Meta has revealed that it has created an AI coding model dubbed CodeCompose. The corporation unveiled the tool during a gathering highlighting its AI infrastructure initiatives. It also includes bespoke chips that quicken the training of generative AI models. CodeCompose, according to Meta, is comparable to GitHub’s Copilot. It suggests code to users in IDEs like VS Code as they type in Python and other languages.
Also Read: Microsoft Power Platform Copilot: No Coding Era Is Coming
How Does CodeCompose Work?
On Meta’s first-party code library, which consists of internal libraries and frameworks written in Hack, a programming language by Meta, CodeCompose was refined. CodeCompose can now propose annotations, import statements, single or multiple lines of code, even significant sections of code, and many more. The system may benefit from nearby code and use comments as signals while producing new code. Furthermore, with 6.7 billion parameters, the largest of the CodeCompose models that Meta trained is comparable to the model that Copilot is built on.
Meta’s Acceptance Rate and Controversies
Meta claims that thousands of employees use CodeCompose every week, with an acceptance rate of more than 20%. The corporation, however, has not addressed the debates around code-generating AI. A class action complaint accuses Microsoft, GitHub, and OpenAI of breaking copyright rules by allowing Copilot to reuse portions of licensed code without giving proper acknowledgment. Legal experts have also asserted that Copilot-style AI can expose businesses to liability if they use the tool’s copyrighted recommendations to improve their production software.
Also Read: Bring Doodles to Life: Meta Open-Sources AI Model
Unclear Training Data
Whether CodeCompose trained on licensed or copyrighted code unintentionally or on purpose is unknown. A Meta representative revealed that they taught CodeCompose using InCoder, and Meta’s AI research section made it public. The publication describing InCoder states that they trained the model using open-source, permissive general code.
Our Say
Developers may write code more quickly and effectively with the help of CodeCompose, an AI coding model from Meta. Meta has not addressed the issues of copyright infringement or potential responsibility with such AI-based tools, despite parallels to GitHub’s Copilot. CodeCompose will likely perform differently once made accessible to the general public. Still, it looks like a promising addition to Meta’s AI-powered technology lineup for now.
Also Read: Google’s Duet AI: A Competitor to Microsoft’s GitHub Copilot
