so say your solving some hard partial differential equations and you modeling it with a two brained neural net.
one in 10d weights between 1 and 0 and biases between 1 and 0
and the other neural net only 4 dimensional with weights as integers and some as real number with some quite high size. the biases being between 1 and 0.
The idea here is to fish out transcendental numbers and integers as part of some algorithm which help solve different partial differential equations and other maths problems.
You would need another AI to reengineer useful algorithms from recorded data of the numerical finds which help to find fast solutions by process data mining ability along with GPT 3 like interfacing as mathematicians need to play a role in finalizing and taking credit for any good finds.
a sort of kinda fishing.