The article also gives an example from quantum physics. Some chaotic or highly numerically unstable problems just need quad precision to not blow up; primarily problems/algorithms (can be the problem itself or the algorithm to solve it that causes the issues) that tend to catastrophically amplify small errors. Matrix inversion is another example, to the point that most algorithms try to do almost anything to avoid explicitly inverting a matrix. But sometimes you just have to.
Do stats often deal with distict probabilities below 10^-300? And do they need to store them all over, or just in a couple variables that could do something clever?
Yes and with very wide dynamic range, though you really try to avoid it using other tricks. A lot of methods involve something resembling optimization of a likelihood function, which is often [naively] in the form of a product of a lot of probabilities (potentially hundreds or more) or might also involve the ratio of two very small numbers. Starting out far away from the optimum those probabilities are often very small, and even close to the optimum they can still be unreasonably small while still having extremely wide dynamic range. Usually when you really can't avoid it there's only a few operations where increased precision helps, but again even then it's usually a bandaid in lieu of a better algorithm or trick to avoid explicitly computing such small values. Still, I've had a few cases where it helped a bit.
Tricks like clever factorization (which a lot of factorization algorithms have their own severe numerical issues e.g. some of the ways to compute QR or SVD), preconditioners, sparse and/or iterative algorithms like GMRES, randomized algorithms (my favorite) etc are all workarounds you wouldn't need if there was a fast and stable way to exactly invert any arbitrary non-singular matrix. Well, you would have less need for, there are other benefits to those methods but improving numerical stability by avoiding inverses is one of the main ones.
Once you decide you need to go over 64 bits, the exact data type is largely a matter of convenience, and I can easily see two doubles being more convenient than either a quad or a 128-bit integer. You have more than 100 bits of precision in such a scheme, and it runs pretty fast.
If for e.g. you see bare summation rather than some sort of sorting by magnitude then expect bad results as numbers get bigger.
I looked up my consumer card on a whim (playing with sprinkling some GPGPU into my code—cupy is actually really easy to use and ergonomic!), and apparently doubles are 1/32 as fast as singles. Yikes!
For example: simulating the three body problem (three celestial bodies of equalish size orbiting eachother). Very small differences get amplified, and any difference at all in how the math is implemented on different systems will result in divergent solutions.