As far as that part goes, you're right. However, someone wanting to make $APP work on raspberry pi 1 level of hardware, and complain to the LLM that electron is so laggy as to be unusable, then I also imagine the LLM is capable of suggestion rewriting $APP with $FOO technology that is the tight low-memory CPU-sipping version of $APP. If TSMC falls and we can't make new hardware, that may have to happen, but until such time that it's forced to, it's not going to.
> However, someone wanting to make $APP work on raspberry pi 1 level of hardware, and complain to the LLM that electron is so laggy as to be unusable, then I also imagine the LLM is capable of suggestion rewriting $APP with $FOO technology that is the tight low-memory CPU-sipping version of $APP
I don't think I've ever seen someone's agentic loop actually decide to do a rewrite. Once it sets off down a path, due to how token generation works, it can't actually backtrack easily.