Show HN: Anki/Duolingo-like app using educational YouTube videos (platoedu.org)

289 points by kirill5pol ↗ HN
Hi HN,

I watch A LOT of educational YouTube videos but wasn't forgetting a good chunk of the details because I was only really passively watching. So I made a tool that generates quiz questions/flashcards from YouTube videos, and uses spaced repetition like Anki or Duolingo to keep it in memory.

Let me know if you find it cool/useful (or terrible ) or if you want to know a bit about the details!

101 comments

[ 3.6 ms ] story [ 150 ms ] thread
(comment deleted)
The 30 minute limit is unfortunate, but otherwise it looks good. Thanks for sharing!
So far it’s basically a heuristic for the video having both captions and the token count for generation be reliably under the limit, but I am working on making it work for arbitrary length videos! I did some tests for 2-3 hour podcasts and it worked pretty well
Who chooses the videos? It seems ... opinionated.

For example, under 'Physics', we have "The big lie about carbon capture', 'Why (toilet) flushing isn't for everyone', 'The scientific basis for miracles', etc.

Yeah it’s not ideal yet the suggestions under each category I just YouTube’s own category labels so often it didn’t give the best results, but it’s something I’m working on!
Perhaps log the videos that users upload and suggest the most common ones?
Yep planning on doing that but didn't have enough videos uploaded to "fill up" the categories before posting the Show HN, so I just scripted some stuff based on the YouTube suggestions from some YouTube channels I watch.

One of my future plans is to actually train a BERT model to limit the video suggestions to something that actually is useful instead of clickbait... I have 2 different accounts on YouTube just so watching random videos on 1 won't pollute the suggestions from the other

Perhaps allow the user to integrate with his YouTube history? Does YouTube have an API for that?
When I saw Anki/Duolingo in the bio I assumed it was for language learning, but this is a great idea!

I too often watch these kinds of videos without really retaining a lot. This is a perfect complement to turn infotainment into time well spent, or at least, less wasted.

Same! Maybe someone could do some language learning on top of all the videos we consume on YouTube?
Languagereactor.com seems to implement some form of that
Yes!! I definitely spend a little too much time on YouTube myself, would you mind sharing what kind of content (categories) you watch in a pm? I’d love to get some better ideas of what kind of material to tune the question generations on
Nice! Would be great to use on longer videos and focus on specific topics of the video.
I love this new Plato app, it’s useful and I educational. The interface is also very clean
What algorithm do you use for spaced repetition?
I have a fairly simple implementation of the SM-2 algorithm, and just making the assumption that if you answered the question correctly then you have “perfect recall”. This isn’t exactly correct but I have been using it myself and seems to still be pretty nice.

But for the next version I want to use something called knowledge tracing to determine an estimated level of recall to then change the spacing

You just saved millions of students life's, A great tool that just solved a problem that existed but no one ever noticed
I want to make flash card form everything I look up this way.
Any particular type of content you’re thinking of? I’m currently working on adding podcasts since they’re pretty similar to YouTube videos, but I’m sure with some tuning I could see if it works for other things
Podcast is definitely one. But I’m not sure how one could get the real salient points out of some of the philosophy podcast I listen to without listening carefully the first time and writing them down.. Another would be for example when I ask Chatgpt how to do something in nvim.
Mind sending me an pm? (my email's in my bio) I'd love to try it out with some of the podcasts/nvim chatgpt responses you mentioned and add it to your account!

  > Any particular type of content you’re thinking of?
When the wife is mad and ranting, can you email me a summary at the end? With a quiz for the important parts.
Now that is a strong need haha, I think will have to be on the Plato premium + plan
Oh god. I did take notes during my last break up.. Maybe I could review those too
I like the idea, but seeing how many new accounts added fake comments makes me a bit suspicious :)
I really dont like the new accounts stigma due to statistical reasons but to each their own!
Is NLU really this reliable yet? I tried making some scripts like this with LLMs, and it seemed to do very poorly. So, I abandoned the effort.
I did quite a bit of that too, I really had trouble getting it to generate good content from scratch but here it's using the transcripts directly.

I'm guessing it only really works well on scripts that are meant to be educational, because there already are "questions" implicit in the transcript of the video because that's the best way to present information when teaching something

This is amazing! I tried a couple of videos and the questions seem pretty relevant and answerable (is there a better word for how a question is worded clearly and the provided answers seem clearly distinct and one of them is obviously correct), which is really hard to do by hand, much less by AI.

I know you've addressed the video selection in the playlists, but I would highly suggest doing something to get it to differentiate "educational entertainment" videos (I notice a lot of Real Engineering and Economics Explained and CGP Grey videos) and actual education videos: primary-source explainers from teachers and subject-matter experts. The information density in the latter is way higher, and I think people overestimate the educational value of the former.

> and I think people overestimate the educational value of the former.

I disagree. Some people might overestimate how in-depth the information is, but the educational value of these videos lies in giving someone a basic understanding of something they otherwise wouldn't have learnt about at all. The lower information density helps making the video easier to understand and thus easier to consume, compared to something like a Havard class.

If you want to learn something in-depth, an actual class, a book, etc. will of course always be better, but if that's neither required nor wanted, the infotainment is just fine.

Yeah, I feel it's not ideal, but at least it's better than doom scrolling, and especially if it's a field that you're not familiar with having some simple explanations is still useful as a starting point
One point is if the end-user understand it is just infotainment and not a concrete guide.For example,People became "experts" on covid thorugh some videos that spread misinformation and became hardcore fanatics.This is just one example of many.
"educational entertainment" videos are way too many, way too popular and binge-able - much more recommended by YT's AI. Actual education are much harder to discover on YouTube.

I have been wanting to build a YT front-end that lets me control how many "new" videos are recommended. New videos are the time-sinks.

Instead this new FE should make me re-watch so I absorb and retain better - maybe thru more Q&A like OP's platoedu or even make me write out some notes. Then I am forced to curate videos and maybe be more productive.

This is a problem I've had too, my current solution is to have multiple profiles on YouTube so whenever I click on a random video one it doesn't pollute my other education heavy account. Also just removing videos helps... but even then YouTube still pushes edu-tainment over harder educational videos.

One of my ideas that's on the backburner is build a BERT classifier to separate between Educational, edu-tainment, and random, then use that to filter suggestions from the ones of people that use Plato

Anyway if you have any good suggestions for better educational content I'd love to add that to Plato over the categories I have now!

Multiple YT profiles is a smart hack! And also, great work on the app!

The YT algo is pretty good - it catches on to what I want to follow and magnifies (ie suggest more content on) that topic. But it never pushes me to educational videos.

I suspect educational videos are best to watch on Coursera. I know people who just open up Coursera and start listening on commutes, etc - instead of infi-scrolling.

The pedagogical (instruction techniques, content structure, etc) aspect in those vids is different. I wonder if there is inspiration for creators/topics from Coursera?

This. I don't think most educational content on YouTube is worth remembering (or the best way to spend your time in the first place).

So I'd be cautious about an app that helps you memorize the contents of said videos. You might end up with a lot of superficial, clickbaity pieces of knowledge.

I invite you to share your own superior knowledge to the masses via your own YouTube videos so we can learn from you. Until then, I’ll learn from what is made available for others. Post back here once you’ve created some better content so we know where to look.
Papers, textbooks, tech talks, university lectures.

That's where you'll find actual knowledge and not in high production value videos which have to be financially viable for their creators.

It's hardly a secret that Youtube has a problem funding long form videos with a certain depth and instead favors clickbaity, short material. No reason to be offended.

As a rule of thumb I'd say everything with a sponsored segment is entertainment but too shallow for education.

> Papers, textbooks, tech talks, university lectures

Perfect list. Tech talks, university lectures (recorded videos) are almost as consumable as YT edu-tainment videos. Papers, books and textbooks are accessible but requires more motivation.

To the parent comment (zadokshi), if YT content is education, why don't the biggest creators make 5-10 videos on a topic, back-to-back? 5-10 is minimum for learning, example Coursera content - I'm not even comparing to semester/yearlong coursework at schools. Because there isn't a demand or incentive for that on YT.

Thank you!!

Yes, this is actually something I've been thinking about quite a bit, I actually built out the playlist feature just this morning because it's easier to "show" how Plato works, but I basically just wrote some scripts to get some good enough videos for the demo

If you have any good channel suggestions I'd love to add them :))

One of the things I have on the backburner for now is building a BERT classifier to decide whether the video is Educational, Edu-tainment, or not educational at all and have a more customizable video suggestion than YouTube has (I actually have 2 accounts on YouTube, just so I can watch some random video on 1 without it polluting my education/learning heavy one)

One thing though, is I actually think both have their merit, while I agree the actual educational content is pretty different, the educational entertainment is a nice alternative to TikTok or IG reels when you just want to mindlessly scroll, I think there still often some useful content there, especially if you don't have any background in the area

> classifier to decide whether the video is Educational, Edu-tainment, or not educational at all

I'm a bit surprised to only see like 3 questions for a 14 minute video of quantum mechanics. For educational videos with very dense information, is there a way to raise the questions per video rate?

Looking forward to see MIT OpenCourseware videos supported. Right now they are too long :D

Working on that soon! That's actually also one of the reasons I'm limiting to 30 min, 5 questions definitely isn't enough for 1 hour video but I have some fixes in the work for it!

Pm me (email in bio) and I can add some MIT OCW videos and turn on support for longer videos to your account

Amazing, thank you. It even works great for music videos, I never appreciated the poetic lyrical context of Pantera before. ))

Small bug, the service requires a youtube.com URL and cannot handle an m.youtube.com URL, as happens when copying from a phone web browser or NewPipe. Perhaps you could support the mobile URL as well.

Thanks, great work!

Well that music video lyrics work is a very unexpected pleasant surprise! :)

Thanks for catching that! Will fix that!

really cool. A while back I've build this database of 1000+ hand-selected educational YouTube videos, so I'm going to try out a few of them to put in this tool :) https://www.edutube.app/
This is awesome! Did you hand select all 1000? I wanted to hand select for the categories to get some better starting recommendations but it was taking too long so I was a bit lazy and just scripted it…
Yeah I literally hand-selected all of them to remain the quality of videos that I wanted
If you have a list of the IDs of the videos (and maybe categories) I can bulk add them! If you want to discuss a collab of some sort feel free to email me!
If this works as I expect it to, it'll be something I've been hoping to see for a long time. Thank your for sharing it with us!
Do you have any specific type of content that you were trying to learn? Right now it's still pretty early/demo stage, but please pm me (email in bio), I'll see if I can tune it better to your use case!
How do you deal with YouTube Terms of Service for extracting transcripts?
International copyright agreements, like the Bern Convention, allow usage of content for educational purposes, as long as you're not replacing the original.
(comment deleted)
This is fantastic! Is there a way to donate? This is the kind of software I want to see in the world!
Pm me (email in bio) and I can come up with some sort of premium plan haha (I'm thinking unlimited time length for videos + podcasts/other material) but very open to suggestions!
I've always wanted to do this for Wikipedia, it could even be a Wikimedia add-on.

However, I have recently transitioned towards becoming better at compiling information quickly rather than spending a chunk of my day memorizing facts that I am not quite sure will be useful.

What do you use to compile information and do you keep track of sources?
It depends on the question. I do not have a list of sources.
Do you trust the information even after forgetting the source and not having the context your wrote it in?
When I say I compile information I mean I have a scratchpad in which I fill in the evidence that is relevant to my current needs. So I'll have the source noted there.

What I'm trying to say is that I rarely go back to the scratchpad of a previous question / problem, but I will approach every question and problem with a new search.

Trying to save information for the sake of it without knowing what future use it is going to have stresses me out, so I find that this fresh scratchpad approach works best for me.

I’ve been doing something similar. If I read a blog post / paper, etc. where I learn a lot on a topic I’m interested in, I will catalogue a pdf of it in Obsidian with a tag and an optional note. This makes it easy to access information locally very quickly and I find I learn a lot more because if I forget something, I open up the resource, read the doc and, come out learning a little more. A kind of convoluted version of spaced-repetition but more passive learning.

Granted, I’m aware this probably won’t scale to many topics but a few years and hundreds of notes later, it’s still working well for me.

I don't even save the sources, I'll just assume I'll always have access to the internet and that I'll be able to find what I need when I'm thinking about it.
so how does this worK? you take the video and extract captions from it? and feed it to GPT? i am sorry, can you clarify?
Pretty much! I'm working on doing some fancier stuff with knowledge tracing but was out of scope for the mvp
This is very cool! I'm trying to do something technically similar by using LLMs to summarize the meeting transcript from youtube (https://parths-newsletter-78dbcb.beehiiv.com/).

Right now I'm doing this manually by copy/pasting into ChatGPT but I want to automate this aspect. I'm not very technical so any guidance you could provide would be helpful :)

Pm me (email in bio) I can send you some scripts and point you in the right direction!
I like the idea. I’m learning European Portuguese, so I added a video but unfortunately got an error that there were no subtitles. The subtitles for the video in question are auto-generated, so maybe that’s the reason? Would be great if I could use this for foreign language studies.
Auto generated captions do work but unfortunately only English for now, it would be pretty easy to add other languages but I just haven’t had time to implement it yet. Feel free to pm me with the YouTube video I’ll try to see what’s the problem (email in bio)
What exactly are you looking to do with YouTube and language? I might have a github repo you could be interested in. It takes YouTube videos, creates transcripts, translations and creates audio anki flash cards from it with audio on the front and text on the back
This would be interesting to try out if it's a public repo!`
it is, ill send a link to your email in your profile