On a primary-key lookup benchmark, the LLM-generated SQLite reimplementation was approximately 20,000 times slower than SQLite.

Format Web Page

Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER How Code Diverges From Prose

"On a primary-key lookup benchmark, the LLM-generated SQLite reimplementation was approximately 20,000 times slower than SQLite."

Hōrōshi benchmarked a primary-key lookup workload involving 100 rows and reported 1,815.43 milliseconds for the Rust implementation against 0.09 milliseconds for SQLite, a 20,171-fold slowdown. The measured result provides direct evidence that software can pass functional tests while failing badly on an important performance requirement.

View Source

Source title Your LLM Doesn't Write Correct Code. It Writes Plausible Code.

Source authors Hōrōshi

Source year 2026

Confidence high

Importance high