Comment by jacquesm

> a trillion-parameter transformer model built using many layers is likely significantly easier to read than one that only has a few million parameters but was constructed using neuroevolution over one gigantic layer.

Why do you believe this?