```
Producing software has become dramatically cheaper. Deciding whether software is worth keeping has not — and that gap is landing squarely on the shoulders of free and open source software (FOSS) maintainers.
That is the central argument of an analysis published by O'Reilly Radar, "When AI Can Code, What Is Left for Programmers to Do?", which reports that AI tools are now in such widespread use for programming that project maintainers are complaining about the burden of sifting through mounds of low-quality, machine-generated contributions.
Writing Became Cheap; Reading Did Not
The economics of contribution have inverted. For a volunteer maintainer, submitting a patch once required hours of effort; reviewing someone else's patch was comparatively cheap. Generative AI flips that calculation: producing a plausible-looking change is now close to free, but assessing it still costs exactly as much human attention as it always did — arguably more, because volume has risen while average contribution quality has not.
Reviewing is also the stage least amenable to automation. It requires judging intent rather than syntax, and it is where security defects, licensing problems, and regressions are caught. Automating the writing, in other words, made the reading more expensive.
The consequences propagate downstream. When maintainers slow, so do patch approvals, security fixes, and release cycles — affecting every team that depends on upstream open source, whether or not those teams ever submit a line of code themselves. For organisations that build on community-maintained libraries, a stalled upstream is a supply-chain risk, not merely an inconvenience.
AI Now Spans the Whole Development Cycle
O'Reilly's analysis notes that AI can assist at every stage of development: drafting the specification, writing the code, uncovering bugs, and repairing what breaks. That breadth forces an uncomfortable question for the profession — if the machine can spec, code, test, and fix, what remains for the human?
The article's answer is that value migrates toward the tasks that were never really about syntax: specifying clearly what should be built, exercising judgement about whether a contribution fits the project's goals, taking accountability for what ships, and stewarding a codebase over years. As code production approaches zero marginal cost, the scarce skill becomes knowing which code deserves to exist at all.
What Development Teams Can Do Now
For development teams — including those in Hong Kong whose products rest on open source dependencies — the practical guidance is consistent with how experienced maintainers already triage volunteer contributions:
- Treat AI output like a junior contribution. Nothing generated goes in without human review, tests, and a named owner.
- Invest in review-enabling scaffolding. Automated tests, static analysis, linters, coding standards, and a strong preference for small, focused changes all reduce the cost of reading a patch.
- Change the metric. Volume of code generated rewards the behaviour that is already overwhelming maintainers. The article recommends measuring time-to-confident-merge instead — how long it takes a change to reach a state where a reviewer is willing to accept responsibility for it.
That last metric is worth emphasising: it makes review capacity visible as a constraint, rather than hiding it behind output statistics. A team shipping thousands of AI-generated lines with a two-week merge queue has not scaled; it has merely relocated the FOSS maintainer backlog into its own workflow.
The broader point for the IT community is structural rather than speculative. Review is where security, licensing, and quality are actually enforced, and it is the one stage where human judgement remains the bottleneck. Automating everything around it will not remove that bottleneck — it will only increase the pressure on it.
Source: O'Reilly Radar.
```
生產軟件的成本已大幅下降,判斷軟件是否值得保留的成本卻沒有——而這道鴻溝,正正壓在自由及開源軟件(FOSS)維護者的肩膀上。
這是 O'Reilly Radar 發表的分析文章〈When AI Can Code, What Is Left for Programmers to Do?〉(當 AI 能寫代碼,程序員還剩下甚麼工作?)的核心論點。文章指出,AI 工具在編程領域的使用已極為普遍,以致項目維護者紛紛抱怨,要從堆積如山的低質素、機器生成的貢獻中逐一篩選,負擔沉重。
寫代碼變便宜了,讀代碼沒有
貢獻的經濟學已經倒轉。對於義務維護者而言,提交一個 patch 過去需要數小時心血;審核別人的 patch 相對便宜。生成式 AI 把這筆帳翻轉了:生產一個看似合理的改動如今幾乎免費,但評估它所耗用的人力注意力,一如既往——甚至更多,因為貢獻數量大增,平均貢獻質素卻沒有提升。
Code review 也是最難自動化的階段。它需要判斷的是意圖而非語法,而安全缺陷、licensing 問題和 regression 正是在這個階段被攔截。換言之,把「寫」自動化了,反而令「讀」更貴。
後果會沿著供應鏈向下傳導。當維護者放慢腳步,patch 審批、security fix 和 release cycle 也會隨之延誤——影響每一個依賴上游開源項目的團隊,無論這些團隊本身是否曾提交過一行代碼。對於建立在社區維護的 libraries 之上的機構而言,上游停滯是供應鏈風險,而非僅僅是不便。
AI 已涵蓋整個開發週期
O'Reilly 的分析指出,AI 可以在開發的每個階段提供協助:撰寫規格說明、編寫代碼、找出 bug、修復損壞的部分。這種廣度迫使業界面對一個令人不安的問題——如果機器可以 spec、code、test 和 fix,人類還剩下甚麼?
文章的答案是:價值會遷移到那些從來就不只是語法問題的任務上——清楚說明應當構建甚麼、運用判斷力評估貢獻是否符合項目目標、為實際交付的產品承擔責任,以及在多年間維護一個 codebase。當代碼生產趨近零 marginal cost,最稀缺的技能便是知道哪些代碼根本值得存在。
開發團隊現在可以做甚麼
對於開發團隊——包括那些產品建基於 open source dependencies 的香港團隊——實務建議與資深維護者處理志願者貢獻的 triage 方式一貫:
- 把 AI 輸出當作 junior 的貢獻。 任何機器生成的內容,未經人手 review、測試及指定負責人,一律不得採納。
- 投資於能提升 review 效率的 scaffolding。 自動化測試、static analysis、linters、coding standards,以及強烈偏好細小而集中的改動,都能降低審核一個 patch 的成本。
- 改變指標。 以生成代碼的數量作指標,只會獎勵那些已經令維護者不勝其擾的行為。文章建議改為量度 time-to-confident-merge——一個改動需要多久才能達到 reviewer 願意為其承擔責任的狀態。
最後這項指標值得特別強調:它使 review 能力作為一項限制條件清楚可見,而不是把它藏在產出統計數字背後。一個團隊交付數千行 AI 生成的代碼,卻有兩星期的 merge queue,那並非擴大了產能,只不過是把 FOSS 維護者的 backlog 搬進了自己的工作流程。
對 IT 社區而言,更宏觀的重點是結構性的,而非純屬推測。Review 正是 security、licensing 和 quality 真正得以落實的地方,也是 human judgement 仍然是瓶頸的唯一階段。把周邊一切自動化,並不能消除這個瓶頸——只會令瓶頸承受更大壓力。
Source: O'Reilly Radar.
