GitHub Copilot: the AI pair programmer for today and tomorrow

Ryan J. Selva, VP of Product at GitHub, is welcomed to the stage. Despite a hoarse voice, he enthusiastically greets the audience of developers and introduces himself and his role in overseeing GitHub Copilot. Selva reflects on nearly three years of work on GitHub Copilot, positioning it as a pioneering product in the large language model space. He emphasizes the research-led approach of the product, focusing on code quality, developer happiness, productivity, and economic impact. Selva also mentions a recent study conducted with external companies to understand the real-world application and benefits of GitHub Copilot in engineering teams, inviting Dan Shockey from Accenture to discuss the research findings.

Dan Shockey from Accenture provides insights into his company's collaboration with GitHub on research regarding the impact of GitHub Copilot. Accenture, a company specializing in helping clients build their digital core and transform operations, has a vast team of software engineers. Shockey explains their interest in AI-assisted software delivery and the decision to enable GitHub Copilot for their developers. The research involved enabling a few hundred internal developers to use Copilot and observing the results. The findings were significant, with developers reporting time savings, increased build success rates, and improved code quality. Surveyed developers found Copilot useful, with half considering it extremely useful. Accenture plans to roll out Copilot to all developers and continue the partnership with GitHub.

The discussion shifts to the future of GitHub Copilot, with a focus on its integration into the software delivery lifecycle beyond the code editor. Harald, a product manager from the VS Code team, joins the stage to demonstrate how they've been working on enhancing the developer experience with GitHub Copilot within the IDE. Harald showcases the capabilities of Copilot Chat, emphasizing its ability to customize to individual coding styles and streamline workflows. He also explains the hybrid approach Copilot uses to provide relevant code snippets, including leveraging GitHub's knowledge graph, local indexing, and VS Code's language intelligence. The presentation highlights the potential for custom models and the extensibility of Copilot through plugins and integration with third-party developer tools.

The conversation delves into the technical aspects of GitHub Copilot's functionality, particularly the use of GPT-4 and other models like Jupiter 3.5 Turbo for a balance of speed and intelligence. The multi-model approach is a result of continuous experimentation to deliver a cutting-edge AI experience without compromising on responsiveness. The speaker demonstrates how Copilot Chat can be used for tasks like code correction and committing changes, showcasing the tool's ability to understand and interact with the codebase. The session also touches on the importance of flow in development and how Copilot's commands and features aim to reduce interruptions and improve efficiency.

The presentation moves on to discuss the concept of custom models in GitHub Copilot. The speaker explains how personalization of AI models can be achieved through fine-tuning, which has been a focus of experimentation over the past year. The process involves using repository data, developer feedback, and reinforcement learning to improve the models' suggestion quality and introduce team-specific biases. The speaker also highlights the importance of working with external partners like AMD to diversify datasets and further enhance the models. Alex from AMD shares his team's experience with fine-tuning a model for their specific needs, including the successful creation of a state machine from a plain English spec, showcasing the potential of customized models in accelerating development cycles.

The speaker, XI, outlines the vision for GitHub Copilot as an ecosystem with a marketplace of plugins, integrating with various third-party developer tools and services. He emphasizes the potential for productivity gains by customizing Copilot suggestions to a codebase and inviting the developer community to contribute to the evolution of Copilot. The presentation includes examples of how extensibility can be achieved by bringing in internal data and building plugins. The speaker also discusses the process for bringing private datasets into Copilot for more tailored suggestions and the development of plugins that can be made public or private. The session concludes with a call to action for developers and integrators to join the GitHub Copilot partner program and contribute to shaping the future of the platform.

The session concludes with a note of gratitude from the GitHub Copilot team towards their members and the broader developer community. The team expresses appreciation for the feedback and support received from developers, which has been instrumental in driving the development of Copilot. They encourage continued engagement and collaboration with the community to enhance the tool and build a future where developers can enjoy using Copilot as much as the team enjoys creating it.