GitHub Copilot technical preview is announced

Background

Language models trained on ordinary text can also be trained on programming code. A model that has learned code patterns can suggest what a programmer might write next. GitHub said that OpenAI's Codex was significantly more capable than GPT-3 at code generation, partly because it was trained on a larger concentration of public source code.^[1]

What happened

On 29 June 2021 GitHub announced GitHub Copilot as a technical preview, describing it as an AI pair programmer. The announcement post was written by Nat Friedman.^[1]

Copilot reads the code being edited and suggests whole lines or entire functions. GitHub said it also helps find alternative solutions, write tests and explore new APIs.^[1]

Copilot was powered by OpenAI Codex, a system created by OpenAI.^[2]

Key details

GitHub said the preview worked especially well for Python, JavaScript, TypeScript, Ruby and Go, and that spots in the technical preview were limited.^[1]

On 10 August 2021 OpenAI released an improved version of Codex through its own API in private beta.^[2]

Copilot became generally available on 21 June 2022 at $10 per month or $100 per year. It was free for verified students and maintainers of popular open-source projects.^[3] At general availability it supported Neovim, JetBrains IDEs, Visual Studio and Visual Studio Code.^[3]

GitHub said more than 1.2 million developers used the technical preview over its 12 months. It also said that in files where Copilot was enabled, nearly 40% of code was being written by Copilot in popular languages such as Python.^[3]

Autocomplete tools that suggest the next word or name in a program existed long before this preview. What was different here was the size of the suggestion: a model could propose a full line or an entire function based on the surrounding code and the comments a programmer had written.

The pair programmer description reflects how the tool was meant to be used. The programmer stays in charge, reads each suggestion, and accepts, edits or rejects it. GitHub's mention of alternative solutions, tests and new APIs points to the same idea: the tool offers options, and the human decides.

A technical preview is an early version offered to a limited audience for testing, so features and terms can change before a general release. The description in this article reflects the preview as GitHub announced it.

Impact

Copilot brought a code-generating model into everyday programming tools, so developers had to judge each suggestion for themselves. The usage figures and the 40% share come from GitHub's own reporting and are not independent measurements.

The figures GitHub reported describe a particular set of languages and a preview that was limited to a small number of users at first, so they are best read as descriptions of that rollout rather than general measures of code quality.^[1]