90% AI 程式碼配額:大型科技公司為何製造脆弱的程式碼庫
The 90% AI Code Quota: Why Big Tech Is Manufacturing Fragile Codebases
Kürşat Yılmaz 批評大型科技公司以 AI 程式碼配額及提交速度衡量工程師,認為這種做法會削弱程式碼庫的可維護性。據 Evan X. Merz 的回顧,Meta 的績效要求涉及至少 90% 程式碼由大語言模型生成,而他的新僱主據報已提高至 100%。文章指出,表面測試覆蓋率、過度抽象、工程師失去系統理解,以及 PR 審查超載,均可能令生成程式碼難以維護。
The 90% AI Code Quota: Why Big Tech Is Manufacturing Fragile Codebases
Imagine being a senior software engineer at Meta, and the baseline requirement to survive performance reviews and avoid layoffs is: having an LLM generate at least 90% of your code.
This isn't dystopian fiction. It was just publicly documented by senior engineer Evan X. Merz, who was laid off from Meta in May. In his recent retrospective, Merz revealed that pumping out 90%+ AI-generated code was the table stakes for maintaining a decent performance rating. Engineers who focused on architectural durability, deep testing, and systemic code quality were pushed out. At his new Big Tech employer, that expectation has reportedly escalated to 100%.
The result? Management dashboards celebrate historic commit spikes and "10X productivity," while production environments crumble beneath an unmaintainable house of cards.
The Four Fatal Flaws of AI-Generated Codebases
Merz's breakdown highlights four critical architectural points of failure that are now quietly infecting enterprise software across the industry:
1. Shallow, Synthetic Test Suites
Modern LLMs optimize for line-level test coverage metrics rather than behavioral edge cases. They happily generate hundreds of lines testing trivial getters and setters, while end-to-end tests merely re-validate happy paths already covered in the codebase. Coverage metrics read 95% green, but the application explodes under the first non-trivial edge case in production.
2. Over-Engineered Accidental Complexity
LLMs dogmatically enforce patterns like DRY to an absurd extreme. Instead of writing a straightforward, readable routine, an agent will splinter a simple task into three separate helper abstractions and generic indirection layers. This rigidity makes production debugging a nightmare for human engineers.
3. Total Loss of Systemic Comprehension
As organizations mandate that engineers outsource thinking to AI models, human developers stop understanding the core data structures and underlying algorithmic flows. When engineers no longer understand their own architecture, code reviews become ceremonial rubber-stamping.
4. The Speed Trap of Unreviewed Pull Requests
When code is generated in seconds, the velocity of pull requests overwhelms human review capacity. Senior engineers skim PRs in seconds. Merz observed top engineers approving pull requests where simple CSS states were replaced by 30 lines of brittle JavaScript mouseover event listeners. Nobody paused to ask, "What are we actually building?"
Velocity Is Not Productivity
Writing 1,000 lines of code you don't understand is not speed—it is taking out a high-interest payday loan on technical debt.
AI models are blind to any architecture that doesn't fit within their immediate context window. Handing a developer who doesn't understand underlying systems an AI code generator doesn't build a cathedral; it just helps them drill through load-bearing walls faster.
When production inevitably crashes, the managers who celebrated commit velocity won't find a single engineer who knows how the machine works.
來源:DEV Community · dev.to