FWIW the MNIST dataset they talked about (handwritten digits) only has 60,000 training samples - and that dataset surely has more variance than printed puzzles

FWIW the MNIST dataset they talked about (handwritten digits) only has 60,000 training samples - and that dataset surely has more variance than printed puzzles from magazines.