Dave Airlie, maintainer of the graphics driver stack in the Linux kernel, announced the implementation of an auxiliary automated patch review system using the Cloud Opus AI platform to review changes in version 4.6. This system was introduced as an experiment to explore the potential of AI in optimizing work processes for Red Hat employees.
The AI review system focuses on changes submitted for inclusion in the DRM (Direct Rendering Manager) subsystem. Results of the reviews are published using mailing lists commonly used in Linux kernel development. To avoid cluttering the main mailing list for DRM subsystem development, a separate mailing list drm-ai-reviews has been created for AI reports.
Patch authors are encouraged to consider AI reviews as initial feedback. However, the analysis of these reviews is not mandatory and authors have the freedom to incorporate the feedback as they see fit. If AI reviews uncover regressions that were overlooked by patch authors, there may be a more assertive approach to utilizing AI reviews in the future.
In the initial stages, Dave attempted to organize reviews in the Cloud using a set of hints provided by Chris Mason to help the AI assistant better understand the context during reviews. These hints were designed to identify regressions and analyze patches by applying them sequentially to the actual kernel code tree and testing the regressions in the final patch.
Subsequently, Dave found it necessary to enhance and expand Chris’s prompts to enable the AI to review a series of patches as a whole and then perform a detailed analysis of each individual patch. The scope of the checks was also expanded to include the drm-next branch. Additionally, Dave developed a toolkit for integration with the Claude AI service and organizing the mailing process.