Poisoned water & anthropic effects

An interesting thought experiment I ran into in class recently:

You have in front of you two glasses. One is poisoned with a lethal dose that will kill you in your sleep, the other is just a normal glass of water — but you don’t know which is which. You randomly select one glass, then drink it.

Which one did you drink?

The only possible way that you can update your beliefs is in favor of “it wasn’t poisoned.” If the water you drank was poisoned, then you will pass in your sleep and never actually make the update in favor of “it was poisoned.” So you know that, tomorrow morning, you either will be updating toward “it wasn’t poisoned” or you won’t be updating at all. But, from Joe Carlsmith: "If you know, now, what you’ll learn later, you can learn it now, too.” In other words: if you know for a fact that you’ll make some update in the future downstream of some evidence you anticipate receiving, then you should just update your beliefs toward that evidence now.

But applied to this situation, it would mean that you should update, now, toward “I drank the unopposed water” — which seems clearly counter to the intuition that you just chose randomly between the two glasses.

See Predicting updating for some more of my thoughts on this in generality, rather than an unusual edge-case.