AI 寫程式的極限在哪裡？從 Vibe Coding 談起 | The Limits of AI Coding: A Vibe Coding Discussion
By Kit 小克 | AI Tool Observer | 2026-03-27
🇹🇼 AI 寫程式的極限在哪裡？從 Vibe Coding 談起
2025 年，Andrej Karpathy 提出了「Vibe Coding」這個概念：不再逐行寫程式，而是用自然語言描述你想要什麼，讓 AI 來寫。一年多過去了，AI 寫程式的能力確實驚人，但極限也越來越清晰。
AI Coding 現在能做到什麼？
先說好的部分。2026 年的 AI coding 工具（Cursor、Claude Code、GitHub Copilot、Windsurf）已經能：
- 快速生成 boilerplate：CRUD 介面、API 端點、資料庫模型，幾分鐘搞定
- 翻譯程式碼：Python 轉 TypeScript、REST 轉 GraphQL，準確率很高
- 除錯與修復：貼上錯誤訊息就能找到問題並修復
- 完成整個小專案：一個簡單的 web app 或 CLI 工具，從零到完成可能只需要一小時
- 寫測試：自動生成單元測試和整合測試
極限一：架構設計
AI 擅長寫「一段程式碼」，但不擅長設計「一個系統」。當專案規模超過幾千行，AI 開始出現問題：
- 無法掌握整個 codebase 的全局脈絡（即使 context window 越來越大）
- 傾向於過度工程化或選擇不適合的設計模式
- 不同檔案之間的一致性很難維持
極限二：邊界情況與健壯性
AI 生成的程式碼通常能處理「快樂路徑（Happy Path）」，但對邊界情況的考慮不夠周全：
- 錯誤處理常常只是 try-catch 包一包了事
- 並發、競態條件（Race Condition）幾乎不會主動考慮
- 安全性考量（SQL injection、XSS）有時會被忽略
極限三：維護性
Vibe Coding 最大的隱患是技術債。如果你不理解 AI 生成的程式碼，未來要維護和修改時會非常痛苦。我見過太多「AI 寫的、沒人能改的」專案。
極限四：領域知識
AI 對通用程式設計很強，但在特定領域的深度知識上有限：
- 金融計算的精度要求
- 醫療軟體的法規合規
- 嵌入式系統的硬體限制
- 高頻交易的效能優化
正確使用 AI Coding 的方式
我的建議是把 AI 當作超級厲害的初級工程師：它寫得很快，但你需要 review 每一行程式碼。最佳實踐：
- 讓 AI 處理重複性高的工作，自己專注在架構和設計
- 永遠要 code review AI 的產出
- 用 AI 加速學習，而不是跳過學習
- 在重要的生產環境中，AI 生成的程式碼要經過跟人寫的一樣嚴格的測試
Vibe Coding 很爽，但不是萬能的。好不好用，試了才知道——但別忘了理解你在用的東西。
🇺🇸 The Limits of AI Coding: A Vibe Coding Discussion
In 2025, Andrej Karpathy coined "Vibe Coding" — the idea of describing what you want in natural language and letting AI write the code. Over a year later, AI coding capabilities are indeed impressive, but the limits are becoming clearer.
What Can AI Coding Do Now?
Let us start with the positives. In 2026, AI coding tools (Cursor, Claude Code, GitHub Copilot, Windsurf) can:
- Rapidly generate boilerplate: CRUD interfaces, API endpoints, database models — done in minutes
- Translate code: Python to TypeScript, REST to GraphQL, with high accuracy
- Debug and fix: Paste an error message and get a diagnosis and fix
- Complete entire small projects: A simple web app or CLI tool, from zero to done in an hour
- Write tests: Auto-generate unit tests and integration tests
Limit 1: Architecture Design
AI excels at writing "a piece of code" but struggles to design "a system." When projects exceed a few thousand lines, problems emerge:
- Cannot grasp the full context of an entire codebase (even with ever-growing context windows)
- Tends to over-engineer or choose inappropriate design patterns
- Consistency across different files is hard to maintain
Limit 2: Edge Cases and Robustness
AI-generated code typically handles the "happy path" well but is often incomplete on edge cases:
- Error handling is frequently just a superficial try-catch wrapper
- Concurrency and race conditions are almost never proactively considered
- Security concerns (SQL injection, XSS) are sometimes overlooked
Limit 3: Maintainability
The biggest hidden cost of Vibe Coding is technical debt. If you don't understand the AI-generated code, maintaining and modifying it later becomes extremely painful. I have seen too many "AI-written, nobody-can-modify" projects.
Limit 4: Domain Knowledge
AI is strong at general-purpose programming but limited in deep domain-specific knowledge:
- Precision requirements in financial calculations
- Regulatory compliance in medical software
- Hardware constraints in embedded systems
- Performance optimization in high-frequency trading
The Right Way to Use AI Coding
My recommendation is to treat AI as a super-talented junior engineer: it writes fast, but you need to review every line. Best practices:
- Let AI handle repetitive work while you focus on architecture and design
- Always code review AI output
- Use AI to accelerate learning, not skip it
- In critical production environments, AI-generated code should undergo the same rigorous testing as human-written code
Vibe Coding is exciting but not omnipotent. You won't know until you try — but don't forget to understand what you are using.
Sources / 資料來源
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言