MIT OpenCourseWare started in 2001 on a premise that sounded faintly mad at the time, which was that MIT would publish the materials from essentially its entire curriculum and charge nobody anything for them. Twenty five years later it is still there, still free, still openly licensed, and still one of the few places on the internet where the phrase free education is not a lead capture form. Whatever else I say about it, that deserves acknowledging first. There are now materials from more than 2,500 courses.
The important word is materials. This is an archive, not a course platform. What you get varies enormously depending on which course you land on and who taught it. At the top end you get complete video lecture series, professionally recorded, with lecture notes, assignments, exams and solutions.
At the bottom end you get a syllabus, a reading list, and three PDFs. There is no warning about which you are about to click on, which is the single most frustrating thing about the site. The great courses are genuinely great and they are the reason to be here. Gilbert Strang's linear algebra lectures have taught more people the subject than any textbook ever has.
Patrick Winston's artificial intelligence lectures are worth watching for the teaching alone, regardless of how much the field has moved on. The algorithms courses, the probability courses, the signals and systems material, all of it holds up because the mathematics does not change. If you are trying to build the foundations that make machine learning make sense rather than machine learning feel like an incantation, this is one of the best places to do it and it costs nothing. The problem sets are the part I would point at hardest.
Almost every free resource on the internet gives you explanation and stops. Explanation is the easy half. The difficulty of a subject lives in the exercises, and MIT publishes the exercises with the solutions, at the difficulty level that MIT students actually face. Working through a real problem set and getting most of it wrong tells you something about your understanding that no amount of video watching will.
Most people skip this and then wonder why nothing sticks. Now the honest limitations, and they are serious. Age is the first one. A lot of the archive is old, and how much that matters depends entirely on the subject.
A 2011 linear algebra course is fine. A 2010 artificial intelligence course is a historical document with real teaching value and no practical currency, and someone who does not already know the field will not be able to tell which parts have been superseded. The site does show dates, and a beginner has no way to judge what those dates imply. The absence of structure is the second.
There is no path. There is no sequence. There is nothing telling you that you should do the mathematics before the machine learning, or which of four overlapping courses is the one to take. Universities provide that scaffolding through advisors, prerequisites and degree requirements, and OCW gives you the contents of the lectures without any of the apparatus that makes them navigable.
People who already have a technical education can construct their own path. People who do not are the ones who most need one, and they are the ones most likely to bounce off. Prerequisites are stated in the way universities state them, which is to say assuming you know what a course number means. A course that says it assumes 18.06 is telling you something useful only if you know what 18.06 is.
The practical effect is that people start courses that are three levels above where they are, conclude they are not smart enough, and stop. They are usually not lacking ability. They are lacking the two courses nobody told them came first. There is no support of any kind, which is the trade you accept for free.
No forum staffed by anyone, no grading, no feedback, no cohort, no deadline. Self directed learning at this difficulty has a completion rate that would embarrass any paid provider, and that is not MIT's fault, it is simply what happens when a hard subject meets no external accountability. If you know you need structure, buy structure somewhere, then use OCW for depth once you have it. I should mention the sibling projects since people confuse them.
MIT Open Learning Library has interactive versions of some material with automated feedback, which is closer to a course experience. MITx on edX has proper graded courses with certificates and prices. OCW is the raw archive with nothing added and nothing charged. My four point four reflects what this actually is.
As a public good it is close to unimprovable and I would not want it changed. As a way to learn AI from a standing start it is difficult, patchy and unguided, and it will defeat most people who try to use it that way. Take the mathematics courses, take Strang, take Winston for the pleasure of watching someone teach well, do the problem sets properly, and get your current AI material from somewhere written this year.