Reporting for duty, SIR!
Unfortunately both Plextor and Taiyo Yuden (the inventor of AutoStrategy) have been rather tight-lipped about the internals. So I can only offer you my conclusions based on my own observations.
I don't think AS is 'learning' in a way that it is getting gradually better as it burns more discs of the same type, as one might think when seeing the link you provided (BTW: Remember that the results shown in the link were done with the first version of AS, as implemented in the PX-716). The reason for this assumption is based on the main problem that AS is supposed to cure: Variance in media quality.
Even with top quality manufacturers every disc is slightly different from another. With low quality manufacturers that difference can be quite huge. This makes the task of implementing a writing strategy that works fine for all discs of that manufacturer an almost impossible task, especially in combination with 16x burning (where even slightest variances become emphasized due to the speed). Hence the invention of AS that can modify the base strategy to suit each individual disc according to its variance. The media used in the link you quoted were CMC-MAG-E01-000 (8x +R) and CMC is well known for having wide margins of quality from time to time.
Now 'learning' in a way of gradually getting better the more media of the same type you burn only works, if the media used are (rather strictly) all very much the same. You cannot fine tune something over several tries if the object of your tuning is changing constantly.
So why does the linked result look like the burn got gradually better? Well, for burning results with AS two main factors come into play: The individual quality of the disc inserted and whatever strategy AS is creating for it.
[Side note: Especially the latter factor can be a real PITA if you are using high quality media. If AS decides to use a just so slightly different strategy from one to the next burn you'll end up with a different error-pattern and eventually worry if your current stock of high quality Yuden discs is maybe junk half way down the spindle. It also makes testing and comparing rather difficult since you can never tell if you're actually seeing the result of the same writing strategy, never mind the quality of the disc used.]
So in the case you linked maybe the first disc was not so good or AS came up with only a sub-optimal strategy or maybe even both; the second was an maybe an OK disc and/or AS worked just fine while the third was even slightly better and AS used the same strategy as in the second case (anyway the difference is to small to be sure). Also in the end the number of the tests compiled is to small to be sure IMHO. It could well be that the next burn would have show worse results again.
Without more knowledge on the internals and more professional testing equipment it's very difficult to say what the true cause for the results is. The folks over at www.cdrinfo.com have started to use professional testing equipment from Datarius. I sure would like to see them taking a deeper look at AS.