lightvector
But that's unrealistic. In real life, the context for how and why there would be a speaker telling you such a thing in the first place can be relevant and affect the probability!
How is this possible? Suppose among all the math-riddle-loving parents of two children who would ask such a puzzle in the first place there are an equal number of parents of B-B, B-G, G-B, G-G, and that each is equally likely to ask you such a riddle when you meet them.
Suppose when asking such a riddle the B-B parents tell you "at least one of them is a boy" (they don't have any girls, so that's the only way they can ask this kind of riddle), the G-G parents tell you "at least one of them is a girl" (same thing but in reverse), while the B-G and G-B parents say one of "at least one of them is a boy" and "at least one of them is a girl" equally at random.
Then, conditioned on being told that "at least one of them is a girl", the probability of another girl is actually 1/2, not 1/3 like the paradox answer claims. To see this, imagine 40 examples of the above puzzle asking taking place. You get 10 B-B parents saying "at least one of them is a boy", 10 G-G parents saying "at least one of them is a girl", and among the 20 (B-G and G-B) parents since they choose randomly, you have 10 saying "at least one of them is a boy" and saying "at least one of them is a girl".
So out of the 20 times where "at least one of them is a girl" is said, there are 10 cases where it's a G-G family and 10 cases where it's a B-G or G-B family, therefore conditioned on being told "at least one of them is a girl", the probability of two girls is actually 1/2.
If there were some gender bias in how the B-G and G-B families might ask the question, or other differences that affect how likely different of these people would be posing the puzzle to you, then the probability could be yet different than either of 1/3 or 1/2.
So there's a difference in being present something as a flat mathematical assertion that you're supposed to take at face value and not supposed to question further (where the probability is 1/3, as the article claims). Versus being told something in real life, where you always need to take into account the context and situation of the speaker, and the probability could be different.
There are real life implications of this too - the big classic one being publication bias / newsworthiness bias. As most people intuitively know by now, it is also often wrong to take the statistical analysis or claims of a particular research study or paper entirely at face value, because there is a bias in the fact that "positive" and "exciting" results are more likely to be reported in the first place, and so statistical outliers that aren't actually replicable are disproportionately likely to be reported (see also https://xkcd.com/882/). And publication bias still occurs with respect to the reporting of results, amplification or not in the media etc, even when the the authors themselves are trustworthy and have done their analysis within the paper in a statistically proper way. So conditioned on you hearing about the result in the first place, it is often less likely to be true (and less likely to replicate in the future, etc) than you would think if you just took the statistical analysis in the paper at face value, even when that analysis was done correctly. The situation in the "sisters paradox" of computing a probability taking a statement entirely at logical face value is rare in real life.
However, I would not dismiss MC-search or expectimax. Whereas alpha-beta should perform well when the current position is in a "good" state, when the current position is in a "bad" or "tricky" state, it almost certainly will perform way worse than search methods that actually take into account the opponent being random.
To consider the extreme example, imagine the position is extremely tenuous and that alpha-beta has proven that if the tiles are placed in the worst possible way in the next several turns, it is guaranteed the position is a loss and a 2048 is impossible. Obviously, it would be better that point to switch to something like expectimax and play so as to minimize the probability of a loss and maximize the probability of recovering, which could still be quite large. Whereas with alpha-beta the only thing you can do is maximize the number of moves until the forced loss, which will often not be the same thing as maximizing the probability of recovery, particularly if most reasonable branches of play lead to a forced loss in the same number of moves.
Similar arguments apply when the position is not a forced loss yet but is still "bad". For example, if the "bad" position is such that your long-term win probability is only 30%, I would expect alpha-beta to pass-up opportunities to fix the position if those opportunities carried any risk of making the position worse, whereas taking those risks is clearly correct if they could fix the position into a probable win and the probability of making things worse was only 1/2 or 1/3.
On the relevant stackoverflow thread, there is another answer much further down that has received far less attention, which is a shame: http://stackoverflow.com/questions/22342854/what-is-the-opti...
I've downloaded and compiled the solver, which uses expectimax rather than alpha-beta, and uses heuristics that are arguably even simpler and dumber, and yet it seems to outperform the OP's solver. It gets 4096 with good consistency and even often gets an 8192, whereas the OP's solver (after a tweak to make it not stop at 2048) can sometimes require a few tries to get a 4096 and only rarely gets an 8192. Granted, there's a large computing power difference, since one runs as native compiled code and one runs as javascript in the browser, but it shows that expectimax, which one would expect to be the most theoretically-sound approach, is indeed at least a comparably good approach, and possibly a superior one.