I think a lot of people are reacting here as if Netflix is moving from the only signal they have being the star and review system to the only signal they have being the up/down system, and thus confused as to why Netflix would throw away the vast majority of their rating ability.
But that's not even close to true. In addition to it technically being up/down/no rating, we've got how long people watch the show for, how the shows they watch form a pattern, how all the patterns of what people watch show global preference patterns, whether they rewatch a show, when they watch what sort of shows, what individual scenes are rewatched vs. skipped... Netflix is swimming in a sea of preference data, not sitting here trying to figure out "Gosh, um, if the user likes this movie 3 vs. 4 stars, uh... what does that mean?"
It makes perfect sense to me to optimize this one-data-stream-among-many to increase user participation and get more bits of information from more people engaging with the simpler system, rather than trying to squeeze bits out of the few people willing to use the star system and the even fewer willing to write useful reviews.
It isn't really even as shocking as it may seem at first. The star system has 6 states, "no rating", 1, 2, 3, 4, and 5 stars. That's 2.6 bits, with some simplifying assumptions [1]. The thumb system has up, down, and no rating; that's 1.6 bits [1]. To make up for the bits, you need only see ~40% increase in participation over the current star system... think that's going to happen?
[1]: The simplifying assumption is that all outcomes are equally likely, but that's not true. I don't have the numbers to run a more complete information theory analysis, but it's not hard to imagine the "no star rating" case is so common that it produces such a small fractional bit that a higher-participation-rate yea/nay/no rating (if you puth this UI in their face, "no rating" becomes much more meaningful, too) straight-up produces more bits of information on average, and is thus simply an improvement even before considering the superior UI experience. I rather suspect this is the case, some very sensible assumptions would suggest this, but I lack the ability to prove this; the "assume all outcomes are equally likely" is at least a concrete case I can discuss.