AI-assisted Code Authoring at Scale: Fine-tuning, deploying, and mixed methods evaluation
This work addresses the problem of enhancing developer productivity through AI code generation at a large tech company, representing an incremental deployment and evaluation effort.
The paper tackles the challenge of deploying an AI-assisted code authoring tool, CodeCompose, at scale for developers, achieving improvements in code generation accuracy by 1.4x to 4.1x over public data models and adoption by 16K developers with 8% of their code generated by the tool.
Generative LLMs have been shown to effectively power AI-based code authoring tools that can suggest entire statements or blocks of code during code authoring. In this paper we present CodeCompose, an AI-assisted code authoring tool developed and deployed at Meta internally. CodeCompose is based on the InCoder LLM that merges generative capabilities with bi-directionality. We have scaled up CodeCompose to serve tens of thousands of developers at Meta, across 9 programming languages and several coding surfaces. We present our experience in making design decisions about the model and system architecture for CodeCompose that addresses these challenges. To release a LLM model at this scale, we needed to first ensure that it is sufficiently accurate. In a random sample of 20K source code files, depending on the language, we are able to reproduce hidden lines between 40% and 58% of the time, an improvement of 1.4x and 4.1x over a model trained only on public data. We gradually rolled CodeCompose out to developers. At the time of this writing, 16K developers have used it with 8% of their code coming directly from CodeCompose. To triangulate our numerical findings, we conduct a thematic analysis on the feedback from 70 developers. We find that 91.5% of the feedback is positive, with the most common themes being discovering APIs, dealing with boilerplate code, and accelerating coding. Meta continues to integrate this feedback into CodeCompose.