GitHub’s OpenAI-Powered Developer Tool with New Capabilities

GitHub's OpenAI-powered developer tool with new capabilities goes live with enhanced Codex

The pair programming plugin for GitHub Copilot for Business, powered by OpenAI Codex, is now publicly available. It comes with an upgraded version of OpenAI Codex and a new real-time vulnerability filter that detects typical security flaws as you code in the editor.

Copilot for Business, an extension of the current Copilot for Individuals, was introduced by the Microsoft-owned code repository service in November. Each user of the program pays US$10 each month. Depending on the number of allocated Copilot seats, the monthly membership fee for Copilot for Business is US$19 per user.

Copilot's code ideas are powered by OpenAI's Codex, which converts natural language into code. With different editors, such as Microsoft Visual Studio, Neovim, VS Code, or JetBrains IDEs, Copilot may be utilized.

According to GitHub, Copilot now has an updated Codex model and additional features. Additionally, it has created a security vulnerability filter to enhance the security of Copilot's code recommendations and aid developers in quickly identifying unsafe coding practices.

The new Codex model, which Microsoft is providing to developers through Azure OpenAI Services together with GPT-3.5 and DALL.E 2, should lead to a bigger proportion of code being authored by Copilot.

According to GitHub, when Copilot for people first debuted last year, on average, 27% of developers' code files were produced by Copilot. Currently, the average for this statistic is 46% across all programming languages that are supported, while it is 61% for Java.

Although Copilot's code acceptance rate is lower, it is still increasing. Developers typically accepted 27% of proposals in June 2022. In September, it increased to 30%, and by December, it had reached 35%.

Along with adopting a newer Codex model, Copilot also received the Fill-In-the-Center (FIM) "paradigm," which improves upon the previous approach of simply considering the code prefix to account for known code suffixes and leaves a hole in the middle for Copilot to fill.

"This gives it additional background information about the code you want to write and how it should work with the rest of your program. We've created many techniques to offer FIM in GitHub Copilot's code recommendations without any additional delay, and it regularly results in higher quality code suggestions "explains Shuyin Zhao, senior director of product management at GitHub.

A "lightweight client-side model" that learns about the user's context has also been included in GitHub's new VS Code addon to lessen the frequency of unwelcome recommendations. According to GitHub, there were 4.5% fewer unwelcome ideas as a consequence.

Large language models (LLMs) are also used by the new vulnerability filter to "approximate the behavior of static analysis tools." Insecure coding patterns, such as hardcoded credentials, SQL injections, and path injections, are among the extremely prevalent security problems that GitHub claims it can identify in real-time and stop.

Microsoft could be able to increase the number of users who utilize GitHub by offering Copilot for people and corporations. According to a recent statistic, GitHub has 100 million members, which is far greater than most estimates of the world's developer population. According to Thomas Dohmke, CEO of GitHub, the developers no longer work for software firms.

"They're a growingly diverse and international community of people working across sectors, fiddling with code, design, and documents in their spare time, contributing to open source projects, doing scientific research, and more," said Dohmke.

"They are individuals working on projects such as the PyTorch project, which drives AI and machine learning applications, as well as software for NASA, hospitals, and the movie industry. They also seek to support family members in communicating and recovering from the disease."