Rohan Nayak: How Pocket FM is Using AI to Reinvent Audio Storytelling
48:49 Share

Bye.

Hey, everyone, and welcome to Generative Now. I am Michael Magnano, and I'm a partner at Lightspeed. The entertainment industry has a billion-dollar question. What makes a story a blockbuster? That's the big question that Rohan Nayak is tackling as the CEO and co-founder of Pocket FM, the global audio fiction powerhouse from Bangalore, India. Pocket FM is using AI to predict blockbusters, localize stories, and act as a co-pilot for writers.

They have 50,000 shows produced with AI and more than 100 billion minutes being streamed to listeners each month. Pocket FM is positioning themselves to 10x the growth on even more formats like web comics and novels. This was a fascinating conversation that really showcases how AI can be leveraged to turbocharge content platforms, and I think you'll enjoy it. So check out this conversation with Rohan Nayak, co-founder and CEO of Pocket FM. ♪

Hey, Rohan. Hey, Mike. How's it going? It's going well. Thanks for having me on this podcast. Yeah, thanks for doing it. I think we're on like opposite ends of the world right now. So great to see you. What time is it there in Bangalore?

It's 8:30 p.m. Okay. Doing this in the morning for me. Yeah, but I walk across two time zones, as you know, so it's not too late for me. Yeah. So wait, so you have teams. We were talking a little bit about this before we hit the record button. So you obviously have teams in Bangalore and then also in L.A. Is that right? Yeah, L.A. and Bangalore. So I keep shuttling between the two. Not a great commute.

but how long is the commute 20 24 hours 24 it's just like a full day yeah i mean i think you just have to figure out what to do on on the plane right so you that that sort of that's what i'm trying to figure out i mean i guess you sleep for part of it what have you realized is that sort of time when you're not connected to the internet i sort of do a lot of my creative thinking and i've sort of realized those flights help me up a bit to you know zoom out

Because when you're plugged in, you just, you know, just messages, Slack, emails. But when you're on the fly, there's no messages. So you can just, you know, zoom out, think about things in a more creative way. It's a fun time for me nowadays. Yeah, it's good for like doing your deep strategy work, probably. For sure. Well, so obviously, you know, Pocket FM being a light speed portfolio company, you know, I'm super familiar with the business and it's been fascinating to watch.

Congrats on all the progress. But for listeners and viewers just tuning in who haven't heard of Pocket FM, I think it'd be great to start with the story. Give us your background and maybe lead into how and why you started Pocket FM and what you guys are doing today. Sure, Mike.

So I've been sort of building content platforms for the last 10, 15 years. You could say I'm just obsessed with it. You know, I consume all forms of entertainment and content from manga, anime, you know, of course, movies and novels. So sort of like I've been building content platforms for a very long period of time, and I just wanted to build something here. And the philosophy for me has been that

Can I build a new platform which brings new artists to life? Because every content platform which gets developed does bring new creators to life in a way. And then that has been sort of my personal vision, so to say. I was working in a startup and I was commuting three hours to work.

And at that point of time, I used to think what am I doing with my life? It's three hours of, you know, three hours every day. But it turns out it does, it did have some value. And, you know, so I was just, I mean, I was consistently thinking, what can I do in content? What more can be done? And here I was, three hour commute. And I'm used to long commutes, as you know, right? And I realized I was getting bored a lot.

Because three hours a day and 15 hours a week, I tried listening to a lot of audiobooks, podcasts, video content,

But, you know, it was, I couldn't consume video content, which was harder, of course, while commuting and audiobooks and podcasts were good when I wanted to learn something new or I wanted to find out some about some topic. But, you know, being an entertainment sort of enthusiast, I wanted entertainment and audio. So, and I was getting bored and then I, you know, I tried everything out. I just couldn't find anything.

And I think then that's when it struck me, the mental model I had for video entertainment was there's a spectrum of entertainment products. There's long-form entertainment, which is your video streaming, Netflix, Disney. And then you have short-form sort of entertainment, which is of course TikTok, Instagram, and so on. And then in between you have Twitch, you have YouTube. There's a spectrum of entertainment, right?

But I think it's about the spectrum of short form to long form. In audio, nothing really exists, right? Music is a different form of entertainment than there's podcasts and audiobooks, which are not necessarily entertaining. It's more infotainment, if that makes sense. Yeah, information. Information, right? And it just...

And from a first-principle thinking point of view, I just couldn't understand why something, let's say like a Netflix for audio or TikTok for audio on both opposite ends of the spectrum doesn't exist. And as a user, I wanted audio entertainment. And it is still strange till date that why something like this never existed in the first place.

But so that built conviction and that if I as a user is facing such a problem, maybe something like this could exist. Being passionate about stories and entertainment,

I sort of, you know, took a plunge there. Let's figure this out. I have no idea what this is going to be. It's a part of a blank. I have no idea. It's a blank canvas, literally, right? Let's just figure this out. So that's how we, you know, we started Rocket. And did you feel like...

that need was not being met by, say, you know, more narrative form podcasts. You know, obviously, you know, in the podcast catalog on Spotify and YouTube and Apple, there are plenty of, you know, fiction or nonfiction, more like narrative-based podcasts.

Did you, so like, was that not meeting the need for you? Yeah, I mean, I tried a few. So I feel that narrative podcasts are still a very small subset of the entire podcast world. I mean, if you unbundle podcasts, right, fiction and entertainment is not a big sort of category. I mean, if you look at the top 100 podcasts in the US, you would find a few, but it's still dominated by information that podcasts and I believe that

you know, entertainment and audio needed a different approach, which is a bit different from like how a typical podcast sort of works out, how it's written, how it's structured. And you need a different platform of thought because entertainment as a whole is huge. It's not a subset of podcasts, right? So that's what the thought process here was. So you needed a different content approach. You also just mentioned you needed a different platform. Was that because you felt like

there were maybe at the time, or maybe you eventually, you know, learn this and doing it, but did you feel that there was a product gap that was missing as well? Like, oh, it's not enough to just listen to this audio, close my phone, put it away in my pocket that you actually needed. There needed to be product differentiation. Did you feel like that was a gap also? Yeah, for sure. I mean, you know, over time what I've realized is that

Building an audio platform is very hard. You have to solve some very fundamental problems like from content discovery, which is harder in audio than video. I mean, in audio in background, you're not exploring the app. Because you're not looking at the screen. You're not foregrounding the experience. Exactly. The foreground experience is not there. So how do you build a discovery? And that has been a challenge for any audio format, even be it music.

And so it needed a different discovery mechanism, which is different from, let's say, how podcasts are discovered today. And then monetization, right? I mean, I believe that we have cracked monetization in audio and specifically in audio entertainment. It needed a different approach to monetization, which was just different from how podcasts were monetized. That totally makes sense why you would need product to solve content.

content discovery. As you know, I also started an audio company and content discovery was a never ending challenge. And fortunately, after we sold to Spotify, we were able to inherit a lot of their content discovery tools. But massive, massive ML challenge.

What about on the content side? What was it about the content that you felt needed to evolve beyond the traditional sort of like narrative podcast structure? Sure. So, you know, when we started building an audience in the platform, we realized we had to find a format. I mean, any new entertainment category needs a new format. For instance, for TikTok, it was like nine by 16.

shot vertical videos for Instagram, the square photos and so on. And, you know, it took us two years. Of course, Lightspeed backed us, you know, right at the beginning. And, you know, we've used to do like, we went through like 10 plus pivots

where the idea was what's the right format for audio entertainment. And over time we realized that audio is a long form entertainment product. It's very hard to make it short form. Like you won't listen to an audio for 15 seconds, right? I mean, because your use cases are longer and you would prefer a long form entertainment use case. And then, um,

We also realized users wanted bite-sized episodes in the sense that it's sort of like 10 minutes episode instead of like a one hour episode episode.

for fiction because you can just consume it whenever you want instead of like a one hour sort of episode length. That was second. And third, we wanted long form content. Think of it like TV shows and audio. It's voice acting, it's sound effects, it's music, it's close your eyes, it's cinematic experience of sort.

And when we, uh, and, and we call it audio series and sort of TV shows and audio, it's 10 minute episodes, bite-sized content, but long form. These are like 500 episodes, thousand episodes. Oh, that's what you mean by long form is the volume of episodes. Yeah. It's all like 10 seasons worth of like, think of it like a TV show having 10 seasons. It's all nice, like 100 hours worth of content. But with one interesting, uh, you know, tweak we wanted, uh,

one app, we wanted the creators to upload one episode every day. And that was a very interesting, you know, I would say modification instead of like a weekly, you know, like say you can have one episode a week, but if you upload one episode every day, it builds a habit, right? You sort of like, you know, you listen to it, it becomes part of your life and you listen to that show every day. So that really improves retention and frequency of usage. So

So that was a very important week. Maybe going back a little bit. So now I understand the content differences you felt were necessary as well as the product differences. How did you go about doing this? Had you ever started a company before? Is this your first time being a CEO of a company? And how do you go from sitting on this commute for three hours every day thinking about this idea to jumping ahead and diving right in and building a company?

I've worked in different startups, but I never started any. I wanted to for a very long period of time. And it's sort of like those things, you're just waiting for an opportunity that you find something where you will build a deep conviction and you just take the plunge.

So for me, it was, you know, I was very clear I wanted to start up and I found something which I'm really deeply passionate about. I mean, I didn't want to start up anything except content. I was very clear about that. You know, I used to think maybe I should do that. I mean, some other sort of ideas which come in, but I've sort of,

realize that over time you tend to do, if you do the things you really love, it just, you know, it just, you get the courage to start up and you go deeper and then you just, you know, stick around for a much longer period of time trying to figure out what would work if you really love it.

So for me, it was like, hey, this is what I want to do. I'm okay doing this for the rest of my life. So I just took the plunge and I'll figure it on the way. I mean, I have no idea what to do. It's not like, you know, I have someone to, some platform to look forward to. Hey, maybe I can, you know, get some learning from what others have done. There was literally nothing at this point, but I loved it. So I thought we should just, you know, I should just take the plunge. And how did you go from inception to where you are now today? I mean, Pocket.fm,

you know, people may not realize is a massive, massive platform with, you know, I don't know what you've revealed publicly in terms of revenue or, you know, hours of content or creators or active users. But I mean, we're talking really, really big numbers. I mean, Pocket FM is a powerhouse. So how did you how did maybe like what's the short story of how you went from, you know, I'm jumping in to where you are today?

It's just been a series of pivots and experiments. I think for us, we are very clear that we, I mean, all of us, we are not, we came from proud and engineering backgrounds, right? So we were like, hey, I'm not a media person, right? So we realized we can always, if we marry art and science, sort of think of it like the way that, how do you make your content decisions using data and technology and now AI, right?

to accelerate your growth and to just get better content, to produce better content for listeners. Every year has been sort of a very interesting journey for us. I mean, we started in September 2018. 2021 is when we really figured out audio series took us like a good two years. When we launched audio series on the platform, this new category of audio entertainment,

Our engagement time shot up to 120 minutes per user per day. It's two hours of binge listening.

And interestingly, what we realized is that users were consuming us throughout the day. Audio series is the only entertainment format which can be consumed 24-7, right? I mean, we are seeing engagement when you wake up, when you commute to work, sometimes during your sort of work hours, when you're free for some time, you commute back from work and then before sleeping, when you don't want to look at a screen.

Right. I mean, when you think about video entertainment, you have to carve out some time to consume, to watch a movie or to watch a show. You just have to, you have to have a couple of hours, right? But in audio, you just don't. You can do it whenever you want. And it's entertaining. It's a content which you can consume whenever you want. It's not information which sometimes you just get an overload of, right? It's not...

It doesn't require a lot of cognitive thinking to consume very lightweight content, if that makes sense. What about video, though? How do you feel about video? I mean, obviously, we've seen Spotify try to aggressively drive adoption of some of their new video products.

YouTube is now the biggest podcasting platform in the world. And obviously, you know, it's all video. How do you think about the opportunity of video for Pocket FM? Or are you so focused on audio, you feel like that's always going to be the predominant format? Yeah, we focused on audio. I believe it's a huge category and it has its own use case, which doesn't compete with video. And I still feel audio is a category that...

can have a lot more adoption if it has different content categories, right? I mean, even in a pocket for audio fiction, like essentially audio series, serialized fiction content, we're just getting started. And I think, you know, we are at right now more than 20 million monthly active users. We went from zero to $250 million in revenue in just two years since we started monetization. Wow.

And, you know, last year we saw over 100 billion minutes being streamed on Pocket. And then this is still very early, Mike. I mean, we're just getting started. We still have a sort of, you know, not a large catalog of content. And because audio and audio series has a chicken and egg problem.

Right. It's not like content for us is out there somewhere and you can just license it and put it on your platform. You have to figure it out with your creators. And how can we get high quality audio content? Users don't care if it's a new category or not. They just want high quality content in each and every subgenre. So I think we're just getting started. But I think we want to be laser focused on audio. But we, of course, want to expand to categories where we believe that

you know, our IPs that we're creating using audio has potential. Like for instance, you know, at a very fundamental level, if you abstract the audio format, we're essentially in the business of finding unique stories,

Right? I mean, because that's what fiction is. I mean, you have to find these unique, unheard stories and make it available to your listeners. Once we have a great idea or a great story, you would not want to stop that just in audio form. You want to adapt that to multiple formats. So we would want to expand formats.

But of course, from a point of view of focus, it's audio. It makes a ton of sense. Okay, so now we've laid the groundwork on what Pocket FM is. And thanks for sharing those numbers. I mean, this is like the listeners can hear now, like this is some real scale.

Let's talk about AI, right? This is an AI podcast. And as I've heard from you in the past, Pocket FM is leveraging AI in some pretty profound ways. And so why don't we get into that? Like maybe starting with content creation. How is Pocket FM leveraging AI for content creation? Sure. So I think if we take a step back, right? So we believe that, I think as a philosophy, we believe that

A generative AI is going to shift the balance of power more towards creators than someone in the middle. And what I mean by that is it's becoming easier for creators to produce high quality content using AI on their own.

And for us, as we just spoke about in the beginning, that for us, it's been very clear that how do you use technology to solve content problems? As a company, we hate gatekeepers. We don't want someone to decide if a show is good or not. We want users to decide that. Of course.

Right. And as a result, we've always believed then how can we create a platform where anyone can produce a high quality audio show? So that has been sort of something we've been working on for a very, very long period of time. But it was just hard without AI. Right. And so the first step that we took that how can we help our writers produce high quality audio content?

I mean, there are a lot of writers who want to, you know, who have great stories. But how will they, you know, find a voice artist, edit the audio file, and, you know, upload it on Pocket FM? We always used to have an option where you can upload an audio show on the platform. But it didn't have traction because it was hard to do. Any user, any user could upload. Anyone. Anyone.

But it didn't have great traction because again, if I'm a writer, I still have to figure out a voice artist and edit the audio file, add background music and so on. But then with AI, finally, it was possible to, if you have a great story, you can adapt that to an audio form using AI voice. You can edit the background music on the app itself. And we worked with 11 Labs for this.

But we launched this product in March, 2024 last year, when anyone on the app now does see that upload option, but then you can now start writing on the app and on the web, on the web, on the web app. And once the episode is sort of, you know, complete, you can, you get an option to select a new voice. And these are really high quality, like, you know,

We have worked with 11 Labs on this in a lot more detail. Very high quality content. And you just click a button. The click of a button, an audio show is live. That's unbelievable. Anyone can do this now. Anyone can do this. And, you know, I mean, till date, we launched this in over a year. We have seen 50,000 AI shows being created.

Just for context, before this, we were primarily a PGCE platform, a professional content platform with just 200 shows. This is 50,000. And, you know, the feedback that we're getting. So it was clear that a product market fit was great from a writer standpoint. But what about a user? I mean, what about the engagement? AI generated content, right, is already contributing to almost like $6 million in revenue.

But that does seem small as compared to $100 million in revenue, $250 million revenue, but that's growing 40% month on month. - Wow, wow. - And this would not have been possible without AI, right? I mean, it just is very clear. - So these 50,000 shows, like talk us through the creation process. Like where are humans doing the bulk of the heavy lifting and where is the AI doing the bulk of the heavy lifting? I'm guessing on things like, obviously, like you said, the background music,

the voices through 11 Labs, but what about like the actual writing process? Like are humans doing the full writing process themselves? Are they getting leverage from AI to like help with ideas? Like talk us through that. - Yeah, so the first version of the product just had voice and sort of, you know, editing capabilities. And then now the second version, which we have been working on is what we are now developing a co-pilot for writers.

Now, one insight here is that most of our top UGC writers who are using AI are actually our power listeners and first-time writers. And there's a very interesting insight here, right? As it's a new category of content creation, someone who has never consumed this category before

logically can't be a great writer because you still need to understand how it is to write for audio, how it is to, if you haven't consumed a high quality audio show, which is, let's say produced by a human, like how would you as a writer know what to write? But what we realized is that when we launched this AI voice product, our content creation exploded,

But then there were challenges where writers were like, I mean, I have a great story, but I don't know how to, the nuts and bolts of writing, right? How do you structure an episode? How do you start an episode? It has to be engaging. When the episode ends, it has to end with a cliffhanger. How do you write great cliffhangers, right? And secondly, how do you design story arcs? Like if you have, imagine like a 500-

500 episode show, how do you design story arcs? How do you take user feedback and change your content? So we realized, hey, this has to be a co-pilot because of two reasons. One, we have episode by episode retention data for a lot of these shows. So we know what works and what doesn't work in terms of, you know, how do you start an episode? How long should be a story arc be? These are like all,

playbooks that we have developed internally. And secondly, we realize we have ton of data. We will have ton of data when let's say a writer is writing on the platform and an AI suggests an output, right? And you like it. And that'll be a feedback loop for the model, right? So any edits would be a feedback loop for the model to improve it.

So we've been working this for... Imagine in ChatGPD, if you don't like the output, you say, "Maybe just change it a little bit." That acts as a feedback for ChatGPD. Are these your own models that you're training? We're calling it Atlas, our co-pilot. We have fine-tuned our open-source models. But I think we had to solve some really hard problems. I think the first problem was

You can't hallucinate in fiction writing. You can't change the relationship between two characters at some point in the future. You just can't do it. You have to maintain context. And even though Lama 4 and all these models have 10 million tokens,

but the context is such a big problem. Yeah. We had to solve it. We had to build a sort of a layer on top of these open source models where we had to, let's say you are the writer writing a story. We, we save all the relationship between entities and characters in a database, which the model then queries to make sure the model is the right direction. That's just one. And then secondly, we'd also now built agentic systems. Think of it like this, right? Um,

There's an output that AI is generating. And then there are agents which detect, hey, is the cliffhanger good enough? If it's not, here are five options. Maybe you should think of a cliffhanger like this. Or is the pacing slow? Because everyone wants fast-paced content nowadays. Everyone wants a drag.

Maybe you can think of this paragraph that you just wrote, condense it into one line. Don't write a full paragraph of content. Something like that. There are different agentic system, one on pacing, one on cliffhanger, one on opening, episode opening. Because if the opening is boring, users will drop off. So think of it like you have relationship between entities which solves hallucination problem.

and you have agentic system for different parts of the different components of the story. In fact, the biggest one sometimes, and if you use GPD or any of these platforms,

the language is not simple. That's not how people talk, right? So we have now created a simplicity agent text model just to make sure the language is simple enough for audio listeners. So I think this, we want to be one of the leading sort of, you know, models for story writing for Copilot because I believe that

A story is sort of the core of entertainment, right? Everything is a wrapper on top of that, if that makes sense. If you have a great story, audio, comics, novel, like, sorry, novels, audio, comics, video, it's all a wrapper or a different form factor of a core story, right? So I think if we solve this really well, that we can enable writers from anywhere in the world to write great stories and then put that in audio,

and just with a button. And maybe at some point we're also working on, if you have a great story, you can convert that to a comic with the click of a button, right? - Yeah, so to like spin out additional formats from your creativity and your IP. - What our aim is to solve the story generation piece really well. Once that is done,

I think we're very close to solving the voice problem as you know, it's working really well. If you're a writer, you get into this pocket ecosystem, you write a great show, we'll help you in writing a great show. The tool will help you write a great show. And I would say there's an interesting insight I want to share on that. We have over 250,000 writers in our community.

But if you think about it, every writer has a great story, but they do need some bit of coaching sometimes on how to write better.

But you can't have 250K coaches, right? I mean, just doesn't start scalable. It doesn't make sense. But now think of it like this. There's an AI editor of thought, which is telling you, hey, did you hear a few thoughts? And this editor has real-time data of retention, of all the edits that our writers have done over the past, let's say, a year as of now. And then over time, it's just going to get better and better.

It's sort of like having an editor, a great, well, the best person editor, right? Yeah. Yeah.

Really, really fascinating. I actually like, I can't think of any other platform where the platform where all the demand occurs also has a creative tool that's sort of this fundamentally tied to AI and where AI gives somebody like this much leverage, whereas like they couldn't do it before, you know, they couldn't write sort of these series and now they can. That's really fascinating. I remember you were also telling me about

a way that you leverage AI. I don't know if you're still doing it where you're effectively leveraging it to help you identify blockbusters. So quickly spinning up content ideas, quick using AI, then quickly testing them with AI. And then where you see real demand, you invest heavily in those ideas and those stories. Is that something that you're doing? Yeah. So, so I think, you know, the, the, the story generation sort of

technology that we're building with AI has multiple use cases. One is, of course, a co-pilot for writers. The second use case, which we haven't talked about, is localization of a story into multiple languages. We touched upon that.

And the third one that you're referring is, let's say for original ideas, right? So we used to do a lot of PGC, we still do PGC content, which is our original sort of content creation. And the challenge in PGC was always that you could, how do you get a blockbuster, right? That's sort of like the billion dollar question in entertainment. You have to predict blockbusters, launch blockbusters every month, every quarter, right?

And for us, the challenge used to be that you can do a lot of pilots. So we have created this AI blockbuster engine. Think of it like this, that

what constitutes a blockbuster. If you fundamentally break it down into, you know, sort of the components, right? So the first principles of analysis, you would see a blockbuster show, which has a lot of appeal, which a lot of people like. It's a large TAM, a large addressable market. And it also has users are willing to pay for it in our sort of world. You know, a lot of users like it. They're willing to pay for it.

But how do you find such shows? So one way for us to do that was launch a lot of pilots.

Launched the pilot on social media. Social media is a very great way to actually get an estimate of do users even like it? Yeah, get feedback quickly. Get feedback quickly. And we started monitoring metrics like completion rates, CTRs, engagement rates. Where, like on TikTok or like YouTube? On YouTube, on TikTok, on Meta, right? So you launch it on all these platforms and get sort of feedbacks of who the audience could be.

And we started seeing some shows, like some concepts when you sort of, you know, write these pilots or make these pilots do it really well. Like, you know, it's sort of like an order of magnitude better than the rest of the pilots, right? And then you also launch these pilots

in app, right, just on app, and also check the conversion rates, like what's your payer conversion, which means are users willing to pay for it? And then you combine these sort of metrics, right? Think of it like a show that has a high sort of click-through rate and high conversion.

And we started seeing, you know, patterns where shows which had high CTRs and high sort of conversion do really well. And some of these shows like Saving Nora, My Empire System have made more than $40 million in revenue now.

- Oh my God. - Just each, each, right? And now the question for us is how do we do these pilots faster, right? And also, can you know the retention of a show even before launching? I mean, when we used to do this with humans, this pilot testing, we used to only sort of, let's say, do a pilot of 50 episodes, right? But the show could be great till 50 episodes and then, you know, not, and just, you know, tank after that.

And this was a genuine problem, to be honest. We tried launching a few shows, which didn't do well because it didn't have great retention.

So we asked to ourselves, we said to ourselves, how can we solve this? Can an AI create a pilot of 500 episodes, right? With AI, again, sort of in the use case of sort of the co-pilot we have created, right? You create a 500 episode show, again, do the same things, launch it on social media, launch it on app, but this time,

I mean, you're testing not just the time of the show, the appeal of the show, the conversion of the show and the retention of the show. And you can do this in a day instead of like a couple of months. What's really interesting about this is obviously you can get to an answer really quickly on whether or not a piece of IP is something worth investing in. But I have to imagine you can do other types of analysis as well, like

I don't know, like A/B testing with different narrative approaches or isolating certain characters or things like, what about that? I mean, imagine with AI, you can do all sorts of very micro tests on the content. Is that something you're exploring? - Yeah, so we've started again just with PGC first, and then we're gonna make it open to writers.

At any point, we have almost like for our top shows, like five parallel versions running. Of the content? Of the content. So like different opening, different, in some cases, different voice. So like five, six different versions of a show. And then you pick the winning version. And of course, you know, it's sort of like A-B testing and content. Yeah. Wow. Does it happen automatically? No. I mean, as of now, I mean, initially we used the, again, the writers, the PGC sort of writers used to do it.

Now with AI, we have, it's possible to change. I think one thing which is very critical for a new show is the opening, the first one hour, the first five minutes, the first one minute, right? And you can, and this is where you can actually do a lot of A-B testing to figure out what's the right start to the show. Because

in today's world, if you're opening or the first one is boring, it doesn't matter if the story is good after that. It doesn't matter, right? So we do a lot of A-B testing, especially in the opening. Now what we're now trying to do with our UGC writers, we're going to give them the same functionality that, you know, here is suggested five, six variations, just, you know,

Again, select which one you want to A-B test with. So the way all of these work, we first do it on our own for internal sort of use case with our own writers, our PGC writers. And then once it's fine-tuned, it's good enough to roll it out. We roll it out to our UGC community. Fascinating, fascinating. What do you deliver back to the creator? Like does the creator need to be able to like understand

understand how to analyze data and run these A/B tests? Or is it really just doing it for them and dumbing it down so that anyone could leverage A/B testing and multivariate testing? It's dumbing it down, right? I believe that how do you make it super easy for writers? One interesting feature that we're just launching in the next couple of months is a chat.

So instead of like, you know, you seeing so many different interfaces and data and so on, the idea is you can just ask a bot. Oh, wow. Yeah, that makes sense. The bot knows the retention data, the knows what your problems could be. So it could be as simple as my retention fell this episode. What could be the potential reasons? Again, this one again, an extension of the story generation sort of

because now that the AI understands the story really well, understand what could be the potential reasons why the show had a retention drop this episode. It could be because, you know, it could be as subtle reason as this episode, you move the plot and focus more on the secondary character and the users just didn't like it. It could be as subtle as that. Maybe just, you know, make sure the next episode you don't do that.

Right? So the chat would be a great interface for writers. I'm thinking back to what you talked about with how it's-- even in writing, it's really hard to maintain consistency, maintain, you know, the right context, wrong characters, their relationships. You know, I have to imagine that you have similar challenges, but a whole different set of them

when it comes to comics and webtoons? Like, how do you make sure that these things maintain style consistency and, you know, draw the characters in the same way? I mean, this is a really hard thing in AI for all these image and video models. So curious how you deal with that. Sure. So, you know, when we thought of launching a webtoon platform, and the reason was very clear that I believe that while building Pocket FM, we have built...

I think we've created a new way to build an entertainment business in the sense that how do you use AI to find out, to produce and discover great shows, to market them and how do you monetize them? These are three different, the way we built out these three verticals are very different from, let's say, how an entertainment business is built today.

And we thought to ourselves that can we sort of export these playbooks, these content, AI-driven content playbooks to another format? And the first one that came to our mind was comics because I've been a sort of, again, a very passionate manga consumer, right? I've been consuming that for the last 10, 15 years. And the most frustrating problem is that you get one episode a week and that episode just takes three minutes to read.

I'm waiting a week for three minutes. That's just frustrating. And the step that takes the most amount of time is the coloring, the illustration, which you mentioned, to making sure there's consistency in style, face, right? But we thought to ourselves, hey, if this, I can, you know, solve this particular step with AI, right?

Everything else remains the same. You have to still write the content, you have to letter the speech bubbles and so on and so forth.

can you sort of shorten that time? So we started building Blaze with the thought that, can we sort of train characters? And we built it on again, from diffusion models. And can you train certain characters using some synthetic data, using some human sketches in the beginning to make sure that you train the models on, let's say, let's say a story has 20 characters, right?

Can you, can you, I mean, we used to do it on our own. Now we're trying to sort of making it, we still do it on our own, but for where we'll open it, open it up to the world in a, you know, in a few quarters where anyone can do it. So you can essentially train characters the way you want the character to look. So you train the character so that you maintain this face consistency. You train it on, we tried, we developed methodologies on how do you make sure one

what should be in your training data for face consistency, for style consistency, for even background consistency, right? And sometimes what happens, the background changes. Could be the same background, but a little bit different.

So how do you train these things so that when you then try to create a comic, it just becomes that you just need to draw a rough out. The way Blaze works is you as a visual sort of as an artist, just draw a rough sketch of what you want the panel to be. For instance, let's say you want, you know, someone sitting on a bench, you'll just, you know, draw a very rough outline.

select the character which has already been trained and just and write a prompt i want this this sort of a background to to you know i mean this sort of a sort of a background let's say and then you click a button right what will happen is as you've done a rough outline you have you've you've i didn't you've you've mentioned the expressions but it's a very rough sketch it's not hard to do right a sort of artist can do it in a very short amount of time

And then what the model will do is it'll sort of use that outline and do image to image instead of text to image, right? They'll do image to image.

And then for the background, we'll do text to image, right? So it's a combined image to image and text to image and generate a panel for you. Yeah, got it. Got it. So you're using image to image to solve the consistency problem. Yeah. But the rough sketch is a very rough sketch. It's not like you have to draw something very detailed. So by the way, with this, artists have gone to from one episode a week, the productivity is three episodes a day. Oh my goodness.

20x productivity. And because of which, what we're seeing in Pocket Tunes when we're launching some of these audio IPs into comics, binge reading is happening in comics where our users are actually all, the average engagement time has gone above 100 minutes. Very similar to audio.

because of the velocity of content, right? It's like binge reading can finally happen in comics, which never used to happen. - Right. So, you know, earlier in the conversation, we talked about how AI has been really helpful in solving content creation and removing bottlenecks. And you went from, you know, hundreds of storytellers to, you know, hundreds of thousands or tens of thousands. Seems like then you'd have a different challenge, which would be, you'd have so much content

you know, you might not know what to do with it. Or more specifically, like maybe, you know, back to some of the things we talked about earlier in the discussion, now you have a new type of discovery problem, right?

How do you solve that? Yeah, I mean, that's the hardest problem that we have right now. The hardest problem we have right now. So we've built a propagation algorithm right now where if a new piece of content comes in, first you let the AI sort of try to first moderate it in a way to make sure it doesn't have plagiarism and a few other checks.

And then you let the AI also evaluate the content a little bit, right? And I think with that, and as I said, sort of with this, the entire stack of story generation, this will get better and it's already getting better. So you evaluate the content a little bit and then you propagate it to a few hundred users.

and then you check the data of these 100 users. How did this particular show perform? Did it have good retention? Did it have good conversion? If it's good, you propagate to 1,000 users, and then 10,000, and then a lot more users, right? So it's sort of like a propagation algorithm. Now, the tricky part here is that, let's say, what's your first 100 users that you show this piece of content to, which may or may not be good?

So over time, the longer-term solution here is the evaluation piece. Can we get better at evaluating the content in the first place? Because we still like a user-first product, right? We believe that we want to be more user-obsessed. If I have to choose between that, I will choose users and their experience over propagating sort of a bad piece of content to them.

So the longer-term solution has to be an AI-led evaluation. So first, basically what you're saying is before you even think about recommendations, discovery, personalization, first it's like, hey, let's look at the content. Let's evaluate it. Let's actually make sure that it meets some quality threshold, that it's not infringing, that it's safe. That's the first step. But what about the second step and then matching it with the right user? Yeah.

Yeah, I mean, for us, that has been in the works for some period of time. We do collaborative. Of course, the collaborative filtering works really well. And also, you know, that's essentially basically you're trying to figure out, you know, users tend to have a few genres they like about a lot more, right? And then what you also do is, let's say there's a show. Usually we tag these shows with multiple labels. So think of it like one show would have close to like 100 tags, right?

And that tag could be, for instance, you know, in science fiction, it could be interplanetary, right? That's a tag, right? And then imagine there are like multiple of these tags that a show is attributed to. So over time, what we get to know from a user sort of consumption pattern is that what labels or what tags are the users liking more of?

And then if a new show comes in, which is hitting some of these sort of tags, we recommend that show to that particular user. Got it. Makes sense. Ron, it seems like you've built some incredible tools for...

content creation, you know, like explosion of content creation now across multiple formats, audio, webtoons. You've got the technology to strip. You've got your own platform and you've got the AI and machine learning technology to distribute this and make sure not only it's reaching the right people, but it's it's sending feedback back to the creators to optimize around things like retention. Seems to me that like Pocket FM is

is sort of perfectly positioned to now go and do this for a bunch of different content formats.

How do you think about that? Is the future of the platform much more formats? How do you go? I mean, you're already obviously performing phenomenally well. Like, how do you go from where you are today and, you know, 10x the growth or 100x the growth? Sure. So I still believe, though, I mean, Pocket FM and Pocket Toons and Pocket Novels, which is the third platform, will have a lot of potential. And we're just scratching the surface for these categories. And AI is just going to help us accelerate growth in these categories, right?

You know, one thing we didn't cover is that because of the fact that now we can localize and adapt one story into multiple languages almost instantaneously, that opens up a lot of growth avenues. So think of it like this, that 83 languages which have more than 10 million speakers.

But content is not produced in 83 languages. It's produced in like a few languages and then dubbed in other languages, which is not the ideal experience for anyone. Now you can beat audio webtoons. You can actually localize

Let's say, I mean, we have now a bunch of shows which were produced for the US, now doing really well in Germany, in France, right, in Mexico. And just think of from a point of view, a longer term point of view, we will have God in 83 languages or more.

And that can happen with not much effort. If you have a great show, how do you localize the show into multiple languages? So just localizing and adapting into different cultures, adapting to different formats. These are huge growth opportunities for us in terms of audio, webtoons, and novels.

Now, when it comes to different formats, I still believe that as AI evolves and if we own the IP and we are allowing creators to produce much better stories, we are open to opening up new formats. But I think these three itself can scale up 10x from here. And I think the language, just locating multiple languages is a big piece of that growth journey.

Awesome. Rohan, for anyone who's just checking out Pocket FM for the first time, what are some shows that you recommend people tune into first? The first is My Vampire System and the second is Lord of Mysteries. I mean, I'm a huge fantasy and a science fiction sort of fan, but do check out My Vampire System, Lord of Mysteries and Saving Nora. Okay. All right. We'll check them out. Rohan, this has been fascinating. I think the way that you're leveraging AI to scale up

content platforms and discovery is frankly unlike anything we've seen on other content platforms thus far. So really, really cool to see you all leading the charge with AI. So thank you so much for your time. Can't wait to do this again. Thank you, Mike. It was a great conversation. Thank you.

Thank you for listening to Generative Now. If you like this episode, please rate and review the show. And of course, subscribe. It really does help. And if you want to learn more, follow Lightspeed at Lightspeed VP on X, YouTube or LinkedIn. Generative Now is produced by Lightspeed in partnership with Pod People. I am Michael McNano and we will be back next week. See you then.