Meta Engineers Train Internal AI Coding Tools by Fixing Code

Mandatory AI Training Initiative

Meta has mandated that its software engineers contribute to the training of its internal AI coding tools. Maher Saba, vice president of Meta’s Applied AI Engineering organization, requested that engineers submit at least one code "diff" each week using MetaCode, the company’s internal AI coding agent. This initiative aims to leverage the daily work of engineers to improve AI performance.

Impact on AI Model Development

The corrections submitted by engineers are directly contributing to the improvement of Meta's AI models. These fixes have already enhanced Muse Spark 1.1 and will be used for post-training an upcoming model internally known as Watermelon. So far, 7,000 weekly active users have submitted over 800 fixes, with Meta encouraging participation through colored badges on employee profiles.

Capturing the Correction Process

MetaCode provides Meta with access to the entire code correction process. This includes the original task, MetaCode’s initial response, the engineer’s correction, and any tests or reviews required for approval. This comprehensive data capture allows Meta to identify recurring problems and understand how human engineers rectify AI-generated errors, without needing to create artificial coding exercises.

Data Utilization and Future Models

Unlike public repositories that typically only show working software, MetaCode records the AI's initial attempt, where it failed, and the subsequent human intervention. While Meta has not detailed how these corrections are prepared or weighted during post-training, the memo indicates a deliberate process where engineers submit corrections when MetaCode makes an error. This data collection is crucial for refining future AI coding assistants.