As LLMs advance so rapidly I think that all the AI slop code written today will be easily digestible by the LLMs a few generations down the line. I think there will be a lot of improvements in making user intent clearer. Combined with a bad codebase and larger context windows, refactoring wont be a challenge.