Consider an extremely high-dimensional feature space [not just genes, but transcripts/isoforms (differently diced-up versions of RNAs with different functions), hosts of active-but-not-transcribed-into-proteins RNAs, and dynamic regulation through interactions] which updates a reinforcement learning algorithm one epoch per generation (~20-30 years), and you'll see why it's far from optimized.

I wonder what other dormant code exists that if toggled would trigger dramatic changes