Meta recently announced their development of a new AI model called CodeCompose that uses generative AI to create code similar to GitHub’s Copilot. At an internal event focusing on their AI infrastructure, Meta revealed that their software engineers already use this program to create code for Python and other languages in their IDEs (Integrated Development Environment) like VS Code. It has been said that this tool has already produced 6.7 billion parameters, compared to Copilot’s 13 billion. CodeCompose assists coders in producing annotations, import statements, and an entire line of code, while also having access to the company’s internal libraries and frameworks powered by Hack, a language created by Meta itself.
Commenting on the tool, Michael Bolin, a software engineer at Meta, expressed that they “are able to integrate CodeCompose into any surface where our developers or data scientists work with code”. The model then goes on to make suggestions based on the surrounding code and code comments.
Meta claims that thousands of employees make use of CodeCompose weekly, and its recommendation acceptance rate has exceeded 20%. However, its use has raised suspicions due to copyright law, which is the case with Copilot as well. Also, using AI generated code runs the risk of producing unsafe or insecure code, as indicated by a Stanford study. Though Meta has promised that the model was only trained on public data from their research team, they have yet to fully address any potential copyright infringement and security concerns.
