Hasty Briefsbeta

Bilingual

Models are worse at reviewing their own code

5 hours ago
  • Models are worse at finding bugs in their own code; each model finds more bugs in code written by the other model.
  • Models produce the same types of bugs they introduce most often, which they also miss during review.
  • Claude models take a breadth-first approach (scope broadly), while GPT models go depth-first (investigate deeply).
  • GPT often identifies bugs in reasoning traces but fails to post them due to conflicting instructions and post-training alignment.
  • Opus produces many comments, including false positives, because it reviews with a human-like focus on intent and potential issues.
  • Greptile built 'Model Inversion' feature: detect the coding agent author and route the review to a different model to catch more bugs.
  • The tension between alignment (following instructions) and usefulness (reporting all bugs) is highlighted, especially with GPT.