I've pretty much resigned myself to only use LLMs as a glorified search and summarization engine when doing research. Which they are excellent at in my opinion.
However, any non-trivial coding problem I've tried always ends in disaster. I can imagine if I was working in a strongly typed language or pushing protobufs between microservices they would fare better.