Peter Norvig uses this 14 November 2012 talk at Stanford University (~34 minute talk; ~22 minutes of discussion - questions just about audible) to reflect candidly on what he learned from making and running the mass online AI course with Sebastian Thrun last year.
The screen-shot below has Norvig's concluding slide, which he uses to support the idea that in the future online learning will i) feel to learners like 1:1 instruction, ii) be organised with cohorts of 100,000, and iii) use analysis of the "big data" flowing from masses of peer:peer interactions to shape formative feedback to individual learners and/or determine how a learner is "routed" through their studies.
For more on how this might work, see Norvig's responses - optimistically to the question that is asked at 39:10, and more cautiously to the question asked at 51:40. See also his answer to the question asked at 53:10 for an interesting insight into how Google itself has been testing the impact of its own online search course (which, like the AI course, attracted over 150,000 learners) on the actual search behaviours of users.
As an aside, it is worth considering Norvig's comments on Carnegie Mellon University's efforts to create a mathematics tutoring system alongside observations made by Dylan Wiliam in Scaling up: Achieving a breakthrough in adult learning with technology a report I wrote earlier this year with Adrian Perry, Clive Shepherd and Dick Moore. Here is an excerpt:
A particular barrier to successfully creating software that helps learners develop their conceptual understanding is the great difficulty in building a solid proficiency model or map of a knowledge domain. For example, a well-funded team of expert researchers at Carnegie Mellon University developed an effective tutoring system (now called the Cognitive Tutor56) for a relatively small proportion of the US equivalent of the Year 10 algebra curriculum. Dylan Wiliam, Emeritus Professor of Educational Assessment at the Institute of Education, University of London, explains: “The Cognitive Tutor has been very well researched and it’s very effective. It’s probably better than 90% of teachers that are teaching this part of the curriculum. But one of the reasons it is so effective is that its focus is on such a very constrained domain. And it still took the Carnegie Mellon team 20 years to work out what are the knowledge structures that are involved in this domain.” In short, whilst the computer science behind the tutoring system is robust and getting even more so, the proficiency models of learners’ cognition are neither well developed nor easy to create.