BrelloAI

X AI ALGORITHM UPDATE 🚨 X updated their open source algorithm yesterday and the early likes matter even more now. Small accounts should focus on: writing posts people hit like on and read till the end (so write to be read) tl;dr: - X's open source code got new scores on Nov 5, the first change in 3 weeks - likes: 0.66 → 0.708 (highest yet) - time spent on a post: 0.006 → 0.0075 - tapping into a post: 0.3 → 0.274 - opening a video: 0.02 → 0 - new rule: AI replies may get flagged under big accounts as "llm slop" and may not reach the person's notifications lesson learned: a tap isn't the win anymore. Someone staying and liking is. full details: - The new numbers Up: likes 0.66 → 0.708, time spent on a post 0.006 → 0.0075, shares 2 → 2.118. Down: tapping into a post 0.3 → 0.274, staying 10+ seconds after a tap 0.4 → 0.385, opening a video 0.02 → 0. "Not interested" got a bit harsher, -52.27 → -53.07. Copy link, replies, quotes, DM shares and follows didn't move. X didn't explain any of this. It's just what the code says now. (I check the open source algo daily with Claude and share updates with you) - Look at where this has gone in a month In early October a like was worth 0.5. Now it's 0.708. Time spent was 0.004. Now it's 0.0075, nearly double. A tap was 0.3 and it's going down. Every change points the same way. X keeps paying more for people who stay. - Why that matters if you're small My read: you can't win on big numbers, but you can win on the people who do see your post. If they stay and like it, that counts for more today than it did last week. And those early likes also decide which new post gets the extra push. - Video and your profile X also switched on two new guesses in its ranking: how long someone keeps watching after they open a video, and how long they spend on a profile after tapping it. I can't tell you what those are worth. The only file that lists them is out of date, so I won't guess a number. My read: opening a video dropped to 0 the same day because X cares more about people who keep watching. So make the first seconds count, and make your profile worth sticking around on. - A new rule for replies under big accounts X runs an AI check on replies under accounts with over 250K followers. Flagged replies can get a label that lasts 14 to 30 days. One of the flags is literally called "llm slop" (AI slop, as I read it). That part isn't new. This is: X's rules now say a labeled reply or mention doesn't reach the person's notifications unless they follow you. Grey badge accounts and accounts X rates as highly credible skip this label. Small accounts get no such pass. I can see the rule, I can't see if every notification goes through it yet. Either way, write your own replies. - What didn't change Your new posts still get their extra push in the first 2 hours and first 200 views, and they still don't get cut first. Here's a recap of what you should be doing: - likes just hit 0.708, the most they've ever been worth - each second someone spends on your post is worth 25% more than last week - staying 10+ seconds after a tap earns 0.385 on top of the tap - reply in your own words, under big accounts especially - make your profile worth a visit, X now guesses how long people spend there - for video, hold attention early, just getting the open now scores 0 - being read to the end also helps X find you new readers - originals from accounts below 50K can jump to around spot 16 - early viewers liking it is what wins that lift - your first 2 hours and 200 views are the window, so publish when your people are online - the biggest single number is still 20, for someone copying your link - quotes, DM shares and replies earn 5 apiece - back and forth between mutuals is worth 20, so follow back your regulars - strangers only get shown your originals - the 1st like and the 32nd like each open another route to new people - a second post of yours in the same feed drops to 62%, so make one count - the video lists want clips over 10 seconds - +4 when a post is likely to earn you a new follower - your pinned post can affect the whole account, so take a look at it What you should NOT be doing: - pasting AI replies under big accounts, "llm slop" is a real flag at X - copy-paste replies, a duplicate label means they never get notified - chasing the tap, it's down to 0.274 and falling - off-topic posts, one likely "not interested" costs -53.07 - a video that only gets opened earns nothing for the open - publishing at 3am burns your 2 hour window - relying on replies and reposts to reach new people, they stay with your followers - getting an insult label, non-followers won't see that post at all - low quality and misinfo labels bring a warning, and non-followers can't like or share - adult media in 3 of your last 5 posts gets the whole account labeled - X rechecks videos for hidden content at 64 and 512 likes - piling your replies under a big account's post can sink your own score - one label on the account and non-followers stop seeing you - a link to a badly rated site in your pinned post risks a 7 day label - reusing your own wording word for word risks the copypasta flag - pairing a low quality link with tags of strangers looks like spam to X - engagement bait gets labeled no matter how big you are - posting several in a row, they fight each other for the same feed - hashtag stuffing does nothing - counting on bookmarks, they're worth 0 - expecting For You to show a post past 48 hours

655

619

970K

191

brelloAI retweeted

Rehan Temkar

@RehanTemkar

6 days ago

Before you let an AI agent change your files, check four things: 1. Can you see what it plans to change? 2. Can you limit which folders it can access? 3. Does it keep a record of what it changed? 4. Can you undo the changes if it gets something wrong? Try it on a copy of a small folder first. Check the results yourself before giving it access to anything important. A good demo shows what an agent can do. Your test should also show what happens when it makes a mistake.

485

brelloAI retweeted

Rehan Temkar

@RehanTemkar

7 days ago

10 GITHUB REPOS FOR YOUR CODING-AGENT TOOLKIT: 1.⁠ ⁠Aider https://t.co/7zxcDRvaOA 2.⁠ ⁠mem0 https://t.co/XQ2z2vupzE 3.⁠ ⁠LiteLLM https://t.co/0qcjWHBwy5 4.⁠ ⁠awesome-cursorrules https://t.co/YCZDzTOnr0 5.⁠ ⁠Langfuse https://t.co/t7W7j4Ba3B 6.⁠ ⁠GPT Pilot https://t.co/vH0TDkbTJy 7.⁠ ⁠Promptfoo https://t.co/famqgZ0hGv 8.⁠ ⁠Stagehand https://t.co/7DRNF8CBEl 9.⁠ ⁠SWE-agent https://t.co/Z6jT1IkuL4 10.⁠ ⁠E2B https://t.co/Z26N86vazx

379

brelloAI retweeted

Stooky

@Stookyit

7 days ago

X ALGORITHM: ATTENTION VS AI SPAM 🚨 X's published code separates ranking from visibility filtering. Mixing them up leads to bad growth advice. tl;dr: - ranking weights aren't a price list for reach - reading time matters, but padding isn't value - AI-related enforcement labels exist - a probable-spam screenshot doesn't reveal which label caused it full details: 1. The numbers need context Published ranking defaults include likes at 0.708, replies at 5 and shares via copy link at 20. These weights combine predictions about what a viewer might do. They don't mean one copied link guarantees more reach than 28 likes. 2. Attention isn't a length contest The repository describes systems that predict reading and time spent. That supports making posts worth reading. It doesn't support stretching one useful sentence into ten paragraphs. Put the point near the top. Use the rest to explain or demonstrate it. 3. Small-account boosts aren't guarantees The published defaults include a new-author boost with follower, impression and post-age thresholds. That is a mechanism, not guaranteed distribution. X runs experiments, and public defaults don't explain every individual result. 4. Video needs the same caution The model includes predictions for video continuation and profile visits. A zero video-open weight in one configuration doesn't mean video is worthless everywhere on X. Show the result early. Give people a clear reason to follow. 5. The AI labels are real Published enforcement rules include: - llm_slop_post: RiskyHighVizReply, 30-day duration - llm_slop_user: SpamHighRecall, seven-day duration Those are different rules at different levels. Neither proves that every AI-assisted reply restricts an entire account. 6. Don't diagnose from a screenshot “Probable spam” alone cannot tell you whether an AI label applied. Manual posting isn't proof of immunity either. And public configuration files don't promise a reset date. What I'd focus on: - specific observations - evidence and real examples - readable screenshots - useful demonstrations - a profile that matches the posts - replies that add something new What I'd stop doing: - generic praise - paraphrasing the original post - forced questions - invented personal experience - padding posts for reading time - treating every weak result as a restriction - presenting algorithm theories as platform rules The goal isn't to imitate whatever the algorithm might reward. It's to publish something worth reading.