The open-source software developer GitHub says that for some programming languages, about 30% of newly written code is being suggested by the company’s AI programming tool Copilot.
Copilot can look at code written by a human programmer and suggest further lines or alternative code, eliminating some of the repetitive labour that goes into coding.
Copilot is built on the OpenAI Codex algorithm, which was trained on terabytes of openly available source code and can translate human language into a programming language. It serves as a more sophisticated autocomplete tool for programmers.
Oege de Moor, VP of GitHub Next, the team rolling out Copilot, commented:
“We hear a lot from our users that their coding practices have changed using Copilot. Overall, they’re able to become much more productive in their coding.”
The company will announce at its GitHub Universe conference today that it will be rolling out Copilot support for all popular programming languages, including Java. “This is going to help bring this technology to a much broader audience,” says de Moor, adding that it is part of GitHub’s effort to “make programming accessible to the next 200 million developers.”
De Moor also added that Copilot has proven sticky with the community’s base (50% of the developers who have tried the product since its launch in July has kept using it). Not unlike OpenAI’s massive text-generating natural language product GPT-3, Copilot is much more effective in augmenting human work than in creating its own code.
Like any algorithm, it is dependent on the quality of its training data. In a study, a group of academics from New York University found 40% of the code produced by Copilot had cybersecurity flaws. However, humans are far from perfect either — by one estimate, the average developer creates 70 bugs per 1,000 lines of code.
Even as Copilot improves, human programmers won’t be out of a job. Demand for software developers grew 25% in 2020, and most programmers spend less than half of their working time actually writing code.