Neural Networks: Zero to Hero review
Andrej Karpathy · youtube
Our Karpathy Neural Networks Zero to Hero review: Neural Networks: Zero to Hero is the best free course for coders who want to see how a language model works by building one. Andrej Karpathy teaches 8 videos, about 14.5 hours in total, from backpropagation to a GPT and its tokeniser. It costs nothing, and there is no certificate.
What does it take?
| Difficulty | 4 out of 5 |
|---|---|
| Hours | About 14.5 hours |
| Format | Video, Projects |
| Maths needed | School-level maths |
| Coding needed | Yes, you will write code |
| Hardware needed | Free cloud GPU (Google Colab or Kaggle Notebooks) |
| Certificate | No certificate |
| Free to audit | Yes |
| Pace | Self-paced |
| Last updated by the provider | Not checked yet |
| Teacher | Andrej Karpathy |
| Length | 8 videos, 872 minutes (about 14.5 hours of video) |
| Prerequisite | Solid Python and intro-level maths |
| Cost | Free on YouTube |
| Certificate | None |
| Notebooks | Videos link Google Colab notebooks |
| Status | Ongoing: the syllabus ends with "ongoing..." |
| Community | Discord channel |
Who should take it, and who should skip it?
Take it if you are
- Coders who want to build a GPT from scratch and see how each part works
- Learners with solid Python and intro-level maths, such as derivatives and the Gaussian
- Anyone who has finished a first deep learning course and wants the ideas underneath
Skip it if you are
- Anyone new to Python: the course asks for solid programming
- Learners who want a certificate: the course offers none
- Anyone who wants quizzes and graded work: only the micrograd video mentions an exercise
What will you be able to do after it?
- Build backpropagation from scratch by writing micrograd
- Build character-level language models, from a bigram model to a WaveNet
- Explain activations, gradients and BatchNorm, and backpropagate by hand without PyTorch autograd
- Build a GPT from scratch, following "Attention is All You Need" and GPT-2 and GPT-3
- Build a GPT tokeniser using byte pair encoding
What's in each module?
| Module | What it covers |
|---|---|
| Video 1: Intro to neural networks and backpropagation | Build micrograd, a small backpropagation engine, from scratch. 2 hours 25 minutes. |
| Video 2: Intro to language modelling | Build makemore, a bigram character-level language model, covering torch.Tensor, training, sampling and loss. 1 hour 57 minutes. |
| Video 3: Makemore part 2, MLP | A multilayer perceptron model, with learning rate tuning, hyperparameters, train, dev and test splits, and under and overfitting. 1 hour 15 minutes. |
| Video 4: Makemore part 3, activations and gradients | Activations, gradients and BatchNorm. 1 hour 55 minutes. |
| Video 5: Makemore part 4, becoming a backprop ninja | Backpropagate by hand, without PyTorch autograd. 1 hour 55 minutes. |
| Video 6: Makemore part 5, building a WaveNet | A deeper, tree-like convolutional model, using torch.nn. 56 minutes. |
| Video 7: Let's build GPT | Build GPT from scratch in code, following "Attention is All You Need" and GPT-2 and GPT-3. 1 hour 56 minutes. |
| Video 8: Let's build the GPT tokeniser | Byte pair encoding, encoding and decoding, and why many odd LLM behaviours trace back to tokenisation. 2 hours 13 minutes. |
What we liked, and what we didn't
What we liked
- It costs nothing: all 8 videos are free on YouTube
- You build everything in code, starting with backpropagation and ending with a GPT
- The videos link Google Colab notebooks, so you need no hardware of your own
- It focuses on language models, and most of what you learn transfers to other areas such as computer vision
- The first video assumes only basic Python and a vague memory of high school calculus
What we didn't
- There is no certificate
- There are no quizzes or graded work: only the micrograd video mentions an exercise, a Colab notebook
- The real time is well above 14.5 hours, because you code along and pause to think
- The later videos expect solid Python, and the course asks for intro-level maths
- The series is marked as ongoing, so the video list may grow
What's the catch?
There is no certificate, and there are no quizzes or graded work: only the micrograd video points you to an exercise. The 8 videos run about 14.5 hours, but the real time is well above that, because you are meant to code along. The syllabus ends with "ongoing", so Karpathy may add videos.
Is it still current?
The course page was checked on 5 October 2026. It lists 8 videos totalling 872 minutes, and its syllabus ends with "ongoing...", so Karpathy may add more. It describes the course as building neural networks from scratch, in code, from the basics of backpropagation up to modern networks like GPT.
What should you take before it?
- Practical Deep Learning for Coders
Practical Deep Learning for Coders has 9 lessons of about 90 minutes, costs nothing and has no certificate. You deploy your own model by the end of lesson 2.
- Hugging Face LLM Course
Hugging Face LLM Course has 12 chapters of 6 to 8 hours each, costs nothing and has no certificate. You learn LLMs with the Transformers library.
What should you take after it?
- AI Agents Course
Build LLM agents that use tools and plan steps, using open-source frameworks and open models.
How we rated it: 9.3 out of 10
| Criterion | Score out of 10 |
|---|---|
| Clarity | 9 |
| Hands-on practice | 10 |
| Up to date | 8 |
| Depth | 10 |
| Truly free | 10 |
Questions people ask
Is Neural Networks: Zero to Hero free?
Yes. The videos are free on YouTube, and there is no certificate. Karpathy also invites learners to join the course's Discord channel.
Is Neural Networks: Zero to Hero good for beginners?
Not for beginners to coding. The course asks for solid Python and intro-level maths. The first video assumes basic Python and a vague memory of high school calculus.
What maths do you need for Zero to Hero?
Intro-level maths, such as derivatives and the Gaussian. The micrograd video assumes only a vague recollection of calculus from high school.
How long does Zero to Hero take?
There are 8 videos totalling 872 minutes, about 14.5 hours. This site lists 14.5 hours, which is video time only, so the real time is well above that if you code along.
Do you need a GPU for Zero to Hero?
No special hardware is needed to start. The videos link Google Colab notebooks, and the GPT video's model was trained on a cloud GPU.
What should you take before Zero to Hero?
Take Practical Deep Learning for Coders or the Hugging Face LLM Course first if you want a deep learning course behind you. You need solid Python either way.
So should you take it?
Take Neural Networks: Zero to Hero if you are comfortable in Python and want to understand what is inside a language model by writing it yourself. It scores 9.3: hands-on and depth are 10 because you build everything from scratch, and truly free is 10, while clarity is 9 and up to date is 8. It offers no certificate and almost no graded work, so you set your own exercises.
Where to go next
- Deep learning, hands-on
A learning path that includes Neural Networks: Zero to Hero.
- Build apps with LLMs
A learning path that includes Neural Networks: Zero to Hero.
- Generative AI for Everyone
What large language models can actually do for you and your team, without writing a line of code. Andrew Ng teaches it in 6 hours on Coursera, certificate paid.
- Machine Learning Crash Course
Google's Machine Learning Crash Course has 12 modules and takes about 13.5 hours by Google's own timings. It is free, with quiz badges and no certificate.
- CS50's Introduction to AI with Python
CS50's Introduction to AI with Python runs 7 weeks with 12 projects, costs nothing and gives a free CS50 Certificate if you score 70% on every project.
Last checked : teacher, video list and durations, prerequisites, notebooks, cost, certificate and series status.
Not sure this is the one?
Answer two questions and we'll point you to the course that fits where you are today.