Q&A with Gabriel Wyner of Fluent Forever

We provide a *noticeably different* path through language learning. Most users have failed at language learning using standard methods in high school or college, which makes them unlikely to try again if the learning methodology resembles what they’ve tried before. No step of the Fluent Forever method resembles what users have tried before, which provides a feeling of novelty for our users. Specifically:

1. Ear Training: We begin by practicing with pairs of similar sounding words (“toma” vs “doma”, “más” vs “moss”) to build new neural pathways. (Note: If you skip this step, it puts the student at a serious, ongoing disadvantage when it comes to remembering words with unfamiliar sounds.)

2. Simple Word-Picture Associations: We don’t teach translations, but rather ask users to associate words with images that they choose themselves. (Note: As it turns out, memorizing translations is a tremendously inefficient/ineffective use of time. The data students actually need to use a word effectively are associations built between new words and images, or new words and previously known words. Translation is, at best, an extra fact a student can learn about a word...but it should never be the end goal of learning.)

3. Content Customization: Users are guided towards specific words to learn, but they’re encouraged to skip words they know or don’t care to learn, and to search for words they specifically want to learn. They choose their favorite sentences to learn from, and choose what content to learn from any sentence. They choose their own pictures to remind them of words or sentences (and can even take their own pictures). Every moment of customization improves learning outcomes and also makes our app stickier, as it increases user investment. (Note: If your content is the same for all of your users, they will lose motivation [“why am I learning the word for Elephant?!”] and it'll reduce your effectiveness (boredom strongly impedes learning] )

3. Spaced Repetition Algorithm: Our app determines when best to test a user on a given word or image or pronunciation based on the user’s progress. This is the best technology we have for rapid memorization.

4. Intelligent Word Order: We group our words into themes (red – apple – food) rather than categories (red – yellow – green). As it turns out, this grouping makes words more than 2x easier to learn/remember.

• Results: Users report early signs of thinking in a new language within two weeks of using the app (i.e., walking through their house and unconsciously naming objects in French). Ear training and pronunciation gains occur within the first week. These early produce committed users that churn far less than competing apps. And then over time (as short as 3 months for power users), we produce users who can speak in a new language, who become tremendously effective sales people via word of mouth.

Duolingo, in some sense, is one of our greatest assets. They've found 300M users, showed them what a gamified language learning app looks like, and then every year, 90% of those customers churn. We have access to 50M of those customers via Facebook ads, and they’re our cheapest converting audience because in many ways, they’re pre-vetted. Duolingo drops our CAC dramatically and provides us with a steady stream of tens of millions of new potential customers every year.

Rosetta Stone has been struggling to stay profitable for years. They don’t have enough market share in terms of user count to be particularly threatening to us, and they’re fighting a reputation as being both antiquated and expensive.

As for Google Translate, we think of them as a useful tool for market segmentation. There will always be a portion of the market looking for basic functional needs (ordering coffee, getting train tickets). We don’t want to spend money on finding these users, because they will not use our app for more than a month, and the better Google Translate gets, the fewer of these users will go looking for language learning apps. Nonetheless, the Online Language Learning market is going to double in size in the next seven years, to $20.2B. This is happening because there’s a substantial (and growing) portion of the market that is looking for *more* than basic functional needs. They are looking for connections with other people, new job opportunities, etc. And these needs will never be met by machine translation tools.

We can save dying languages with just a couple of remaining speakers. To explain, I’ll need to talk a bit about where we’re headed.

Within the next year, we intend to release our Tutoring Platform, which will let students work with a tutor (via text or video chat) to create their own custom sentences, which can then be memorized with our app. Students who use this platform will learn twice as fast as our average users, because customized content is generally twice as memorable as generic content. Over the course of reaching fluency in a language, a student will produce 1000-1500 new sentences with their tutor, which will then get shared with the rest of our user-base. For languages that we officially support (Spanish, German, etc.), these sentences will get added to our 1875-sentence database, and give us ever increasing coverage and content within those languages.

But this works for languages that we don’t support just as well. Five students will produce enough content (~5000-7500 sentences) to allow any future students to reach advanced levels in that language. And this applies to popular languages with low marketshare and poor resources (Hindi, Greek, Czech) just as well as it applies to dying languages in need of restoration (Indigenous languages). All you need are a couple of surviving native speakers and a handful of students willing to learn from them, and those interactions will produce enough content for any number of students to learn that language to advanced levels without direct interaction with native speakers.