> A CS student developing novel learning-based image processing methods has no need for knowledge of obscure gdb commands and C++'s pointer and dependency voodoo.
I guess you could argue that AlexNet was the product of one very motivated CS student who was very proefficient with C++ and CUDA. This particular skill set had a very deep impact on the computer vision community. Today, how ever, you don't need it because all that has been abstracted away. But someone had to lead the way.
Definitely! I'd say we need more people like that (although for obvious security reasons, I'd personally prefer the use a safer lang than C/C++). But we also different people who don't care about performance or integration, but have unique ideas that are worth exploring. Let them experiment and quickly iterate using the easiest most high-level tools we have, then let the GPU magicians rewrite and optimize their work for production.
The best progress happens when you have both of these groups, not just one. As well as some application development plebs like myself (jk) to allow end users to benefit from all of that progress.
I guess you could argue that AlexNet was the product of one very motivated CS student who was very proefficient with C++ and CUDA. This particular skill set had a very deep impact on the computer vision community. Today, how ever, you don't need it because all that has been abstracted away. But someone had to lead the way.