This story has been reported around the web, and especially on several blog sites of academic researchers in the QIP field. While those blogs do contain some interesting criticisms, I've seen a number of cheap-shots that some bloggers are taking.

I guess that one reason for this is that D-Wave has some pretty hyped up material, with the more measured statements tucked away in their own blogs. For example, the opening sentence on their site, www.dwavesys.com is
...but what they currently have is far from a commercial device, even if it works as advertised, and therefore D-Wave is not (yet) a source for such things. I guess this is about marketing the company's image, but to academics that kind of hype is beyond the pale.

On a more scientific level, one thing that has annoyed a number of QIP researchers is D-Wave's imprecise use of language regarding "solving" NP-complete problems.

Background: Problems classed as NP are difficult, in the sense that they get harder FAST as the size of the task increases. The most famous example is The Traveling Salesman problem, which when posed as follows is NP-complete (see Wikipedia.org for more).

Imagine a map with a number of cities, and a set of roads that connect up these cities. Some city-to-city trips are more expensive than others, e.g. because the road is longer and needs more fuel. Suppose a traveller must visit all cities, and has a route in mind that requires a certain budget; is there a cheaper alternative route? Of course, for a small number of cities (say, 4) even a human could work this out in reasonable time. But as the number of cities (i.e. the complexity of the map) increases, it rapidly becomes intractable, not only for humans but for supercomputers too.

Now the "complete" part of NP-complete indicates that this problem is equivalent to many other problems, things as far afield as simulating biologically important molecules. Those other problems can be rephrased as a Traveling Salesman task. If you could find a way to quickly solve large Traveling Salesman problems, you could quickly solve the others, too. Such a thing would be an immense step forward.

But here is the crucial thing: because NP problems are so hard, simply speeding them up a bit isn't really that much of an advance. Sure you could say "we are creating a machine to solve NP-complete problems", and (since even humans can do small problems!) go on to say "our machine will do this faster than any conventional computer". But this isn't enough to REALLY "solve" NP-complete problems in the sense meant by academics -- that's because your speedup will only buy us the ability to handle a few problems that we couldn't previously tackle; as we look toward bigger tasks (bigger maps in Traveling Salesman) we'll soon find that the rapidly increasing difficulty is beyond even our new, speedy computer. To really solve an NP-complete problem, in the sense an academic would mean, you need to invent a device for which an NP-complete problem no longer has this horrible scaling up in difficulty. Then, doubling the problem size (number of cities for the salesman) may only modesty increase the time required, say by a factor of two. The problem, when run on your new device, isn't "hard" anymore: we can solve bigger problems just by allowing the computer proportionately more resources.

I think that D-Wave's site tends to give the impression that they have achieved (or will achieve) this strong sense of the phrase "solve NP-complete problem". Whereas I suspect that what they will really be able to offer is the more modest kind of speedup. Customers could tackle a few tasks in the narrow band between the tasks that a conventional computer can reach, and those that their devices can handle (before that explosive increase in difficulty makes new larger tasks impossible even for them.) Some of Geordie's comments seem to indicate this, and I'd be interested to hear him say so explicitly.

Of course such a machine could have a big market, but not the world-changing vast market that would come with a device that made NP-complete problems into easy problems.

For the investor, the (tricky) question seems to be the following. How many real-world problems ...

(a) are beyond the reach of conventional computers (so we exclude specific NP problems that happen to small enough to solved classically), but

(b) are within the reach of quantum computers (so we exclude all potential problems that are too big EVEN for quantum computers), AND

(c) are not adequately solved by approximate algorithms that DO run fast on ordinary machines.

The last one is non-trivial since (as I gather) many many problems that are NP-hard in their exact form, can be solved adequately for real world applications by fast, imprecise algorithms.

I'm not aware that that is true in general, although I know that there are some NP-hard problems for which this is true. But I was using the term approximate loosely and maybe I should have said imperfect instead, to include heuristic solutions. Take a look at http://en.wikipedia.org/wiki/NP-Complete#Imperfect_solutions in the section on Imperfect solutions.

In point (c) I was really alluding to the fact that, where hard problems occur in the real world, people have of necessity already found ways of dealing with them. The case of optimizing circuit layouts for chip design, for example. Now generally the solutions must be non-optimal, but is the amount of improvement that would come from having a quantum computer worth the costs of such a machine? Maybe if you are Intel, and you will sell many millions of copies of the new chip, it is. But I'd speculate that in many cases, the current solutions are close enough to optimal that people aren't too concerned about the defect.