If you have ever stared at the backpropagation equations and felt like the symbols were hiding the actual idea, this is the series that pulls the idea back out into the open. Grant Sanderson builds a small handwritten-digit network and then walks you through what each layer is really doing, what the weights and biases mean, how gradient descent nudges them, and how backpropagation distributes the blame for an error backwards through the network, all carried by the kind of animation that makes you go oh, that is all it is. The genius of it is that he never lets the maths become the point, the maths is always in service of a mental picture, so by the end you are not memorising formulas, you are imagining a surface being rolled down or a vector being nudged, and that intuition is the thing that makes every later course easier. He has also added chapters on attention and transformers that explain, at the same patient pace, roughly how the models behind ChatGPT actually process language, which is worth the watch on its own.
The catch, and it is an obvious one, is that this teaches understanding and nothing else, there is no notebook to run, no dataset to wrangle, and no assignment to check your learning, so on its own it will not make you able to build anything. It is also genuinely short, a handful of videos, so think of it as the conceptual spine you wrap a proper hands-on course around rather than the course itself. Pair it with something like fast.ai or Andrew Ng's specialisation and you get the best of both, the picture and the practice, and honestly I think everyone learning deep learning should watch this series at least once.