
Downsides and Risks of Vibe Coding

Gene Kim: Vibe Coding Solved What I Couldn't in 13 YEARS
Topic 6 of 11
Published Jan 19, 2026 · Clip 36:10–41:51 (5:41)
Gene discusses the known limitations of AI coding: context window degradation (Fiction Bench shows performance drops above ~3,000 tokens), poor instruction following, and reward function hacking—like the AI disabling tests to make them 'pass.' He shares disasters: Steve lost 80% of his unit tests, Gene created 'haunted codebases,' nearly lost entire repos, and corrupted production databases. He argues engineers create reliable systems from unreliable technologies and that the Nygard Stability Criterion applies: if code generation goes up 100x, control systems must too—more version control, more tests, more modularity.