Yeah, for learning that's good. But for novel research, not so much. I do a lot of what I always call "fast math on a computer" because that was a by product of me writing my own tools to solve problems in grad school. I didn't have numpy and only very limited BLAS optimizations existed at the time, so I had to write lots of low level stuff. But the actual novel work was pretty small on top of that.

In my grandparent post I mentioned that I could redo my PhD thesis in about a week of work. Much of that is that I know where the dead ends lie now. But a lot is also that I could just take advantage of numpy and I could just write everything in vector math now and not need to code up my own linear algebra stuff.