Dogfooding and the Chip on My Shoulder
Dogfooding AI for language learning
I feel it is useful to provide some context for how I arrived at this project. I apologize for the navel-gazing, but I’m hoping it’s useful to see where I’m coming from.
Language learning is a passion of mine. I’m passionate about learning in general, but language is particularly interesting in how it can rewire the brain so that one can view the world through a different conceptual lense.
I have found that most language materials assume someone is learning within an academic context or for career advancement reasons. Finding ways to persist and advance in language learning outside of these settings remains a major unaddressed pain point in the language learning market. For the past few years, I’ve been experimenting with ways to address this problem with AI-powered tools. I’ve been dogfooding various AI-based workflows for language learning as I learn my fourth language - Portuguese. I’ve found two successfuly strategies that inspire some of what is presented in this proposal.
Integrating learning into media consumption habits
One strategy is to integrate language learning into one’s existing media consumption habits. For example, if someone enjoys romance reality shows and is learning French, they can watch Love is Blind: France. If they enjoy true crime podcasts they can listen to such podcasts in French. If they enjoy music they can use Spotify to learn to sing the lyrics of a French musician. If they love Harry Potter books they can read the books in French.
This strategy works because if the habit already exists and can be modified to become a learning habit, it is easier to be consistent about learning. Learning a language is frustrating because one is constantly aware of their lack of ability, while it is difficult to be aware of one’s progress. Finishing a book or a season of a show, or performing a song, can provide that valuable feeling of progress.
Several apps provide language learning tooling around media consumption habits. For example, LanguageReactor provides tooling around streaming services. Pimsleur provides audio lessons that one can listen to on their commute or while doing housework, as one might do with a podcast or audiobook.
There are media consumption habits that are not addressed by existing tools (e.g., audiobooks, TikTok feeds). Modern AI tools (LLMs, voice-to-text transcription, video description with VLMs, RAG systems, image/video generation, etc.) are providing new opportunities for building this type of tooling.
Online tutoring
The second strategy is using paid language tutoring platforms like Italki, Prebly, and Verbling, as well as language exchange platforms like HelloTalk and Tandem. Firstly, it is easy to slot in a 30-60 minute online tutoring session in a post-pandemic world where adult professionals regularly schedule and attend video meetings. Secondly, for learners not pursuing career advancement or academic objectives, having conversations with new people on these platforms is in itself rewarding. This is one of the biggest insights from the research presented in this proposal – talking with people has value beyond just the incremental benefits of conversation practice.
The chip on my shoulder: Replacement Bias
They say good entrepreneurs have a chip on their shoulders. My chip is the tech culture’s obsession with using AI to replace humans in meaningful human interactions, rather than using AI to enhance those interactions. For example, I think “AI girlfriend” like Nomi.ai apps are bad, not because talking to agents is bad, but because the marketing and design of these apps are premised on the idea that the value of a romantic partner is based only on how well they can gratify you.
But beyond the moral considerations, this replacement bias is an opportunity to build tools that enhance human interactions. Indeed, some apps are trying to be Nomi.ai for language learning. Even DuoLingo has gotten into the game – their “Max” tier allows you to do video calls with its character Lily.
To be clear, I think there may be value to audio to video calls with agents (perhaps as a rehearsal for a conversation). But DuoLingo has made it clear that it is fully bought in to the replacement narrative with this product. This blinds them to the opportunity to provide the reward that learners recieve by talking to human tutors.
There is an opportunity to support human interactions on these tutoring platforms with AI-tools. Indeed, research I conducted in India demonstrated that sellers on gig economy platforms in general–who you might suspect would be replaced by AI–actually tend to be AI power users. You’ll see that trend reflected in the tutor interview insights presented in this proposal, where many tutors use ChatGPT to generate content for use with their students. Of course, the challenge is that tutors don’t usually want to pay – it is the independent tutors, not the platform tutors, that pay for tooling for teaching. But there may be an opportunity to provide tutors with tools that create more value for paying students, and then to capture that value.