This lives here now.

X algo daily, Oct 30 (`9106e9f`): no change to For You ranking. The settings were refreshed today at 6:59 PM ET. What it means for you: • The ad blender switched from `multi_risk` to `slot_spread`. That changes where ads land in your feed, not how your posts score. • Under the Hood now lists specific post IDs under each label. The code ships with this off, but it's live in the app. If a post of yours got labeled, you can now see which one. • `EnableVideoCarouselHighRiskFilter` is now on. The carousel itself is still off, so it does nothing for now. • The withheld-post appeal text was rewritten, with links for the UK, EU, Australia, Brazil and Japan. • Added but switched off: a page-decode reward head (0), debias for copy-link and DM shares (debias is still off), and new ad-model training heads. Bottom line: check Under the Hood to see which of your posts are labeled. Replies from mutuals and copy-link shares are still the biggest prediction levers. Full breakdown in the pinned post.

X algo daily, Oct 29 (`7fef3ba`): no change to For You ranking. The param sync is still from Oct 28. What it means for you: • The video weight was renamed from `ReactWithVideoWeight` to `PostVideoWeight`. It's still 0.0, so video posts get no extra score. • Carousel scoring was reworked: there's a new `VideoCarouselWeightedScoreWeight` (0) that would add the main For You score, and the 30 extra candidates are now a param. The carousel is still off, so nothing changes for your reach. • Notifications got new visibility rules for mentions and replies (`NOTIFICATIONS_WRITER_V2_MENTION`/`_REPLY`). They reuse the existing drops (PDNA, spam, low quality, takedowns) and only decide whose mentions and replies show up in notifications, not who sees your posts in For You. • Under the Hood report: appeal switches were added (off), along with FOSNR labels and a Brazil link for appealing withheld content. These are labels only, with no filter change. Unchanged: Fav 0.66, Reply 5, Quote 5, copy-link share 20, mutual reply +15 on originals, the ×0.75/×0.50 haircuts, the 2,800 fair-share trim, the spam thresholds (66/180k), cold start, and debias (still off). Bottom line: replies from mutuals and copy-link shares are still the biggest prediction levers. Full breakdown in the pinned post.

X algo daily, Oct 28 (606b254): nothing changed how For You ranks or filters your posts. What it means for you: 💠EnablePageLogging is now on. Home records the scores of every page it serves, but it's only recording, so your reach doesn't change. It's the groundwork for training on whole pages, which matters if the page-decode reranker ever turns on. 💠Ads brand-safety went back to v1 (EnableAdsBrandSafetyVerdictV2 off). That only changes which posts ads appear next to, not your organic reach. 💠Carousel videos won't count as "already seen" for the main feed. The carousel is still off, and so is a new video-open × immersive-dwell term (weight 0). 💠The visibility rules were regrouped into named sets, with the same rules in the same order. DNA and FOSNR are untouched. Bottom line: post the way you did yesterday. Full details in the pinned post.

X algo daily, Oct 27 (5ed7963): ranking math is unchanged. No changes to weights, haircuts, cold start, spam thresholds (66/180k), DNA or FOSNR. Live changes: Brazil2026ElectionFilter is deleted from For You, so the 25 runoff accounts no longer get removed for non-followers. The BRAZIL_ELECTION_LEGAL label is unchanged. Home no longer blurs or hides posts based on report-heuristic NSFW/gore labels alone. A reaction video now gets the violence blur only when the original has the high-precision gore label. Gated off: a page-decode reranker, a top-64 exploration shuffle (temp 0.2), page logging and a high-risk drop for the carousel. Full breakdown in the pinned article.

This is further evidence that X wants to make sure shared photos and videos earn for the original poster.

x-algorithm 2f25214 (param sync 2026-10-26T16:00:39Z) No live scoring change. Details: • VMRankerSendDebiasInputs false→true. Home now ships per-viewer debias inputs to the ranker. But UserDebiasEnabled=false, strength 0 → no weight moves yet. Watch this one: when it flips, predicted engagement gets recentered against each viewer's own baseline. • Experiment19 Phoenix ranker added as shadow scoring at 30%, alongside Exp18 at 30%. Comparison only, not served. Shadow requests include recent searches; live feed still has that off. • Abuse enforcement: separate very-high-follower skip removed (only ever exposed a mock value). Follower-based skip now runs through one daily threshold+ceiling check. Real thresholds still private. • Unchanged: reply-spam pageRank 66 / Grox 180k, DNA, FOSNR, OCR. Plumbing: • Ad lookup timeout 1200→1300ms • VF safety-label loader rebuilt, staging VF service removed, new verdict-testing tool. Rules identical. • Post-length counter rewritten, parity test included • Search: author name/handle match + has_video/has_image fields • Phoenix training: generative-recs + semantic-ID retrieval experiments removed. Two-tower retrieval tuned (2h cold-start window, Home-only variant), ads purchase-value heads. Not live.

It's important not to misunderstand what this is saying. X doesn't pay more in these niches, it's that more verified people hang out in those niches.

New X algorithm update last night (Oct 23). Short version: nothing changed in how your posts are ranked. Likes, replies and reposts count the same as before, and so do the spam and reach filters. What did change: 1. Ad tracking is back on. X turned its ad-tracking switch back on. It only affects ads, not your posts. 2. A spam cutoff is now secret. X's moderation system goes easier on accounts it trusts. The number it uses to decide who counts as trusted used to be in the public code. Now the code shows a fake number, and X says the real one is kept private so people can't game it. There's no sign the real number changed. 3. Small backend work: tighter security on X's moderation tools, and groundwork for future ad features. Also from Thursday's update, which I missed last time: • Fewer junk follow alerts. X now blocks follow notifications from spam, impersonator, suspended, blocked and muted accounts. • A safety brake on moderation. If too many people appeal one of X's automatic rules and win, that rule can be slowed down or paused. • Your searches aren't shaping your feed. The switch that would let your recent searches influence For You is turned off. Bottom line: post like a real person and nothing changes for you.

New X algorithm update (Oct 23). Here's what it means for you: 1. Flagged posts get a warning and lose buttons. If a post is labeled low quality or misleading, Home now shows a notice on it, and you can't reply, repost, like, bookmark or share it. 2. "Only followers can reply" now really works. If you don't follow the author, your reply gets limited. 3. Popular videos get a closer look. On posts with 64+ likes, videos flagged for review are now scanned frame by frame, including videos inside quote posts. 4. Tougher on spam bots. New checks catch accounts spamming posts or replies. The worst get suspended. The rest get a 30-day spam label and have to prove they're human. What didn't change: how likes, replies and reposts count toward reach. Normal posting is unaffected. Bottom line: if you post like a real person, nothing changes for you. Spammers and misleading posts just got harder to spread.

Handing an AI agent your bank password to handle financial drudge work isn't just a security headache, it's a liability trap. Under US banking rules (Reg E), if you hand over your login credentials, whatever that bot does generally counts as "authorized." If the AI hallucinates a transfer, fraud protection likely won't save you. Schwab is already blocking unauthorized bots dozens of times a week. Expect banks to lock this down fast: read-only APIs for bots, human verification for moving money.

New X algorithm code dropped overnight (989fb91). What it means for you: The "fair share" change I flagged yesterday is now ON. When X has too many posts to rank for your For You feed, it now trims every source evenly (people you follow, popular posts, recommendations) instead of cutting whatever lands at the end of the list. Posts from smaller sources, like the accounts you follow, are less likely to get crowded out. Reaction videos now take on the labels of the post they react to. If the original is adult or graphic, the reaction gets blurred too. If the original account is private, the reaction is hidden. Nothing else changed in your feed. Likes, replies, reposts, time spent and the spam rules all count the same as before.

New X algorithm code dropped last night (7f2ca2b). What it means for you: Nothing changes in your feed today. Likes, replies, reposts and time spent on a post all count exactly the same as before. Spam and visibility rules are the same too. Most of the update is behind-the-scenes cleanup to make the system faster and more reliable. One thing to watch: X built a "fair share" option that would give posts from people you follow, topic picks and recommended accounts an equal cut when there are too many posts to rank. It's switched off for now. If it turns on, recommendations would be less likely to crowd out posts from the people you follow.

Checked this against the open-source drop that landed ~90 min later (74d49a4, param sync Oct 19). What's actually in the code: - No dwell weight changed. ContDwellTimeWeight is still 0.006 (since Oct 13). Any "new dwell prediction" would live in the Phoenix model checkpoint, which isn't public. The drop does include Phoenix training changes, but nothing shows it serving yet. - The "read to the end" piece is a new long-dwell cluster (DENSE_WITH_LONG_DWELL_HOME_EXCURSION) running in SHADOW mode: computed and logged, not used to pick posts. So: groundwork, not a live change to your timeline yet. The live For You weights, spam thresholds, and DNA/FOSNR filters are all unchanged.

Rocky Top got some snow overnight, but we are still singing after sending the Tide home with their 3rd loss in 3 trips to Neyland.

Two X algo dumps in the past 24 hours: bc3948b (param sync 2026-10-15T16:00:34Z) and 862a123 (param sync 2026-10-16T16:00:32Z). Short version: no ranking or visibility numbers moved in either one. Weights, the OON ×0.75 and topic ×0.50 haircuts, cold start, DNA, FOSNR, the spam thresholds (66 / 180k) and the Brazil filter are all the same. What did change, in plain terms: Popular posts now have their own label (ForYouPopularPosts). They used to be counted as Phoenix retrieval. Same reach, they're just tracked as their own source now. Reposts fail closed. If X can't fully check the original post, the repost is dropped rather than shown on a guess. Safety lookups were split into must-have and best-effort. If a lookup fails, only the posts it failed for are affected, not the whole batch. Video carousel prep: the dedupe defaults are now on (by author, by semantic ID at 3 levels, by embedding at 0.72 cosine), so if it launches you won't get a row of near-identical clips. It also yields to Who-to-Follow. But EnableVideoCarousel is still false, so you won't see it yet. Abuse enforcement moved to a new settings format. The confidence floors are unchanged at 0.95, 0.85 and 0.70. Net for posters: nothing changes how your posts score or who sees them. This is X tidying up the plumbing and getting the video row ready.

New X algo dump: 92f0918 (param sync 2026-10-14T16:00:44Z) Ranking: nothing moved. All engagement weights, the OON (×0.75) and topic (×0.50) haircuts, cold start, and the DNA/OON doors are unchanged. Visibility: no thresholds moved. Spam pageRank exemption is still 66, the coordinated-spam floor is still 180k, and FOSNR, DNA and the Brazil filter are untouched. The VF engine was rebuilt as declarative rule specs, and every surface still runs the same 109 post rules and 21 author rules. New behavior: VF now fails closed. If the data it looked up isn't complete enough for that surface's policy, the candidate is dropped instead of judged on partial data. A per-post policy trace shows exactly which rule fired. Ads: EnableAdsAuthorBrandSafetyFallback false→true. A post without a Grok brand-safety label now inherits the author's SFA/NSFA label, and it's treated as NSFA if the author has none. This only affects which posts ads sit next to, not organic reach. Quick refresher on the 21 author rules, since they're now spelled out: all of them drop the post, and none blur or downrank. The Following tab only applies account status plus blocks and mutes. Account labels like spam, compromised, read-only, impersonation and NSFW pull you from For You even for your followers. DoNotAmplify and Abusive only hide you from non-followers.

New X algo dump: 17c96bf (param sync 2026-10-13T16:00:30Z) Ranking: FavoriteWeight 0.6→0.66. Second bump in two days (it was 0.5 before e1a0375). NotInterestedWeight −47.52→−52.272, about 10% heavier. Topic-filtered For You requests now skip the DPP diversity rerank. Regular For You keeps it. Visibility: Brazil2026ElectionFilter named-account list cut from 2,795 to 25 (the Oct 25 runoff candidates). Still removes their posts, reposts, quotes and replies unless you follow them. The geo label filter is unchanged. Adult checks on Article media now run as the viewer. A no-access or failed lookup counts as no media instead of an incomplete verdict. Unchanged: reply/repost/quote/share/dwell/video/block/report weights, OON and topic haircuts, cold start, FOSNR, spam thresholds (66 / 180k), DNA. Net: likes keep climbing and "not interested" hurts more. Make posts people like, not ones they wave off.

New features. More coming for the web screens.

What changed in X's For You feed today (from the open-source code, Oct 13): Likes matter a bit more, and so does time spent actually reading a post. Just tapping open a video now barely counts. Posts people stop and read win; clips people open and bail on lose ground. If you choose to hide videos, you'll now also stop seeing quotes of videos and replies underneath video posts. Bottom line: write things people actually stop and read. Likes and real reading time are worth more now, and quick video taps are worth less. The pinned article has been updated.

The X Subscription went live tonight and the switch is already breaking. The big one: "Cursor's unification API is unreachable" at the last step of checkout. Others report getting billed for the new plan while the old Premium+ or Cursor Ultra keeps charging. Annual Grok refunds go back to your card separately, not as credit. The part I haven't seen anyone flag: unifying appears to move you onto the bundled Cursor tier, not Cursor Pro. If that's right, you lose ChatGPT and the other outside models that only come with Pro. Your current plan keeps working. Wait for the bugs to clear and check what Cursor tier you land on before you switch. When you do switch, do it via a browser, not the app. This is your chance to get away from 30% extra app fee.