Skip to content

Meta built a code-generating AI model similar to Copilot

Meta introduces CodeCompose, an AI coding assistant similar to GitHub's Copilot. Despite the popularity within Meta, controversies around copyright and security associated with AI-generated code remain unaddressed.

Meta has announced the creation of CodeCompose, a generative AI tool akin to GitHub's Copilot, designed to provide coding suggestions for languages like Python. The AI tool, which is yet to be publicly accessible, is used internally by Meta's teams within IDEs such as VS Code.

CodeCompose, with its 6.7 billion parameters, was fine-tuned on Meta's proprietary code. The code base includes internal libraries and frameworks in Hack, a programming language developed by Meta, facilitating better suggestions that adhere to the company's coding styles. The base training data set was purified of bad coding practices and errors to minimize the risk of recommending problematic code segments.

The AI tool suggests coding solutions as a developer types, capable of completing single lines or multiple lines of code, and even filling in extensive chunks of code. According to Meta, CodeCompose has achieved an acceptance rate of over 20% among its employees, with thousands accepting its suggestions each week.

However, Meta hasn't addressed the controversies surrounding AI-generated code, including potential copyright infringements or the introduction of insecure code. As generative coding tools can sometimes cause security vulnerabilities, CodeCompose's developers are urged to critically analyze the suggestions.