I just wanted to point out that "from scratch" is not really well defined. There is always some arbitrary line. I just found it interesting to discuss and think about where to draw this line exactly. Obviously it's never really from scratch, i.e. you don't reinvent the hardware level, for example. Or you don't start with teaching quantum physics. So you start from somewhere.

And I was wondering whether auto-diff is maybe something which could also already be the starting point, or also matmul. Reimplementing an efficient matmul on CUDA is not easy, and might distract from the main deep learning content. But it depends also where you want to have the focus on.