Full disclaimer: Claude Code was used heavily, but I experimented with fun stuff in addition to the cpu emulation, like a fully procedural asset production with aesthetic checker script.
The massive data gathering part should only be part of the learning phase of the system imo, once it get a good model of reality it should infer useful knowledge information from few data, like an expert.
I guess the point of view is that if a department is well running, it means it is overressourced. So you reduce the ressources until it's breaking point, just enough for it to not fail. A jaded service manager told me it was part of its official training: if the clients was too satisfied that meant that human ressources were wasted on them, so he had to spin plates between clients. I guess it was economically optimal.
I think that engineering progress made while building those machines are maybe more relevant for practical technical development than the discovery they make.
The problem isn't the cheaper MRI. The problem is the expert that needs to interpret the results. Detecting millions of cancers that don't actually exist doesn't help anybody.
This is a problem domain AI is good at. Have AIs do first-pass, then when they flag something an actual doctor reviews it. Then if they concur it goes to your doctor, who knows you, who can review it.
It's always cited as an example, and it's highly impressive yes, but it's also highly focused, with a very partial feature set, I'm not sure that type of coding is applicable to mass scale software.
