In this issue:

• Duolingo's Efficient Frontier

• Amazon's Upside

• Copilot

• The Microsoft Legal Strategy

• Take Rates

• The Infrastructure Layer

Duolingo's Efficient Frontier

Duolingo, which just filed its S-1, can reasonably pitch itself as an education company. The main product teaches languages through bite-sized translation and transcription exercises: learn a few key words through pictures, and then start translating one sentence into another with a tap-and-drag interface:

But Duolingo can just as easily be modeled as a case study in how app developers drive continuous interaction and monetization. The company started with a language-learning project, and continuously A/B tested it until it ended up with a game. It's Wittgenstein-as-a-Service. The core interaction is basically a themed puzzle game, and like most such products its success hinges less on the quality of the game itself and more on the thousands of incremental hacks that keep people playing and spending. ("Thousands" is not an exaggeration: Duolingo discloses that it performs more than 500 A/B tests each quarter.) Among the tricks:

• Duolingo breaks language-learning into discrete exercises (which take a few seconds), lessons (a minute or two), levels of multiple lessons, and units of multiple levels. So there's always a small enough chunk to make opening the app rewarding, and a goal to reach.

• Duolingo interrupts usage to congratulate players on winning streaks, like getting five questions right in a row.

• The app has multiple in-game currencies, like "gems" (for in-game purchases), "experience points" to track progress, hearts (which let free users get more questions wrong before they have to quit). Like other free-to-play games, they let users purchase these in-game currencies with real money, and let users convert between the currencies. Duolingo mostly monetizes by selling subscriptions, which get rid of the "hearts" mechanic and have a few other features. That represents 72.6% of revenue, while in-app purchases of game currency are just 1.2%.

• Duolingo earns the usual virtual currency issuer seigniorage by giving in-game currency to users who recruit their friends to play the game—why spend money you can't print, like dollars, when users are happy to accept gems and hearts?

• The app tracks users’ "streaks," or consecutive days of play. This seems to be a very big driver of continued usage; there are online anecdotes of users who have 1,000 or even 2,000-plus day streaks, and it gives the app a regular excuse to send push notifications.1 Per the S-1 "Over 50% of our daily active users have a Streak longer than seven days. And approximately one million learners have an active Streak longer than 365 days, meaning they have used Duolingo daily for at least one year." Edit: The app as a whole had 9.5 million daily active users in Q1, so over 10% of their daily sessions come from users who use the app literally every day. A previous version had the number of subscribers instead of DAUs, and thus overstated the importance of streaks. Apologies for the error. Colgate, Gillette, Starbucks, and Procter & Gamble are all businesses that are built on daily habits (and outside of the business world, the religions that mandate some kind of daily practice seem to show better retention than the ones with a more relaxed once-a-week-if-that schedule).

• Duolingo is constantly offering small tweaks to the reward function: bonus points, a way to freeze streaks so they continue even if the user misses a day, and extra-challenging problems.

• They use leaderboards to give users people to compete with. An app with enough users can always create a subset that's pretty close in skill, which is motivating for the users in question and great for the app—when everyone is about equally skilled, the big differentiator is effort, and for Duolingo "effort" either translates into subscription revenue or ad views.

Interestingly, Duolingo plans to be one of the first companies to distribute some shares through Robinhood when it goes public. I have been in favor of Robinhood users getting more access to IPOs, not because it will make them rich, but because it protects them from themselves. Valuation-insensitive buyers will tend to push stocks up regardless of how richly they're priced, which has led to some large IPO pops in the last year. Duolingo is perhaps the ideal case for this—their user demographic of people who have a real-world need and have chosen to satisfy it in the most gamified possible way no doubt overlaps with Team Diamond Hands.

The tradeoff for Duolingo is that while learning can be fun, there are lots of ways to make an activity more fun by sacrificing the learning part. Since Duolingo is teaching language at scale, for example, it needs to rely heavily on multiple-choice questions: synonyms can make a given answer subjective, so the right approach to possible answers is to list very distinct words, but that means that a partial translation provides some clues to fill in the missing word; multiple-choice answers to sentence-long translation questions asymptotically approach Minesweeper, not translation.2 Meanwhile, a lot of the exercises involve a fairly literal translation, not one that alters the form too much: if there's one right answer, its sentence structure will be closer to what the original language uses than to how a native speaker in the other language might phrase things.

This tradeoff also applies to the question of how Duolingo will keep retention high over time. At some point, the most active users have learned everything the app has to teach them. There are two broad approaches to content-limited retention: add more content for a given language, or convince users to try a new language. Users naturally churn out as they progress: some of them reach the limit of their retention, at least if they're using Duolingo alone, and others are using the app just to get by; the amount of French you need for a week in Paris is a lot less than the amount you'd need if the State Department is sending you there for a year. But if Duolingo can hook users on the mechanics of their language acquisition approach, they can keep people playing. Some users report learning multiple languages in parallel. Since each language has some natural demand already, the marginal cost of switching a user from Spanish to French or from French to Klingon is lower than the cost of giving them more French.

For many apps, those long-term users are the most valuable, but the most important task is to retain first-time users. For most app categories, 30-day retention is under 10%, and even second-day retention is in the 20-30% range excluding categories that skew to one-time-ever use. Duolingo has gotten its one-day retention to 40%, up from 12% in 2012. This is convenient for anyone doing due diligence on the app; since that retention has a multiplicative effect on the rest of the business, it matters a lot, and basically forces the company to put all its retention-maximizing tricks on display upfront. Which means that the app is probably a bit less gamey than it looks to the median user; it has to be gamey in order to get those users.

Their stated mission, of course, is not to maximize DAUs and average revenue per user—it's to reduce inequality by making language education accessible to everyone. They explicitly cite economic inequality as a driver of different educational outcomes in the S-1. It's interesting to consider Duolingo's long-term effects on income inequality. English is by far the world's most popular second language, and while there are many ways to learn a language, Duolingo's approach certainly seems efficient. But making it easier to learn English means that the market for English-language content is bigger, while the market for other kinds is (relatively) smaller; the rewards for writing a popular English-language song or novel will rise commensurately, and the rewards in other languages will shrink in proportion. So, from a consumer's standpoint Duolingo produces more equality: more people have access to the biggest pool of knowledge, not to mention career opportunities and travel. But from a producer's standpoint, Duolingo makes the world a more competitive and thus unequal place; if more people are competing in the Anglosphere rather than in domains with fewer speakers, it's a lot harder to be #1, but a lot more lucrative, too.3 Per their numbers, 54% of active users are learning English, and 37% are English speakers learning another language; 9% neither use nor are learning English.

Making English a more global language is a business opportunity for Duolingo, and one that indicates that the growth hacks they use aren't overwhelming the educational service they provide. The company's fastest-growing revenue source in percentage terms (and second-fastest in dollar terms) is from English fluency certification tests, which are accepted by 3,000 schools as proof of English proficiency. Duolingo cites some very interesting data indicating that their courses actually outperform classroom instruction:

In 2020, we conducted a formal study to evaluate Duolingo’s effectiveness versus traditional university language courses. We measured Duolingo learners’ results on French and Spanish listening and reading proficiency tests from the American Council on the Teaching of Foreign Languages (ACTFL) against those of US university students who had completed four semesters of French or Spanish. To ensure a fair comparison, we included only US-based Duolingo learners who: (1) had no prior knowledge of Spanish or French before starting Duolingo, (2) had only completed the first five Units of their chosen Duolingo course, and (3) were not using any other materials to learn besides Duolingo. We found that the Duolingo learners earned proficiency scores comparable to those of the US university students at the end of their fourth semester of French or Spanish. Moreover, the Duolingo learners were able to attain this level of proficiency in about half the time as the university students.

This is positive for the company's economics, and for their social impact: they start with an addictive game, that has an educational component, but the most engaged users end up developing real-world skills—and Duolingo can profit from giving them a credential alongside this. The company still faces the fun-vs-educational tradeoff, but it can push out the efficient frontier enough to get more of both. The Duolingo story, at a user level and as a company, ends up being the story of American higher education in reverse: it starts out being mostly trivial entertainment that's designed to soak up time and money, but turns into a way to learn useful skills and get a broader education.

Elsewhere

Amazon's Upside

The WSJ has a piece on how Amazon increasingly asks for warrants from companies it does business with ($), a pattern I've noted before. The WSJ piece has much more on the breadth and scope of these deals: not just air freight, but wholesale groceries and other businesses. Amazon’s warrant holdings totaled $2.8bn last quarter. If it's not because Amazon is especially bullish on air cargo, is there another reason?

One possibility is that it mimics the economics of the rest of Amazon's retail operation: merchants who sell on Amazon end up spending a lot of their revenue on Amazon: a cut of transactions, delivery and storage costs, and increasingly advertising. When sellers operate within the Amazon ecosystem, Amazon can infer their profitability by giving them lots of ways to competitively spend money; when Amazon does business with a company outside its system, they don't know how much upside they're leaving on the table, and warrants are a way to capture that.

(Disclosure: Long Amazon.)

Copilot

Github has launched copilot, a coding tool that guesses what a function should do based on its name and inputs. There is a style of writing code that consists of 1) thinking up useful functions, naming them, and leaving them blank, and then 2) filling in the details one at a time. Usually the first part of this process requires most of the creativity and forethought, while the rest requires a lot more alt-tabbing between a code editor and Stack Overflow. So, for one programming approach, this boils the process down to the high-value essentials—figuring out what needs to be done—while outsourcing how-it's-done to an AI trained on the many previous projects that had to do some subset of the same thing.

This is a nice combination of two very long-term trends in programming, one technological and one social. On the technology side, the general trend in programming has been to outsource more and more of the details to either a) the language, b) libraries, or c) APIs that essentially give access to somebody else's libraries. Implementation details are important to get right—once—but higher-level constructs let coders focus more on what they're trying to accomplish than on which bits get flipped how. The social trend is similar: documentation quality has gone up, and StackOverflow has cogent answers to a huge variety of questions. Some of the answers are terse, if they arise from a simple mistake, but some of them are attempts to model the mistake and help users fix the underlying concept. Copilot is a blend of both of these. At one level it's just saving the user a click.4 But it's also a language-agnostic way to implement the general trend towards high-level languages and widely available code snippets.

The Microsoft Legal Strategy

Microsoft is the biggest of the big tech companies not to be directly targeted by Congress or the FTC. After its antitrust issues in the 90s, the company has gotten more careful, and the areas it dominates don't seem as pressing to regulators. The company seems to expect this tide to turn soon, and is adjusting defensively and offensively. On the defense side, Microsoft has been hiring more lawyers. On the offensive side, Microsoft is asking for more restrictions on executive branch data requests and letting a legal and PR ceasefire with Google lapse after six years ($, FT). For most big tech companies, increased regulation is a net downside, even if there are specific ways it can help them—GDPR seems to have relatively benefited the largest ad companies, for example, but size-based restrictions will naturally hurt them. Microsoft might be positioning itself to capture upside; it has less at risk than other big companies, but could still benefit if laws change.

(Disclosure: I am long Microsoft.)

Take Rates

It's always tricky to compare take rates between marketplace businesses, because the true take rate is a share of gross margin, but the available data is based on share of revenue. If a company sells purely digital goods, it can support a high take-rate. If it's selling tangible products that are expensive to store and ship, there's a limit. But that limit is probably not 3%, so Shopify is working on capturing more value with its audience network, which seems to be a way for merchants to share data and use it to identify likely buyers of their products. This is another way to replicate Amazon's on-site advantage—if there's a correlation between propensity to buy can openers, likelihood of watching a particular movie to the end, and taste in novels, Amazon is one of the only economic actors that can know this and take advantage of it. Shopify doesn't have quite that breadth of user interactions, but it can create a lot of value for merchants by using their collective data to improve individual sellers' results. And, since the resulting product is a digital good rather than a physical one, there's room for a heftier take rate.

The Infrastructure Layer

Software is a complement to material abundance, because the richer a country is the more value there is in organizing and accessing goods, rather than producing them. But it can also be a good solution to some kinds of deprivation: a Ghanaian startup, Jetstream, has raised $3m to build software that helps African importers and exporters track goods throughout the supply chain. Inefficiency and low labor costs create a lot of slack in a system; US logistics is often capacity-constrained (especially now!), but when the problem is widespread inefficiency and slow-turnaround paperwork, a software layer that can route around this can increase the shipping system's output even if its total capacity doesn't grow up.