This course has sold enormously and has a high average rating, and I want to explain why both of those facts are true while still thinking it is a weak course. Eremenko and de Ponteves are good presenters. The delivery is warm and enthusiastic, the pacing is comfortable, and the course is structured so that beginners feel progress quickly. That is a real skill and it explains the ratings, because course ratings measure how a course felt rather than what it taught.
The intuition sections are the strongest part. Before each architecture there is a conceptual explanation with diagrams, done without equations, and these are genuinely decent. The explanation of how a convolution works, or what a recurrent connection is doing, will leave a beginner with a usable mental picture. If the whole course were that, I would rate it higher.
The problem is the coding methodology. The approach is template driven. You are given a code template, the instructor walks through it line by line, you type along or fill in blanks. At the end you have a working model.
What you do not have is the experience of deciding anything. You did not choose the architecture, did not diagnose why training failed, did not debug a shape mismatch, did not tune anything. Those experiences are where deep learning competence actually forms, and this course systematically removes them. The mathematics is avoided rather than made accessible.
There is a difference. A good beginner course simplifies the mathematics so you understand what backpropagation is doing without deriving it in full. This course largely skips it, which means the model remains a box with knobs. When something goes wrong later, and it will, you have no basis for reasoning about why.
The dating is a serious issue. The course has been updated over the years and substantial parts still reflect practice from when it was written. Framework versions, API patterns and the general workflow have moved on. Learners regularly hit code that no longer runs and spend time on version debugging that teaches them nothing useful.
The topic selection is odd for a modern course. Self-organising maps, restricted Boltzmann machines and autoencoders get significant coverage. Autoencoders remain relevant. SOMs and RBMs are essentially historical, and time spent on them is time not spent on transformers, attention, transfer learning or anything that resembles current practice.
A deep learning course in 2026 that covers Boltzmann machines and not attention has its priorities from another decade. The most damaging outcome is the confidence. Someone finishes twenty three hours of video, has a certificate and several working notebooks, and reasonably concludes they know deep learning. They then attempt something on their own dataset and discover they cannot begin, because every step they took was scaffolded.
I have seen this pattern repeatedly and it is worse than learning nothing, because it wastes months before the person realises they need to start again properly. The alternatives are not expensive. Fast.ai is free and gets you building real things with real understanding. Daniel Bourke's PyTorch material is free and far more current.
The deeplearning.ai specialisation costs a subscription and actually teaches the mechanisms. Karpathy's Zero to Hero is free and is the best thing in existence for understanding what is really happening. All of them are harder. All of them are better.
My three point two is for competent presentation, genuinely good intuition sections and a low barrier to entry, marked down heavily for a template approach that prevents real learning, avoided mathematics, dated material, obsolete topic choices and the false confidence it produces. If it is five pounds and you need something gentle to start, take it and then take something real.