Duolingo cofounder explains crowdsourcing to the MIT crowd

When Luis von Ahn was 12, he dreamed of becoming a “gazillionaire” by building a chain of free gyms where the kinetic energy people expended on exercise equipment would be captured and sold to power companies.

“I thought this was a genius idea. We’re going to have a set of free gyms, where we don’t have to charge people because as they are exercising they are generating value for us. And because the gyms are free, we are going to take over the whole world of gyms,” von Ahn, an associate professor in Carnegie Mellon University’s computer science department, recounted at the recent MIT Platform Strategy Summit.

As he later discovered, not only was his big idea unoriginal — many people had entertained the same notion — there remain two big reasons it doesn’t work.

“The first is that people are just not very good at generating power, so you can’t make much money off it,” he said. “But second, and most importantly, gyms make their money from people who pay and don’t go.”

A native of Guatemala, von Ahn, 35, is famous for being one of the pioneers of crowdsourcing or, to use his term: human computation. He defines human computation as a system “that combines humans and computers to solve large-scale problems that neither can solve alone.” As a graduate student at Carnegie Mellon he worked on CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart). The security tool, in wide use today on the Web, does double duty as a crowdsourcing tool, using our typed-in transcriptions to turn images of hard-to-read text in old books and magazines into an accurate digital record.

For this work and other projects, von Ahn has become a millionaire, if not the gazillionaire of his pubescent dreams. He was awarded a $500,000 MacArthur Fellowship “genius” award in 2006. Three years later he sold his company, reCAPTCHA, to Google reportedly for more than $25 million.

Duolingo platform

Von Ahn currently commands the spotlight as co-founder and CEO of Duolingo, a learning/translation/crowdsourcing platform that combines the lofty aim of teaching people a foreign language for free via an interactive app  — with crowdsourced labor. As learners acquire language skills, they are asked to translate portions of foreign news articles into English as a way to test their skills; Duolingo then sells these translations to the likes of CNN and Buzzfeed. The site has more than 100 million users, von Ahn said.

“More people are learning language on Duolingo in the United States than in the whole U.S. School system,” he said, a fact that would probably not surprise what’s left of the foreign language school teachers in the nation’s public schools.

Indeed, according to von Ahn, of the 1.2 billion people in the world trying to learn a foreign language (his figure), most are trying to learn a language — English — as a means of getting better work, but can’t afford to pay for classes.

“This is the irony. Most people that are trying to learn a language are doing so to get out of poverty, but it seems you need $1000 to get out of poverty,” he said.

Duolingo aims to level the playing field, von Ahn said, and is part of a larger goal to make education more widely available, regardless of one’s ability to pay.

How to make something like this pay is something he’s been thinking about since he was 12. “The question was, ‘How do we apply the idea of the gym to language learning?'” he said. And the answer is to get users to give you something of value in return for learning. The platform matches up what users are learning to the material CNN and other clients want translated.

Game theory and A/B testing

Three years in, the platform keeps growing and keeps adding languages (e.g. Irish; the number of people learning it on Duolingo now exceeds native speakers, von Ahn said). Schools are starting to adopt Duolingo. Like other platform companies, the company does not rely on traditional marketing to promote itself but has gone to great lengths to figure out how to hook people and keep them coming back.

Language learning is divided into separate skills — the ability to talk about food or animals; the ability to do plurals — and then game theory is applied to encourage people to keep acquiring skills. Every skill level has a “strength bar,” a sophisticated mechanism that goes up as language proficiency increases and decreases when mistakes are made or the student slacks off, disappearing altogether if one ignores ones lessons for too long.

The company has used a lot of mechanisms to lure back Duolingo dropouts, von Ahn said. But the one that has worked the best is sending an image of Duo, the company mascot, crying — crowd tested, of course. “We A/B tested how many tears and how large the puddles of tears should be,” von Ahn said.

Crowdsourcing language acquisition

To figure out how to build a system for learning language, von Ahn and his co-founder read a lot of books (including French for Dummies) on how best to teach a language, but realized the field is rife with conflicting theories. Rather than settle on one method, von Ahn said he and his co-founder “took what we could” from a variety of approaches. As the user community grew, they tested empirically what worked and what didn’t, using A/B testing to improve the heuristics – for example, when is the optimal time to teach plurals.

Whether this back-to-the-future apprentice model has staying power remains to be seen. In the meantime, it is using its current crowd of language learners to not only sell translations but also to perfect itself.

“DuoLingo keeps getting better and better over time, as we are experimenting,” von Ahn said.

Let me know what you think of this post — email me or find me on Twitter @ltucci