“how the memory contents of a NTM are initialized may be a defining factor in the success of a NTM implementation”

Does it mean that one can expect better explainability in the future from these models?