This seem very unsurprising to me. Neural networks are good at being "good enough" at "soft" problems where the answer isn't cut and dry even for humans. Strict

This seem very unsurprising to me. Neural networks are good at being "good enough" at "soft" problems where the answer isn't cut and dry even for humans. Strict logical rules where there's one right answer are not a good domain to try to solve probabilistically.