Neural Networks: Zero to Hero

What is Neural Networks: Zero to Hero?

Neural Networks: Zero to Hero is Andrej Karpathy's video series on building neural networks from the ground up, published free on YouTube with the code in public repositories. It starts with backpropagation derived by hand, then implements micrograd, a tiny autograd engine written in plain Python, so the gradient machinery stops being something a framework does for you. From there the series builds makemore, a sequence of character-level language models that progress from bigram counts to multilayer networks, and finally works up to a transformer and a small GPT implemented line by line. Nothing is imported to skip a step. Each lesson is a long recorded coding session where the model is written in front of you, mistakes included. The result is unusually good intuition about what a language model actually computes, at the cost of covering a deliberately narrow slice of the field.

Who is Neural Networks: Zero to Hero for?

Take this series when you have used language models, possibly built on top of them, and want to understand what is underneath instead of trusting the abstraction. It is the right material before you attempt anything involving training, fine-tuning, or debugging model behavior at a level deeper than prompt wording. It also suits people preparing for machine learning interviews where being asked to explain backpropagation or attention is likely. You need comfortable Python and a willingness to type along rather than watch, because passive viewing gives you very little here. It is not a survey course and will not teach you computer vision, tabular modeling, or how to deploy anything, so pair it with broader material if you need coverage.

How does Neural Networks: Zero to Hero compare?

fast.ai works top down, getting you to a trained working model in the first lesson and explaining afterward, which builds practical capability faster across more problem types. Zero to Hero refuses that shortcut on purpose and goes bottom up through one narrow path, so you finish with less coverage and much sharper intuition. DeepLearning.AI sits between them with structured lecture sequences and a broad catalog of applied short courses, better when you need to learn a specific technique for a project than when you want mechanism. Hugging Face NLP Course assumes the transformer exists and teaches you to use and fine-tune it within an ecosystem, which is the practical complement to what Karpathy builds by hand. Many people watch this series alongside one of the others rather than instead of them.

Compare with: fast.ai, DeepLearning.AI, Hugging Face NLP Course

Neural Networks: Zero to Hero pricing

The catalog lists this as free. The lessons are published on YouTube with no account, subscription, or trial involved, and the accompanying notebooks and repositories are openly available. There is no certificate and nothing is sold alongside it. The models built in the series are small by design and most of them will train on a laptop, so unlike heavier deep learning courses you are unlikely to need paid compute to follow along and reproduce the results yourself.