ACR
Back to index
OtherSelf-paced, around 10 to 15 hours of video, far more if you code along properly·Free

Neural Networks: Zero to Hero (Andrej Karpathy)

4.9

The single best resource I know of for actually understanding how neural networks and language models work, and it is completely free. It is demanding, it expects you to code along, and it is worth every hour.

What We Liked

  • Taught by Andrej Karpathy, who helped build the systems he is explaining
  • You build everything from scratch, including a working GPT, no hand-waving
  • Completely free on YouTube with the code openly available
  • Builds the kind of deep intuition that most paid courses never reach

What Could Be Better

  • Genuinely demanding, you need solid Python and real patience to keep up
  • No certificate, no cohort, no support, it is you and the videos
  • Coding along properly takes far longer than the runtime suggests
  • Not the place to start if you have never written code before

Detailed review

I do not say this lightly, but Neural Networks: Zero to Hero is probably the most valuable thing on this entire site, and it costs nothing. It is a series of long, unhurried YouTube lectures from Andrej Karpathy, who was a founding member of OpenAI and led AI at Tesla, so when he explains how these systems work he is explaining things he personally helped build. That pedigree matters, because instead of describing neural networks from the outside he builds them in front of you, in plain Python, one line at a time, starting from the absolute basics of backpropagation and ending with a working GPT that you have written yourself. The reason this works so well is that Karpathy refuses to hand-wave.

He starts with micrograd, a tiny automatic differentiation engine you build from scratch, so that gradient descent stops being a magic phrase and becomes something you understand mechanically. From there he builds makemore and walks through language modelling step by step, and by the later videos you are implementing the components of a transformer and assembling a small GPT. Nothing is imported and left unexplained. If you genuinely follow along and type the code yourself, you come out the other side with an understanding of how modern AI works that very few paid courses will give you, because most of them stay at the level of using tools rather than building them.

The honesty I owe you is about difficulty. This is not casual viewing. The videos look approachable because Karpathy is calm and friendly, but the material is real and it expects you to be comfortable with Python and willing to pause, rewind and wrestle with things until they click. Coding along properly turns a fifteen hour series into many more hours of actual work, and there is no certificate, no cohort and no one to ask when you get stuck, just you, the videos and the code.

If you have never programmed, this is not your starting point, you should get some Python and a gentle ML primer under your belt first. But if you have that footing and you are willing to put the hours in, I would point you here before almost any paid program. It is rare to get teaching this good from someone this qualified, and it is hard to believe it is free.

[ final ]

The verdict.

If you are serious about understanding modern AI rather than just using it, and you are willing to do the work, this is the resource I recommend above almost anything paid. Complete beginners should build some Python and basic ML footing first, then come here.