Your LLM Isn't Writing Correct Code — It's Writing Code That Looks Reasonable

The author benchmarked system SQLite against an LLM-generated Rust rewrite. Even though it compiled and passed all tests, primary key lookups were ~20,000x slower. The takeaway: define acceptance criteria before you talk about AI productivity.