Last month, X's developers shared the source code behind the platform's 'For You' recommendation system, with Elon Musk describing it as a step toward openness. Musk commented that the algorithm has flaws requiring significant enhancements, yet the public release allows observation of ongoing efforts to refine it. He added that no competing social platforms undertake such disclosures.
Although X stands alone among major social networks in open-sourcing parts of its recommendation engine, experts argue the disclosed material fails to provide meaningful insight into the platform's operations moving forward.
Similar to a prior 2023 disclosure, the latest code release is heavily edited, as noted by John Thickstun, a Cornell University computer science assistant professor. Thickstun expressed concern to Engadget that these publications create a false impression of openness, suggesting possibilities for external review or monitoring that simply aren't feasible.
Following the code's unveiling, X users quickly shared extended discussions on its implications for creators seeking greater exposure. One thread, seen by over 350,000 people, suggested engaging in discussions and uplifting the platform's energy to gain favor. Another, with more than 20,000 views, promoted video content as a key to success. A third advised maintaining focus on specific topics, warning that changing subjects reduces visibility. However, Thickstun advised skepticism toward these viral tactics, stating they exceed what the released code supports.
Certain minor elements do reveal aspects of post recommendations, such as excluding content older than 24 hours, but Thickstun described most of it as impractical for those producing content.
A major structural change from the 2023 version involves using a large language model akin to Grok for ranking. Ruggero Lazzaroni, a PhD candidate at the University of Graz, explained that the earlier system computed scores directly from metrics like likes, shares, and replies. The updated approach instead estimates user interactions based on the model's predictions of engagement likelihood.
This shift increases the system's inscrutability, according to Thickstun, as key decisions now occur inside opaque neural networks trained on proprietary datasets. He noted that such processes are increasingly inaccessible not only to outsiders but even to the engineers maintaining them.
Compared to the 2023 release, the new version omits details on interaction weightings that previously influenced rankings, such as equating one reply to 27 retweets or an author-responded reply to 75 retweets. X justified these omissions citing security concerns.
Additionally, the code provides no details on the training data for the model, limiting potential analyses or bias checks. Mohsen Foroughifar, a Carnegie Mellon University assistant professor in business technologies, emphasized the need for visibility into training datasets, noting that biased inputs could perpetuate biases in outputs despite other safeguards.
Studying X's recommendation system would offer substantial research value, particularly for Ruggero Lazzaroni, involved in an EU project developing alternative social media algorithms. His simulations of platforms require detailed replication, but he found the X code insufficient for accurate emulation.
Lazzaroni pointed out that while the code exists to execute the system, the essential model weights or parameters are absent, preventing full operation.
Access to the full X algorithm could inform broader applications beyond social media, including AI chatbots facing similar issues. Thickstun observed that problems in social recommendation systems mirror those emerging in generative AI interactions.
Lazzaroni, focused on modeling harmful online behaviors, criticized AI firms for prioritizing engagement over accuracy or user well-being to boost profits, echoing social media dynamics that enhance revenue at the expense of societal or mental health outcomes.