Note their goal is to demonstrate the *tech*, not actually solve a problem.
Their key points seem to be.
a) It's *programmable* (like a real computer), provided you can figure out how to describe the problem in "qubit" terms. Learning to think *in* those terms is likely to be easier for the yoof (if this takes off and does not become yet-another-technology-of-the-future, always was and always will be).
b) They have fabricated a chip (8 elements, each element of what seem to be 4 input qubits and 4 output qubits) full of quibits, which *mostly* work
c) You could have an *array* of such chips. It's *scalable*.
d) They are superconductive. Page 13 onward of the supplementary material describes it. It appears to operates at 20mK !. The process uses 4 Niobium layers and TiPt for resistors. Lines & spaces are 0.25 micrometres and the SQUID junctions are 0.6 micrometres (so a fair bit of room for reduction). This seems the closest anyone has got to actually *selling* a superconducting computer.
They claim it searches s solution space of 2^81 entries (which is pretty clever) partly by cycling the temperature.
Sadly a really hot cup of tea will result in a massive out of range error :(.
BTW A lot of AI seems to have been done on protein structure analysis (you've got X-ray images). This seems to have helped knock down initial analysis from *years* to hours. I don't quite see why heuristics cannot be used to constrain the search space and carve it up into multiple (but much *smaller*) search spaces to deliver a confirmation.