The day Copilot arrived in general availability on June 21, 2022, was a milestone for the AI-assisted coding tool. After a year in technical preview, the product had come a long way from its experiment origins. Developers were overwhelmingly enthusiastic about the commercial product.
What Copilot learned from its technical preview is a key factor in its success. From June 2021, the preview gave GitHub valuable insights into how developers interact with AI code suggestions. This data helped refine the GA model and user experience. As a result, Copilot's general availability saw better completion rates for common languages and improved test generation compared to the early preview.
One notable improvement was in Python, where Copilot's completion rate increased by 15% compared to the preview. For Java, the improvement was around 10%. These numbers came from a combination of user feedback and GitHub's own analysis of Copilot's performance. The company used tools like TensorFlow and Kubernetes to monitor and optimize the model's performance.
GitHub priced Copilot at $10 per month to be accessible to individual developers without the need for corporate procurement. This pricing strategy reflects GitHub's approach: individual adoption drives enterprise adoption. When a developer uses Copilot personally, they're more likely to bring it to their team, making it a purchasing conversation.
OpenAI Codex is the AI model behind Copilot, derived from GPT-3 and fine-tuned on code. The model was trained on public code repositories on GitHub, raising questions about licence compliance that were already in legal proceedings at the time. GitHub and Microsoft chose OpenAI over an internal model due to the acceleration-at-the-cost-of-control trade-off.
In terms of trade-offs, GitHub had to balance the benefits of using a third-party model like OpenAI Codex against the risks associated with IP ownership and data control. For instance, training the model on public code repositories meant that some code snippets might be derived from GPL-licensed projects, potentially introducing licensing complexities. GitHub addressed these concerns through a combination of legal reviews and internal policy updates.
Enterprise response to Copilot's general availability was initially cautious. Security and legal teams raised concerns about IP ownership, suggested code derived from GPL-licensed repositories, and code containing company-confidential patterns being sent to an external API. These concerns were legitimate and took time to address. By the end of 2022, most large enterprises had either approved Copilot for development use or had policies defining its conditions of use.