Agreed. When I've written very low level code where there are "odd" constraints ("this function must never take a lock, no system calls can be made" etc) the LLM would accidentally violate them. It seems sort of obvious why - the vast majority of code it is trained on does not have those constraints.