Computerphile has been running since 2013 and the format has never changed: an academic, a sheet of paper, a marker, and fifteen minutes explaining an idea. Brady Haran produces it, the same person behind Numberphile, and the production philosophy is that a good explanation needs a good explainer and not a graphics budget. It works. The channel is the closest thing YouTube has to sitting in a university office hour with someone who genuinely enjoys the question.
The AI and machine learning catalogue is substantial. Neural network fundamentals, backpropagation, convolution, attention, transformers, diffusion models, tokenisation, reinforcement learning, and a great deal on the theoretical computer science underneath all of it. The strength across all of these is conceptual precision. A Computerphile video will not hand you an analogy that falls apart under pressure.
It will give you an explanation that survives contact with the real thing, which is rarer than it should be. Rob Miles deserves separate mention. His videos on AI safety, on the channel and on his own, are the best introduction to alignment that I know of in any format. Specification gaming, instrumental convergence, mesa optimisation, the difficulty of writing down what you actually want, reward hacking.
He explains these without doom, without hype, and without assuming you already agree that the problem is real. He argues the case. Anyone working on AI systems should watch that series, and it takes maybe four hours total, and it will change how you think about objectives and evaluation in ways that apply to ordinary engineering work and not only to hypothetical futures. The structural limitation is that this is a channel, not a course.
There is no sequence, no prerequisite ordering, no progression from one video to the next. You watch what the algorithm shows you or what you search for. That is fine for filling a specific gap and useless as a way to learn a subject systematically. Someone who watches forty Computerphile videos will know a great deal about computer science and will not be able to implement any of it.
Depth varies. Some videos go properly into the mathematics and others stay at a level a curious teenager could follow. Neither is wrong and the inconsistency means you cannot predict what you are getting. The comments and the video description sometimes signal which, and often they do not.
The archive age is a real hazard in AI specifically. Videos from 2015 about neural networks are still conceptually sound about what a neuron does and describe a field that has moved enormously since. A newcomer who lands on a pre transformer video about natural language processing and takes it as current will come away with an outdated picture. Check the upload date.
This is not the channel's fault and it is a thing to watch for. There is no code. Almost nothing here is implementable as presented, and the point is understanding rather than building. Pair it accordingly: Computerphile for why something works, Karpathy or fast.ai for making it work.
The two together are stronger than either alone, and quite a few people find that a conceptual explanation before an implementation makes the implementation land properly. Where I would start: the Rob Miles safety series first, then whatever concept you are currently stuck on. It is an excellent place to go when a textbook chapter has not clicked and you want the same idea from a different angle. My four point one is for consistently excellent explanation, remarkable longevity, and one series that is genuinely best in class, marked down for the absence of any curriculum and for an archive where the old material needs a date check before you trust it.