In recent years, AI developer tools have progressed from research projects to critical parts of the developer toolset. Today, there are over 1.8 million paid subscribers to GitHub Copilot, Google reports that 50% of their code is written with AI assistance, and TabNine estimates that 5% of all code is written by AI. These numbers will only increase as these tools continue to improve and gain further traction among developers.
Researchers in the machine learning, programming languages, and software engineering communities continue to push the boundaries of what is possible, but turning promising research results into tools that are useful at scale is hard. It involves determining the right use cases, designing the right UX, evaluating and improving quality, and many other factors.
In this talk, I will discuss how to bridge the demo-to-real-world gap via lessons learned from building several AI developer tools at GitHub, including the original GitHub Copilot, Copilot for Docs, and Copilot Workspace.
Eddie Aftandilian is a Principal Researcher at GitHub Next, GitHub’s advanced R&D team. He was one of the original members of the GitHub Copilot team, for which he led quality measurement and improvement. He led the Copilot for Docs project, an early retrieval-augmented generation system that can answer questions about private technical documentation. Currently he leads the Copilot Workspace project, an agentic system that helps developers solve issues and address feature requests across an entire source code repository.
