But, if I have understood correctly on a quick read, they also claim transformers have pretty low expressive power. In particular, they claim they are limited to star-free subregular languages, whereas RNNs can recognize any regular language/simulate finite automata.

This doesn't imply you can't get aid from a LLM to e.g. implement a function that has a formal specification (an application I think is very promising), but surely it has some profound implications on how much of a large system can be understood by a LLM at once, without supervision.