Neural networks are Turing-complete, so in theory a large enough one should be able to express any logic. So the main limitation here is the training process, b

Neural networks are Turing-complete, so in theory a large enough one should be able to express any logic. So the main limitation here is the training process, but the way it’s done seems very flexible, so I think the main limitation at the moment is just model size, which is limited by hardware which is still improving exponentially (though it has been slowing down slightly)