|
In this issue: - The Duolingo Disaster That Wasn't—Duolingo looks like exactly the kind of company that would have suffered disproportionately in the great growth drawdown of 2022. But unlike some of its high-flying peers, the company has more than recovered. As it turns out, their model is uniquely suited to some high-ROI ads, and they've found new ways to price the product to get the most revenue possible from different usage patterns.
- Chatbots—The FT launches a chatbot.
- Alpha Capture—The hedge fund industry looks more and more like every other industry.
- Coreweave—When rapidly-depreciating assets make borrowing to fund growth a better bet.
- Preannouncements—China is releasing some positive macro news early.
- AI's ROI—One constraint on using AI: figuring out what it's worth.
The Duolingo Disaster That Wasn't
Suppose you were coming up with a list of investment themes that had the most painful reversal after 2021. It would look something like this:
- Growth used to merit a much bigger premium; high-growth/high-multiple stocks saw lower valuations, so they trimmed their spending and saw lower growth.
- Companies that were a bet on The Great Indoors—whether that meant office furniture or streaming media—did great in the early stages of the pandemic, but turned out to be experiencing pulled-forward demand.
- Anyone whose business involves ingesting strings of human-readable text, and returning those strings, faces disintermediation risk from AI. Why ask a site like WebMD, Investopedia, or Stack Overflow about your specialized question when ChatGPT can give you a more precise answer and address follow-up questions.
- Perk-heavy companies have had to curtail those perks. This has been good for margins but bad for morale. A company with a generous vacation policy—two weeks off and a weeklong work-free company-sponsored retreat, for example—would have more trouble adjusting than companies with skimpier benefits.
- It's harder than it used to be to underwrite a bet on globalization, and the convergence between poor countries and rich ones.
With all that, it wouldn't be surprising if shares of Duolingo had lost most of their value post-2021. And for a while, they did; at one point shares were trading at one third of their peak valuation. But the stock has since recovered, and now trades 14% above its 2021 high. It's actually outperformed the S&P since it went public.
What happened? Or what didn't?
At a high level, Duolingo tapped the brakes in a few separate ways. Their marketing spending started to slow right when the tech economy slowed; in 2021, incremental marketing spending was 27% of revenue growth (i.e. they added $89m to the topline and reinvested $24m in additional marketing). In 2022, that ratio was 7%. Marketing was up some quarters, down others, and grew 13% overall, while revenue rose 47%. It took longer to hold back on R&D spending. Incremental R&D spending as a share of incremental revenue dropped from the 20s in 2021 to the mid-teens in the last year.
This doesn't mean that they've eliminated their marketing spending, of course, just that they're looking for the most cost-effective tool they can find. Including, but not limited to, a Super Bowl ad. Normally, Super Bowl ads are a negative sign that a company has exhausted all targeted marketing opportunities. And the Duolingo ad has that superficial feel: it's a six-second clip in which a cartoon owl farts out another cartoon owl, displaying a tagline. But it's a great ad that's native to the context in which it's consumed rather than to the medium itself—the natural reaction for the intended audience of the ad is to either ask "What's a Duolingo?" or, if they're really clued in, to turn to someone and say "That ad wasn't for the same 'Duolingo' you said you like, was it?" If the audience for an ad will include people who love the brand and people who don't know much about it, and will put them in the same room, the more weird and polarizing the ad is, the more defensive the brand's fans will be. So it's a way to buy six seconds of expensive ad inventory in order to engineer a million parallel word-of-mouth endorsements. (They did something similar with an ad that ran in theaters before showings of the Barbie movie, coupled with brief product placement that's only noticeable to people who are actively using the app, which they say they didn't pay for. Apparently the Duolingo format is to have actively engaged customers explain references to the app in shared media experiences.)
The result of this shift shows up in the aggregate performance. Looking at their Correctly Adjusted EBITDA margin, they went from negative 22% in late 2021 to negative teens in 2022, to negative mid-single digits in most of 2023 and a positive performance in Q4.
US-based investors get a skewed view of what Duolingo is for: English is the most popular language on the app in 122 countries, and #2 in fifteen—including the US. They have a low-priced English test offering, which is 8% of revenue today but was a big source of growth in the past. Ads are another contributor, at 9% of revenue; they were growing fast for a while, but various policy changes by assorted trillion-dollar companies have ensured that the minimum scale for a truly great online ad business is only attainable by companies of a similar size.
The main sources of revenue and growth today are subscriptions (76% of revenue, decelerating this year after a great 2023) and in-app purchases (7% of revenue, but growing faster than the rest). These two revenue lines present an interesting behavioral laboratory. What sells subscriptions is a user's commitment to repeatedly using the app. Duolingo is constantly testing new ways to do that—notifications, winning streaks, etc. Those tricks all switch from motivating to demotivating when users miss them: break an 800-day language-learning streak, and you're stuck below your record until 2026. So they have an in-game currency that can be spent on things like freezing a winning streak, and those in-game currency units can be purchased for out-of-game US dollars. At that point, the goal is to optimize for changes from pure retention to a kind of tactical non-retention; a user missing a day can be revenue-accretive if it prompts them to pay for a streak freeze, but revenue-negative if they actually quit. The optimal user is similar to the optimal subprime credit card holder: someone who keeps the time equivalent of a high-interest revolving balance that they never quite manage to pay down, but don't default on, either.
One recent source of growth for them is their family plan: it's $10/month compared to $7/month for their standard plan. But as with other cases where companies heavily discount when they bundle, the bet is partly that the incremental user on the account wasn't going to pay anyway, but also that having more users associated with an account means having more people who want to veto any plan to cancel it. Bigger bundles do cap their theoretical upside, but in practical terms they're a good deal (especially since they trade off lower potential subscription revenue for a higher surface area for selling in-app purchases).
Risks remain, of course. AI was and remains a threat to Duolingo, but so far seems to be better for costs than it is bad for revenue: they cut 10% of their freelance translators, replacing them with AI. One reason it's working well for them is that AI is so general-purpose that it's hyper-sensitive to individual use cases, in the same way that there are many industries that are downstream from electricity, transistors, or the car (and not just in the direct sense; modern fast food, and the supermarket model for groceries, and suburbia itself all make sense only in a world where every household has a vehicle). Duolingo is packaging a general-purpose tool—human-level machine-based translation—into a format that combines some aspects of a good grade school teacher and a slot machine: constant encouragement, lots of specific feedback, and a continuing effort to engineer 100% daily attendance.
Duolingo ends up being a story about leverage. AI models do improve the world's ability to seamlessly translate text, but they don't yet offer a substitute for being able to converse live in a foreign language. Deglobalization lowers the ROI of learning English in order to do business with America, but raises the ROI of learning English in order to move here. And a company that focuses more on product than on marketing, and that consequently makes its marketing extremely weird, is a company that has the option to get a high ROI on ads by investing in minimalist surrealism and expecting its legions of satisfied customers to explain what's going on.
Diff JobsCompanies in the Diff network are actively looking for talent. See a sampling of current open roles below: - A company building the new pension of the 21st century and building universal basic capital is looking for a product designer with fintech experience. (NYC)
- A crypto proprietary trading firm is actively seeking systematic-oriented traders with crypto experience—ideally someone with experience across a variety of exchanges and tokens. (Remote)
- An alt data provider is using LLMs to scrape and is in need of a backend engineer—experience building with LLMs required, web scraping experience preferred. (Remote)
- An investment company using AI to accelerate investment in esoteric asset classes is looking for a product engineer with Python and Typescript experience; preferably someone with a track record of building on their own (Bay Area, remote also a possibility).
- A data consultancy is looking for data scientists with prior experience at hedge funds, research firms, or banks. Alt data experience is preferred. (NYC)
Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up. If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority. Elsewhere
Chatbots
The Financial Times has launched a new chatbot allowing a small subset of users to ask questions, which are answered based on the FT corpus. This is something The Diff has experimented with as well ($). What's notable about these is that there are really two specific use cases:
- As an enhanced search engine, especially one that can provide "conceptual synonyms" that are beyond the scope of a simple thesaurus.
- Asking for new content in the voice of the existing content.
The latter is much, much harder (which is good news for anyone who writes for a living), and seems to have less to do with training on specific data and more to do with training on as broad a corpus as possible in order to very accurately determine what a given writer sounds like. The easiest-to-duplicate elements of someone's style will be the subconscious linguistic tics and repeated mistakes; getting LLMs to not just produce insight but produce a specific kind of insight is, for now, a challenge.
Alpha Capture
Here is a not-uncommon occurrence at big multistrategy funds: a portfolio manager running a relatively small amount of money. Their best idea happens to be a large-cap company in which the overall fund could make a larger investment—a careful statistical analysis of the manager's previous picks shows that this one has an unusually good risk/reward. So, either by algorithm or ad hoc, the firm doubles-down, or 10xes-down, on that portfolio manager's bet, in a separate account. This general approach, known as alpha capture, is getting more common, and seeing a growing number of permutations ($, Business Insider). Some funds are doing pure alpha capture, with analysts and no portfolio managers; some people are doing alpha capture for third-party funds (offering them capital if their picks work out well). Every industry tends to specialize over time as it grows—both Henry Ford and Elon Musk insourced production for many components because they simply couldn't find an acceptable third-party option, but over time both industries have developed more of a supplier ecosystem. Asset management should be no different: researching individual picks, managing a portfolio, and running a firm all require different skills, and it's unlikely that those are so correlated that anyone in the world can be best at all of them.
Coreweave
GPU-focused cloud computing startup Coreweave is planning to raise at a $16bn valuation, up from $7bn last year. They're part of an interesting category of companies where the financing options are part of their unit economics. Affirm, for example, can underwrite a transaction not just based on whether a given user has good credit, but whether there's market demand for users in that category. Opendoor can look at its assets as both a source of future capital gains from sale transactions and a source of collateral for immediate financing. GPUs turn out to be a very financeable asset, in part because their useful life is short enough that lenders have a margin of safety—demand for AI has to reverse very fast for an allocation of new Nvidia chips to be a bad deal. That means that a company like Coreweave combines the capital moat of a high fixed cost base with the capital efficiency of a company whose growth can be funded by something other than equity. But that valuation is the confluence of several factors Coreweave doesn't control, like Nvidia's economic interest in keeping AI fragmented ($, Diff). So what this really amounts to is a deal with political risk, not from nation-states but from a company whose policy could change at any time.
Preannouncements
One aspect of companies' earnings guidance is the strategic early release of financial data. This takes two forms: when performance is falling short, it's better for risk-adjusted returns to break the news up into incremental smaller announcements, and when shareholders aer nervous without a good reason, it makes sense to preemptively announce that results will be just fine. China is experimenting with this, announcing macroeconomic statistics hours to days in advance of their official release. One thing this does is make it a little less safe to be bearish on Chinese equities; even if a typical approach is to de-risk ahead of market-moving data, if that data shows up off schedule there are better odds of inducing a temporary short squeeze.
AI's ROI
The WSJ has a roundup of companies' efforts to put a number on the return on their AI investments ($, WSJ). As with many other technology transitions, implementing a preexisting model has fairly predictable returns, albeit with some friction from getting employees to actually use whatever new technology the company is paying for. More speculative ideas—remaking the org chart rather than speeding up the pace of answering customer service inquiries—necessarily has a less predictable return. And there isn't a good rule of thumb for which technologies benefit their creators more than their customer: jet travel was a subsidy to Las Vegas, air conditining was an essential ingredient in the economic models of Singapore and Dubai, and in both cases the consumer surplus was also substantial. It's always possible that the AI dividend will be one of those widely- and unpredictable-distributed economic gains, not one that any one company will be able to monopolize.
|
