We use LLMs to automatically generate quizzes, practice questions, and real-time feedback from any educational video or resource—turning passive watching into active learning. Here’s a short demo: https://youtu.be/alO7FaorHOY.
Improving education has always been tricky. Bloom’s 2-sigma problem (showing that a high-quality personal tutor is far more effective than conventional methods) has persisted, even as technology has advanced.
We met at MIT as CS majors and have always been passionate about education. Over the years, we’ve become teachers and experts in subjects like chess, algorithms, math, languages, and ninja warrior. A common theme was that we both heavily relied on YouTube to learn.
YouTube has incredible content for learning pretty much anything, but it’s buried in a lot of distractions. Also, passively watching videos is far less effective than taking notes, asking questions, and doing practice problems, which is what we aim to do with Miyagi Labs.
Our solution is essentially a multi-step function that takes in a YouTube playlist (or list of any resources) and outputs an entire course with summaries, questions, answers, and more. The pipeline is roughly: video/resource —> transcript/text —> chunks —> summary and question —> answers to questions, with some other features along the way.
We mostly use prompting and different models at each step to make the course as useful as possible. Certain topics require more practice problems vs. comprehension, and we’d use reasoning models for highly technical subjects.
We launched about three months ago and currently have 400+ courses and partnerships with some businesses and awesome creators. Some of our popular courses include 3Blue1Brown’s linear algebra course, a botany course on plants and ecology, and YC’s How to Start a Startup series.
Our product resembles classical MOOC-style course platforms in terms of UI, but is more interactive. It’s really easy to ask a question or receive custom feedback compared to a static course on Coursera. It’s also comparable to AI tutor sites, but we try to build more of a community and require less activation energy as a learner. We’re basically betting that AI can hugely improve education, but that students still want to learn from their favorite creators and want baseline shared resources for standard topics that are then augmented with personalized features.
You can try it here: https://miyagilabs.ai (no login required for most courses—but if you sign up you can also create your own course).
We’d love your feedback on what kinds of videos/resources you’d like to learn from, what’s missing from current learning tools, and if you know any creators or educators who would like to collaborate. Happy to hear any feedback and answer any questions!
This concept is really cool and solves big challenges around content creation. Obviously, it adds new challenges around pedagogy, licensing, and ads. The last part is a big no no for blue chip edtech platforms.
1) Section Lectures on the left side need to be cleaned up, instead of just a numbered list. Seeing 30+ lectures off rip is a bit daunting, especially with no labeling, sectioning, etc. I'd imagine feeding a model a list of all the lecture titles, then having it structured should work?
2) You're doing too much on the bottom section.
You need to incorporate all those tabs into the single Ai tutor, which can run whatever tools required (maybe notes/discussion can be a small additional indication). No one's going to be using the Flashcards section, and it's calling probably the same LLM as the AI tutor, so might as well combine them.
For the quiz, maybe when the video ends or the user wants to continue, the Ai Tutor goes into "quiz mode" forcing the user to attempt or pass the quiz (depending on the settings?).
Think of this like Cursor but for Education. Cursors powerful agent can handle/do so much, you're not using 3-4 different fields.
Oh and have it on the right side instead of transcript, so it's right there in users faces instead of having to scroll down.
Not completely sure about the AI tutor points though. I don't think the standard AI chat interface is the ideal form factor for the average person trying to learn something, and there's value in having pre-generated content that users can see instead of having to actively go to the tutor.
Also, a lot of people do like using flashcards specifically to learn! Granted, our current implementation is pretty barebones so it's not super useful yet. And definitely agree that things can be cleaned up quite a bit.
And no, I don't consider "they get extra views" a valid answer. Especially if you expect to make the windfall you want to make off their labor.
You mean like how all art is created? Nothing is original... everything we create is a sum of past inspirations, human or not.
I’m curious why you didn’t use multiple choice for the exercises? I feel like those would be easier than typing out full answers and be closer to MOOC style homework. Maybe have a longer written question at the end of a section.
The exercises work pretty well, I like the highlighting red wrong vs. green right. It does feel a bit like the MOOC-style discussions. The tutor doesn’t just tell you the answers which is cool, but something about talking with the tutor feels a bit flat. And the flashcards weren’t very helpful for the course I picked.
I could see myself doing some courses like this with some more gamification. Being able to filter by course provider (Ycombinator, or MIT) would be cool too.
Anything specific we could improve about talking to the tutor? Definitely will add some of those features and gamify better.
Again, very cool idea. I'm going to try some of the nuclear courses later this week.
Best of luck!
Blows my mind that 1:1 tutoring dwarfs the impact of other factors such as socioeconomic status, reinforcement, assigned homework, classroom morale, etc (at least according to the researchers).
Does anyone know if this thesis has been replicated? Or if these results hold in modern times (original study was 40 years ago)?
[1] https://www.science.org/doi/full/10.1126/scirobotics.aat5954 [2] https://www.tandfonline.com/doi/abs/10.1080/00461520.2011.61... [3] https://journals.sagepub.com/doi/abs/10.3102/0013189X2091279...
It's super common for students to not understand material or express that they don't, but it's just not possible to drill into each specific student's particular knowledge or skill level in the classroom environment.
On the social and behavioral side, many students who struggle in a classroom environment transform into model students when taught with both the care and privacy of 1:1.
For me, I feel it's a combination of hyper-personalized instruction plus compassion in a relationship where it feels safer for the student to accept the value of improving at something without social pressure or embarrassment.
For example, taking a video about building garden retaining walls and generating detailed system prompts for Q&A with the expert in the video.
I reference ~home improvement or tool videos and often comments contain points of wisdom or even corrections of mistakes (errata) on videos that are otherwise good. For example, setting up a hand plane and ways to mark a board you're working on.
Do you use video comments in your context? I've (manually) scraped content on educational videos and built prompting to assess signal and incorporate what are likely important errata in LLM context.
> video/resource —> transcript/text —>
For this step in your pipeline, are you multi-modal? I mean, are you using the LLM to interpret what is shown in the video itself? How is that content used?
Do you have any sense for allowing people to generate educational content off arbitrary videos?
We may make that an option though, since we also offer other resource types (pdf, slides, docs) -> course.
This maybe doesn't matter as much in the content you're processing now, but see my reply to sibling comment here, where it matters a lot for these niche videos where the creator did not mention or did not know about some important detail of a process or task.
If I have personally qualified a person as an expert on building retaining walls for landscaping and want to have your platform process a particular video, would you allow this?
I'm not necessarily interested in the topics you're providing resources around. I want niche topics. Like roofing or how to properly set up a traditional hand planer.
I am interested in individual, often amateur contributors of niche knowledge on youtube, not channels designed to educate in the first place.
Prof. Steve Brunton's YT channel is a treasure trove of material for you folks, with course-like playlists for controls, data-driven engineering, and dynamical systems: https://www.youtube.com/@Eigensteve/playlists
He should be a featured creator, much like 3b1b is for math!