LLMs tend to rise to the level of the complexity of the codebase. They are probabilistic pattern matching machines, after all. It's rare to have a 15 year old repo without significant complexity; is it possible that the reason LLMs have trouble with complex codebases is that the codebases are complex?

IMO it has nothing to do with LLMs. They just mirror the patterns they see - don't get upset when you don't like your own reflection! Software complexity is still bad. LLMs just shove it back in our face.

Implications: AI is always going to feel more effective on brand new codebases without any legacy weight. And less effective on "real" apps where the details matter.

The bias is strongly evident - you rarely hear anyone talking about how they vibe coded a coherent changeset to an existing repo.