> Many rely on that, in fact, since otherwise they'd be asking people to come up with a publishable (once published, perhaps) finding, under pressure, on the spot, in maybe an hour.
In much of science the hard part is asking the right question rather than coming up with the solution, so figuring out these algorithms is not equivalent to making publishable research.
Fair criticisms. I tend to prefer asking/being asked system design questions, which in nature allow for a wide range of experiences to excel at.
What they aren't the best at, although more realistic, is coming up with an objective measurement of quality. Algorithms , at the cost of realism, do measure things quite nicely.
This might be fine if the "very similar problem" is something you'd likely have encountered in your work, but often they questions are drawn from a pool that most people in this line of work see rarely if at all, and if they do it's likely to be some very small subset of the questions, so dedicated study of the remainder of the pool still puts one at a large advantage, regardless how useful it is in doing your actual work.
They're measures of "how bad you want it" (how much of your time you spent memorizing stuff you don't actually use to prep for the interviews) and/or how recently you took an algorithms course. And maybe those are things worth measuring, I dunno. Maybe the absence of strong enough signals on either of those is important enough that it makes sense to use them to reject people who are otherwise very capable of doing the actual work.