A recent benchmark starkly illustrates a critical limitation of Large Language Models (LLMs): they optimize for plausible code over correctness. Comparing SQLite's efficient primary key lookup at 0.09 ms to an LLM-generated Rust rewrite taking 1,815.43 ms shows the problem isn't a minor typo. The reimplementation is 20,171 times slower on a fundamental database operation. This isn't just a performance hiccup; it's a systemic failure where the code compiles, passes tests, and claims features like MVCC and a C API, yet fundamentally breaks core functionality.

The author, a practitioner with extensive LLM integration experience, argues this highlights the need for users to define acceptance criteria before generation.