Slow Burning Torch

Akshay Mestry

National Louis University

Apr 18, 2025

Speed is for production, serving machines while slowness is for humans, for understanding

So, it all started on a quiet afternoon in early January of 2025, just after I wrapped up xsNumPy. I was still basking in the glow of having finally built NumPy (at least a fraction of it) from scratch to demystify how array operations worked. It was a project born out of curiosity. I didn’t want performance, nor did I care. I just wanted to understand. And as I was pushing the final commit to the repository, I felt two things very clearly:

• Relief, because rewriting xsnumpy.ndarray and debugging the broadcasting logic from first principles is as mentally taxing as it is rewarding

• Restlessness, because once you’ve peeled back the curtain on one big library, you start seeing others in a whole new light

Rewriting NumPy had shown me the power of writing slow code to understand fast abstractions. Now, I wanted to take that same approach to something I’d been using for over half a decade, PyTorch.

Rewriting PyTorch Badly, On Purpose

Now, I’ve been working with PyTorch since 2018, back when torch.nn.Sequential felt like magic and torch.Tensor.backward() method seemed to compute gradients by sorcery.