My own experience with leveraging LLMs has been one of an efficiency boost. As I have around 15 years of software architecture and development experience, I have yet to come across an instance where an AI is used to do something I can't do, and instead is used to do something that I could have otherwise done myself, albeit much more slowly.
This has had a great effect on my workflow. I am still able to do high-level architectural planning, determine use-cases and usability parameters, etc. When I have those pieces figured out, I can use an LLM (in this instance Claude Sonnet or Opus 4.6) to do the actual generation of code, which I can then review and correct as I see fit. I have not (and will never) used an LLM as a replacement for my "higher brain functioning", but when it comes to the "code-monkey" aspects of my work, it does them far faster than I can (and typically with a healthy respect for naming conventions, code patterns, etc.). I still get to enjoy the fun critical-thinking-laden aspects of my job, but the simple "regurgitation of learned code words" is offloaded to an all-too-willing counterpart.