Because we don't currently know a problem like this that both has a quantum algorithm we can run on this type of device with expected exponential speedup and has a fast classical verification algorithm. That's exactly the author's point/he has been advocating for quite a while the importance of researching such an example that would be better to use.