CS188 is the course I point people to when they say they want to understand AI from the ground up rather than just use a model someone else trained. It teaches the part of the field that does not go out of date, how an agent searches a space of possibilities, how it plans against an opponent with minimax and alpha-beta pruning, how it makes decisions under uncertainty with Markov decision processes and reinforcement learning, and how it reasons about probability with Bayes nets and hidden Markov models. The reason it has lasted is the Pacman project series. Instead of toy exercises you build the actual logic that drives an agent through a maze, first with uninformed and heuristic search, then adversarial search, then value iteration and Q-learning, and the autograder gives you that satisfying click of knowing your implementation is correct.
Doing the work yourself is what turns these from abstract ideas into things you genuinely understand, and it is hard to overstate how much better that is than watching someone explain them. The lecturers built much of this material, so the explanations are clear and the framing is the one the field actually uses. The honest limitation is scope. This is classical AI, so deep learning and the large language models everyone talks about now are touched on rather than taught in depth, and if you came expecting to build a chatbot you will be surprised.
You also feel the absence of course staff when you self-study, because a stuck project is yours to untangle. But none of that changes the verdict. The foundations here make everything else you learn afterward make more sense, and I would happily send any serious beginner through CS188 first and a modern deep learning course second.