If you can define a problem, you can probably find a way to build a ML model that can solve it with enough examples. It's the problem definition that it can't d

If you can define a problem, you can probably find a way to build a ML model that can solve it with enough examples. It's the problem definition that it can't do. Also, I have seen some research e.g. on telling whether things are the same or different in an image. Like draw two arbitrary shapes side by side and build a same / different classifier. Last I heard, it doesn't extrapolate to new shapes, i.e. it is not actually learning "sameness".