What are you looking for when reviewing LLM generated code? (opens in new tab)

For work I often have to review multi thousand file changes in PR mostly generated by AI/Llama/claude code. It's mainly rust with many unit tests, so there is a lot tooling helps verify the code already to help a lot. But I struggle with these reviews, finding me to just glance over at times unable to comprehend the changes in any higher level capacity (e.g. architecture). Which I think is mainly due to so many changes of these thousands of lines happening that I lack any coherent model of wh...