GPT learns to pass tests, not to engineer: the RL reward-mismatch behind ugly code

xiaohu · x · 2026-10-03

Xiaohu translates and comments on a viral critique of GPT's coding quality: the model learned to "pass the task," not to "do engineering well."

Three core arguments:

Reward mismatch: RL optimizes easily measurable signals — tests pass, no errors, task done — while architecture, maintainability, module boundaries and readability are hard to quantify. The model learns that passing tests equals success, even with rotten internals.
Optimized to think less, output less, finish fast: OpenAI's inference-efficiency push produces compressed wall-of-text code with poor decomposition, while real engineering requires upfront design (requirements → data structures → modules → interfaces → edge cases → tests → refactoring) that costs far more reasoning tokens.
Training horizon too short: training only sees task → output → tests → reward, but real projects face new requirements, bugs, refactoring and collaboration months later. A hacky login feature passes today, then can't extend to OAuth, multi-account and permissions — and "what happens in three months" never feeds back into the original generation.

Conclusion: models keep getting better at fast delivery, but not necessarily at engineering.

More from coding & agent