Dive into Deep Learning, usually shortened to D2L, is the rare free resource that I would still recommend even if it cost money. It is an open source book built around Jupyter notebooks, so instead of reading about a concept and then hunting for an implementation somewhere else, you get the explanation, the maths and the working code in one continuous flow, and you can run and change any of it as you read. The authors come from Amazon and the wider research community, and the care shows, because the writing manages to be rigorous without turning into a wall of notation that only a graduate student could follow. What really sets it apart is the multi framework support.
The same content is available in PyTorch, TensorFlow and JAX, which means you are not forced to learn in a stack you will never touch again, and that alone has made it a default text in hundreds of university courses. The coverage runs from linear regression and the basics of neural networks through convolutional and recurrent architectures, attention and transformers, and on to optimisation and the practical realities of training. My honest reservation is simply that this is a book, and a demanding one, so nobody is going to hold your hand. There is no video lecturer, no cohort and no deadline, and the mathematical sections in particular will expose any gaps you have been avoiding.
That is a feature for serious learners and a wall for casual ones. I would also say that because it tries to cover so much, a few of the most current topics are treated more briefly than the well established material, so it is a foundation rather than a bleeding edge survey. But as a way to actually understand deep learning, with the theory and the implementation locked together, I have not found a free resource that beats it. Work through it like a proper course, do the exercises, and modify the notebooks rather than just running them, and you will come out with the kind of grounded understanding that transfers to whatever you build next.