That way you get sort of the best by both worlds, you get a (limited) programmability and you get easier observations than you would have access to just by for example designing a molecule encapsulating the problem you want to solve, as you would have to observe it "doing its stuff" somehow.
Just because it can't crack crypto doesn't mean a large-scale version of a setup like this can't be valuable. The field of quantum chemistry needs to use huge supercomputer clusters to simulate stuff like this.
Having a proper quantum computer with enough fully programmable bits to demonstrate quantum supremacy would be valuable, but (as the author states) the meaning of quantum supremacy is tricky, and we currently have achieved "quantum supremacy" only if we use a very limited, "cheating" definition of quantum supremacy.
Where does Wilczek claim that? The only sentence I can see where he arguably comments on the programmability of Google's device is "the computation that Sycamore performed is very specialized". But this is before he brings up the carbon example, and he never says the programmability of Sycamore is comparable to the programmability of carbon.
Indeed, Sycamore is programable. It's obviously not capable of arbitrary-length computations, because the noise level is too high to be fault tolerant, but it has a certain accessible state space in which it can be arbitrarily configured. Here's Scott Aaronson if you don't believe me (emphasis mine):
> [Question:] Even so, there are countless examples of materials and chemical reactions that are hard to classically simulate, as well as special-purpose quantum simulators (like those of Lukin’s group at Harvard). Why don’t these already count as quantum computational supremacy?
> [Answer:] Under some people’s definitions of “quantum computational supremacy,” they do! The key difference with Google’s effort is that they have a fully programmable device—one that you can program with an arbitrary sequence of nearest-neighbor 2-qubit gates, just by sending the appropriate signals from your classical computer.
> In other words, it’s no longer open to the QC skeptics to sneer that, sure, there are quantum systems that are hard to simulate classically, but that’s just because nature is hard to simulate, and you don’t get to arbitrarily redefine whatever random chemical you find in the wild to be a “computer for simulating itself.” Under any sane definition, the superconducting devices that Google, IBM, and others are now building are indeed “computers.”
https://www.scottaaronson.com/blog/?p=4317
Therefore, I think abdullahkhalids's interpretation is right: Wilczek's carbon analogy is inapplicable.
The more I think about the distinction people are drawing, the fuzzier it gets in my mind.
The author [1] is trivializes the notion of computational machine. Calculating using a classical/quantum computer vs doing the experiment, differ in at least two important respects.
(1) Computers are fully programmable while experimental setups are usually specific to calculating a choice few parameters. The universality of various computational models implies exactly this: a well designed physical machine can simulate any other physical system, assuming you have the correct model of the system. A carbon-spectra calculating experimental setup does not have this universal programmability.
(2) A computer often yields the correct answer with a lot fewer resources, again assuming your model of the simulated system is correct. This is because, a well designed computation can ignore unneeded part of the model, which the experimental system must include. In fact, computers are precisely useful where we can get such an advantage, and useless where we can't. As the author points out this is not possible for some carbon atoms, and it also is not possible for medical drugs. But it is possible for a wide variety of other systems - say rockets to the moon.
[1] a Nobel prize winner, so take my words with a grain of salt.