I think the methodology is interesting, but you can probably design a similar, and probably better loop for "diving" into a topic and going through examples than is presented.
I have to imagine they just asked Claude to dive into a topic and generate example problems along the way.
As for the content... a lot feels like knowledge overload and concepts are introduced without explanation or "why". It basically says "here is a training loop" and never answers my immediate question of "...for what?" It also introduces random concepts like setting the seed that don't look like they're even in the final training loop it provides.
As a comparison, Pytorch itself has docs that go over the entire training loop as well. And it explains why you're loading the data its loading. With examples that run instead of being scattered on.
PyTorch already does a huge job at keeping DL not verbose. It captures the complexity to allow you to make a whole GPT model with it, running on your GPU, and keep the code readable. Yes, there are a lot of breaking points, but at some point, DL is supposed to be non trivial, it handles math concepts most people don’t understand. That is why Data Scientists have a job ;)
5 comments
[ 4.0 ms ] story [ 21.9 ms ] threadI have to imagine they just asked Claude to dive into a topic and generate example problems along the way.
As for the content... a lot feels like knowledge overload and concepts are introduced without explanation or "why". It basically says "here is a training loop" and never answers my immediate question of "...for what?" It also introduces random concepts like setting the seed that don't look like they're even in the final training loop it provides.
As a comparison, Pytorch itself has docs that go over the entire training loop as well. And it explains why you're loading the data its loading. With examples that run instead of being scattered on.