AI for code generation has seen significant breakthroughs over the last few years. Large language models such as OpenAI Codex, are now capable of producing large chunks of coherent code.

This is unlocking a new generation of developer tooling, where tools like GitHub Copilot augment a programmer’s capability by pairing with them inside their IDE.

But there is a long way from a model that predicts code to a transformative AI pair programmer. We discuss where the challenges lie, and how they can be overcome.

Albert Ziegler is a Staff ML engineer at GitHub Next. After his PhD in Pure Mathematics and Computability Theory, he’s turned to applying machine learning to the software development process, researching coding productivity at Semmle and then joining the ML-on-code group at GitHub. He has been the Copilot project’s resident machine learning expert since its inception.