The true test challenges should be how far an AI can minimize a given fucked up codebase and keep full functionality.

I also think that writting large codebases into a sort of functional transformer tree as information compression stage would allow them to easier reason about large code bases by having a large lossless overview with minimal token usage.