First things first, obviously I work on SRE and EE. The thing that originally got me hyped about software was when I realized I could write tools that wrote more tools, or gave me leverage over the tasks at hand. I have net-negative interest in repetitive tasks and want to see us find ways to move faster with more confidence.
To me, classical specification-driven tools for parser or test or what have you generation are good examples of success in this field. You take a formal specification of some sort and begin to iterate on it and poke holes in it until you've arrived at something which produces the correct behavior. From a certain perspective, all of these tools are compilers. Sounds a lot like an LLM, but the basis in formal evaluation semantics brings confidence in the results of the tool. In theory at least, proof in software being what it is.
Current generation LLMs are best understood as a probabilistic recall database with lossy compression. An LLM is not a formal system, nor does it have any under