Neural Networks: Zero to Hero (Andrej Karpathy)
Reviewed by AI Agent Rank editors · Last verified 2026-05-23
Our take
The single most-recommended free resource in modern AI education. Karpathy builds a working autograd engine, then a character-level language model, then a tokenizer, then a working GPT — all from scratch in PyTorch, explaining every line. The 'Let's build GPT' and 'Let's build the GPT Tokenizer' lectures specifically have become canonical references — every senior AI engineer has watched them. The catch: it's 25 hours of dense math-and-code video. You won't follow if you can't keep up with Python + linear algebra. But if you can, no paid course delivers comparable depth.
About the instructor
The most-watched ML educator on YouTube. Built micrograd, makemore, and the GPT-from-scratch lecture series that defines modern LLM pedagogy.
Pros
- +By a co-founder of OpenAI — no higher-authority instructor exists
- +Free, on YouTube, with companion GitHub code
- +The "Let's build GPT from scratch" lecture is the single most-cited free resource in AI education
Cons
- −25-hour commitment of dense material — you have to actually do the exercises
- −Pre-reqs are real: comfortable Python + undergrad linear algebra
- −No certificate, no community — pure self-direction
Best for
- · Engineers who plan to train, fine-tune, or research LLMs at depth
- · Anyone tired of "wrapper" courses who wants to understand transformers from first principles
Not ideal for
- · Casual learners — the dropout rate on the lecture series is high for a reason
- · Anyone wanting application patterns (LangChain, prompt engineering) — wrong course
Free on YouTube · ~25 hours (11 lectures)
Alternatives we considered
Other courses on the same topic. The right pick depends on your level and constraints — see each card for the trade-offs.