1) You make two and see by how much they differ after a certain time. (Further reading, see Allan variance.)
2) As with all the base units, we must 'define' the second in terms of something physical, which we can measure, so that we can use this abstract idea in the real world. This real-world embodiment is imperfect, and it is an engineering challenge to make something which better approximates the idea. For illustration, consider the kilogram, which is defined by a lump of metal in Paris. In principle, chipping a bit off this block makes everything else weigh more in terms of kilograms, but we immediate recognise this as crazy and we can imagine a better physical embodiment of the ideal kilogram (indeed, efforts are under way to do just this). So it is with the second: the caesium clock is the best we've got so far, but it's just a physical embodiment of the ideal second, and we can strive to make a more accurate (with accuracy defined as in (1) above).

It's an exciting idea, and it's streaks ahead of 'traditional' microwave transition atomic clocks. These do not represent the state of the art, however, for which one should look at the experimentally demonstrated ~9e-18 accuracy by the Wineland group at NIST http://arxiv.org/abs/0911.4527v2 ; http://www.nist.gov/physlab/div847/grp10/ , or the Strontium ion clocks at NPL (Teddington, UK) Essentially, the higher the frequency, the more clicks you get in a certain time, and the more accurate your clock can be (the smaller an error one missed click would represent). The caesium atomic clock is about 10 GHz (1E10 Hz). Strontium is in the optical, so a few 100THz (1E14). Aluminium ions are at about 1PHz (1E15 Hz). This new proposal with Thorium is around 7.6eV, which is about 2PHz, so not a million miles away from the current, demonstrated, state of the art. Also... orbit of the neutron around the nucleus isn't a fair description of a magnetic dipole transition, which would more accurately be describes as a flip in the direction of the neutron's spin axis. :)

Full text available on the arXiv, for those without a subscription to PRL: http://arxiv.org/abs/1112.5068

This reminds me of a device by Sandia National Labs of a micro-electromechanical steam engine. Sandia's device uses resistive heating to vapourise the water and capillary forces retract the piston.
Anyone in-the-know care to comment on the relative merits and the relative scales?

Add in shaking, and you're essentially describing simulated annealing.
Without the shaking, you'll quickly converge to a local minimum. With shaking, you explore the possibilities nearby and, provided you shake it just right, you eventually converge on the global minimum.
Simulated annealing is a really common approach when you have lots and lots of variables; in this example, the free parameters are the locations of each of the spheres. The authors even use this in their paper as a check.

I actually have it patented. :)
(WO2006134552) FLEXIBLE DISPLAYS AND USER INPUT MEANS THEREFOR.
(Although I think Philips let the patent lapse, and I think they stopped doing anything about it years ago.)

We're both loosing and gaining it here on earth, and it's not really practicable to make it. :)

'Green lasers' are typically diode-pumped solid state lasers with a frequency-doubling crystal. It's not easy to make green light directly.

Oh. So it is. Nevermind. :)

Cambridge Consultants demonstrated something similar a few years ago. It's called Sprint and there's a great big picture of it here.

Gunwharf Quays in Portsmouth, UK has been doing this since about 2008 (not that this makes it in any way ok).

Some blog
BBC News video

You bring the light from a pulsed laser to a very tight focus inside a photoresist -- the same type of chemical used in standard photolithography. When this photoresist absorbs light with a wavelength of, say, 400nm, it cross-links to become a fairly solid plastic. In normal photolith, you'd illuminate a controlled area with 400nm light.

In two-photon polymerisation, you start with light of, say, 800nm, and you rely on two photons being absorbed at the same time, which together have enough energy to do what a single 400nm photon could. The key here is that, since the probability of this two-photon process depends on the square of the intensity, rather than linearly as in the case of normal one-photon processes, then you can localise it much better: with a tight focus, the chance of polymerising a ~100nm region near the focus is pretty much unity, while the chance of polymerising something away from the focus is pretty much zero. You then move that spot around inside the a blob of photoresist on a microscope slide.

Have a look at Nanoscribe GmbH for a commercial device, with images of some things they've made.