Written in jest and yet using CV at scale on GPUs initially used for LLMs make sense.

Yet... why do so? As I wrote just minutes ago in https://lemmy.ml/post/41546700/23280257 there is already very high quality signal that requires nearly no compute : your wireless trace. Google/Apple and your mobile provider or ISP and thus the government hosting it already know 24/7 where you are, how active you are, etc solely from your 5G/4G signal. Well OK for activity it's with the IMU but the point is this is basically computationally free.

You move around,

• your mobile phone scans for 5G signals,
• login in a nearby tower via its SIM/eSIM
• and voila, you are there. It's basically few requests on some databases and it's instantaneous.

compare this with

• your identity with facial features (lots of photo) is store in a large DB
• there is no known location so a network of thousands if not millions of cameras have to be queried to try to match your facial features again the last frames, so that's ~gigabytes of data to send somewhere or query all those cameras with setup locally
• there is a match! then repeat this locally for the next cameras, maybe just hundreds
• light change or hoodie on, no match, restart process

this is ridiculously expensive to run. I'm not saying it can't be done (it's been done and it's not hard to setup) but... WHY would one do so when the first setup works more reliably and is orders of magnitude cheaper?

Obviously both can be combined but also both can be bypassed conveniently and extremely cheaply (leave your phone home, wear sunglasses and a hygiene mask) so even though a realistic scenario I'd argue it's not rational to not just rely on what already works for the vast majority of situations.