影片進階EN2.7 萬 次觀看
The Death of the Code Review: What the Data Actually Says — Laurie Voss, Arize AI
來源 AI Engineer
看影片(在新分頁開啟原站)連到 AI Engineer
摘要
探討 AI 自動編寫程式導致人工程式碼審查面臨新瓶頸。研究顯示開發者使用 AI 寫出 741% 程式碼卻僅產出 30% 軟體,且人工審查效能在超過 400 行時急劇下降。影片指出,雖然自動審查工具如 GitHub Copilot 和 Cursor 已普及,但缺乏可靠的「可合併性」基準,導致模型難以可靠地判斷程式碼品質。
This video explores the new bottleneck in AI-generated code: human code review is becoming less effective as agents write more code. Research shows developers using AI write 741% more code but ship only 30% of software, and review effectiveness drops sharply after 400 lines. While tools like GitHub…
重點
- AI 自動編寫程式速度大幅提升,但人工程式碼審查效能因行數增加而急劇下降。
- 缺乏可靠的「可合併性」基準,導致自動審查工具難以可靠判斷程式碼品質。
- 人工審查正從直接閱讀程式碼轉向設計和調優審查系統。
章節
依話題轉折切分,標題由 AI 產生
- 00:00The new bottleneck: human review
- 01:37741% more code, 30% more software
- 02:37Generation is no longer the bottleneck
- 03:22Why "review harder" fails: the Cisco study
- 04:52Stop reading code? Loops and OpenAI's zero-human-code product
- 06:42Do passing tests mean mergeable? METR's SWE-bench study
- 08:01FrontierCode: 88% vs. 29
- 09:11Mergeability as the next training signal
- 11:06Automated review today: GitHub Copilot and Cursor
- 13:16Review is fusing with repair
- 13:51CodeRabbit, Greptile, Graphite and the acceptance metric
- 15:00Can you skip the human? Carlini's C compiler
- 16:00Bun's Zig-to-Rust port and 13,044 unsafe blocks
- 17:30How OpenAI rebuilt review as a system
- 18:45Dex Horthy: "Please read the code"
- 19:30Context the tests can't see
- 19:55Where the human checkpoint survives
- 20:40Who reviews the reviewers? Prompt injection
- 22:04Production: the last reviewer standing
- 22:44Code review is being rebuilt, not killed
- 23:29What to do today: build a review harness
提到的工具與公司
- GitHub Copilot
- Cursor
- SWE-bench
- METR
- FrontierCode
- OpenAI
適合誰看
程式開發者、AI 工程師、軟體架構師。
摘要依據
- 講者
- Laurie Voss
- 依據
- 自動字幕
為什麼排在這裡
- 人氣
- 0.83
- 新鮮
- 0.99
在主題頁與搜尋結果裡,名次由相關、人氣、新鮮三個分數決定;這一頁沒有搜尋的關鍵字,所以沒有相關分數。排序怎麼算
相關內容
- Agentic Code Review文章 ・ Addy Osmani(部落格)
- AI-Generated Code Is Already Competing With Human Code文章 ・ AIE Talks
- EP407 – 人工審查跟不上 AI 產出速度?如何不讀程式碼也能確保軟體品質? 這集一次拆解給你聽!Podcast ・ AI懶人報 ・ 13 分鐘(在新分頁開啟原站)
- Code Review in the Age of AI文章 ・ Addy Osmani(部落格)
- GitHub’s COO Explains Why AI Hasn’t Replaced DevelopersPodcast ・ The Every Podcast(AI & I) ・ 28 分鐘
- AI Code Review That Understands Your PR's Intent影片 ・ AI Native Dev ・ 6 分鐘(在新分頁開啟原站)
摘要由 AI 根據原文產生,可能有誤;完整內容請看原站。看影片(在新分頁開啟原站)
