Edtech Companies 🎒📱 Riding The GPT Wave 🌊🤖
originally drafted August/September 2023
I’m hoping to take a look at some aspect of growth models leveraged by some of the most intriguing organizations and individuals in the world, starting with an area near and dear to my heart: education.
My primary focus will be on the concept of growth models and engines.
I’m a founder, growth and product-leader who has been through lean and abundant times, peace and war times.
I understand that growth can stall or plateau, it can look great one month and terrible the next (until you build a solid foundation of PMF and growth).
Navigating the messy middle of growth requires taking a hard look at the data and results (are users signing up but failing to convert to paid and/or retain), comparing it to your hypotheses/assumptions, and figuring out the core decisions that will unlock growth.
Sometimes that’s about the underlying product strategy (or model).
Sometimes it requires taking a hard look at the business model.
Sometimes, it means revisiting your financial model and assumptions
Sometimes it requires raising funding.
Sometimes it takes validation to double down on a specific decision or strategic direction. Or just intuition and guts.
This Week: The Early Days of AI
While many aspects of LLMs (whether base models or especially assistants / agents) are still largely experimental1 , ChatGPT becoming the fastest 2nd fastest app to 100 million users has suddenly made the power of generative AI and LLMs known to much of the world, and made it imperative to move rapidly.
🍱🐅 to avoid having your lunch eaten
(as Chegg has experienced with its user base and been (even if unfairly) punished by the markets, with a recent small recovery thanks to its recently announced partnership with Scale AI)
🌊🤖 to “ride the Wave of AI” (or more precisely, GPT)
We’re likely still in the early innings, but it seems the prior developments in AI and machine learning (computer vision, GANs, deep reinforcement learning, RNNs, CNNs, attention, and LLMs) have finally reached a point of usefulness where we’re poised for explosive growth in use cases and applications, infrastructure, both for incumbents and startups.
Most people agree, the leap from DALL E to DALL E 2 or GPT 2 to GPT 3, 3.5 and 4 was from nice party trick and oh that seems intriguing to ok, this shit is just plain useful and helps me unlock even more productivity and creativity in my work (no where is this more clear than with developers for whom their expertise allows them to prompt, review generated code, and increasingly (with the help of code interpreters as plugins or integrations, actually push code to production faster than ever).
I love this visual and framework from Elad Gil on AI adoption curves and where we find ourselves right now, with fast mid-market incumbents riding wave 2 of AI (and especially LLM) adoption curves.
And… edtech is no exception.
In edtech, some of the fastest movers of the mid-market incumbents include teams that already had strong AI/ML teams, features, infrastructure and were well positioned to ride the latest wave.
Three Edtech Companies Walk Into A Bar An Order A Drink…
We’re going to look at Duolingo, Quizlet and Khan Academy, who I think can justifiably be called fast mid-market incumbents:
Duolinog is post IPO.
Quizlet is pre-IPO with one likely coming
Khan Academy is a large, well known education nonprofit - small in total revenue compared to Duolingo and likely Quizlet, but are no slouch coming in at nearly $60 million in revenue and expenses in FY 2021.
In addition to size, all three launched AI Tutor/Chatbots within a few weeks of each other in March, at the same time Open AI launched ChatGPT Plugins (and it was not coincidental, at least not for Duolingo and Khan whose announcement coincided with OpenAI’s).
Quizlet Q-Chat announcement, March 1st, 2023
Duolingo Max announcement, March 14th, 2023 (OpenAI’s announcement)
Khan Academy’s Khanmigo announcement on OpenAI, March 14th, 2023
OF NOTE:
None of this analysis or thinking references any non-public information, and I have not spoken to anyone working at these organizations. (Although if you want to discuss and share I’d be open to it!)
🤳📱 If you’re someone who has used any of these AI Chatbots, I’d be love to hear what you think of the experience - feel free to share more in the comments (or DM me)!
🦉💰 Duolingo: Show Me The Monetization
Duolingo has been at the vanguard of auto-generated content (per one of my favorite analyses of the company courtesy of nibnalin), and this has always separated it from many other edtech companies (and even tech companies in general).
It has built a flywheel atop:
This auto-generated content and a differentiated, gamified product experience.
The learner usage data has network effects - the more users that use the product, the more data Duolingo gathers, the more it can use this data to improve the content, product and learning.
Core Growth Loops: viral and paid
Business Model: B2C, % of Freemium that conver to Paid Subscription; with some B2B via schools and moving into new areas like assessment (its English test competing with the likes of TOEFL)
Because Duolingo has been built with machine learning at its core, it is more accurate to say Duolingo has been riding this wave for quite some time.
Nonetheless, Duolingo still saw an opportunity when ChatGPT was released, both with the improved underlying technology (GPT3/3.5) and a new product UI/UX that took the world by storm.
But, while adding a tutor certainly can have impact on acquiring new users (and riding the GPT wave specifically), I don’t believe acquisition that was the primary goal:
because Duolingo’s core product is free2 and
because they already have a strong AI/ML core foundation.
because it was only launched in limited languages and markets:
Duolingo Max was only launched initially for Spanish and French for English speakers on iOS, and only available in the U.S., Great Britain, Ireland, Canada, Australia, and New Zealand.
I think one the main goals is to actually see if it’s possible to monetize a new learning modality of practicing language.
This also solves for one of the longstanding known weaknesses of Duolingo, it can help a lot with vocabulary/basics, but is weaker when it comes to learning grammar, syntax and for actual usage.
Additionally, charging a higher fee of $29.99 per month (vs. $12.99 per month for Super Duolingo) ensures you are getting users who are serious about using it.
I’d be curious as to what kind of conversion they have experienced, as well as their MAUs and DAUs (engagement) and retention curves for their first cohorts now that it’s been nearly 6 months since launch. And the kinds of issues and product improvements they’ve made since (and limitations they’re running into as well). I’ll be looking out for any mention of this in their next Quarterly Report.
They didn’t disaggregate it in their last 10-Q (for results up until 6/30/2023), but that’s understandable both because 10-Q’s usually don’t offer that much color to the results, it was such a short period since launch (just over a quarter), and it covered such a small base of languages and regions that it probably made a negligible impact from an overall perspective.
For this strategy to work for monetization longer term, the experience has to feel magical and work (toward improving ones language skills), such that those who convert to paying continuing engaging and retain.
Finally, I intentionally don’t think this was a defensive move because while you can go use ChatGPT to work on language learning, the modality is just so different. The public markets have also correspondingly increased Duolingo’s valuation (vs. say Chegg, where its content and textbooks are significantly less differentiated from a product, UI/UX perspective, and it has admitted to losing users due to ChatGPT.)
🆀💬 Quizlet: Q-like to Chat and Stay Awhile? (engage, convert and retain)
Quizlet has built its company largely atop user-generated content, with its core original product being the flashcard sets students would create. It has since expanded well beyond this and also built a machine learning/AI team starting around 2015 and building products/features around figuring out how to surface, judge quality, even monetize (through partnerships and publishers) content shared through the platform.
Quizlet has also built a host of learning tools to supplement the flashcards and enable better learning, studying, review, and like Duolingo, more data helps power its flywheel of product and learning experience.
Core Growth Loop: content, specifically UGC and data
I believe they also increasingly have publisher or company generated content from themselves and parnters.
Business Model: B2C, % of Freemium that conver to Paid Subscription with some B2B (group purchases of subscriptions)
Quizlet has over 700 million sets that have been manually created by teachers and students (per the August additional AI tool announcement).
Q-Chat was launched in Beta to US students 18 and up (likely driven largely by OpenAI’s own age requirements) with limited access to free users (up to 5 rounds), and unlimited access to users paying for Quizlet Plus.
While the core flashcard/content sets and educational resources of Quizlet are likely higher quality and the experience obviously make for a stronger holistic product experience, I’m still concerned that the modality might not be as different (as Duolingo) from ChatGPT and thus I believe that defensiveness and potentially keeping users might from alternatives might have pushed them to speed (and they got out ahead of OpenAI’s own Plugins announcement).
Defensiveness aside, a chatbot built atop their vast amount of quality learning resources can help maintain or improve engagement and retention by giving users more control and flexibility about how they engage with those learning materials besides just via chat. I think then this move has three components for Quizlet:
Let users try the experience for free to experience what it’s like and (hopefully) the magic of it (commonly used with product-led growth motions): try-before-you-buy.
Users who find it useful but limited by the number of rounds convert to paid subscribers. (at $7.99 per month or $35.99 per year, reducing the monthly cost to ~$3.00 per month)
Those who subscribe continue to engage and retain.
Some additional anecdotal experience using the tool described here.
My hunch feels further validated by Quizlet’s latest launch and announcement on August 8th, appropriately titled: Welcome to Quizlet’s AI Study Era: Studying will never be the same. There are a few tools, but really to me the interesting ones are:
Magic Notes: automatically turn your notes into outlines, flashcards, practice tests and more (instead of manually creating the educational sets and resources)
To quote the announcement: And it gets even better. Magic Notes generates additional course materials and relevant concepts, such as sample essay topics, and simplified summaries, so that students have everything they need to master their studying.
Magic Notes is really a content transformation tool GPT and other LLMS are well suited to, but being able to turn into flashcards and practice tests makes using Quizlet better than using ChatGPT on its own. The integration and solving a big challenge students used to have should improve retention and engagement (and potentially even accquisition).
This is also a useful anecdote of a journalist being given access to Magic Notes to take it for a test drive, and their experience.
Expert Solutions: combines Quizlet’s millions of expert-written explanations with AI to help make breakdown difficult problems and scale feedback —> in addition to the explanations that already exist, students can go deeper into key concepts, see alternate explanations or practice problems, and personalized AI-powered tutoring all built within the experience.
Likewise building AI on top of Expert Solutions as the core helps address another challenge with ChatGPT and other LLMs —> their ability to hallucinate information and make mistakes in calculation and/or logical reasoning makes this another smart move. Using explanations effectively as part of a fine tuning and RLHF (reinforcement learning with human feedback) process, AI makes it easy to offer alternate explanations when the core content is already validated. But it remains to be seen how well this works in practice.
This announcement shows Quizlet leveraging AI tools, but further differentiating Quizlet by integrating them into the learning ecosystem of content and tools, and building atop their UGC and data flywheel. This quote from the announcement feels like it validates my view (emphasis mine).
For students looking for a deeper understanding of study materials, there’s Q-Chat: Quizlet’s AI tutor. Since launching in March, we have had millions of conversations with students on Q-Chat and it has become a core part of the study experience.
As with Q-Chat, “free users will have limited access to the features, while Quizlet Plus users can gain unlimited access.” (per an article from ZDNET). These new features (Magic Notes, Memory Score, Quick Summary, and AI-Enhanced Expert Solutions) will be available “to students 16 and above in the US, UK, and CA for the back-to-school season.
🌱👋 Khan Academy: Let’s Meet the New Kid On The Block (acquisition)
Khan Academy is a company that started famously with Sal Khan teaching his cousin over Skype (Zoom didn’t exist yet, wild!), and then realizing he could make these videos so he could ‘teach’ 1 to many. Fast forward, and Khan Academy has created some really high quality learning materials and learning experience around its platform, from learning coding to algebra, calculus, and physics to history.
Core Growth Loop: viral (word-of-mouth)
Business Model: nonprofit (donations, grants) with some earned revenue (accounting for just over 15% of revenue in FY 2021)
Khan Academy launched Khanmigo in Khan Labs (anyone remember Google Labs?) on the same day as Duolingo (and in parallel with OpenAI’s launch of Plugins for ChatGPT).
Personally, I’m a big fan of the way Khan launched and that they’re a first mover, because as a nonprofit I fundamentally believe in the values alignment and desire to drive impact in teaching and learning. Because GPT and LLMs are still experimental (with known challenges), this was intentionally rolled out and framed as a beta, a product under Khan Labs.
Like Duolingo, Khan did set a ‘price’ and was smart to ask for a monthly contribution/donation ($9/month or or $99/year) to access Khanmigo:
It ensures users are motiated and more likely to engage with the tool if they sign up and pay.
It creates the potential for a recurring, SaaS-like subscription revenue stream, but as a nonprofit.
It was positioned and targeted at parents (see screenshot of an email invite I received when I was given an opportunity to get off the waitlist and begin using Khanmigo). While I imagine Khan Academy has done work to target parents, Khanmigo does feel well suited to go direct to parents and students (as well as B2B to school districts, which Newark has been piloting with, for now, admittedly mixed results, NYT Article from June 2023)
First, I believe Khan is building as an early-mover so that they can create impact.
BUT, I also think it’s their latest announced partnership with Instructure that unveils their longer-term ambition — which is to drive acquisition and expand even more deeply with teachers and a newer use case.
Again, the partnership with Instructure and being a part of their marketplace is still framed as R&D (and part of Instructure’s “Emerging AI Marketplace” in beta), but to quote the launch announcement:
Instructure and Khan Academy will partner with educators to design solutions that will amplify the power of the teacher and drive increased learning outcomes for students. The vision is to provide Canvas users with an integrated Khanmigo experience supporting essay feedback and grading, lesson planning, rubric creation, and student differentiation. This partnership aims to address widespread concerns educators have raised about ChatGPT and the problem of cheating on student writing assignments. We aim to support students through Khanmigo’s tutor capabilities to help students learn versus helping them cheat. This partnership is the first time Khanmigo will be deeply integrated into an industry leading LMS.
Note that this is about helping educators with four use cases:
essay feedback and grading
lesson planning
rubric creation
student differentiation
Khan included a teacher tool with its launch of Khanmigo, but partnering with Instructure provides (an eventual avenue) for distribution where the tool is accessible by all the teachers who use Instructure’s LMS Canvas.
Let me know what you enjoyed about this piece. If you want to see more like it, if there are other companies you think I should cover.
If you work at one of these companies or organizations and want to challenge my opinion or share more nuance or depth, please reach out! Or if you’ve actually used these features, let me know too!
Per Andrew Karpathy during his talk on the State of GPT at Microsoft Build at the end of May, a mere 3 months ago but ages ago in the fast moving world of LLMs. Nonethless, this is still one of the best talks on the state of play.