Mostly because, when you're doing something that cares about environmental overhead and compiler efficiency, it tends to involve the kinds of algorithms that benchmarks test with. In other words, benchmarks are designed to provide exactly the information that is useful to the people that need benchmarks. You're absolutely right that isn't the case for web, phone, and applications programmers, and it's true that those people shouldn't pay attention to benchmarks, but there are people out there that care quite a bit.
For example, I write code for humanoid robots. If my code takes more than about half a millisecond to respond to sensor input, my robot will fall over and possibly hurt itself (thousands of dollars of damage) or me (...no thanks). That means that the language and environment I use actually matters - I don't get to have a garbage collector or a high-level runtime VM, to start with. And what algorithms am I running? K-means, n-body, matrix multiplication and inversion, and constraint satisfaction AKA sudoku. As I said, the benchmarks are designed to suit the people that need them.
Personally, I am very fortunate to not think about nothing but CRUD apps all the time. Working with games, all the fun things like sort speed, path finding, etc are real world problems :)
Not that long ago my main programming tasks were: * plotting some kind of Julia sets and analyzing its properties * high precision computations (>150 significant digits) of invariant manifolds
A lot of people I know are world-class in... N-body stuff.
It is hard to quantify those through and stackoverflow posts and github repos are an insufficient metric
Lots of people have to dig into clustering/mining/stats of the data we gather to figure out how to pull money and from whom.
As an admittedly fairly poor analogy, in programming you generally still have to assemble the car first. Silly things like laying out controls, moving data from/to the db, etc.
Having a really easy to assemble car is still going to take so much more time than the time you spend driving from A to B. So it's much more important. Add on top that you're going to be constantly asked to disassemble it and reassemble it to change the wheel size, etc. because accounting forgot to say that you'd be driving off road.
This was the point that I completely disagree with. When somebody gives me a new programming language (Julia for example), I'm going to want to know the numbers just so I have that as a baseline understanding. Afterwards sure, I want to know all the bells and whistles the language gives me over C.
Most of us write software that moves bits from an API or DB to a screen. If we're lucky we translate a couple of those values to a color or something.
Why are benchmarks so focused on code no one writes? How many successful YC companies aren't CRUD apps?
Type the problem you have into google, write your stuff in the language that has the most concise stackoverflow answers on how to solve that problem.