X pushed new numbers on November 5, and within a day posts were going around saying early likes matter more now. I went through the code line by line. Here's what actually changed and what didn't.
What changed on November 5, 2026?
X changed seven ranking weights in home-mixer/params/param.rs, one of the two files in its published code that hold the feed's weights:
- Like: 0.66 to 0.708
- Share: 2.0 to 2.118
- Time spent on the post: 0.006 to 0.0075
- Click: 0.3 to 0.274
- Time spent after a click: 0.4 to 0.385
- Opening a video: 0.02 to 0
- Not interested: -52.272 to -53.07
The reply weight didn't move. It's still 5.0.
Were any new predictions switched on?
Yes, two. The same commit turned on predictions X's model can now make for every post: how many seconds a reader keeps watching after opening a video, and how many seconds they spend on your profile after tapping it. Before November 5, the setting for both was off.
What X's published weight files don't give either one is a value. So there's no honest number to report for them yet, and I'm not going to guess one.
Which of the two weight files does your feed use?
X keeps its weights in two files. The other one, vm-ranker/params.rs, belongs to the ranking service, and it still has the like at 0.5, the share at 2.0, the click at 0.3, and a video open at 0.07.
Those aren't the live numbers. Both services load their live weights from the same X settings file, features/home-mixer/main/rust_home_mixer.yml (vm-ranker/config_sync.rs), and home-mixer/params/param.rs says at the top that it's mirrored from that file. The values written in the ranking service's file are only what it falls back on when a setting is missing. So the November 5 values are the ones X runs: a like is 0.708.
Do early likes count more now?
Not according to the code. Some posts on X said this update means likes in the first few minutes now count more. I couldn't find anything that says that. Each weight multiplies X's prediction of how likely a reader is to take an action, as X itself explained in August. The weights don't look at when a like happened. The only new code about early and late engagement is in the ads model, and it's switched off.
Some posts also compared today's 0.708 with a like that was "worth 0.5 in October." Those two numbers come from the two different files. The 0.5 is the ranking service's fallback value, and it didn't change. The live like went from 0.66 to 0.708.
What's the bigger picture?
Put this next to the September 29 change and the October model updates, and the direction is the same every time: less value on the tap, more on the reader who stays.
Every current value is on The numbers. The book's next drop on November 25 covers all of it.
What else moved in the same push?
Four more weights changed in that same file on November 5:
- The click weight went from 0.3 to 0.274.
- The weight on time spent after a click went from 0.4 to 0.385.
- The weight on time spent on a post went from 0.006 to 0.0075.
- The not-interested weight went from -52.272 to -53.07.
Source: X's published code, commit 021afa4.
What to do about it is in the book. Updates say what changed. The book explains how the feed works and what to do, and it's revised every month from X's code.
See what it costs