Why? Three reasons:
First, a 'dirty little secret' of the US software industry is how much of the learning from the beginning of electronic computing to the present has been just from individuals teaching themselves from books, e.g., K&R on C, Lippman on C++, Ullman on database, Sedgewick on algorithms, and on-line, e.g., Microsoft's MSDN site, StackOverflow, etc., essentially independently without courses, lectures, problem sessions, credits, homework, tests, etc.
Or since just K&R, ..., StackOverflow, etc. have been responsible for so much learning so far, then 'the bar is low' and on-line courses should do even better.
Second, in my experience in technical subjects, pure and applied math, mathematical physics, some topics in electronic engineering, e.g., surrounding the fast Fourier transform (FFT), and software, with a class or not, nearly all the learning (in my case) took place from study, alone, in a quiet room, from good materials just on paper. My 'educated guess' is that on-line learning can't replace such learning but can help stimulate more of it.
Third, for a researcher in applied math and software, and also likely some other fields, one of the main 'work items' is to take recent books and papers and work through them much as working through advanced course materials in such subjects.
Back to my case, in the fields I worked hardest on and did the best in, math, physics, and software, starting in the ninth grade, through my Ph.D., and in my career to the present, I did nearly all the work with relatively little contribution from teachers. E.g., in plane geometry my teacher was the most offensive person I've ever known in education, and I slept in her class and refused to admit doing her assigned homework. Instead, I worked every non-trivial problem in the book including the more advanced supplementary problems in the back where she never made any assignments. Then, after working all those problems, no wonder, on the state test in the subject, I did fine: I came in second best in the class; the guy who beat me also beat me by a few points on the Math SAT -- we were 1, 2 in the school. Net, my approach to learning plane geometry worked fine.
I never took freshman calculus. The college I started at wanted me to take some math that really was just a review of what I'd covered in four years of math in high school. So, a girl in the class let me know when the tests were, and I showed for those. The teacher said I was the best math student he'd ever had. Meanwhile, I got a good calculus book and started in, worked hard, and did well. For my sophomore year I went to a much better college and started on their sophomore calculus and did fine. Yup, never took freshman calculus.
When I went to graduate school, I took a problem with me and had an intuitive solution. My first year had some good courses (one was just terrific, from a star student of E. Cinlar now long at Princeton) and gave me what I needed to turn my intuitive solution into a solid math solution; I did that in my first summer, independently; and that was the research for my Ph.D.
I continued that way in my career: E.g., in a software house working for the US DoD, I saw a problem in a specification, got Blackman and Tukey, 'The Measurement of Power Spectra', and read it carefully enough to see what was wrong with the specification and how to fix it. Right: without courses, lectures, problem sessions, ....
As far as I can tell, nearly all the technical content on HN, StackOverflow and other Internet fora is from people who taught themselves in similar ways with little or nothing in courses, lectures, problem sessions, .... And that's part of what researchers have to do and is just part of getting tenure as a research professor.
So, since so much work is being done by essentially independent study now, just by not making things worse on-line courses should be able to look successful.
But I see some problems with the on-line materials I tried:
(1) The video quality just sucked. I couldn't read the board. That meant I couldn't copy what was on the board and study it. Bummer.
(2) The sound quality was not good enough.
(3) The course materials, e.g., on paper or in PDF files, were from not good enough down to just missing.
(4) Sadly the quality of the course content was too low; apparently the main reason was the desire to make the course more 'appropriate', that is, 'easier', for 'the common man in the street'. But, omitting material 'waters down' the course content and, really, for a good student, requires that they fill in the gaps for themselves -- bummer.
E.g., I looked at the course by Stanford professor Ng on 'Machine Learning'. What I saw were weaknesses (1)-(4) above. For more, (A) a lot that he was doing was maximum likelihood estimation but with far too little explanation and justification; so, I would have had to have run off and studied maximum likelihood estimation on my own. So, again I was on my own to do some independent work, trusting Professor Ng that somehow maximum likelihood estimation was better than it has long seemed in the statistics community. (B) He mentioned the 'maximization' to be done via following gradients, and that is an overly simplistic and not very promising approach to maximization -- the standard, first problem is that spend nearly all the computing time moving in directions nearly orthogonal to the direction really should be moving in.
So, from (A) and (B), I concluded that for a good course I would have had to have taken his lectures just as topics to be investigated, gone to good materials elsewhere, collected good details, and written my own text. That's his job as a professor, not my job as a student. His field, 'machine learning', didn't look worth that much work for me now. I've done some serious work in several cases of applied math that could be called 'machine learning' as much as his material, and I'm left without much respect for his material.
I looked at the course 'Probabilistic Graphical Models' taught by Daphne Koller. Since I very much liked a course by a star student of E. Cinlar, maybe I should like Kollar's course. Sorry, I didn't -- the quality looked too low. Better quality from Stanford? Sure, K. Chung, H. Royden, D. Luenberger (his 'Optimization by Vector Space Methods' is a beautifully, even elegantly, done one mile long applied math dessert buffet), D. Knuth.
For courses in 'how to code', that is, introductory material in software, gotta be kidding! 'Coding' alone is easy; it's just, pick a language, learn the basic syntax, and write if-the-else, do-while, call-return, allocate-free, etc. It's easy but doesn't take one very far. So, don't get very far with thousands of Web pages of documentation at MSDN on .NET, ASP.NET, ADO.NET, administration of SQL Server, IIS, Windows Server, etc. or the equivalent in the Linux world.
I have two broad conclusions:
First, traditionally in academic material in technical subjects, the author and the student 'reached' to each other, and they connected at a well written textbook. So, the author went far enough toward the student to prepare a good text -- and the best texts are terrific. And the student reached far enough toward the author to make do with little or no more than a good text. It's how I learned plane geometry, freshman calculus, theoretical, applied, and numerical linear algebra, everything I learned about statistics, most of what I learned about advanced calculus, signal processing and the FFT, stochastic optimal control, artificial intelligence, ..., and everything I learned about software. E.g., it's heavily what worked for me.
The on-line community will have to face the fact of this 'reaching'. In effect, on-line learning requires the professor to do more work of a kind that promises not to be well rewarded by tenure and promotion committees. So, net, so far a student should still reach mostly for one of the best texts, on paper or PDF.
Second, the situation will 'settle out': The professors will come to understand what minimum quality is needed, and the students will realize that the learning is not just a spectator sport, is not like watching a movie, and still requires nearly all the traditional work from a book or PDF file. Then the courses will get better; the students just looking to watch a movie won't sign up; and the course completion rates will increase.