I agree in all of what you are writing. These are the hard problems of machine learning that we have not solved yet.
One thing that might be interesting in this context is that iterative algorithms (even complex ones like iterative image reconstruction) can be made trainable (similar to RNN training).
Kerstin Hammernik showed that iterative CT Recon can also be made trainable and allows to learn custom sparsifing transforms:
I think, algorithms like these can help us towards general problem solvers as I already speculated here:
Artificial Intelligence — A Bitter-Sweet Symphony in Modelling