When Models Start Writing the Code for the Next Model: What Developers Should Watch

O'Reilly Radar recently published a commentary, "Who Builds the Next Model?", arguing that models are no longer only writing applications — they are beginning to take part in building the models themselves. The piece is analysis rather than breaking news: the figures below are reported as O'Reilly states them and have not been independently verified here, while the interpretations that follow are this publication's own.

Three examples sit side by side in the source. Anthropic says Claude writes more than 80% of the code the company merges. At Google, an AI agent called AlphaEvolve found a way to make Gemini train faster. At OpenAI, an "automated research intern" carries out research tasks under human direction. Put them together and a pattern appears: models are participating in building the next generation of models.

O'Reilly's reading of that pattern is notably restrained. The 80% figure measures code output, not the weight of architectural decisions. AlphaEvolve's proposal still had to be validated by people before it was adopted. The research intern sets no agenda; it executes what it is told. In other words, autonomy has boundaries and accountability stays with humans. This is not evidence of models sustaining themselves — it is a leading indicator.

The three observations below are this publication's extended analysis, not points made in the source article.

The apprenticeship pipeline is thinning. Junior engineers once built judgement by writing large volumes of routine code — handling edge cases, learning the standards of code review, spotting code that looks like it runs but is quietly wrong. That is precisely the layer models are absorbing. Once routine implementation is automated, the route to seniority narrows with it: what gets replaced is the practice material, not just the job.

The review burden gap in open source. Well-resourced labs can pair their internal agents with dedicated verification teams, giving each machine-written change a human check; volunteer maintainers face an endless stream of machine-generated patches with no such backstop. Review time is becoming the scarcest resource. For Hong Kong's Linux and open-source communities, checking patch provenance and reconstructing the context of a change may be shifting from an occasional chore to part of routine maintenance — worth watching.

The feedback loop needs to be broken deliberately. If model-written code flows back into the training corpus for the next generation, the diversity of ideas may quietly narrow. Introducing unconventional solutions on purpose, and keeping exploratory work off the automated track, may be what prevents that self-reinforcing homogenisation.

Our view is that the source's conclusion is equally pragmatic: what becomes genuinely scarce is judgement, verification and explanation — saying why one approach was chosen, identifying where a model is confidently wrong, and standing behind the decision. For Hong Kong technical teams planning a transition or hiring, that is a more useful thing to weigh than any single statistic about how much code AI writes. The Hong Kong angle here is our inference from the source's argument, not a condition O'Reilly reports.

(Source: O'Reilly Radar, "Who Builds the Next Model?"; all figures are as reported in that article and have not been independently verified.)


當模型開始為下一代模型撰寫程式碼:開發者應留意什麼

O'Reilly Radar 最近發表了一篇評論文章〈Who Builds the Next Model?〉,認為模型已不再只是撰寫應用程式——它們開始參與建造模型本身。本文屬評論分析而非突發新聞:下列數據按 O'Reilly 所述引用,本刊未經獨立核實,而隨後的詮釋則為本刊觀點,並非原文明言,讀者不宜將以下論述全部歸於 O'Reilly 一方。

原文並列了三個例子。Anthropic 稱 Claude 撰寫了公司合併(merge)程式碼的八成以上。在 Google,名為 AlphaEvolve 的 AI agent 找到了加快 Gemini 訓練的方法。在 OpenAI,一名「自動化研究實習生」在人類指示下執行研究任務。三者合而觀之,一個模式浮現:模型正參與建造下一代模型。

O'Reilly 對此模式的解讀相當克制。八成這一數字衡量的是程式碼產出,而非架構決策的權重。AlphaEvolve 的方案在被採納前,仍須經人類驗證。研究實習生不會自行設定議程;它只執行指令。換言之,自主性有其邊界,問責始終留在人類身上。這並非模型自我延續的證據——只是一個先行指標。

以下三點為本篇的延伸觀察,並非原文明言。

學徒式養成路徑正在變窄。初級工程師昔日透過撰寫大量例行程式碼來建立判斷力——處理邊界情況、熟悉 code review 的標準、辨識那些看似能運行、實則暗藏錯誤的程式碼。而這正是模型正在吸收的層次。一旦例行實作被自動化,晉升資深的路徑亦隨之收窄:被取代的是練習材料,而不只是一份工作。

開源專案的 review 負擔落差。資源充裕的實驗室可讓內部 agent 與專職驗證團隊協作,為每一項機器撰寫的變更提供人類把關;義工維護者面對的卻是源源不絕的機器生成 patch,背後並無此等保障。review 時間正成為最稀缺的資源。對香港的 Linux 與開源社群來說,核查 patch 來源、重建變更脈絡,可能正由偶爾為之的麻煩事轉為日常維護的一部分——這點值得持續觀察。(此段為本刊就原文論點所作的引伸,非 O'Reilly 所述情況。)

回饋循環需要刻意打斷。若模型撰寫的程式碼回流至下一代模型的訓練語料,觀點的多樣性可能在無形中收窄。刻意引入非主流方案,並讓探索性工作脫離自動化軌道,或正是防止這種自我強化同質化的方法。(此段為本刊延伸觀察,非原文明言。)

本刊認為,原文的結論同樣務實:真正變得稀缺的,是判斷、驗證與解釋能力——說明為何選定某一方案、指出模型在哪裡自信地錯了,並為決定承擔責任。對香港正在規劃轉型或招聘的技術團隊而言,這比任何單一關於 AI 寫了多少程式碼的統計數字都更值得衡量。上述香港角度是本刊從原文論點作出的推論,並非 O'Reilly 所報告的狀況。

(來源:O'Reilly Radar,〈Who Builds the Next Model?〉;所有數據均按該文所述引用,未經獨立核實。)

新聞來源 / Original News Source