I am fascinated by the Stanford online courses in machine learning and artificial intelligence. My first inkling of them came when quite a few of my students started taking the artificial-intelligence class. Olin is very small, only about 400 students, so I realized that these online courses must be large. But I almost fell over when I saw that enrollment varied from 66,000, at the low-end, to 160,000.

**Sebastian Thrun**, who co-taught the artificial-intelligence course to 160,000 students, is now leaving Stanford teaching in order to teach courses to 500,000 students for free. What an inspiring goal!

My only complaint, due to my habit (or handicap) of thinking about every number that I come across, is with the comments on the number of perfect scores. Here are the data: Of the 160,000 students, 248 students got a perfect score, and none was among the 200 Stanford students. These data have been widely reported with the implication that the Stanford students were less successful than the online students.

As a Stanford graduate, I needed an alternative model. Thus, imagine a pool of 160,000 students, among whom 200 are at Stanford. Now strew 248 perfect scores among the 160,000 students randomly.

What is the probability that no perfect score lands on a Stanford student?

This probability turns out to be almost exactly 0.73 (a reasonable approximation to it is 1 – 200*248/160000, or 0.69). Thus, assuming only that all students are equally likely to get perfect scores, it is quite probable (a 73-percent probability) that no Stanford student will have a perfect score. It’s not that they are any different from the other students; there just aren’t enough of them.

Which brings us back to Thrun’s reason for teaching outside the university. From his talk at the DLD 2012 conference (at about the 20-minute mark in the video):

Having done this, I can’t teach at Stanford again…It’s impossible…there’s a red pill and a blue pill and you can take the blue pill and go back your classroom and lecture your 20 students. But I’ve taken the red pill and seen Wonderland.

I wish Thrun every success making knowledge and understanding so widely available.

It’s about time someone brought modern mass production techniques to the inefficient world of academia! If people like Sebastian Thrun are able to create knowledge in some sort of factory or “mill” it will go a long way to solving our problems with higher education affordability.

Thank you for sharing this, absolutely fascinating! I just signed up for one of the classes. In my mind, this is a wonderful project and I hope it will continue for a long time.

The class was far too easy. Around 25% of class got above a 97%. Also, the “perfect” 248 got a highly contentious question correct on the final, that was later proven by the community to be erroneous. 100% scores in this class should not be taken too seriously.

I don’t know whether that statistic (25% getting over 97%) means the course is too easy. It might. It might also mean that lots of students learned a lot. One of my happiest moments from teaching last semester was a student who said upon handing in the final exam, “Thanks for the class. I had a strong physics background before the class, and thought it was easy. But when I took the final I realized that I could not have done most of the questions before the course, and I can do them now.”

I don’t doubt that a lot of people learned a lot, but grades certainly become meaningless as a signal of understanding when the score distribution looks the way it did.

That’s amazing. If I was taking this class and not paying anything for it or getting any credit for it, I would get through half of it then quit.

Thanks for sharing. This is how the price inflation in the higher education will be solved. Now the high priced institutions have to think hard about bringing the value to the money they ask. I doubt they can compete with ZERO. So they have to copy what Apple is doing – otherwise, they would be swept by this commoditization of the higher learning.

What would be the high value college be providing in 2060 – (Access to network, Personal coaching – what else?)

While I am also intrigued by this concept, (and kind of delighted) I would also want to see further study into the results of this kind of academic course. For instance, give a relatively even distribution of onsite to online students, how to the two types of student’s retention of the material compare over intervals, say 1 month, 6 months, 1 year.

One factor that may determine the likelihood of a perfect score in an undergraduate course is the experience level in the particular field of the student. Because almost all of the undergraduate students who are actually attending Stanford will have a minimum of work experience this leaves the field of non-Stanford enrollees to all types of age ranges where many of the people taking the test may actually be professionals in the fields of Machine Learning or AI. So, the median age of a Stanford student may be 20, where the non-Stanford may be in the late 20′s or even 30′s. Might be an interesting path to look down.

Thanks for sharing. I had signed up for this class a couple of months ago and I received a mail saying that the course is postponed. But when I read this post, I visited the site and I was able to login

They have made great strides. Back in the early 90s I took an “online” course through Standford with CDs and tutors available from time to time.