ai models are all about quality of input and output

neural networks are nothing more than INPUT TO OUTPUT MAPPING

given input a

receive output b

we are trying to train that a results in b

net(a) -> b

net(1+1) -> 2

net(2*2) -> 4

given enough examples

it learns math

given enough examples

it learns anything

the accuracy to generate correct output for certain input

is very much product of quality of training

when training data contains things like this, it results in ai that does not function correctly:

net(1+3)=3.5

models should be trained only on data that is verifiably 99.9 % correct

human language is difficult to verify

math is almost 100 % verifiable. excellent training data