More like an open question. Can we optimize digits? Clearly if you look at an expression like 2+3=5 there is no hint in the shape of 2 and 3 as digits that lets you predict the shape of the digit 5. Looking at 4 there is no evidence that it is twice 2 You can't even tell, visually, odd from even.. Digits are essentially just arbitrary squiggles. Can we do better? Given a set of constraints encoded as a loss function and a representation of digits as say 20x20 black and white images can we use a machine learning algorithm to come up with squiggles which respect the constraints? The constraints are the key. For instance you could ask that the digit for k+h looks similar to a combination of the digits for k and for h. Dumb constraints will get you dumb digits, cool constraints maybe will produce cool digits.