X open sources its “core recommendation system”, which powers For You, on GitHub, and says the system “ranks everything using a Grok-based transformer model”
In 2023, the website then known as Twitter partially open sourced its algorithm for the first time.
Developers and researchers can inspect the published core recommendation system on GitHub, including X's stated use of a Grok-based transformer model for ranking.
Grok becomes more explicitly tied to X's primary distribution mechanism, not only to AI features previously introduced on the platform.
Second-order effects
Scrutiny can shift from broad claims about algorithmic openness to whether the released code reflects live ranking behavior—a gap made consequential by the data and system complexity behind recommendations.
Creators, researchers, and competing platforms gain a concrete reference point for assessing how a model-led ranking stack is structured, even if the release alone cannot reveal all production inputs.
Third-order effects
If this approach is sustained, recommendation transparency may increasingly mean disclosing model architecture alongside code, while the data, tuning, and deployment controls remain the harder governance questions.
The release points toward AI models becoming shared infrastructure for both content creation and content distribution, concentrating strategic importance in the model layer.
The trend: Social platforms are embedding proprietary AI models deeper into feed distribution while using open-source releases to make parts of that infrastructure inspectable.
@XEng The code is out. The secrets are gone. Here is how you go viral: Bookmarks are Gold: A “Save” is now a 50x multiplier. If your post isn't worth referencing later, it's invisible. The 75x Rule: A reply + your response is the strongest signal in the code. Ghosting your commen…
@XEng The only things you need to know: Every account has a base score that's applied before a tweet is evaluated. Verified accounts: +100 automatically Unverified accounts: up to +55 That base score is then multiplied by: Your follower-to-following ratio (only meaningful if it's
@XEng Cursory review of the algorithm code: - algorithm predicts engagement signals (likes, reposts, replies), and naturally more engagement is higher visibility, so early engagement matters - content from accounts followed is prioritised, out-of-network is ML predicted - transfo…
@XEng Is the new open source algorithm fair? Or even good? Here Grok analyzes my posting behavior and says that I'm doing everything wrong if my goal is to get reach: https://x.com/... It suggests that I will need to shift toward reply-bait hooks, stronger emotional pattern
If you're curious to explore the new algorithm: https://deepwiki.com/... Notably there are no new constants (base scores) published in the repo. But doesn't seem to stop people from claiming to know all of them anyway! https://deepwiki.com/...
I read the code so you don't have to. here's the breakdown: 1. reply to your comments. the algo weighs “reply + author response” 75x higher than a like. ghosting comments = strangling reach. 2. links kill visibility. put them in bio or pinned. never in the post. 3. watch time is …
@XEng So, we have to boring one trick ponies thanks to SimClusters. If we speak outside of whatever lane the AI has put us in then our posts go nowhere. That's boring. That's discouraging. And no matter how hard you work, trolls have full control over this entire platform. [image…
We know the algorithm is dumb and needs massive improvements, but at least you can see us struggle to make it better in real-time and with transparency. No other social media companies do this.
Assuming this is actually used, these appear to be the parts of a tweet that factor into the algorithm. Grok calculates these through god knows what. There's no transparency there. Note that whether a tweet has a link or says “shop” or anything like that seems to have no effect. …
ML diagram of twitter's algorithm overall nothing too surprising. no content features, no follower count, and isolation between potential candidate tweets shame about algos all being out-of-network discovery though. it'd be cool to see tweets from people i like and follow.. [imag…
🚨X JUST DID WHAT NO OTHER PLATFORM HAS THE GUTS TO DO... X released the source code for its new 𝕏 algorithm. Every major platform has hidden behind “proprietary technology” excuses for years. X just opened the black box. The code reveals a sophisticated system powered by [image]
Transparency... Maybe the algorithm is open-source, but there are points, like the payout system, which are still a black box. I am not necessarily asking for open sourcing it (among others), but we need clarity at least.
The smartest play would be to switch the algo off. Just go back to human networks. Right now it is engineered for dopamine addiction and exploitation and content creation slavery.
I feel like the algorithm needs to be 90% posts from those you follow with sporadic similar posts sprinkled in. If you want to see other stuff, search and follow people that tweet about it. That's how most social media started, and any further control makes it feel unnatural.
I appreciate the transparency here, I really do. But all we're asking is you go back to whatever the algorithm and platform were doing circa 2020 - without the overt censorship. And keep community notes.
Kudos for transparency! But now getting serious: 1. Now can we see the parts that are hidden for “security” reasons? 2. Can we get database parameters for throttling? 3. Does that database itself have a custom algorithm for serving, where an IT Ops person can easily