Eyewatering Expense of Tech Junk

Around 2011, I ended up working in Nashville on a "music recognition system". The company I worked for was a part owner of the Shazaam music recognition system patents. They initially use the system to monitor airplay of songs on the radio. It was interesting to be able to work through all that code.  I fixed a minor bug in the software which caused it to crash periodically, but then the company mothballed the system. It was easier and less expensive to pay people overseas to do the same work by hand.

Anyway, that system relied on "fingerprinting" techniques which were developed, I guess, in the 1990s. One application was panoramic photo stitching. The music recognition system was actually an image processing technique and used the same approach as image stitching algorithms. Anyway, by 2010, audio was "easy", that is well within the processing capability of plain-old x86 server hardware so one server could handle many simultaneous searches of a database. Panoramic photos were "easy" too and were in the processing capability of even a smartphone.

An interesting trend today is many ultra-brute-force super dumb algorithms are appear to be viable on hardware like GPUs and FPGA accelerator cards, but that shit is super expensive at the moment. For example an FPGA with a cutting edge FPGA is around $10,000. Capable GPU cards are similarly expensive. Additionally the GPUs consume so much power that they are having trouble cooling the boards adequately. Right now there's a problem of connectors melting, for example.

One example of a brute force algorithm is so called "Gaussian Splatting", which can be used to reconstruct a 3D model of a scene from multiple 2D photos; the demos are almost like magic. However, it's extremely computationally intensive. Similarly various radio networking and RADAR applications are stupidly computationally intensive. The neural network pattern matching and machine learning techniques are also brute force approaches.

This is a novel scenario in tech world. The fingerprinting algorithms of the 1990s and early 2000s were the long standing pattern where clever and efficient techniques created new applications. They also followed the pattern of "mass production" leads to lower cost and very low per use resource consumption.

Brute force solutions to problems were generally dismissed as not close to viable for real world use. They're too expensive. I think this is still true, but there's a lot of investor money and PR bullshitting that obscures this common sense perspective.

When there is a lot of funny money, many things can look viable and get support even if they're totally stupid. In the 1920s, for example, people thought blimps were a viable mode of transportation because it kind of worked and there was enough surplus cash to "invest" in blimp companies, but obviously it wasn't close to viable. Similarly, I doubt a datacenter on every corner chewing up 1000 households of electricity is viable and a technician isn't going to run around to a million data centers swapping out burned up GPU boards.