I imagine this is worst at "AI native" companies that started and never had a basis of understanding in their code bases. I suspect that over time, any basis we had will erode more and more. I already feel that in my code base. I feel like the only way to actually fix that is either accept that you will never have full understanding anymore, and hope the models will get good enough to keep the lights on, or to use models to get understanding and then make or suggest changes that are informed from that understanding. I don't know how to do that yet, though.