Linux memory management developer Lorenzo Stoakes, an engineer at Arm who has contributed to the kernel since 2014, is stepping back from some of his reviewership roles, citing the noise generated by the flood of kernel patches coming from AI and large language models, according to a report from Phoronix.
The move is a partial step-back rather than a departure: Stoakes remains active in the kernel and continues his work in memory-management development. What he is relinquishing, as reported, is part of the review workload — the layer of the process where submitted patches are read, questioned, and either accepted or sent back.
That distinction matters. Review, not authorship, is the scarce resource in the scenario Phoronix describes. AI tools have collapsed the cost of producing a plausible-looking patch, but they have not reduced the cost of verifying one. Anyone can now generate a submission in seconds; only someone with deep, subsystem-specific knowledge can decide whether it is correct, safe, and consistent with the surrounding code. As submission volume scales with cheap generation, reviewer bandwidth stays roughly fixed — shifting the bottleneck from writing code to trusting it.
Memory management is among the least forgiving places for that bottleneck to form. It is a subsystem where a subtle mistake can mean corruption, deadlock, or performance regressions far from the fault's origin, and where the number of people qualified to review with confidence is small and cannot be expanded quickly. Volunteer reviewers also tend to disengage when queues fill with machine-generated noise; the incentives that keep experienced reviewers doing unpaid, painstaking work erode when much of the mail is low-signal.
The pattern is not new to this cycle. Kernel maintainers have publicly voiced concerns about AI-generated submissions through recent development cycles, and long-standing commentary on open-source sustainability has flagged maintainer burnout and the concentration of critical review work in a handful of individuals as structural risks. Stoakes' announcement is a concrete instance of that risk materialising in a core subsystem.
Policy responses remain unresolved at kernel scale. Proposals range from explicit disclosure norms for AI-assisted contributions, to automated filtering of likely machine-generated submissions, to stricter barriers before a patch ever reaches a reviewer's inbox. Each has drawbacks, and the kernel's size and decentralised, per-subsystem maintainer model resist any single uniform rule. Phoronix notes that Stoakes cited the noise from AI-driven patches as his rationale; it does not suggest his remaining contributions are being wound down.
For IT professionals and open-source contributors more broadly — including those involved in local Linux user groups and distribution packaging work — the structural lesson is straightforward: artificial intelligence has made contributing cheap, but curation and trust remain human and finite. When the review layer thins out, everyone downstream of it waits.
據 Phoronix 報導，Linux 記憶體管理開發者、自 2014 年起為核心作出貢獻的 Arm 工程師 Lorenzo Stoakes，正逐漸退出部分審閱者職務，理由是來自 AI 及大型語言模型的核心 patch 洪潮所產生的噪音。
這次是部分退下而非完全離開：Stoakes 仍活躍於核心社群，繼續從事記憶體管理開發工作。據報導，他放棄的是部分審閱工作量——即流程中負責閱讀、質疑已提交 patch，並決定接納或退回的那一層。
這個分別十分重要。在 Phoronix 所描述的情境中，稀缺的資源是審閱，而非撰寫。AI 工具大幅降低了產出看似合理 patch 的成本，卻沒有降低驗證 patch 的成本。任何人現在都能在數秒內生成一份提交；但只有具備深入、特定 subsystem 知識的人，才能判斷其是否正確、安全，以及是否與周邊程式碼一致。當提交量隨廉價生成而擴張，審閱者的處理能力大致維持不變——瓶頸遂由撰寫程式碼，轉移到信任程式碼。
記憶體管理正是最不容許此類瓶頸形成的領域之一。在這個 subsystem 中，一個細微錯誤可能意味著損毀、死鎖，或在遠離故障源頭處出現效能回歸；而有足夠資格自信地進行審閱的人數不多，亦無法迅速擴大。志願審閱者在隊列充斥機器生成的噪音時，亦往往會脫離；當大量郵件都是低訊號內容時，維持資深審閱者無償、費神工作的誘因便會消散。
這種模式對本輪開發週期而言並不新。核心維護者在近期開發週期中已公開對 AI 生成的提交表示關注，而長期以來關於開源可持續性的討論，亦早已指出維護者倦怠，以及關鍵審閱工作集中於少數人手上屬結構性風險。Stoakes 的宣布，正是這項風險在一個核心 subsystem 中具體成形的個案。
在核心的規模層面，政策上的應對仍未有定案。建議方案各異，包括要求 AI 輔助貢獻作明確披露的規範、自動過濾疑似機器生成的提交，以及在 patch 送達審閱者收件匣之前設立更嚴格的門檻。每一項方案都有缺點，而核心的龐大規模，以及去中心化、按 subsystem 分設維護者的模式，都令任何單一統一規則難以推行。Phoronix 指出，Stoakes 以 AI 驅動 patch 所帶來的噪音作為理由；報導並未暗示他其餘的貢獻正被逐步淡出。
對 IT 專業人士以及更廣泛的開源貢獻者——包括參與本地 Linux 使用者小組及發行版打包工作的人士——而言，結構性的啟示相當直接：人工智能令貢獻的成本變得低廉，但甄選與信任仍然由人承擔，而且是有限的。當審閱這一層變薄，下游所有人都只能等待。