I am very interested in this, and I have personally built manual workflows to do Youtube video -> rip audio->transcript->llm context.
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?
Really cool idea! Some improvements I'd recommend with the ultimate goal being "getting users to learn the subject at hand".
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.
I was at Coursera for years and pitched this exact thing multiple times internally! So excited to finally see it being built. Congratulations on the launch!
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.
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!
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.
Please don't, or at least don't without a looooooooooot of behind-closed-doors trial and error. There are few things more off-putting then an AI try-hard "i am a quirky hooman with quirky hooman personality traits".
Can you extend this into language learning content on YT? I think that would also have amazing utility. As a biologist, so happy to see Crime Pays but Botany doesn't on here. Thanks for the awesome tool. I will be using it.
I think this is a great idea. I’ve learned so much on YouTube but it’s always been in small chunks and very task oriented. I imagine there’s a lot of content Which covers broad topics that I don’t come across.
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.
Just wanna say that this is one of those magical ideas that I'd never personally think of, but when I see it like this, it makes perfect sense! So cool.
Poker is interesting. I think these videos do work in our current course generation process. However, I do think some subjects like poker need custom tooling around the course to really make the learning experience great. For example, access to solvers or actually playing a hand on a table is a part of the course experience as well. Chess is another one that falls in this special bucket imo. Some of this tooling is on the roadmap!
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)?
The article states that Anaina and Burke separately conducted their tests, but social robots [1] have been shown to be effective in individual tutoring. Human tutoring is not always better than a well-designed computer program [2]. There have been issues with how studies interpret their effect on group size / scalability [3].
I work in edtech and one of my teams is content creation, so pretty excited about this space but also very aware of the challenges and massive amounts of hype and over promise / under deliver. To assess I tried to generate a short (< 10m), one-video course from a YT video I've previously watched on a topic I'm an "expert" - after an hour all I see is the embedded video, the transcript and "generating content" dialog.
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.
Lol yea not just poaching, we do revenue sharing (signed deals with a bunch of top creators). They get the majority of all revenue from courses.
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?
Yep, if anyone didn't want their videos to be on our site, we would take it down.
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.
> I don't think that's true? We're embedding the videos, which is allowed.
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.
> Yep, if anyone didn't want their videos to be on our site, we would take it down.
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.
It’s an 8 minute video…and even shorter at 1.5X that will take me longer to summarize than you to watch.
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.
The videos are the intellectual property of the creator, and YouTube has the rights to distribute and make money off of it for hosting it for you to billions of users. What's the problem? The creator can take their content somewhere else or host it themselves on their website
The people who spend hundreds of hours carefully creating content for their viewers care [0].
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.
They put their videos out for public consumption. Not behind a paywall. Once its out there, they lose control of how people interact with it. Should cliff notes and other study guides be banned or regulated?
I don’t find Cliff’s notes to be similar at all. They represent standalone short-form content written by authors that is a purchasable option alongside more in-depth options written by other (or at times the same) authors.
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.
Amazing approach! Is learning a second language too different from the types of courses Miyagi was designed for, or do you see a potential for that category?
Thanks! Definitely some potential, we actually built a language learning tool for a few days early on (but decided that it was too crowded of a space to start in).
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.
I was actually thinking about building this because I watch a lot of YT videos in other languages (best way to do travel research is to search the destination using the local name and getting local videos).
I don’t think they’re attached to YouTube too rigidly. (Well, I hope at least.) In theory this should work with any platform that provides subtitles. But I think if YouTube falls, or blocks their API access, they would just start hosting the videos themselves.
Yea YouTube is only one format we support. Users can also upload pdfs, mp4, docx, pptx, etc. And we already do support video hosting ourselves. It wouldn't be great if YouTube decided to part ways with us, but we'd be just fine.
For official courses, we go over the generated course with the creator to vet the content. Generally they're pretty impressed but have a few things they'd like to change/add before publishing.
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.
Is the course creator being impressed the most important metric? Are there other more concrete metrics you are able to use to determine quality from the perspective of a student?
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)
At this stage it seems like a good metric since it's the creator's content. We're adding a feature for question feedback from users, so they can like/dislike/report questions, but very open to other metrics if you have any ideas.
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.
- 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.
Yes for the first, technically yes for the second? The user can go in and change the content as well (i.e. if it's a teacher generating a course for students). But not sure what other human supervision + quality control methods you're referring to that we could implement.
Users can upload slides (ie. docx or pptx) and create a course from them - give it a try! For videos, we don't currently process any frames from the video just the transcript, but this is on the roadmap.
it's a natural extension after you use the LLM to generate the content. We do these in our content creation - and I assume learners use an LLM to answer them :)
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.
I am very interested in this, and I have personally built manual workflows to do Youtube video -> rip audio->transcript->llm context.
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?
Really cool idea! Some improvements I'd recommend with the ultimate goal being "getting users to learn the subject at hand".
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.
I was at Coursera for years and pitched this exact thing multiple times internally! So excited to finally see it being built. Congratulations on the launch!
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.
This is really cool!
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!
We'll reach out and hopefully add some courses! Thanks.
This is a fun concept, and I love the name!
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.
Thanks! We do have multiple choice questions now (agreed) but some of the older courses were generated when there were only short answer.
Anything specific we could improve about talking to the tutor? Definitely will add some of those features and gamify better.
Maybe give the tutor some personality or persona (having it speak as the instructor). I’m probably off base with that suggestion, though.
Again, very cool idea. I'm going to try some of the nuclear courses later this week.
Best of luck!
Please don't, or at least don't without a looooooooooot of behind-closed-doors trial and error. There are few things more off-putting then an AI try-hard "i am a quirky hooman with quirky hooman personality traits".
Can you extend this into language learning content on YT? I think that would also have amazing utility. As a biologist, so happy to see Crime Pays but Botany doesn't on here. Thanks for the awesome tool. I will be using it.
Yes! I did something similar with daily exercises at https://app.fluentsubs.com/exercises/daily
I think this is a great idea. I’ve learned so much on YouTube but it’s always been in small chunks and very task oriented. I imagine there’s a lot of content Which covers broad topics that I don’t come across.
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.
Just wanna say that this is one of those magical ideas that I'd never personally think of, but when I see it like this, it makes perfect sense! So cool.
Poker, specifically Texas No Limit Hold'em, is widely taught on Youtube.
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
Poker is interesting. I think these videos do work in our current course generation process. However, I do think some subjects like poker need custom tooling around the course to really make the learning experience great. For example, access to solvers or actually playing a hand on a table is a part of the course experience as well. Chess is another one that falls in this special bucket imo. Some of this tooling is on the roadmap!
Thanks, we'll reach out. We have a poker course from MIT (https://miyagilabs.ai/course/mit15s50) but yea these seem more practical & engaging.
For anyone else interested in Bloom's 2-sigma, here's the original paper (1984): https://web.mit.edu/5.95/readings/bloom-two-sigma.pdf
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)?
The article states that Anaina and Burke separately conducted their tests, but social robots [1] have been shown to be effective in individual tutoring. Human tutoring is not always better than a well-designed computer program [2]. There have been issues with how studies interpret their effect on group size / scalability [3].
[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...
Would be nice ie to see this product with focus on elementary school age content.
I work in edtech and one of my teams is content creation, so pretty excited about this space but also very aware of the challenges and massive amounts of hype and over promise / under deliver. To assess I tried to generate a short (< 10m), one-video course from a YT video I've previously watched on a topic I'm an "expert" - after an hour all I see is the embedded video, the transcript and "generating content" dialog.
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.
Sorry we're running into some rate limits with course generation but will be fixed soon. Valid points—will respond in a bit.
Does it work on YouTube videos that have transcripts disabled?
No, it won't work for those.
You probably know this, but gemini is great to generate transcripts. I did a quick browser extension for that: https://github.com/za01br/yt-subtitle-extension
Congrats on the launch!
Do you have any revenue sharing program with the content creators? Or are you just poaching them?
Lol yea not just poaching, we do revenue sharing (signed deals with a bunch of top creators). They get the majority of all revenue from courses.
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.
They probably haven't decided yet
That would mean poaching lol
Are the people that create the content okay with this?
Yes. Any content that we monetize we are revenue sharing with the creator. We already have more than 5 partnerships with creators.
Do creators have the option to opt out?
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
Yep, if anyone didn't want their videos to be on our site, we would take it down.
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.
I do think the fact that your product is likely to drive views makes this less of a concern than what YT is doing.
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.
[flagged]
I don't think that's true? We're embedding the videos, which is allowed.
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.
> I don't think that's true? We're embedding the videos, which is allowed.
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.
> Yep, if anyone didn't want their videos to be on our site, we would take it down.
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.
Not to be ironic, but... is there a summary of that video? It's a bit long and he doesn't seem to get to the point for quite a while.
It’s an 8 minute video…and even shorter at 1.5X that will take me longer to summarize than you to watch.
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.
Thanks, I appreciate that!
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.
The videos are the intellectual property of the creator, and YouTube has the rights to distribute and make money off of it for hosting it for you to billions of users. What's the problem? The creator can take their content somewhere else or host it themselves on their website
3b1b is a monetized partner?
Association with that brand would be very valuable.
Not yet! We don't monetize his content (it's not behind a paywall). But we are talking with him :)
Who cares?
The people who spend hundreds of hours carefully creating content for their viewers care [0].
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
They put their videos out for public consumption. Not behind a paywall. Once its out there, they lose control of how people interact with it. Should cliff notes and other study guides be banned or regulated?
I don’t find Cliff’s notes to be similar at all. They represent standalone short-form content written by authors that is a purchasable option alongside more in-depth options written by other (or at times the same) authors.
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.
Marketers, among others
Amazing approach! Is learning a second language too different from the types of courses Miyagi was designed for, or do you see a potential for that category?
Thanks! Definitely some potential, we actually built a language learning tool for a few days early on (but decided that it was too crowded of a space to start in).
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.
I was actually thinking about building this because I watch a lot of YT videos in other languages (best way to do travel research is to search the destination using the local name and getting local videos).
How worried are you about platform risk?
I don’t think they’re attached to YouTube too rigidly. (Well, I hope at least.) In theory this should work with any platform that provides subtitles. But I think if YouTube falls, or blocks their API access, they would just start hosting the videos themselves.
Yea YouTube is only one format we support. Users can also upload pdfs, mp4, docx, pptx, etc. And we already do support video hosting ourselves. It wouldn't be great if YouTube decided to part ways with us, but we'd be just fine.
How do you validate you’re not generating garbage, and thus teaching people nonsense?
For official courses, we go over the generated course with the creator to vet the content. Generally they're pretty impressed but have a few things they'd like to change/add before publishing.
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.
Is the course creator being impressed the most important metric? Are there other more concrete metrics you are able to use to determine quality from the perspective of a student?
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)
At this stage it seems like a good metric since it's the creator's content. We're adding a feature for question feedback from users, so they can like/dislike/report questions, but very open to other metrics if you have any ideas.
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.
So in less promotional words:
- 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?
Yes for the first, technically yes for the second? The user can go in and change the content as well (i.e. if it's a teacher generating a course for students). But not sure what other human supervision + quality control methods you're referring to that we could implement.
Neat idea! Do you do anything with the video itself? Understand the visual content or extract details from slides?
Users can upload slides (ie. docx or pptx) and create a course from them - give it a try! For videos, we don't currently process any frames from the video just the transcript, but this is on the roadmap.
Does the list of resources simply need to be a list of links to video objects?
They can be links or actual files such as mp4 videos.
Great idea! Automated quiz generation seems like a nice use case for LLMs.
it's a natural extension after you use the LLM to generate the content. We do these in our content creation - and I assume learners use an LLM to answer them :)
Congrats on launch
The tech looks cool.
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.
See: https://news.ycombinator.com/item?id=43973153
Nice work! Really cool.
Great concept!
Good Stuff!
This is awesome! Congrats Tyrone & Guang!