Launch HN: Miyagi (YC W25) turns YouTube videos into online, interactive courses
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!
This is, I think, particularly important for kids in most parts of the world as a majority of the internet pedagogical content is English-based. Or for people like me, that struggle with that language when talking about topics outside the tech industry.
Congrats for your project, I'm sure we share the same positive view about the future of learning :)
I already use LLMs to quiz me on books that I'm reading. The current book I've been studying is "Computer Networking: A Top-Down Approach (8th edition)." Obviously since it is a text book, there are homework problems and other discussion points. However; it's also fun to ask grok/chatGPT to quiz me on previous chapters and correct my answers.
There's plenty of conference talks posted on YouTube that I've watched over the years. Next talk I watch I will probably test out being quizzed by LLMs.
Here are some of the very best in the category, it would be really cool if you partnered with any of these.
https://www.youtube.com/@hungryhorsepoker
https://www.youtube.com/@CarrotCornerPoker
https://www.youtube.com/@PokerCoaching
Something I’ve been doing more and more lately is asking chatgpt to create a detailed description of a topic which can be read aloud for whatever duration I plan on driving. This works exceptionally well - even for short 5 minute drives.
I wonder if the same can be done for video-based content. Sometimes I’m short on time but still want to learn something.
i opened a random lecture (MIT 14.01 Lecture 19: International Trade: Welfare and Policy)
a- i tried an exercise without watching the video, and then clicked on 'watch' next to the question, i expected it would take me to relevant portion of the video instead it played it from beginning.
b- i used the ai tutor and asked 2 random questions, instead of giving me an answer it prodded to provide more detail. this might be an intentional choice to ensure learning, however i personally prefer chatgpt's style where it gives me some information and then provides more angles to pursue.
i also asked the tutor about batman, here's the response for reference:
tell me about batman That's a good start, but your question is very broad. When you say "tell me about Batman," are you looking for information about the character's origin, his role in comics, his significance in popular culture, or something else? Try to narrow down your question a bit.
For example, are you interested in: - Batman's backstory and how he became a superhero? - The different versions of Batman in comics, movies, and TV?
And yea probably thinking to do a guiding mode (as you said, often better to have the tutor guide you through step-by-step instead of giving you the answer right away) and a straightforward answer mode for the tutor. Or if you have any other ideas.
For self-created courses, it's generally been quite accurate and we're playing around with some eval metrics to make it as good as possible, but it's definitely a concern.
I am curious if you are using any methodologies from the digital learning space like knowledge tracing to help ensure that learners are actually retaining knowledge and improving over time or knowledge mapping to understand the gaps that might exist in your content?
Do you maintain your own skills taxonomy? Are you tagging your questions or assessment events with knowledge components or skills of any kind to understand what you are testing your students for?
All of this is really cool, I’m just curious at what level you’ve gotten to on some of this because there is a very fine line in online educational content between making the students life more difficult and actually helping them learn, especially when you get into auto-generating content, and especially if you aren’t following solid principles to verify your content. (I work for an online education company and particularly in the space of training LLMs and verifying their outputs for use in educational contexts)
Yep—also in the process of adding learning paths for certain subjects, so you can go from an introductory course to more advanced topics and fill in gaps in understanding. Agreed: our mission is to help students actually learn in the best way possible, we have individual courses now to start out but the goal is to integrate the learning experience.
Very curious to chat about what you guys do, and if you have recs for any literature in the space that we should look at.
Also happy to chat more, I find this space really interesting!
- For official courses the creators are doing some quality control and do necessary fixes. - For self-created courses there is zero human supervision or quality control.
Is that correct?
Recently I've gotten a lot of benefit from much of Alex Hormozi's thoughts. Yesterday I grabbed a 4-hour video, ran it through a free transcript generator, dropped it in Claude, and asked for an outline.
Unsurprisingly, Claude struggled to provide a complete outline. Like, it did a lot! but kept leaving parts out. I was able to prompt Claude to fill in more, good enough for me.
Anyway I dropped it into Miyagi Labs, waited 30 minutes while it said the course was creating. I kept it open in a tab and kept an eye on it. Eventually I tried to open a new tab into Miyagi (maybe I'd see it under My Courses?) but I was no longer logged in, and can't log in.
Also only the top right login button gives the option to use non-gmail.
Sorry if I broke something! I love the idea!
UPDATE: " This course failed to generate. Please try again or contact us."
I really like a lot of the components of your idea, but the execution is underwhelming. Right now it feels like you're providing middling tools for too many components without nailing any of them. Alternatively I could watch the YT video at all ready has a transcript, take notes in any tool, and ask questions to any LLM; the piece missing is context, so that's where it feels like you should focus.
Re: assessments; it feels like you're being distracted here; I'm not convinced that's how your natural target market learns in this modality. We generate quizes in our product, but it's typically used in the "internal compliance" segment - think mandatory training like food safety for food preparers - not the external (typically adult) self-improvement market (which is huge!). If you're going to do asessments you need a lot of non-AI boilerplate around tracking, validation and certification/credentials. My two cents: quizes in your app are a cool demo feature with little real value.
Not sure you can draw many conclusions on an experience you didn’t have. My own two cents
Congrats on the launch!
More serious comment.
It seems like there is less space (in our mind, and time) for yet another app. If you have this, vs youtube for course, people will still flock to youtube. Why don't you add the value that you are giving here by directly integrating or enhancing youtube. A chrome plugin or a youtube browser?
I really like the motivation behind this product. Learning is hard.
Imagine you would need, another ai pipeline that poses as the consumer and applier of the knowledge, instead of a direct processor ai of content information as it currently seems.
Let us know if you’d want to extract from long videos. You might find this useful: https://docs.vlm.run/guides/video-ai/guide-video-transcripti...
For instance we worked directly with Crime Pays but Botany Doesn't & Faculty of Khan etc. to get official courses that they also had input in, and 3Blue1Brown is on board with us having his content on our site.
I’m still coming up to speed on the full scope of what your product does, but I’m curious what you’d say to someone like pal2tec, who has some fairly strong and what I feel to be reasonable views about the impact of content summarization [0].
Getting direct buy-in and sharing revenue is great. But it’s not clear to me that this is the only thing that creators care about, i.e. are you still summarizing content you’re not monetizing without creator buy-in?
- [0] https://m.youtube.com/watch?v=ULUSS1-G3do
Just watched the video, I don't initially agree with his take completely but do totally respect the viewpoint and think a payment split to the creator whenever someone summarizes the video makes sense.
Yes we do offer the option to summarize content without creator buy-in, although it seems a bit different since we're also augmenting the content with questions etc. which should drive users to watch the video even more as opposed to skip it and just read the summary.
But you're right it's not perfect. If we ever have creators who don't want their stuff on our site we'd totally respect their wishes, but that hasn't been the case right now so this seems like the best thing to do.
From a creator’s point of view, I think the concern would be about how true this remains as the product grows/evolves.
But as long as there’s an opt-out, that seems like a reasonable approach.
Also to be clear we have partnerships for all the featured courses. This refers to if a user creates a course based on some videos.
Are you not still making derivative content of the work without the copyright holder's permission? A judge might not care that much whether the video is embedded or not.
Do note that this behavior of "opting creators into a program without their consent, justifying it via revenue share, and CYA with a 'they can opt out if they want to!' shield" is still... awful optics.
The whole Brave scandal (https://news.ycombinator.com/item?id=18736888) is a good case study on how laypeople will perceive this. It's not popular at all.
But in summary, YouTube is rolling out AI summarization features on some content without giving creators any say in the matter.
Concerns include:
- Low quality summarization of high quality content will devalue the content, and in many cases is just a worse version of the content
- Impact to watch time on the channel can impact channel success over time
- YouTube is not doing anything to compensate creators for reducing watch time such as sharing revenue from viewers who primarily interact with the AI summary
But I think he articulates this much better than I did. Much better to watch the video.
FWIW, unfortunately, I think the problem is a two-headed one, and maybe reversed for viewers vs creators. Creators want as many people to see their work as possible. But viewers have to sift through a graveyard of 95%+ junk videos to find the 5% worth watching. AI (or Google/TikTok/etc. in general) acting as gatekeeper in between isn't great, but not having any metrics/summaries/descriptions for videos would be even worse.
In this particular case, I get that this particular creator might've had a point to make, but the description and summary were so cheekily written (to make a point, I guess) that I had no idea what it was about.
The creators who I do follow typically make long-form educational videos with a lot of nuance; I wouldn't want to rely on even the best-written human summary for those. But there are many, many videos for which I'd prefer a 1-sentence summary over 3 minutes of intros and jokes, a 45-second sponsorship, and a gradual dramatic buildup before getting to the point.
Not sure what the long-term solution is.
This is totally fair. I watch quite a bit of his educational content on cameras so I already trusted him enough to watch past the preamble. When this thread came up it was the first video that came to mind since I’d recently watched it. I can totally see how it’d be a bit less effective for someone not already familiar with his stuff.
I also agree with the general theme of your last paragraph. For the most part, I’ll avoid channels that are primarily fluff, or skip through the preamble if it’s a creator that I begrudgingly follow because they make good - if unnecessarily fluffy at times - content.
Also not sure what the long term solution is. I do tend to believe Google’s approach with this YouTube feature isn’t it.
For sake of argument, let's say that this feature causes a 20% reduction in video views.
This feature is part of YouTube Premium, meaning that YouTube is making money on it, but in its current form the creator is not. So in essence, YouTube has chosen to take the creator's content, create derivative content based on it, and make money off of that derivative content while removing some portion of the creator's revenue. In most contexts, this would be described as theft, and I think that's a fair word to use here even if I'm sure the T&C covers it somewhere.
> What's the problem? The creator can take their content somewhere else or host it themselves on their website
You don't see a problem with a move like this? Obviously creators can move elsewhere, but it's a hostile move on YouTube's part nonetheless.
Association with that brand would be very valuable.
The referenced video is from a photographer who has some pretty strong and reasonable thoughts on this - specifically the features YouTube itself is experimenting with.
Depending on the nature of the AI product, it has the potential to completely sideline creators.
Not saying that’s what Miyagi is doing and it sounds like they’re actually working with creators on this which is good. But the broader point is that such tools need to be thoughtfully implemented.
- [0] https://m.youtube.com/watch?v=ULUSS1-G3do
If Cliff’s notes were actually just AI summaries of specific books generated by an unrelated entity and presented in a way that allowed the reader to avoid purchasing the underlying content, that’d be a very different scenario.
In the linked example, YouTube is essentially doing the latter. The product launched in this thread sits in a greyer area I think, but still raises some questions about content ownership and how creators will react to these new kinds of tools and modes of consumption.
Whether or not it’s strictly legal is a different conversation than whether or not creators feel comfortable with these emerging options.
> Once its out there, they lose control of how people interact with it.
Sure. But they also have every right to choose to put it behind a paywall if new tools change the calculus that originally made publishing it publicly make sense.
And copyright problem, i think people shouldput on the plataform instead of web-scraping like
Learning languages seems a bit different in that there's more focus on repetition compared to comprehension questions, but there are certain topics (like grammar concepts) that could work well in our current structure. Also there are some really popular YouTube channels for learning any language, so we definitely see a potential to augment those videos to more accurately & effectively learn.
Streaming platforms can vary quite a lot in how they choose to distribute subtitles. I've worked with scraping subtitles from both Youtube and Netflix and I will say that these platforms distribute subtitles very differently!
And of course, if they decide to host videos on their own platform, they can just import subtitles from srt/vtt/ass which should be easily possible with some open source library.
Most of our courses are built for a slightly older age range, but we'd consider rolling out more beginner courses that are geared towards younger children and have more safety tests. If you have any thoughts let us know!
But it does seem that your platform ingests video content without the permission of the person who creates these videos? The value of your platform is driven by the people creating the videos. You say that you do revenue sharing, and that you have done 5 partnerships. But you have 400 courses, so what about the other 395?
Putting it as kindly as I can: this is ethically fraught. Really, did nobody in the room point this out? You do not come off looking like a partner here.
You need to make this opt-in, not opt-out, and specify revenue sharing terms up front. Those terms need to be generous. The people who produce video content are producing the majority of your product's value. Opt-out, of an ambiguous revenue sharing agreement, is not enough.
https://app.youlearn.ai/?via=sou