Elon Musk: X Platform Algorithm Can Open 'Under the Hood' Page
Elon Musk retweeted X's open-source account, stating that a more readable "Under the Hood" page has been launched at user request, and JSON can still be downloaded. The entry point is the platform's under the hood report. He claims this is an easy way to understand how X's algorithm works. This tool is paired with the "For You" recommendation sorting code released in August: the code explains how tags affect visibility, and the report shows which tags were applied to one's account and posts last month.
Eligibility requires an account to be at least one year old and to have posted no fewer than 10 posts in the previous month. In August, it was piloted with a random test group, and starting September 19, eligible users can view it, with new information on whether their reach was limited or hidden by country due to local laws. The GitHub repository xai-org/x-algorithm is released under Apache 2.0, covering model configuration, filtering, and core sorting, expanding in size by about 10 to 15 times compared to previous open-source releases. Product Vice President Keith Coleman stated that external researchers can now train and run Phoenix scoring locally. Sorting predicts various behaviors, with the weight of shared links analyzed to be approximately 40 times that of likes; replies and retweets have additional discounts, and standalone posts are easier to generate than those buried in threads.
Musk admitted the algorithm needed a complete overhaul after Andrej Karpathy criticized the information flow as "toxic, like Reddit" earlier this year. Open-sourcing and personal tagging transform the accusation of "shadow banning" into a self-checklist. The repository also reveals filtering based on national laws, for example, accounts related to the Brazilian elections are removed from the information flow when not followed. Non-technical users are guided to submit JSON and the repository to a large model for interpretation. The system summarizes tags daily, and the service layer aggregates periodically.
Mechanically, users buy the right to self-check the reasons for their reach limitations, while the platform sells the old model of denying shadow limiting without providing evidence. The event-driven approach replaces pure JSON with web reports, lowering the threshold for viewing. Attention shifts from guessing weights to verifying tags and legal orders; benefiting are accounts that can compare their traffic with the code explanations, while pressured are content strategies that rely on opaque downgrading, as well as jurisdictions that must be named in reports. Open-sourcing the code does not mean real-time synchronization of online weights; reports are monthly aggregates, not single-post real-time debuggers.
Source: Public Information
ABAB AI Insight
X breaks down the recommendation system into two parts for users: Phoenix tells you "how the system ranks," while Under the Hood tells you "what tags you have received." Musk positions the product as an algorithm manual in one sentence, avoiding further disputes over "whether there is shadow banning." The 10-post threshold keeps the tool for active producers, while the one-year account age blocks new accounts from gaming the reports. The legal reach limitation entering the report transforms government orders from backend tickets into visible fields for users, also shifting compliance costs from the PR department to the product page.
The capital path is trust exchanged for retention. Both advertising and subscriptions rely on the duration of "For You" recommendations; a toxic algorithm will drive away high-value users. Open-sourcing allows researchers to run Phoenix themselves, effectively outsourcing part of the R&D review; if pull requests are accepted, external engineers work for free. The weight of shared links is much higher than likes, indicating that the product seeks external dissemination rather than internal hearts. Creators will adjust their formats according to weight: independent long posts, fewer buried replies, and more guidance to share links.
Similar to YouTube's restricted tag explanations, TikTok's creator academy, and the EU's required recommendation transparency reports: all use downloadable data to replace verbal denials. The difference is that X places the core of sorting on GitHub and allows comparison with personal tags. The industry phase shifts from "algorithms are not to be spoken of" to "algorithms can be checked but weights are still modified by the company." Those who can understand the mapping of tags to code will write less about natural traffic drops as political persecution.
Structural judgment belongs to the overlap of regulatory changes and technological replacements. Pricing power partially shifts from the platform's explanation of reach limitations to a user-verifiable tag list. The mechanism is that the recommendation system relies on tags for filtering; when tags are opaque, all drops in traffic appear as suppression; once tags are visible, the debate shifts from whether there is downgrading to whether a particular tag should exist. Legal orders entering the report turn national filtering from a hidden switch into a statistically measurable product feature.
ABAB News · Cognitive Laws
After the algorithm is open-sourced, reach limitations must first provide tags, then excuses.
The weight of shared links is higher than likes; the platform seeks external dissemination.
Once legal removals are written into personal reports, jurisdictions become product features.