Torvalds Let Claude Code Write 8 Kernel Patches: 3x Reads on One Laptop, Not Merged

说明与用途

Linus Torvalds 使用 Claude Code 创建了八个补丁，将 Linux 页面缓存中的小读吞吐量提升了三倍，但这些更改尚未合并。

问题

Developers need a clear process for integrating AI‑written code into production kernels while maintaining accountability and reproducibility.

结果

The AI‑generated patches tripled small‑read throughput on a single laptop, demonstrating the potential of agent‑assisted kernel development, but the changes are not yet merged and include regressions.

工作流

• 1. Author a branch name indicating AI authorship (e.g., noref-filemap-read-with-llm).
• 2. Generate code, tests, and a draft cover letter with Claude Code under human direction.
• 3. Run benchmarks on the author’s hardware and record results with precision.
• 4. Publish regressions alongside wins and maintain an open‑problems list.
• 5. Add an "Assisted‑by: LLM" trailer to commits and sign off manually.
• 6. Review agent‑written changes in small, isolated commits before merging.

证据

On his laptop, eight threads reading one file went from about 10 million to 30–33 million reads per second

来源

The cover letter credits Claude Code, Anthropic's coding agent, with the code, the test programs, the review behind the list of open problems, and the first draft of the cover letter itself

来源

The kernel's AI Coding Assistants guide asks contributions to carry an Assisted-by: LLM tag

来源