The paper introduces CodeCompose, an AI-assisted code authoring tool using generative capabilities and bi-directionality to suggest entire statements or blocks of code when building software. The paper discusses the unique challenges that arise when deploying such tools in large-scale industrial settings and presents metrics that reveal the tool’s impact on Meta’s internal code authoring experience, with 4.5 million suggestions made over a 15-day window. Quantitative metrics show a 22% acceptance rate and 8% of code typed by users coming from suggestions made by CodeCompose, while qualitative feedback indicates a 91.5% positive reception. CodeCompose is also shown to encourage in-code documentation and aid the discovery of new APIs.
https://arxiv.org/abs/2305.12050