Q 1337
The Problem:
The public is unaware of breakthroughs in mathematics and computation in regards to statistics and machine learning and do not realize that Facebook, Google, and more are not just selling your data to advertisers but using it to train sophisticated Neural Networks capable of predicting the immediate future(and more).
The Who:
Components often have multiple uses a capacitor not only smooths a waveform but it can also generate a minute voltage through being stimulated by a kinetic force (sound waves for example) which can then be graphed and analyzed. Newer devices are constantly listening for 'key words' but in reality all the information has to be processed, it is trivial to encode that information on carrier waves or store it on the devices drive encrypted until a network connection can be found. (think of what you could justify data wise with 'check for updates' type features) Q must have the spec of the entire program to claim 24/7/365 monitoring, something most of us hadn't necessarily considered (the ones who did are probably dead).
The What:
Facebook, Google, etc are 'feeding' these computer programs all the information the public is willingly giving them. DO NOT UNDERESTIMATE MACHINE LEARNING (But don't be afraid either, more below) with the big datacenters and modern (publicly available) computing architecture they have the capacity to run extremely sophisticated algorithms. New 'TPU' (tensor processing units) are 3D RISC arrays that pump out 120+TFLOPS per card. GV100 series NVLINK has 300MB/s bandwidth between GPU units(TPU's are on GPU's). DARPA has also poached the talent from other projects to include photonics(resistant to EM side channel attacks), Unified Computer Infrastructures(FPGA + GPU + CPU + TPU + QC(presumably)), and more importantly and entire ARMY of mathematians to write software for these sophisticated computing tools. DO NOT UNDERESTIMATE STATISTICS.
The When:
Most 'successful' technologies are first developed by DARPA and then rolled out to the public in one form or another by a major company that gives you just enough to justify using it but not enough to actually make it a viable tool. Google claims to have developed their TPU in 15 months, riiight. Put it this way in 2014 you could have a 64x64 (4096 cores) 1Ghz RISC Array (a 2D 'TPU') and today you can have the luxury of a 9x9x9 RISC Array for the price of a small car (a 3D 'TPU'). A Tensor Processing Unit allows the acceleration of Machine Learning applications nearly 16x of what the most powerful GPU can do with very little energy. The when has been since the age of the telegram, the M.O for state level actors has never had to change.
The Where:
All of the datacenters are suspect, Q seems to be referring to specific ones but any large datacenter can be used for these purposes. If you have 25k to pay AWS and a bunch of training data, a day later you can have a 99+% accurate non-invasive cancer detection platform. All it takes is training data and the public has up to this point been very willing to put every detail of their life on the internet.
The Why:
It's possible to compute and statistically derive a great many things based on observations that cannot be observed by a human. DO NOT UNDERESTIMATE STATISTICS.
The How:
There is nothing that can be observed that cannot be measured. Anything that can be measured can be used to derive useful information. MASINT + Machine Learning = Potential beyond imagination.
Not everything is all gloom and doom though.
Once FB, Google etc are [here] the training data can be deleted and the neural networks they've created unethically can be used in a manner that can benefit humanity forever.
1) Realtime Universal Translators
2) Realtime Health Analysis
3) Realtime Plan/Risk Analysis
4) So much more, Potential beyond imagination
"How is POTUS always 5 steps ahead"
Welcome to the Machine.
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
https://deeplearning4j.org/lstm.html