Yahoo Finance

Luis von Ahn on open models and Duolingo's motivation moat

Luis von Ahn· Co-founder and CEO at Duolingo
·~11 min·English·Yahoo Finance
Open SourceInferenceAI CompanyBusiness Strategy
TL;DR

Duolingo swapped frontier models for open weights across its product — cutting a video call from about 30 cents to under a penny — because its real moat was never the AI, but keeping people motivated to come back.

01The Cost Cut

From 30 Cents to a Penny

Duolingo moved its AI video-call feature to open-weight models and cut the cost of each session from about 30 cents to under a penny, and says it does not think it lost any quality on this simpler task.

And the main way we did that is by moving most of this to open weight models.

— Luis von Ahn, Yahoo Finance
Key Insight
The striking part is not the saving, it is the admission that for a bounded task — a simple spoken conversation — a frontier model was overkill. Duolingo had been paying frontier prices for capability that use case never needed.

02The Full Migration

Soon, Almost All of It Runs Open

The switch is not limited to the flashy video call. Duolingo is moving the AI behind its stories, cartoons, and generated lessons to open weights too, until open models carry the vast majority of its AI usage.

The vast majority of our usage of AI will be in openweight models pretty soon.

— Luis von Ahn, Yahoo Finance
Key Insight
Once one high-volume workload proves open models are good enough, the default flips. The question stops being can we move this to open and becomes why is this still on a frontier model.

03The Decision Rule

Frontier for Coding, Open for the Rest

Von Ahn's advice to other executives is not go open, it is test per use case: frontier models still win clearly for hard work like coding, while open models are excellent for high-volume, simpler tasks.

I would definitely encourage people to test this.

— Luis von Ahn, Yahoo Finance
Key Insight
The real lesson is that which model is best is the wrong question. Capability and cost only matter relative to a specific workload, so the winning strategy is a portfolio of models, not a single vendor.

04The Real Moat

A Motivation Machine, Not a Lesson App

Asked whether chatbots or translation earbuds make Duolingo obsolete, von Ahn is unworried: anyone can deliver lessons, but the hard part, getting people to come back every day, is what Duolingo actually sells.

you don't need AI to learn a language. The reality is that you could learn a language before Duolingo.

— Luis von Ahn, Yahoo Finance
Key Insight
This reframes the whole open-model decision. Duolingo can treat AI as a swappable commodity precisely because its defensibility lives in behavior and habit, not in the model, so making the model cheaper only widens the margin on the part competitors cannot copy.

05Why Demand Grows

AI Can Translate; People Still Want to Learn

Von Ahn notes that computers have crushed humans at chess since 1997 and the game is more popular than ever: hobbyist learners do not care that AI can translate, and people learning English for work need real fluency that earbuds cannot fake.

Computers have been better at chess since 1997 and chess is bigger than ever.

— Luis von Ahn, Yahoo Finance
Key Insight
The chess analogy quietly separates can a machine do it from do people still want to. For a hobby, superhuman AI is irrelevant; for a job, a laggy earbud is not fluency, so AI competence does not automatically erase human demand.

06Growth Strategy

Audience Now, Dollars Next

Duolingo deliberately grew daily active users faster, about 29 percent, than paying users, about 14 percent, this year, planning to close that gap next year with monetization levers like the games industry's hardest trick: the first paid dollar.

it is well known that the hardest thing to do is to get people to spend the first dollar.

— Luis von Ahn, Yahoo Finance
Key Insight
Sequencing audience ahead of revenue is a bet that a bigger engaged base is worth more than near-term conversion, and borrowing the first-dollar framing from free-to-play games signals Duolingo sees subscriptions as a psychology problem, not a pricing one.

07The Human Element

The Owl That Took On a Life of Its Own

Duolingo added a mascot to help motivate learners and never expected the owl to become a global internet meme; its reminders lean deliberately passive-aggressive, a tone von Ahn says works the way a nagging parent does.

we have found that passive aggressive works.

— Luis von Ahn, Yahoo Finance
Key Insight
The telling detail is that Duolingo did not manufacture the meme; the internet did. A brand with enough personality to mock is a brand people pass around for free, and a tone that annoys you a little is one you remember, which is why the deliberately passive-aggressive owl stuck rather than faded.