Duolingo: The app teaching humans new languages while teaching machines how to learn
For all of our other failings, the human brain still remains the most flexible and powerful piece of computing hardware on the planet. And there are a huge number of them out there.
Unfortunately, much of that computing power is wasted on trivial pursuits like addictive-but-pointless smartphone games, rather than tackling some of the more profound problems that face humanity.
The idea of subtly harnessing vast amounts of otherwise wasted human processing power and using it to a worthwhile end has been a hallmark of Luis von Ahn's work. His projects involve engaging vast numbers of people in games that seem simple, but actually have useful by-products.
Read this
For example, his 'ESP Game' was, on one level, a mildly diverting game where players had to agree on the same words to describe an image. However, it was also a way of harnessing the human brain's innate talent for image recognition, which was then used to help categorise some of the mass of images found online.
Similarly on one level ReCaptcha, probably the best known of von Ahn's projects, helped users of online services to prove they were human by keying in a word displayed on the screen - something a spam bot might not find it that hard to do. However, users also had to type in a second word, often a scruffily-printed one, selected from an old book that a computer charged with digitising the text would struggle to process.
By using ReCaptcha, humans get to prove they aren't robots, and help to digitise books along the way (Google is now making ReCaptchas out of hard-to-read text such as house numbers in Google Street View).
Language teaching app Duolingo, von Ahn's latest project launched with Severin Hacker, takes this idea of elegant reuse even further. It's a language app that has designs on remaking teaching.
Launched two and a half years ago, the free app now has 70 million users and is the most downloaded education app in both the Google Play and Apple App Store; the vast majority of its use is on a smartphone.
One reason for its success is that it aims to feel more like a game than a language class.
Another is that Duolingo's curriculum is tuned by relentless A/B testing to make it as efficient as possible: the company even conducted A/B tests to find the optimum number of virtual tears its owl mascot ought to cry when users make a mistake.
"Because we have so many users, we can tune it based on data - that's something that offline schools haven't been able to do at the scale that we do it. If we want to know whether we should teach you plurals before adverbs or the other way around, we do an A/B test with our users so for the next 50,000 people that sign up - half of them we teach plurals before adverbs, the other half we teach adverbs before plurals and we measure which ones learn better," von Ahn said.
This not only helps the students by getting them to learn as quickly and efficiently as possible, it also helps the company to build a system that understands learning.
"We are reusing mental energy for something else... the way I see it is we are observing people learn and, in this case, we are trying to figure out how to teach better. Once you are able to observe tens of millions of people learning, then you can start to figure out ways to teach better than ever before," he said.
Duolingo's scale puts it in the unique position, he argues, of being able to observe how people learn in minute detail: "We really can tell this person is taking half a second longer to answer any question that starts with that word - that's the sort of thing that without technology you would never be able to work it out."
Many education companies base their business model on charging a small number of customers a lot of money, said von Ahn, whereas Duolingo aims to reach a large number - 300 million users. "So we don't need to charge $100 per person. If we are somehow able to make a dollar off of each one then we're good so we are in a different segment where we are trying to reach as many people as possible."
Another example of the dual use that runs through von Ahn's work comes in one of Duolingo's revenue streams: translating articles. Students translate articles into their native language as part of their studies - many students translate the same article which creates an accurate translation within four hours for customers such as CNN. Duolingo also makes money by charging for tests through which users can prove their language proficiency.
Unlike many software companies, Duolingo is motivated by a desire to improve education rather than just make money, von Ahn says.
"Most people learning a language are poor people trying to learn English to get out of poverty. At the same time most of the ways to learn language, especially through software, have been very expensive. So we decided to do something to teach languages entirely for free."
That's not to say that money is irrelevant.
"It's a bit weird that we are not entirely motivated by money, but we do have $40m of investment and we need to make sure that our investors are happy, and that it's also sustainable. Because we have so many users our servers are very expensive and we have to pay salaries. If you look at companies like the Facebooks of the world, they make their money off having so many users and from each one they only have to make $5, so that is closer to the model that we have."
To reach its 300 million user target, it will not only teach languages but other skills: it is already looking at teaching literacy for both for children and adults.
"Most of what we are building, the machine learning that we are building, can be applied to other things and literacy is probably the next thing we are going to do," he said.
Earlier this month, the company launched Duolingo for Schools, which gives teachers a dashboard to track their students' learning - for example, which lessons they have completed, and how many points they have scored. The company said that two weeks after going live, it has seen 20,000 teacher accounts registered, and already plans to add more detailed analytics for each student.
"Internally we already have all these analytics for every one of our users. Whether they're good at conjugating verbs but bad at adjectives, we know all of these things and we're going to start exposing that to teachers. The idea is they will have much better analytics on each student to allow them to do much better teaching," he said.
In the next few months it's also going to allow teachers to modify the Duolingo curriculum to fit their own. "Our curriculum is well tuned, which is why many teachers are starting to adopt it. But we know that many won't be able to because what you learn in school is decided by politicians. So we are going to start allowing them to change our base curriculum to theirs so their students learn in the order that the teachers need to teach," said von Ahn.
But in an age of simultaneous translation technology, why would anyone bother learning a language anymore?
There are still a lot of reasons, said von Ahn. "When you are talking with [a computer translator], it's such a weird interaction that you don't ever really feel you are talking to [someone]. Until the day we have some implant in our brain that just lets us speak another language, I think people will still continue wanting to learn a language."
Read more
- Six Clicks: Androids Apps on Chromebooks
- Smartling lands $24 million for cloud-based translation software
- Twitter brings Tweet translation to iOS, Android and Windows 8 phones
- Lost in translation? Smartling would like a word with you