diff also changes the aggregation type and enables a home-excursion scoring setting. The table is a record of changed coefficients, not a complete description of the recommendation system. (Full diff)
Does a like literally add 0.708 points?
No. The code’s own explanation is unusually direct:
“Each weight multiplies the predicted probability of that action … or a continuous value e.g. watch time — the weights do not multiply raw engagement counts.” (Source comment)
In plain English, the system estimates whether a particular viewer will take an action. The coefficient scales that estimate as part of ranking. It isn’t a cash register adding 0.708 every time someone presses Like. Continuous outputs, such as watch time, are also different from probabilities. (Prediction and personalization explanation)
That is why dividing one coefficient by another cannot tell you that a reply “equals” a fixed number of likes. And a higher dwell coefficient does not establish that every second watched earns a corresponding increase in views. The source explicitly warns against converting these ratios into engagement-count equivalents. (Code warning)
Did this commit introduce a rule hiding AI replies?
This commit does not establish a new Nov 5 AI-reply ban. The internal llm_slop_post signal maps to a RiskyHighVizReply label in this snapshot. But the identical mapping already exists in the parent commit. It wasn’t introduced by this update. (Nov 5 snapshot; parent snapshot)
The claimed rule dropping these replies for non-followers is absent from this historical snapshot’s tweet rules and notification rule sets. Reading a later version of the repository cannot establish that the rule existed here.
A label mapping also doesn’t tell us how accurately a classifier identifies AI writing, who was affected, or whether a particular restriction ran in production. Those would require evidence beyond these files. We are not claiming universal suppression of AI-written replies.
What this means for you
If you post on X: don’t turn this table into a points-based posting plan. It shows a change in how certain predictions are weighted, not a promise of more distribution for a particular format. The code itself cautions against treating engagement as reproducible ranking arithmetic. (Source explanation)
If you mostly read: this commit cannot diagnose why your personal For You feed looks different. Its public defaults are evidence about code, not a record of the settings and predictions used for your account. It also does not establish a new feed-reset feature. (Scope: the commit)
If you see a claim about an algorithm update: ask which revision it cites. A current file and a historical commit answer different questions. Here, the immutable diff supports the seven-number change; it does not support a newly launched AI-reply ban.
Where can I check the weights myself?
Start with the immutable commit, then open the pinned parameter file. Read the comment above the coefficients before interpreting the table.
These links preserve the historical snapshot rather than silently switching to today’s defaults. That makes the claim checkable: seven coefficients changed; predictions are not counts; a repository update is not proof of a universal rollout.
More Tech coverage from AltTabDaily.