Jeremy Howard's fast.ai course has been the loudest counter-argument to traditional ML pedagogy for almost a decade. The pitch is simple. Most courses spend weeks on calculus and probability before you ever touch a model. fast.ai puts you in front of a working image classifier within an hour of starting and then teaches you everything underneath as you go.
I have watched this work brilliantly for some learners and fail completely for others. If you learn well by doing first and asking why later, this course will feel like a revelation. If you need to understand exactly what a gradient is before you trust the code, the early lessons will feel like cargo culting. Stick with it long enough and the theory does arrive, but a lot of people drop out before that point.
The first part of the course covers practical applications across computer vision, tabular data, collaborative filtering, and NLP. By the end you have built and trained models in all of those areas. The second part is more advanced and goes into the actual implementation of diffusion models, which is some of the most genuinely cutting edge teaching available for free. Part 2 in particular has earned its reputation as one of the few resources that demystifies stable diffusion at the level of someone who could rebuild it.
The fastai library itself is a point of friction. It is opinionated and abstracts a lot of PyTorch away. Some people love this because it lets them focus on the modelling decisions rather than boilerplate. Others, particularly engineers who want to deploy models in production, find it gets in the way.
My honest take is that you should treat fastai as a teaching tool rather than a long-term production framework. Use it to learn the concepts, then drop down to PyTorch for real work. The course videos hold up well even though some are now several years old, because the underlying ideas have not changed nearly as much as the AI news cycle would have you believe. The forum is one of the best learning communities on the internet, although it has slowed since Jeremy and the team moved focus to other projects.
If you are willing to do the work, this is still the single highest leverage free resource in deep learning.