Sergey Levine is one of the better lecturers in reinforcement learning and it shows most in the second half of this course. The early material on backpropagation, initialisation, normalisation, convolution and optimisation is competent and standard. Where the course becomes distinctive is when he gets to the topics he actually researches, and the RL lectures here are clearer than most standalone RL courses I have watched, partly because he has spent years working out how to explain policy gradients to people who already know supervised learning. The framing of the course is worth flagging because the title is honest about it.
Designing, visualising and understanding is the point, so a fair amount of time goes to what the network is doing rather than how to get it trained. Adversarial examples, visualisation of learned features, and the general question of whether your model learned the thing you meant get proper treatment. That emphasis was slightly unusual in 2021 and has aged well, since interpretability turned out to matter enormously once these systems started making decisions people care about. The problem with recommending it in 2026 is the date.
Spring 2021 predates almost everything that defines current practice. Transformers are covered because they existed, but the treatment is as one architecture among several rather than as the thing that swallowed the field. There is nothing on scaling laws, nothing on instruction tuning or RLHF, nothing on mixture of experts, nothing on the practical reality that most people now fine-tune or prompt rather than train from scratch. If you came here expecting a current LLM education you would leave confused about why the field looks so different.
There is also a slightly awkward detail on the course page itself. It links a public YouTube playlist and then says plainly that this link is not intended for students taking the course, who get an internal playlist instead. I appreciate the honesty and it does mean outside learners are getting the second-best version by design. Combined with no access to the assignments, which is where deep learning courses do most of their teaching, you are getting maybe sixty percent of the course.
Who should watch this. Someone who already has the classical machine learning foundation, wants the deep learning concepts built up properly rather than through a framework tutorial, and is willing to get the last five years from elsewhere. The RL lectures alone justify the time if that is a direction you are heading, and honestly I would watch those before I watched a dedicated RL course. Four point two.
Strong teaching, real breadth, free, and five years behind the front of the field in a field that moved further in five years than most move in twenty.