On a summer day four years ago, a Stanford University computer-science professor named Andrew Ng held an unusual air show on a field near the campus. His fleet of small helicopter drones flew under computer control, piloted by artificial-intelligence software that could teach itself to fly after watching a human operator. By the end of the day, the copters were hot-dogging—flipping, rolling, even hovering upside down.
It was a milestone for the field of “machine learning,” the same area of artificial intelligence that lets Amazon recommend books based on a shopper’s previous habits and helps Google tailor search results to a user’s behavior. Mr. Ng and his team of graduate students showed that artificial-intelligence software could control one of the hardest-to-maneuver vehicles and keep it stable while flying at 45 miles an hour. That same year, Technology Review, published by the Massachusetts Institute of Technology, included Mr. Ng among the top 35 innovators in the world under the age of 35.
Today Mr. Ng is an innovator in an entirely different setting: online education. He is a founder of the start-up Coursera, which works with 33 colleges to help them deliver free online courses. After less than a year of operation, the company already claims more students—1.3 million—than just about any educational institution on the planet. Mr. Ng likes to say that Coursera arrived at an “inflection point” for the idea of massive open online courses, or MOOC’s, which are designed so a single professor can teach tens of thousands of students at a time.
What do self-piloting helicopters have to do with the growing movement to transform education online? A day spent with Mr. Ng here at Coursera’s offices, with the aim of getting a sense of the company’s culture and the ideas that make up its DNA, helped answer that question.
It turns out that the links between artificial-intelligence researchers and MOOC’s run deep. Another AI visionary (and Stanford faculty member), Sebastian Thrun, started the online-course provider Udacity soon after developing a self-driving car for Google. And Anant Agarwal, head of the nonprofit edX project run jointly by MIT and Harvard University, has long directed MIT’s Computer Science and Artificial Intelligence Laboratory. Coursera’s other co-founder, Daphne Koller, is also a Stanford computer-science professor and a renowned artificial-intelligence researcher. (She and Mr. Ng became friends when collaborating on research and found they shared an interest in reforming education through technology.)
Early efforts at online education focused on replicating traditional classrooms online and, as a result, were just as labor-intensive as the old way. Professors or teaching assistants still needed to grade each assignment, and class sizes were limited to small groups, just as on physical campuses.
To a scholar accustomed to teaching robots to fly, though, the problem of online learning looked like an opportunity to install an autopilot.
“There’s a certain way of thinking that many AI researchers have—it’s the idea of automation,” Mr. Ng explains, his lanky frame folded onto a couch in a conference room. He speaks in a quiet voice colored by a British accent—he was born in Britain and grew up in Hong Kong and Singapore—and his understated manner makes you forget that his teaching videos have been viewed hundreds of thousands of times. He is sometimes recognized as a kind of celebrity on the streets near Stanford.
“A lot of AI successes have been about automating the routine things that do not call on the highest levels of human creativity,” he says, noting that spam filtering and recognizing faces in photographs can now be done deftly by software.
After teaching at Stanford for several years (he’s now on leave), Mr. Ng felt that grading was eating up too much of teaching time. Computers, he thought, could step in and grade complex assignments, not just multiple-choice exams.
“I actually enjoy working through problems with students,” Mr. Ng says. “What I don’t enjoy is grading 400 homeworks. And so our thinking was to automate some of the grading so it frees up more faculty time for the interactions.”
He put his ideas into practice about five years ago, when he started Stanford Engineering Everywhere, which offered MOOC’s before anyone had heard of them.
Last year he and Ms. Koller saw the growing interest and decided to start their own company to support MOOC’s, in part to bring courses from a range of institutions into one system. As Ms. Koller remembers, “If we wanted to make the best education available to students, we needed to enable multiple top institutions to provide education to the world, and this could only be done effectively as an independent entity.”
The biggest benefits from mixing AI and education might come later, when Coursera develops systems that can hunt for behavioral trends among all those online students. For instance, in Mr. Ng’s recent online course on machine learning, which attracted more than 100,000 students, some number-crunching detected a possible flaw in his recorded lecture.
“When 2,000 out of 100,000 students submitted the wrong answer, it was a very clear signal to me that I had done something wrong,” he says. Something in his explanation must have thrown students off-track. “So what we were able to do is create a custom error message, so that when the 2,001st student submitted the same wrong answer, they get a custom message that says, ‘Consider the order of these two steps.’ And then once students got that, we found that students were able to much more quickly get over the conceptual error.”
That doesn’t mean computers become the instructors, though. “It’s not robot teaching,” Mr. Ng insists. “It’s important that pedagogy is in the hands of expert professors, and it’s their course. What I’d love to do is understand the data so as to provide extra information to the instructor.”
That view hasn’t stopped some professors and college leaders from worrying about the implications of all this automation, of course. One reason Coursera is getting so much attention is that it’s a well-intentioned idea that threatens to reshape the higher-education landscape. It’s easy to imagine MOOC’s leading to future reductions in faculty, or even the end of smaller colleges, which might have trouble competing with free alternatives.
The first thing visitors see when they open the door to Coursera’s offices is a Ping-Pong table. There’s no reception desk, so visitors have to skirt past the table to get to the main hallway. One wall is dotted with dozens of pink and yellow Post-it notes, stuck there by employees. The notes are grouped under two headings: the “Promise Wall,” listing courses under construction, and the “Ideas Wall,” with scrawled phrases proposing new features. One note reads “Buy me a pizza,” which apparently describes a way for students to send gifts to other Coursera students. (As a condition of visiting the office, The Chronicle agreed not to publish the other ideas.)
There are few signs that an online classroom for some of the world’s most famous universities, like Stanford and Princeton, is being built here: Employees seem too busy working to properly move in. Mismatched Ikea bookshelves stand in one corner, half-filled with computer books; one unoccupied office provides parking for bicycles. A couple of beanbag chairs rest in one room—a requirement for Silicon Valley start-ups, it seems—and a free catered lunch is delivered to employees daily, but there are few frills.
That low-key approach is a good sign, believes Charles Severance, a clinical associate professor of information at the University of Michigan at Ann Arbor, who taught a course on Internet history and security on Coursera’s platform this summer. “I have been to enough start-ups that succeeded and failed and know what the good ones look like,” he says. “The ones that have dog-play breaks and espresso and fancy chairs in the cafeteria before they ship a 1.0 product are setting themselves up to fail.”
Everyone here shares office space with at least one colleague. That goes for Mr. Ng as well, who shares with Ms. Koller. The two serve as co-chief executives, and their desks are arranged so that they sit back to back. Employees say they often hear the two arguing—in a friendly but intense way—about where to put the energies of their team of 17 programmers.
“Often we debate things like product priorities,” says Mr. Ng. “Should we be focusing more on social features, or should we be focused more on grading of mathematical expressions—those kinds of decisions.”
The two do agree on the company’s prime directive: “Always do what’s best for the students.” It’s Coursera’s equivalent of Google’s mantra: “Don’t be evil.” For Coursera, it means choosing features that will help their nonpaying customers rather than features that might be better for the bottom line, says Mr. Ng. This philosophy is apparently explained to investors before they sink money into a company that so far has zero revenue.
“Our investors seem to be fine with that,” says Mr. Ng. The company has raised more than $18-million in venture capital and $3.7-million from the California Institute of Technology and the University of Pennsylvania, which chose to invest in the project. “Look at Craigslist. Most of Craigslist is free, and only a small part of Craigslist brings in revenue. But that sustains the rest of the business.”
So far Coursera’s main plan for bringing in money is to sell certificates to students who complete courses, and help match top students with companies looking for employees. “We need to keep it sustainable,” he says, “but other than that, I really think it’s not about making money. It’s about educating students.”
This is not the first enterprise Mr. Ng has been involved in—he worked with some former students to start Zunavision, a company that helped insert advertisements into online videos, and which he said “wasn’t successful.” That experience—and years of consulting for Google and other major technology companies—prepared him for Coursera, he says. Stanford allows professors to take up to two years off to pursue spinoff ventures.
When he talks about Coursera’s online platform, Mr. Ng sounds more like an engineer than a businessman. The scientific method frequently informs even mundane design choices. In a so-called “A/B test,” he had the course platform show some students a video lecture in which the professor used a black background when sketching out his ideas on a digital tablet, while another group was shown the same video with a white background. “Students seemed to prefer the white background,” says Mr. Ng. “In a bright environment, you can see fine detail better.”
It’s hard to imagine a typical classroom instructor sweating the color of the chalkboard. “Online ed is such a new medium that we almost have to do this,” Mr. Ng argues. “It’s almost like we’re starting from a clean slate, so we almost have to invent the stuff from scratch.”
On his laptop, Mr. Ng shows a visitor an early version of his online course “Machine Learning.” He criticizes flaws that he has since fixed in the latest version offered on Coursera. “I developed a lot of that sitting alone in my apartment,” he says.
In one lecture video, a cluttered living-room shelf is visible in the background. Amid the books and files is a remote-control helicopter.
