Nope, you need the reference phase to still be coherent with the observed object (temporal and spatial), so the interference is only possible between two parts of the same wave (of light), different in space (think two pinholes in a plane through which you collect the light) and/or in time (think delay line, let one part of the line you collected run a longer distance). The first is the famous Young's double slit experiment, the second is the Michelson interferometer.

Also, for reference : frequency of the visible EM fields is in the order of 300THz (300,000GHz).

This telescope operate in the radio bands (sub-millimeter) and not in the visible. That's why it is easy to make interferometry over very long base line. In the visible domain this is very tricky to realize over a couple of 100m (such as with the VLTI).

You can think of it as completing piece by piece the Fourier transform of the image you want to observe. Every pair of telescope gives you a measurement in the so-called UV plane (spatial frequencies). The furthest the observations point are (the telescopes) the smaller details you can get. Except this is only valid if you can measure the amplitude and phase of the electromagnetic radiation (or find a way to reconstruct it in some way). This is easy in the radio bands. But this oscillation is just too fast with visible wavelength and thus, we can not record and adjust offline, we have to interfere the waves right away...

Nope, they are coded with the relation color = abundance of atomic component. Colors are a stimulus, they do not exist outside of all of our brains. What is real is the wavelength, and that, for instance, the transition of an electron from the 3rd to 2nd layer of the structure of the Hydrogen atom will emit a photon at 656nm, which we call red.

Nothing to do with the FITS format. That's the same type of information all RAW formats have : unprocessed data, as close as possible to the signal coming from the sensor after quantization, with ideally no processing, offset or other adjustments performed.

Our vision is also subjective, it permanently adapt to lightning and ambient color conditions. There is no such thing as a true image representation. Especially in the mentioned case (a magazine), where it is desirable to have an image which pops the eye rather than a blob of washed out colors.

So what's the news here?

Do you have some references?

That is why people want to use compressed sensing for this. Then you never have to deal with the full bandwidth, as you're directly acquiring compressed data. Then, you also do the inference in the compressed domain.

The nightmare is that you might have to do the compiler job and generate an optimized SPIR-V IR from your language.

What about :
!s.empty() && s.back()=='t'
Or, before C++11 :
!s.empty() && (*s.rbegin())=='t'

For this one, the C++ implementation makes more sense than the other too : you can immediately store, test and work on the iterator and not wasting your time checking if the element is in the container, then find it again to work on it. If you don't need to use it, you can use std::count (because your container might contain it multiple times) as :
std::count(myContainer.begin(), myContainer.end(), theValue)>0
Or if you know that the element can only be present once, you should be using a std::set as container which already has a count function which only returns 0 or 1.
mySet.count()>0

std::string is still not very high level. Good parsing almost requires regexp methods which C++ did not have until recently (in the standard).
Otherwise, the first solution given in your link is verbose but does the job.

Do you have any other good examples?

No, the driver does not run a virtual machine. It has to recompile the bytecode in whatever-type-of-assembly the targeted GPU is taking.