starl3xx.fun / algo / 011
Issue № 011 (commit 3b6919f)
A second rise for the like, the video open weighted at zero, a quarter more weight on time spent on a post, time away from Home as a new Phoenix input, and a Grok weight at zero
Every Sunday I diff the open-source X ranking code, xai-org/x-algorithm, against last week’s mirror and write down what actually moved. This issue covers 9106e9f to 3b6919f, which is 5 commits, dated 3 November, 4 November, 5 November, 6 November, 7 November.
Everything below is a default in the code, not a claim about production. The repo’s own README says the two can differ. Every number is copied from the diff; where I’m inferring what a change means for posting, I say so.
If you change one thing this week
- Keep writing for the reply and the copy link. A like went from
0.66to0.708, so a reply now weighs about 7.1 times a like, not 7.6, and a share via copy link about 28 times, not 30. The gap narrowed again; it did not close. - Make a video worth watching, not only worth opening. A video open went from
0.02to0.0, so the tap adds nothing at the default, and the weight on predicted dwell time went from0.006to0.0075.
Everything below is the evidence for those two, plus what else moved.
What moved
1. The like rises again, and six more weights move with it, all in one commit. In param.rs, FavoriteWeight went from 0.66 to 0.708, about 7 percent heavier, its second rise after issue 008 took it from 0.5. ShareWeight, a share by any route other than a DM or a copied link, went from 2.0 to 2.118, so this week two of the nine weights on the /algo ladder moved. Off the ladder, ClickWeight went from 0.3 to 0.274, ContClickDwellTimeWeight from 0.4 to 0.385, and NotInterestedWeight from -52.272 to -53.07; the video open and the dwell time are item 2. All seven moved in one commit, 021afa4, dated 5 November, the same commit that switched on the new Phoenix input in item 3. A post’s score is a plain sum of each predicted action times its weight (compute_weighted_score in scoring.rs), so against a like, a share via copy link at 20.0 falls from about 30 times to about 28, a reply or a quote at 5.0 from about 7.6 to about 7.1, and “not interested” from about 79 times to about 75. The generic share now weighs almost exactly three times a like. The ranking service’s own copy in params.rs took none of the seven: it still has the like at 0.5 and the share at 2.0, so the two copies now disagree on eight defaults, up from five. Home-mixer’s copy is the one it sums itself, and the one served when the service has not loaded its config, which the service’s open-source startup flag leaves off, or when the call to it fails (vm_ranker.rs). Expect to read “likes matter more than replies now.” A like is still the lightest action on the ladder, and a reply between two accounts that follow each other still weighs about 28 times a like. Takeaway keep writing for replies and for the link someone copies to send on; the like gained about 7 percent on everything else, which narrows the gap and leaves the order of the ladder as it was
2. A video open is weighted at zero, and time spent on a post gains a quarter. In param.rs, VideoOpenWeight went from 0.02 to 0.0, so at the default home-mixer adds nothing for the predicted chance that a viewer opens a video. This series has watched it fall: 0.07 in issue 001, 0.02 in issue 008, and now nothing. The service’s copy still says 0.07. ContDwellTimeWeight went from 0.006 to 0.0075, a quarter heavier, and nearly double the 0.004 it held before issue 008. It multiplies Phoenix’s predicted dwell time on the post, a length of time rather than a probability (apply(scores.dwell_time, weights.cont_dwell_time) in scoring.rs), so it does not compare with a like the way the ladder does. The yes-or-no dwell weight, DwellWeight 0.05, did not move, and VqvWeight, the video quality view, is still 0.0. Expect to read “X is burying video.” Nothing in this range filters or demotes a post for carrying video: a video post earns its likes, replies, shares and dwell like any other post, and what it lost is the credit for the tap that opens it. Takeaway make the first seconds worth staying for; the tap that opens your video earns nothing at this default, and the time a viewer then spends on the post earns more than at any point this series has read
3. Phoenix now reads how long each viewer stayed away from Home, and the predictions it adds are weighted at zero. In param.rs, PhoenixAggregationType, which picks how a viewer’s engagement history is assembled for the Phoenix ranking model, went from "DENSE_WITH_LONG_DWELL" to "DENSE_WITH_LONG_DWELL_HOME_EXCURSION". Issue 001 reported the move to the long-dwell variant. Before this range the new variant was set only for shadow traffic, on one experimental cluster, Experiment16Fou (ShadowTrafficPhoenixClusterAggregationTypes, unchanged). In Phoenix’s training code, the home_direct_packed config in xrecsys.py gained enable_home_excursion_history: each action in a viewer’s history now carries two more inputs, the seconds that viewer went on watching video after opening one from Home and the seconds they spent on a profile they opened from Home, each capped at 300 seconds and log-scaled (recsys_feature_prep.py). The same config trains the model to predict both durations on Home, conditioned on the video open and the profile click. EnableHomeExcursionScores went from false to true, so phoenix_scorer.rs now copies those two predictions onto every post it scores. Nothing weighs them yet: home-mixer hard-wires the three weights that would read them to 0.0 in value_model.rs, and the service’s VideoContinuationWeight, UserVideoContinuationWeight and ProfileVisitSecsWeight are all 0.0, as they were when issue 007 reported them. So what reaches ranking at these defaults is the input: for each thing a viewer engaged with, the model now also sees how long they were away from Home afterward. Expect to read “X now rewards profile visits.” The weight on the predicted time spent on a profile is zero; what changed is what the model reads about the viewer. Takeaway nothing to chase yet; the time someone spends on your profile or in your videos after leaving Home is now part of what Phoenix reads about that viewer, not a term in your post’s score, which is inference from where the switches sit
4. A new weight for asking Grok about a post, at zero. GrokAnalyzeWeight is new at 0.0 in both param.rs and the service’s params.rs. It multiplies a new Phoenix prediction, grok_analyze_score, the chance that this viewer taps the button that asks Grok to analyze the post (CLIENT_TWEET_CLICK_GROK_ANALYZE in recsys.proto), and it joins the sum in scoring.rs. weights.rs counts it with the positive weights, beside the like and the reply, not with “not interested” and the block. At 0.0 it adds nothing. Expect to read “get people to ask Grok about your post.” At this default that tap is worth nothing to your post’s score. Takeaway nothing changes at this default; if the weight is ever raised, a post people want explained would gain, and this series will report the number when it moves
Also in this range, and not worth an item each. Ads placement went back: AdsBlenderType went from "slot_spread" to "multi_risk" on 3 November, reversing the switch issue 010 reported (multi_risk_blender.rs), and the new EnableAdsHistoryXdcCacheWrite, true, writes each viewer’s served-ad history to the peer datacenter’s cache as well as the local one, kept for AdsHistoryCacheTtlSecs, 300 seconds (served_ad_history_cache_side_effect.rs); all three decide where ads go, not how organic posts rank. The video carousel from issue 007 is still off (EnableVideoCarousel false). It gained a click-through model that would score its videos from the predicted video open and each video’s embedding, fetched from the ranking service (EnableVideoCarouselCtrModel false, video_carousel_ctr_model.rs), a drop for videos the viewer already saw in a carousel (EnableVideoCarouselSeenFilter false), a drop for a medium-risk brand-safety verdict (EnableVideoCarouselMediumRiskFilter false, video_carousel_filter.rs), and VideoCarouselVideoContinuationWeight at 0.0. Shadow scoring now samples two more experimental Phoenix clusters, Experiment22Fou and Experiment23Fou, at the same rate as the first two (ShadowTrafficPhoenixClusterRates), and serves nothing from any of them. The Under the Hood trace from issue 007, still off (EnableUnderTheHood false), now shows each post’s three largest weighted terms and each one’s share of the post’s total, dropping any under a tenth (MIN_HEAD_SHARE in value_model.rs), where before it listed every raw prediction (under_the_hood.rs). Ten home-mixer filters were rewritten onto one shared helper, FilterResult::partition in filter.rs, and the rules they apply did not change; author_socialgraph_filter.rs still drops the same muted and blocked authors. Visibility filtering was reorganized at length again, with each data source split into its own file and the notification rules expanded; this issue did not establish which of those changes are new behavior.
The standing numbers
For anyone new here, the ladder moved this week at two rungs, the like and the generic share. From param.rs, each one against a like; the service’s params.rs declares the same values for the rest and still 0.5 for the like:
- share via copy link:
20.0, about 28 times a like - reply between two accounts that follow each other:
5.0plus a15.0boost, also about 28 times - reply or quote:
5.0, about seven times - like:
0.708, the unit
Nothing older than 48 hours enters the For You feed at all: MAX_POST_AGE in config.rs is 48 * 60 * 60 seconds. A reply or a repost from an account a viewer doesn’t follow is dropped outright by oon_retweet_reply_filter.rs at its default, so replies earn their weight from your existing followers rather than from new reach. A quote carries the same weight and does travel out of network. And where the ranking service applies them, two multipliers follow the sum: the repeat-author factor, which keeps 0.625 of your second-best post and 0.4375 of your third-best, and OonWeightFactor 0.75, on a post from an account the viewer does not follow and on a reply or repost from one they do.
Caveats
These are defaults in open-source code, overridable per experiment at runtime. The ranking model itself (Phoenix) is trained, not rule-based; the weights blend its per-viewer predictions and don’t describe what it learned. Reading weights as tactics is inherently lossy. I’d rather you know that than not.
Sources: param.rs, scoring.rs, params.rs, vm_ranker.rs, xrecsys.py, recsys_feature_prep.py, phoenix_scorer.rs, value_model.rs, recsys.proto, weights.rs, multi_risk_blender.rs, served_ad_history_cache_side_effect.rs, video_carousel_ctr_model.rs, video_carousel_filter.rs, value_model.rs, under_the_hood.rs, filter.rs, author_socialgraph_filter.rs, config.rs, oon_retweet_reply_filter.rs, all at 3b6919f.