Twitter has uploaded code fragments to Github that are responsible for recommending tweets on the social network
Twitter © pixabay.com
This will allow users and developers to evaluate the algorithm's performance and suggest improvements. Reuters reports that the code was uploaded to two repositories on the Github website and includes various parts of Twitter's source code, including the recommendation algorithm for users.
The code is open under the AGPLv3 license. The implementation used the programming languages Scala (53.8%), Java (29.7%), Starlark (6.3%), Python (4.7%), C++ (2.4%) and Rust (1.5%). The code associated with the machine learning models used on Twitter is published in a separate repository (the models themselves are not published for security and privacy reasons).
The company emphasized that the repositories do not contain code related to advertising recommendations.
This move was made at the direction of Twitter CEO, Elon Musk. He stated that the transparency of the code will increase users' trust in the social network and will contribute to the rapid improvement of the product. It will also help address common concerns of users and legislators, who have begun actively scrutinizing how social media algorithms select content recommended to users.
Musk also stated that Twitter will update its recommendation algorithm based on user suggestions every 24-48 hours. He held an audio chat session on Twitter with other company employees, during which users could ask questions and give recommendations about how the social network's code works.
One participant in the session asked why Twitter's code classifies users as Democrats or Republicans. A company employee responded that this is an old function that does not play an important role in the recommendation system and that they plan to remove it.
Elon Musk said that soon, in addition to the information already published, he would reveal "literally all" the source code associated with issuing publications to users. The entrepreneur wrote about this on his page.
