Leveraging YouTube Resources with AI

Introduction

The scale of YouTube and the curation challenge

about the estimated number of videos on YouTube. Google doesn't publish these numbers. It's currently estimated to be about 15 billion.

And a couple of cool quotes about the point I've just made that YouTube doesn't curate. They let you do that for them.

Our solution: Playlist Builder

So we've built a platform to do that.

Product overview and demo setup

On the left side, you can see what it looks like. And I won't spend time with that because I'm about to demo it.

User experience and business model

It's an intuitive ad-free interface.

Our business model is based on subscriptions. That's including to play nice with Google.

So we're not trying to advertise on top of YouTube ads. We basically feed traffic back to YouTube.

Key capabilities and roadmap

We've got a steadily expanding set of search features.

We have an MVP now, which is free. You can try it out at playlistbuilder.ai.

And we will be introducing a paid version with additional features.

Technology and IP

The data pipeline which gets you the curated YouTube videos is pretty clever. We've got a patent pending for that, and we'll see where that goes.

About Learnershape

Just a few words about the company behind this. It's called Learnershape. We've been around for a little while.

We pivoted a couple of times and our current focus is on Playlist Builder.

Open-source contributions and customers

We also have some open source tech, which is at this GitHub link you can see here. And we had some good, basically it's open source for taxonomy independent content recommendation.

We've had some good customers for that, which you can see on the right.

Pivot to a product-first strategy

And for the last of those Northeastern University listed at the bottom, we designed Playlist Builder and got some great feedback for it and basically decided to direct the company primarily towards Playlist Builder rather than doing projects, which were relatively high priced, but individual long sales. Now we're focused on a high volume product.

Live demo

So I'm going to switch to Playlist Builder.

Signing in and workspaces

and this is our landing page which is on squarespace then you get here which is the app and we at the moment require sign in with google because it links with youtube so to be able to use the youtube related features you need to sign in with google we've considered a free version a login free version but for the moment we don't see the sign-in as too big a barrier so i'm signing in with uh that learner shape named account it's a brand account which means everybody on your team can have access to it

This is what you get when you've done a few searches in Playlist Builder. We work on the concept of workspaces, and these are each of those that you, each rectangle that you see is a workspace.

Creating a new workspace

We'll create a new workspace. I, let's see, I usually ask for somebody to choose something, but let's see, we will do,

wildlife of Sussex. I was just out walking in the Sussex countryside today. So Sussex is an English county south of London. It has very interesting flora and fauna.

And we'll put in Sussex Wildlife, nature.

We can choose the video length. I'm not gonna restrict that. We have some options on how many videos we return. I'll choose 50. And then we create the workspace. It should take about a minute to do that.

Exploring prior results

While we're doing that, I'll show you what the results look like.

Flat earth theories is what the audience chose at the meetup last week in London and, The Included tab, you'll see how that works, is some of the early ones where it came up with some good things on flat Earth theories, which this was the description I typed in during the demo.

The Earth is round, but some people think it's flat. Learn about it here. And it pulled up these videos.

And there's a lot more why the people think the Earth is flat. Is the Earth actually flat? Maybe it is.

In any event, Let's see how we've done.

Refining results and building a playlist

If we go back to my workspaces, now we have 11. We have Wildlife of Sussex, rock pooling with the Sussex Wildlife Trust. Seems like a pretty good one.

The search isn't... isn't perfect, so Sussex dialect doesn't sound great, so I'm going to exclude that one.

Henfield, so we'll look at one village there and maybe this one. So we've included three videos and then we can either here by clicking this or go to the included tab, which is the videos that we like, we can create a playlist of these videos.

Ad-free viewing and video details

I should circle back and show you that as you're doing this, you can watch the video. You get ad-free watching in this little window, YouTube allows that. And you can see the description of the video without having to click through to it, some stats about it.

How selection works

And you see the selection criteria we used. We use a mix of AI, not generative AI, it's a convolutional neural network with some statistical methods to choose the videos.

Publishing playlists to YouTube

So now we're ready to create a playlist. We can have the title and description be anything, but I'm just going to use this button to copy the same information.

And then we can create a video with any of the three statuses that are possible in YouTube, fully public, unlisted, which means it's public, but you can't search it, or private. I'm going to create an unlisted video.

That's very quick. It takes about a second or two.

Once we click on the video, now I'm back on YouTube. I'm no longer on the Learner Shaped site.

And so we can now use all the YouTube tools to do things like move the videos around, or we can add other videos from YouTube.

Combining searches and iterating

We can also go back to the Playlist Builder site, choose a different set of videos, and create a new playlist. And we can combine other searches on the Playlist Builder site to create playlists that combine multiple searches.

Product status and use cases

That is our advanced MVP product.

And it's really probably useful for a wide variety of applications. But we think there's a lot of potential for it to be used in a wide variety of applications.

Target markets

We're targeting educational applications, particularly universities and learning and development organizations of small to medium-sized companies.

Q&A

So I will stop there and take any questions.

Not sure Joe has come back in, but I can, as Alexandra did, I'll read from the chat. Kumsal has said, interesting service, less about the technical part. That was a question for Alexandra.

I'm sorry, yeah. No worries.

I don't know it. So I think we'll give people another second or so to work out some questions for you.

Ad-free playback: how it works

I have a question, which is, how did you manage to get the videos ad-free on Playlist Builder? There's no particular trick to that.

The window that displays the videos is a standard YouTube widget. And YouTube allows those widgets to display videos ad-free.

So if you click through from Playlist Builder to the video on YouTube, you'll get ads if you don't have YouTube Premium or you don't get ads if you pay for YouTube Premium. It's a standard thing that Google makes available for thumbnails.

Very, very cool. Very cool.

Okay. Any more questions about Playlist Builder in the chat? Okay. I've got one more, which is like...

Transcripts and deeper relevance

It's less of a question, but do you see any role for like transcriptions and scripts from these videos and what applications you might think about using for that? Yes, is the short answer. The longer version of it is that's probably the most common question we get.

1The biggest aha moment for me building this is when our pipeline involves the standard YouTube API and we hit it multiple times. And then we get back from the API, the full, you can't, nobody is allowed aside from Google to search the full metadata. But once you get the videos back, you can do anything you want with the, you get the metadata for the videos.

You can, watch the video, you can create a transcript. We've definitely thought about that, and our search relevance, as you could see from the demo, is pretty good, but we think we could get it even, you know, really pinpoint relevance if we watch the whole video with an LLM.

Cost and feasibility considerations

The trouble with that is it's more expensive in terms of compute. You know, what you've just seen, it runs a convolutional neural network, but that's pre-trained, so the computer's pretty cheap. To run an LLM on everyone would be more expensive. So we haven't done that yet, but as token prices come down, we think it's something we'll consider seriously.

Yeah, that's real. That's something indicative of a lot of people are experiencing in this space.

Community feedback and MVP access

You haven't got any more questions to the audience, but you've got a new participant on the platform who's created their own playlist. Well, that's a big reason why we do these because we're in MVP mode now.

You can, as David who said that, got in for free as anybody can and do a number of searches. So please, we're looking for people to sign in, try it out.

I'd love to hear feedback from you. I particularly like negative feedback because we're trying to figure out what's wrong with it. Positive feedback makes me feel really good, but negative feedback tells me where to go. So negative or constructive rather than negative.

People say, thank you, David, for saying it's great, by the way. But thank you very much.

Stay in touch

I'm on LinkedIn.

Conclusion

Thank you very much, Murray.

I think we'll probably move on to our next speaker now, John, if I can see a back return to the listeners.

Finished reading?