Let AI implement a quite complex feature. I didn't have time to continuously review, so I used the Vibe Coding approach to let it add e2e tests by itself, then had another AI review and iterate feedback. In the end, the report was perfect, all functionalities were implemented, and there were no errors in testing.

Feeling great, I submitted a PR for Github Copilot to review. The result was that it created two versions of the core logic at the lowest level—one version implemented the code but didn't import, while the actual effective version was another mocked version😅.

The next challenge for Vibe Coding is how to verify the results to prevent being fooled by AI.