Comment by vidarh

This is not logically sound at all.

We can replicate a great many things we don't fully understand.

Text to speech is a good example. Early text to speech were format based - many of them tried to physically emulate the human vocal tract. And it gets passable results, but the best current results comes from throwing away that and not try to understand and model precisely how we talk, but instead "just" apply machine learning approaches to create a model. The result replicates human speech vastly better, despite us explicitly "giving up" on understanding precisely how to model the underlying system.

The same is true for a huge amount of other control problems, where we often achieve far better results at replicating something when we don't try to understand or assess exactly how something came to be or works, but instead design systems to learn by example.

We still may see benefits from trying to achieve a deeper understanding, but there is little to suggest that there is some universal law that we need to understand something to replicate it.