In 2009 we hosted a one day event we called “Hacking Education.” We brought together a diverse group that included game designers, principals from schools, and even homeschoolers. We were fortunate to have the late Sir Ken Robinson give opening remarks which set us off on a course of open exploration that day. We have since invested in an ever growing portfolio of learning companies: Duolingo, Codecademy, Skillshare, Quizlet, Top Hat, Outschool, Fiveable, Sora, and Brilliant. Roughly ten years later we held a second Hacking Education event in 2020 to exchange what we have learned from these investments. Since then there has been an explosion in the capabilities of Artificial Intelligence, which has kicked off a lot of new discussion and prompted us to hold an AI Learning Summit earlier this year.
There is a clear ideal end state: Every person in the world can have access to an individual tutor on any topic they want to learn about. This doesn’t preclude also having group activities which have their own importance, such as developing social skills or building team cohesion. There is overwhelming evidence that access to an individual tutor results in dramatically better learning, something known as the 2-sigma effect based on studies that found a two standard deviation improvement. Going back in history before the emergence of schools the children of the elites were all trained by individual tutors. Famously, Alexander the Great was taught by Aristotle. Even today, at a time when schools are dominant, parents often invest heavily in additional tutoring.
The amazing promise and potential of AI is to make access to individual tutors a reality for everyone. Here is how Luis von Ahn, the founder of Duolingo, put it: “I want the poor person in Guatemala to be able to learn with very high quality, [t]he only way I know how to do that is with A.I.” This will of course take time as at present running large models is expensive. And models are, despite the recent rapid progress, still in their infancy with significant issues, such as hallucination. Still even today one could use AI to give human tutors the capacity to interact with a far larger number of students in a “human-in-the-loop” system. For instance a tutor could be looking in on the chats happening between students and bots and course correct these as necessary.
There are many other fascinating issues to be addressed in building such systems. What are the subjects best suited to deploying AI initially? What is the role of empathy, for example when a learner is frustrated? Is that something AI can have also? Potentially even better because it might not get annoyed at a lack of progress? What is the role of illustration and animations? Can these be created in part or entirely by AI? What is the right path: adding AI to existing experiences or starting from scratch?
While incumbents are investing heavily, we expect that there is a lot of room also for AI native experiences created from the ground up based on the emerging capabilities of the new foundation models. Here are some startups we have become aware of: Tutor AI, Teach Anything, Synthesis Tutor, and Foondamate. We are sure there are many more. All of this falls squarely within USV’s Thesis 3.0 which has broadening access to knowledge as a core component. We are excited to engage with teams that are building towards such a vision.