Three separate signals from three leading AI labs, taken individually, are familiar news. Taken together, they describe a shift in what software engineering work actually consists of — and the pattern is set out in an analysis piece published by O'Reilly Radar titled "Who Builds the Next Model?"
The first signal is the most concrete: according to Anthropic, Claude now authors more than 80% of the code the company merges into its own production systems. The second comes from Google, where an AI agent called AlphaEvolve discovered a method for making Gemini train faster — an improvement no human engineer had proposed. The third is OpenAI's so-called "automated research intern," an agent that executes research tasks but only ever under explicit human direction.
O'Reilly's reading is that these are not three unrelated milestones but a single trajectory: AI is increasingly building the systems that produce the next generation of AI. What looks like a benchmark story is actually a labour-share story. Anthropic's figure is a production metric from a company deploying its own model on itself, and it reframes "AI-assisted coding" as human effort migrated rather than eliminated — away from typing and toward specification, review, architecture, and prompting.
Two thresholds sit inside that trajectory, and they are not the same one. AlphaEvolve represents autonomous discovery: an improvement generated without a human having conceived it. OpenAI's research intern represents execution under direction, with initiative deliberately withheld. Which of the two a given team permits — and where the boundary gets drawn — is the negotiation most engineering organisations will face next.
The practical consequences land squarely on job shape rather than job count. When a majority of merged code is machine-authored, the scarce skills become the ones machines do not supply: deciding what to build, judging whether output is correct, and being accountable for it. Tooling literacy — the ability to drive, evaluate, and constrain these systems — becomes a hiring signal in its own right, independent of language fluency or algorithm depth.
For IT teams in Hong Kong, the accountability question is already live in any organisation that must explain who is responsible for code no one on the team wrote by hand — a familiar governance problem in new clothing. The 80% figure sharpens it into a disclosure gap: no industry-wide equivalent of the labs' self-reported statistic exists, and few procurement or audit processes currently ask how much of a delivered codebase was machine-authored. This is a question any IT team shipping software in Hong Kong may eventually have to answer, though no local regulator is known to require such disclosure today.
There is also a pipeline problem the piece raises for the open-source community specifically. Entry-level and routine maintenance work has historically been how new contributors learn a codebase and earn trust. If that work disappears into automated flows, the supply of future maintainers erodes alongside it — a slow-moving risk for projects that depend on volunteer depth.
Two caveats frame all of this. Anthropic's number is self-reported by a vendor with a commercial interest in the story, and it counts merges, not effort: whether 80% machine-authored code means less total human work or simply more prompting, correcting, and reviewing is not addressed. Access may also compound — teams without frontier-model capability could fall further behind those that do.
The open questions are worth treating as follow-up reporting rather than settled conclusions. Until employers and vendors disclose these figures routinely, the burden sits with developers to ask how much of the code they are shipping was written by a machine, and who signs for it.
來自三家頂尖 AI 實驗室的三項獨立信號,若單獨來看,都是熟悉的新聞。但把它們放在一起,則描述了軟件工程工作的實質正在發生的轉變 — 而這一模式已由 O'Reilly Radar 一篇題為《誰來構建下一個模型?》(Who Builds the Next Model?)的分析文章所勾勒出來。
第一項信號最為具體:據 Anthropic 稱,Claude 目前撰寫了該公司合併到自家生產系統中超過 80% 的程式碼。第二項來自 Google,其名為 AlphaEvolve 的 AI agent 發現了一種令 Gemini 訓練更快的方法 — 這一改進並無任何人類工程師提出過。第三項是 OpenAI 所謂的「自動化研究實習生」,一個執行研究任務的 agent,但始終只在人類明確指示下運作。
O'Reilly 的解讀是,這並非三個互不相關的里程碑,而是同一條軌跡:AI 正日益構建出產生下一代 AI 的系統。看似一篇基準測試(benchmark)的故事,實際上是勞動份額的故事。Anthropic 的數字是一項生產指標,來自一家將自家模型部署於自身的公司,它將「AI 輔助編程」重新定義為人類工作的轉移而非消失 — 從打字轉向規格設定、審查、架構設計與 prompt 撰寫。
這條軌跡中包含兩個門檻,而它們並不相同。AlphaEvolve 代表自主發現:一項無須人類構想便生成的改進。OpenAI 的研究實習生則代表在指示下執行,主動性被刻意剝奪。一個團隊容許兩者中的哪一種 — 以及界線劃在哪裡 — 將是大多數工程組織接下來要面對的談判。
實際影響直指工作形態(job shape),而非工作數量。當大多數合併的程式碼由機器撰寫時,稀缺的技能變成機器無法提供的那些:決定要構建什麼、判斷輸出是否正確,以及為此承擔責任。工具素養(tooling literacy)— 駕馭、評估和約束這些系統的能力 — 本身成為一項招聘指標,與語言熟練度或算法深度無關。
對香港的 IT 團隊而言,問責問題在任何必須解釋「團隊中無人親手編寫的程式碼由誰負責」的組織中早已生效 — 這是新瓶裝舊酒的治理問題。80% 這個數字將其尖銳化為一個披露缺口:目前並不存在與各大實驗室自我申報統計數據相當的全行業指標,而現有的採購或審計流程也很少會問:交付的程式碼庫中有多少比例由機器撰寫。這是任何在香港交付軟件的 IT 團隊最終可能必須回答的問題,儘管據悉目前沒有本地監管機構要求作此披露。
文章還特別為開源社區提出了一個人才管道問題。入門級和例行維護工作歷來是新貢獻者學習程式碼庫、贏得信任的途徑。若這類工作被自動化流程吞噬,未來維護者的供應也會隨之流失 — 對依賴志願者深度的項目而言,這是一個緩慢演變的風險。
有兩點保留意見框定了上述所有內容。Anthropic 的數字是由一家對此敘事有商業利益的供應商自行申報的,而且它統計的是合併次數,而非投入的工作量:80% 由機器撰寫的程式碼究竟意味著整體人類工作減少,還是僅僅意味著更多 prompt 撰寫、修正和審查,文章並未討論。獲取能力的差距也可能累積 — 不具備前沿模型能力的團隊可能落後於具備能力的團隊越來越遠。
這些未解問題應被視為後續報道的題材,而非已成定論的結論。在僱主和供應商例行披露這些數據之前,責任落在開發人員身上:追問他們交付的程式碼中有多少由機器撰寫,以及由誰簽字負責。
