GitHub launches Copilot in technical preview

GitHub opened a technical preview of Copilot, an editor extension that suggested lines and whole functions as developers typed. It was powered by Codex, a GPT-3 descendant that OpenAI had fine-tuned on public source code; the accompanying paper, published a week later, introduced the HumanEval benchmark and reported that the model solved 28.8% of its problems on a first attempt.

Copilot was the first large language model that a large population used daily for real work. Its reception established several arguments that recurred for years. Developers reported genuine productivity gains on boilerplate and unfamiliar APIs, alongside confidently wrong suggestions that were harder to catch than obvious errors. Security researchers at NYU found that roughly 40% of Copilot’s completions in security-relevant contexts contained vulnerabilities. And because the model had been trained on public repositories, including copyleft-licensed ones, its output raised an unresolved question about whether generated code could infringe the licences of the code it learned from — a question that reached court in a 2022 class action against GitHub, Microsoft and OpenAI.

The product also demonstrated a commercial pattern. Copilot moved from preview to a paid subscription in June 2022, making it one of the first language-model products with per-seat pricing and a measurable business, and it gave Microsoft an early argument for the OpenAI partnership it had begun funding in 2019.

More than the model, it was the interface that carried. Inline, opt-in, always-suggesting completion became the default shape for AI in professional tools, and the word “copilot” became Microsoft’s company-wide brand for the category.