In a recent event centered on Meta's AI infrastructure progress, the company disclosed the development of an AI-powered code generation tool named CodeCompose. This innovative tool shares similarities with GitHub's renowned Copilot offering. Though Meta has not yet made CodeCompose publicly available, the company says its internal teams already utilize the tool for receiving programming suggestions in Python and other languages when working in integrated development environments (IDEs), such as VS Code.
Michael Bolin, a software engineer at Meta, stated that the underlying model of CodeCompose is founded on the company's public research and tailored to suit internal use cases and codebases. Bolin also mentioned that CodeCompose could integrate into any surface, allowing developers and data scientists to work with code more efficiently.
The largest CodeCompose model that Meta has trained possesses 6.7 billion parameters, slightly more than half of Copilot's parameter-filled model. Parameters function as crucial parts of the model, which are learned from historical training data and define the model's competence level concerning a problem, such as generating text.
CodeCompose was fine-tuned using Meta's first-party code and internal libraries and frameworks written in Hack, Meta's in-house programming language. By doing so, CodeCompose can integrate these components into its programming suggestions. The tool's base training dataset has been meticulously cleaned of errors and poor coding practices, such as deprecated APIs, to reduce the likelihood of it recommending a problematic piece of code.
In everyday use, CodeCompose recommends annotations and import statements as developers type. Furthermore, the tool can complete single or multiple lines of code, even filling in large chunks of code when needed. Bolin claimed that CodeCompose harnesses the surrounding code and code comments to provide more accurate suggestions.
According to Meta, thousands of its employees use and approve suggestions from CodeCompose every week, with the acceptance rate exceeding 20%. However, the company has not yet addressed the controversies surrounding code-generating AI tools. Platforms like Copilot have faced copyright issues, accused of regurgitating licensed code without proper credit. These concerns raise questions about whether CodeCompose might be susceptible to similar troubles.
Another vital topic is the potential of generative coding tools inadvertently introducing insecure code. A recent Stanford study discovered that software engineers using AI-generated code systems are more likely to create apps with security vulnerabilities. Although the research did not examine CodeCompose, it is reasonable to assume the same risks would apply.
Bolin emphasized that developers do not have to follow CodeCompose's suggestions and that security was a major consideration in developing the model. He added that the company was enthusiastic about the tool's progress and that their developers would benefit from developing such a solution in-house.
Tools like CodeCompose could potentially be integrated into platforms like AppMaster, one of the leading no-code / low-code application development platforms. AppMaster focuses on delivering powerful tools for creating web, mobile, and backend applications while improving the development process in terms of cost-effectiveness and speed.
