Andrej Karpathy's Neural Networks: Zero to Hero
karpathy.ai/zero-to-hero.html
A free video lecture series by Andrej Karpathy building neural networks from scratch in pure Python, from backpropagation fundamentals through GPT-level language models.
Why it is useful
The clearest explanation of how neural networks actually work at the implementation level, taught by one of the leading researchers in the field. Building a GPT from scratch in numpy removes the magic from transformer architectures and gives you genuine understanding rather than surface familiarity. Particularly valuable before diving into fine-tuning or building on top of existing models.