Neural Networks: Zero to Hero (Andrej Karpathy)

Andrej Karpathy's hands-on series that builds neural networks and language models in code from an empty file: 10 long videos, about 19 hours. It starts by building micrograd to explain backpropagation, then the makemore character-level language models through bigrams, an MLP, activations, gradients and batch norm, manual backprop and a WaveNet-style model. It goes on to build a GPT from scratch, explain and build the GPT tokenizer, and finally reproduce GPT-2 (124M), with a talk on the state of GPT. Best for developers comfortable with Python who want to truly understand how ChatGPT-style models work.