Why AI writes code better than it reads it
Early Greptile thesis: codebases are graphs, not file bags—reading at scale needs structure-preserving index, not naive embeddings.
Source: AI is better at writing code than reading code. Here’s why.
Still-on-topic foundational post: generating a script is easy; understanding an interdependent graph of functions is hard. Context windows can’t hold real company repos; naive vector chunking fights code’s non-linear DAG shape.
Their claim: index as a directed graph of classes/functions with edges for calls/inheritance, propagate updates to neighbors, make that graph semantically searchable. That’s the difference between “chat with files” and “review like a senior who command-clicks.”
Sets up why full-codebase reviewers exist as a category beside diff-only bots.
Why it matters for code review
Explains the architectural reason agent PR review needs repo intelligence. Without graph-aware retrieval, AI review inherits the same skimming limits as overloaded humans—just faster.