A Program That Writes Computer Code as the Programmer Types

GitHub, the site where most of the world's software is stored, is testing a tool called Copilot that suggests whole lines and functions of code based on what a programmer has written so far. It was built with OpenAI and trained on public code.

GitHub — the Microsoft-owned website where programmers keep and share their work, and which holds most of the world's open-source software — began testing a tool today called Copilot. As a programmer types, it suggests what comes next: the rest of the line, the rest of the function, sometimes an entire block of code, based on a comment describing what the code should do. The programmer accepts the suggestion with a keystroke or keeps typing.

It was built with OpenAI, using a version of the lab's GPT-3 writing program trained on the billions of lines of public code stored on GitHub itself. In that sense it works the way a text-message autocomplete works, but for programs, and at the scale of whole functions rather than words.

GitHub is careful about the claims. Copilot's suggestions are sometimes wrong, sometimes subtly wrong, and occasionally insecure, and the company says a programmer should review everything it produces the way they would review a colleague's draft. The tool is available for now only to a waiting list of testers.

Two questions came up within hours of the announcement. The first is whether a tool that learned from other people's code is, in effect, copying it — GitHub says the suggestions are generated, not retrieved, and that verbatim copies are rare. The second is what it means for the job of programming if a machine can write the routine parts. Neither has an answer yet.