Would that experience get eroded though? LLMs let me perform experiments (including architecture/system level) quickly, and build stress test/benchmark/etc harnesses quickly to evaluate those experiments, so in the time you can build human intuition with one experiment I've done 10. I build less intuition from each experiment, but I'm building broader intuition, and if I choose a bad experiment it's a small cost, but choosing a bad experiment and performing it manually is brutal.