But the problem isn't "how to do this in 48KB of ram" these days it's "how do I do this on a distributed compute cloud with high level services".

Neural networks were a hypothetical approach, nowadays you can rent a ML pipeline for a couple of bucks and download prebuilt models.

So taking an approach from that era (for example C or UNIX) has obvious legacy baggage and isn't optimal in modern context but is still used because of inertia.

Also at this point it's less about your brain and understanding and more about being able to search and mash together. It's not just having a fancy IDE, internet and search fundamentally changed how you learn and solve problems.