In this episode Divya talks to Soumyadeep, recently named Forbes 30 under 30 in Asia for Consumer Tech, about learning visual styles as he builds his AI models for his startup, Dashtoon, breaking down computer vision as a technology and the role of media in driving tech.
You can check out dashtoon here : https://dashtoon.com/
Defining what a style is turns out to be incredibly hard. It starts with the character — the face shape, the eyes. The moment you make something round, it suddenly becomes younger.
The color palette is another part of style. A revenge romance can't be made in a bright, Marvel-ish style. The genre demands a darker theme.
Hi, I'm Divya. Hi, I'm Kahran. And this is... Thinking. I'm thinking.
Welcome to this episode of Thinking on Thinking. Today I talk to Somini about a lot of different things, everything from AI and creating content through AI to how AI works, what is GPU, what is CPU and later into discussion about how all of this is going to change the content landscape.
We hope you enjoy. Hey, I'm Swami Adip. I'm a co-founder at CTO at this company called DashTube where we are trying to make something like YouTube for comics. The idea is that anyone can come and make a comic,
publish to our app and then monetize through that and people pay to consume their content. What are you doing these days? What is exciting that's happening in your life? So, till date we have been making content in an X style
and very recently we decided we are going to move to a more conventional comic style, which also interestingly enables far more kind of variation in the style. Firstly, one of the things I'm doing or fall morning I spent on was defining what a style is. So, what is a style? That is interesting. It's very hard but I think couple of things that we came up with that
it starts with the character. Is it a round face? Is it something in the mix? Is it a realistic like a human variation? Like an anime is a very conical character face usually. And then there is the eyes part. Is it a small eyes, lateral eyes, round eyes? The moment you make something round, it suddenly becomes younger age,
things that I discovered. So, that's one part of style. The other part of style that I discovered is the color palette, which is very obvious. Like what color palette has been used for example, something that... So, we made a style which is a very American-ish, marvellous style with round eyes and all of that. And the content team that we have said we make a revenge romance.
We can't make it in this. And then I was like, how make sense? So, okay, let me figure out what does revenge romance style looks like. And then understood a bit like there is a darker theme. It is a little more... The color palette is on the darker side, but you still want the ability to have characters which are varying.
So, that is another part, like color palette. Do you guys also think about like background? And so, like there are some elements that are rendering elements that you're already speaking, right? Like style and color and all of that. But like, do you guys think about background and spaces and all of that also?
Yeah, in the hierarchy of things that we can... I'll come to that. So, after color palette, the next thing that we noticed is actually strokes. Like how thick the strokes are, how thin the strokes are, how fine-grained the strokes are. Is it like a smudge kind of a stroke? Is it an actual stroke? Like a... Those are the next things.
And then we decided that we call something universe. For example, I also have to reduce this style definition to something that I can do from an AI model standpoint. So, for that, I need to understand that, like what is the universe? For example, interestingly, one of our models, we noticed that whenever you type a telephone, very simple, telephone or a phone, it will give you that rotary phone. It'll give you that.
And then, like my content team keeps saying, I can't make modern content with this or I have to make so much effort. And the moment you type any background by default has a retro-ish background in that model. And then we are like, so that's what I call the universe. So now we also know that we have to make a universe that is modern. When someone says phone by default, it has to be an iPhone at least or a smartphone.
So, that is interesting because you can't directly tell the AI model any of those things, right? Like you can't say, oh, this is a phone or this is not a phone. Or like when I say building, have these kind of assards and not like, you know, this other kind of facade. Yes. Right.
So you're like, I have to first translate, then I have to re-translate it so that the model can also understand. Yeah. So correct. First, I need to understand and then I need to align some 20 people to be able to do this because now that I know that what the world means, I will have to, we find new models to be able to do this, right?
Now I need to create a data set out of this. And this data set generation is like, I have to find 3000 images. And if I can't find them, what I do usually, I just increase the top of the funnel. I will throw 10,000 images and get a person or some way to just prune out images, which don't pass my taste barrier. And these are the metrics of my taste.
So now I need to have those people who are looking at every image understand this language, universe, character shape, eye shape, color palette, all of this. And this is so subjective and so tricky. It took me like 30 minutes to align my leadership. Then it has to go down that, that operational expertise. It's interesting, right?
Because like for an artist or somebody who's been trained in art, they would be like much more familiar with the vocabulary, not just the words, but also like, you know, what are the things to think about? So like, you know, stroke weights, stroke nature, how flowy is it? Is it organic shapes? Is it angular shapes? All of that, right?
Like we would be much more used to it. But like you've not done visual work before this at all, right? If computer vision is called visual work, then yes. So what's like, if you think about yourself from like maybe two years ago to today, like how would you say your ability to look at has it changed how you look at things? Yeah, definitely.
Like just outside in the world, like, you know, you're watching a movie and are you thinking about it differently? I'm noticing more art for sure. If we define it as art, I'm noticing more nuances in like how people portray things. How Andhra Kashiya portrays the story versus how Sanjay Leela Bansali portrays the story. I'm noticing them.
The amount of content that I watch and I have watched since I have grown up, I've grown up in a family where I used to watch movies every Friday and Saturday in a theater, right? I have those things registered, but now I can see them also, right? You're classifying your data set internally. Correct. Yeah.
Like now I know that, okay, kuch kuch hotha hai was this way because it was this way. Because now I can think of it that way because I have that vocabulary. Of course, I was noticing them and since I have watched so much data is there. So like that's definitely something that I have started noticing a lot. That is one very, very recent.
Like last three days, I have been just fighting people on what just to align that this is the style we have to run with this. We can't push it. But style is just one aspect of content, right? The core aspect at the end of the day, what we have noticed is that the user gets about the story. The style is a medium to tell the story.
So it's not like a decision that once taken, it can't be changed or something like that. So like just have to run with something. But even for that, like, you have to decide. What are other people sort of like, I want to say philosophical alignments on the visual aspect. I'm sure everybody has different point of view, but still like, you know, what is it?
Because everybody has some sort of idea of like, you know, this is where narrative fits and this is where like the visual fits. And like, this is how we want to as a brand because like, regardless of how many stories you guys have, still you want to build your identity also. Right. Like somebody looks at your content and they're like, oh, this is from dash two. I think that's, that's not a super important thing at this point. We are very style and stories and characters.
These are fmrl things. We will get sheep shifting, experimenting new things. There's nothing that we think this is what our brand is. I would like to believe that what the brand we want to be is that we just offer so much variety that that is the differentiation. That is the brand that I want to build.
But I just want to offer variety. But that's not possible. Right. Like even if you look at like, let's say something like Instagram where people put all kinds of content, especially because these days I'm like using Twitter and Instagram and threads. Like not a lot, but like, you know, I'm using all three equally. I can tell that like there's just a vibe difference threads in Twitter, despite being text medium, despite being short form text medium.
There is still like a vibe difference. People are talking about everything and anything, but just like threads, people are sharing more stories. They're more chill and Twitter. Everybody is trying to make you angry. Yeah, that's true.
So when I mean everything, I mean, okay, here's how I stack rank this this way, right? You can go to webtoon and read comics. You can go to Netflix and watch movies. What is it that dashed on or any new first content platform should be able to give you, right? It is the ability to do things that wasn't possible.
And we can bring them down into many parts. One of the things is just production speed, which is something that like we are doing. So that is one. Second thing is that the ability to do kind of styles that was just like not possible at the same production speed. Right? You could not do a, let's say, let's look at some game like style, right?
Who's talking to Shiba, right? That style at the same production speed of a comic is just impossible at some point of time before AI, right? That is something that we would want to unlock. Right? That's when I mean that we want to be the brand that just unlocks things that people have not seen before.
Like most of our content right now, if you ever open the app, it has a very gamey or a 3D style to it. Because we thought that that is something people have not seen. Now we are deciding that, okay, let's go deeper into the 2D genre and not into the 3D genre because heck, why not? It's just like, there is no math here. I'm just trying out things.
Yeah, a lot of it is just taste oriented. And like, you know, I feel like at least for me personally, that's where media industries are like a little more earnest and honest. And like all the other industries try to, like most industries try to present themselves as essential, but like almost nothing is essential. Right? Like if somebody decides that, oh, I want to go to space, like that's also not essential.
It's also a taste based decision. Like it fascinates me. I want to go underwater. Like again, it fascinates you or I want to make cars. Like everything is one of those things, right?
But I just feel like media is much more honest about that fact that I like, yeah, I'm doing it because I want to do it. Not because like this is the quote unquote correct thing to do. Yeah, it's more taste based. So of course it has to be because when I want to do. I think everything is taste based.
I just think that like, you know, people in media are a little more honest about the fact that it's all taste based. I mean, okay, like let's say we know a lot of people who are making like AI products, right? Almost all of them are chasing their own curiosity and interest. Like they're not really chasing sure there is market gap. Like there might be some demand all of that.
I'm not disagreeing with that. But that is true for a Marvel movie also. That is true for like an animated movie also like this some market demand for these kind of things, right? Like, I don't know. Like I just find it interesting that like just going from tech to media, I have found it interesting that at least like, you know, people here are more willing to own up to the fact that this is a whim.
While like, you know, everybody in the tech and startup world tries to say, no, this is the logically correct thing to do. It's like, bro, it's not like there's no logically correcting. Yeah, interesting. Isn't that true for most consumer products? It is a taste based.
I think we have to look at the world is consumer products are far more taste than enterprise product. Enterprise product is a far more need of some kind. It's a different vocabulary of certain like scale and all of that. But I still think it is a taste based thing like wire processes the way they are in big organizations because somebody does that. Oh, this feels better to me.
And then they came up with like, you know, motivated reasoning to prove those things. It's like, you know, half of it might be logic, but orientation or rather I would say like, you know, the sort of magnitude part of it can be logic, but the directionality part of that vector is always taste. Like, where do I want to orient myself? Makes sense. I think, yeah, true.
Yeah, I guess, yeah, everything at least the ideation definitely comes from some taste view of the world. Then people can rationalize through data through process through emotion, whatever they whatever works for them. Interesting. So like, let's say we are five years in the future, you guys are doing really well, you have like, you know, a lot of audience, all of that. What kind of stories would you want to tell?
And I don't see myself or even dash tune as a platform where I control the narrative. Like, I don't want to be there. Ideally. But like, what kind of stories would you want to tell? Like, it enables everything and if money were not a matter, what would you want to tell?
I don't know. Like, firstly, I don't personally think that I want to decide what stories to tell. I want people to tell their stories and find their own niche audiences. That is how I would look at it. Like, personally, like if you say that if I am a creator, what stories I want to tell or what I mean, then like, like, educational content is what I would like to create as a creator.
Interesting. One of the greatest things that I have felt value for myself is when someone can distill really complex things in ways that some people can understand really, really well in less time or less effort. Right. I would want to do that. Like, C. Blue, one brown is a great example and with some lens like C. Blue, one brown and this primer does it through some data simulation type of a lens.
I'm okay looking at it from like a person teaching kind of a lens also, but they have to be a great teacher. And I think there is a lot of skill and depth needed to do this. And most people don't do a good job. That is something that I would want to do if I personally become a creator, explain and computer networks in a way that most people don't understand and most people should understand. Explain very recently I was explaining like GPUs, like why GPU versus CPU to someone in a way that most people have not thought about it that way.
It builds a mental model that lets you expand what a GPU can do. What is the explanation? You can't like make so much context and then like not get the explanation. It's not, it's not that complex. The simple thing is that, okay.
The mental model is that what a GPU does is it throws throughput to obfuscate latency. That's it. That's it. More explanation. So the idea is that everything in the world, like remove GPUs, everything in the world where we use anything that increases throughput is to obfuscate latency.
Like in a restaurant when you go, right, you put four people there so that your wait time reduces. What you're doing actually is doing this. You're increasing throughput to reduce latency for the end user. That's exactly what a GPU does. GPU is something that is going to do 1000 jobs, the same job very fast for the end user to reduce latency.
Right. That's why GPUs tend to have many more cores than like a CPU might. Yes. But the moment you ask GPUs to make a lot of decisions, it is going to slow you down. The GPU is designed for throughput.
It is not designed for decision making. It's almost like a CPU is the CEO and like GPU is more operational oriented. Yes. Yes. And operations and decision making CPU GPU is that lowest level of worker whose entire job is to take like in a cooking game context, take a beat from place one to place two.
That's it. And they're going to they're optimized to do that really, really fast and there are 10,000 of them. It's like assembly line. Like as you were describing it and just made me think of like a similar and you break the task down in many small pieces and then one part is just executing the same task over and over. And that's exactly what programmers don't get when they are offloading things to GPU.
You can't give GPU things to do which needs context switch. The moment you do that, GPU is going to do a terrible job. So what does that even mean in case of like a processing unit that there is a context switch or there's a decision making like what does that mean? So for example, matrix multiplication, right? Like I think that's the most common example, right?
If you ask a GPU to do matrix multiplication, it will do really well. But if you ask the GPU to do matrix multiplication, when the matrix is eigenvalue or some some computation is greater than 10, then you do matrix multiplication. Otherwise do vector addition. That's a hypothetical right? If you ask this thing, this this if then else thing to do in a GPU, it will do a terrible job.
It can't do it because GPU is optimized to do both parts of this branch. The moment you ask it to do one part on this condition and other part in that condition, you're blocking both of those parts. Because that guy, that guy there who's taking the meat from one person to another, that person has been asked to do this job. And then if there is a condition, they're going to stop this job, do the next job, right? And he's still waiting there.
If you ask the if then else to happen on the CPU and then have two different guys, two different GPUs doing these, each of them job, your throughput is going to be better. This is core mental model of GPU programming. This is something that we were explaining to recently explained to someone. And I also learned while I was explaining that, okay, this is a mental model that if every programmer would understand, we would have much more efficient way of programming with GPU CPU. Who codes on a GPU?
Like who does a GPU coding? Who do you mean? Like who person? Like what person? I mean, any AI engineer is going to...
Okay, so like while AI is very popular and I've heard a lot about it, like a lot of it is still black box, right? Okay, so like let's say I'm building a startup, I have like, you know, I'm taking chat GPT and like I'm running calls and stuff. Why do I need an AI engineer? Why can't a normal person just do it? You don't need an AI engineer if you're not yourself, either fine-duning something or inferencing something.
Let me explain what both of them are. Yes. Or training. Let's say training. Yes.
Fine-duning is subset of training. Training is the process which created the model that chat GPT is. What does that mean? It means that there is some architecture, not getting... Let's make architecture the black box.
There's some architecture, some pretty smart people have designed. And we have proven that you give X data to it. It is able to do a good prediction on why data outside the distribution of the X data. This is what the model is. Now doing this loop of throwing the data at it and making sure that this it is performing up to some performance benchmark that we believe in, right?
Or we need it. That is training and that is the training part of it before the creation of the model we train it. Inferencing is that, okay, this model is ready. Now, wherever a user comes or whatever, we send the some more new data to the model. It will inference the output that is the inference part.
These are the two parts. Both of them use UPS. So if any of these tools, if you have to do, you need to understand UPS well because in each of these, like inferencing is easier to understand. While inferencing, inferencing is nothing but you give the input to a model. The model is nothing but a complicated network of matrix multiplication of some kind, right?
And that gives the output, right? That is what a model is, right? That is inferencing. Now, when you're designing this model or inferencing, there can be some decision making that needs to be made depending on the business logic. That is where the understanding that what decision should a GPU do, what decision should a CPU do is the inferencing pipeline that the engineer is writing.
That is the person who needs to understand this. And of course, there are frameworks that obfuscate this thing for the engineer also of some kind. Then it is the framework designer, like a PyTorch, which is a framework, like a library, which will obfuscate this thing also, but it gives you the tools to do this talking with the GPU. Okay, so how is like doing a API call to something which already exists, right? Like there are many models that already exist.
And how is that different from inferencing? Is it? Like API call means that you are not doing any GPU. It's the person serving the API is doing the GPU thing or inferencing. You are asking another server, another person that, okay, dude, this is my input.
Give me the output. It is like me asking Instagram when I'm opening Instagram. It's the same thing that I am asking the Instagram server to give me my feed. That is also an API call. Right.
So like here, the decision making is on somebody else's head. Correct. Yes. But if I host my own model, then you are doing the inferencing yourself, at least inferencing yourself. Correct.
And then you have to understand the integrities of the inference better or more. So then like, let's say I used, so for example, I used Confi UI on my system to run something, right? Right. But I don't know any of this clearly. I'm just learning right now.
But like I was able to run it and I was able to inference. Correct. In all of this, what I talked about many levels of abstraction that have been built so that different people can optimize different parts of it. Right.
At the core level, at the base level, it is CUDA, which is Nvidia's proprietary software who has abstracted out GPU programming. Okay. The level above that is PyTorch, which is what I said, who is abstracted out understanding of the CUDA, like the person working on that level doesn't need to understand CUDA, but
they can still find train models, fine-tune models and do every CUDA thing without knowing it. Above PyTorch, there would be more libraries. Like in the case of Confi UI or image, there's something called diffusers, which is a library which has abstracted out image specific things, image model specific things.
So that I don't need to care about like how to manage the weights, how to manage the inference saying how to handle PyTorch level things and decision making, shifting memory, shifting memory, shifting, not all of them. Diffusers does that. Over diffusers, then there is, there would be some like a back end UI or code, even we
work in that level. Dash-tune works in that level, Confi UI works in that level. Right. That they've made an interface over diffusers so that you can do some operations. Right.
Then there is another level which is where Confi UI, which has made it super flexible. Right. You can move around things and see this operation, that operation, this operation, you can make a node-based graph. Dash-tune is different.
We are saying you can't do so much flexibility. You do just this part. Like one specific workflow of Confi UI of some one or two is what Dash-tune is. Did you guys need to tell the models that you're using to not ever generate realistic people?
Because I assume that like the base diffuser model actually might be much better at generating real people and real images than not. Correct. Yes. So, we want to use the fine-tune, which is why we use fine-tunes and fine-tuning is as
a different set of nuances of different kinds of fine-tune. But which is why we fine-tune. Like we want consistency on the kind of images the model gives. Like when my user says that I want a style, we call it style. I want this style, which is like a three-day gaming style.
We have given it a name. We call it Euler or whatever. Right. Like we will make sure that we are fine-tuned so good that you will always get that consistent style.
That is why we fine-tune. Because most of the image generation tools in the world, because you're doing single image generation, you don't care as much about consistency across generations. But for storytelling, we care a lot. So, we spend a lot of time to get that.
This is what we will nail. Right. So, that is something that that is why we fine-tune. That's the only reason otherwise it shouldn't be. Had you done like AI stuff before?
Like had you done AI work before like you started Dash 2? I mean, the definition of AI was different. But yes, like most of the time I have spent in my life has been in computer vision for robotics. I used to build aerial robots for competitions in college or then soccer playing robots and in all of them also other than leadership part or operations part.
My core expertise was always computer vision, but computer vision has evolved a lot. Like it started with very simple things like edge detection, color detection, like that was what? Line follower and all of that. So line follower is a application or something like edge detection or a color detection.
Right. So we would do very simple things like this screen or this camera field. You can see this much amount of red here. So let's find out where the red is and take a centroid and do things like that and build applications on these things.
So the level one of computer vision, which was the state of the art when I started is 2012. Then came some feature extraction type of things where people made some models like classifiers as to okay patterns that can recognize some of some kind. Edge is also a pattern of some kind, but people came up with far more complicated patterns.
Like for example, people would start coming up with face detection. So it was a pattern, but it was a feature type of a thing. It was nothing, but they had trained a model on a lot of faces saying this is a face. It's like a classifier. Right.
This is a face. This is what a face looks like and you would do what they were horrible. They were never that it used to do that thing, right? Like two dots in the eye and one at the nose or something. Correct.
Yes. So, uh, yeah, facial features is that one thing. Then people define neural networks and neural networks and convolutional neural networks came, then we got all better than now we could identify what a dog is. This is all in the last 12 years.
Correct. Yeah. Yeah. Like CNN's became mainstream. I think somewhere in 2015, maybe 12 actual to 15.
Yeah, but very recent, right? But computer vision had had hit a roadblock after a point that none of this was happening really well. There is a competition called image net, which was the state of the art of identify things and all of that.
And then came transformers, which is what changed everything. Transformers started making a lot more things. Uh, transformers with neural networks started making a lot more things far, far more, uh, applicable. And then, and then we are still discovering things of applications around transformers
here in the generative space, like LLM was the first use of transformers and then diffusion models does not use transformers. Now there are diffusion transformers, which is mixing some of them and they are seeing some more wins, but very early days, very, very early days. But that is like, I came from that background.
That's when we started the question that I used to do this edge detection, color detection type of things and figure out, okay, this is where my robot should run to and figure out some smart things around that. That is the start of the computer vision career. From there now it has come to creating comics.
So but still some amount of computer vision, it has changed completely. Like we don't look at feature extraction at all these days. Like who cares? Yeah. Like now also everybody in their phone, if they want can have like a detector of any
kind, right? Like, oh, here is this photo and then I asked you, what tree is that? And then it'll tell me also what tree. Yeah. I wanted to build a startup which was doing Google photos because I had just discovered
that this is possible with some computer vision. Then there is a startup called hyper which is a computer vision company from my batch, my batch people only they build this product and Google launched this product and killed them. Right?
Like if something is that valuable, like and you can and Google has access to the photos is pretty cool. So that, but that was the technology that came then that you can make something like Google photos and we did that. Someone did that.
So this is 2017. This is 2017 when Google photos. We are still evaluating that Google photos is a good idea. And now it's like part of our life. I've always felt like, you know, I'm one of that rare generation people where like I've
seen too many technological revolutions in my lifetime. But that's true for a lot of us. I mean, not me. I guess I wasn't there during the dot com bubble. I was still too young.
I mean, like, you know, there was like computers and then there was internet and then there was phones, like smartphones. I mean, and then your entire app universe that sort of came out of it. And then like mobile internet and like the sort of penetration that it had wherever I was talking to a friend yesterday and she was like so annoyed.
She's like, how's he's going to deliver food in 10 minutes? And she's like, this is only possible in India because the income discrepancy is so high. And this is so useless. It's just interesting that like, you know, at some point you were seeing like a certain kind of life and then as you grow within like 30 years, you're like so much has changed.
And now this AI stuff is like, oh, okay. All of that before is trivial. Yeah, exactly. Like it just keeps and to give your argument that income distribution, I think if people want things fast, even if there is no income discrepancy, I also think like in one of the
reasons why these kinds of things are more feasible in India is like it's a unique use case. We are very high density like entirety of Bangalore's population is probably the amount of population that many European countries have which like, you know, like, yeah, sure, you know, you wouldn't be thinking, oh, this business can't work in Germany or like this
business can't work in France. I think I usually flip this by thinking that like people do want things fast. I think that's human behavior. Right. People want, like when they want, they want gratification ASAP.
The question is someone will innovate even in Germany to give them 10 minute delivery if it is possible, if people realize the need. India realize the need early because it was easier in India than in Germany because income discrepancy because then high density society, it is easier in India. But I think we will have this in US food.
Someone will solve for it because it is human behavior. If people want things fast, then you give them fast. Yeah, but like there is enough literature and we know enough about it. That instant gratification is not necessarily good in the long term. It is that if you give it, people will incline towards it.
I don't know. I think that a large part of why we are leaning more and more towards it is also because of fracturing social connection because like, you know, like all of that research around instant gratification is also dependent on that, that like, you know, oh, will a mouse take the cocaine or regular water and it's like, oh, it always takes cocaine water, but
not when you throw it in a pool with other mice. Now it's gonna like, you know, hang out with people. And I think like a lot of that is, I don't know the way we are going. A lot of instant gratification need is also because people don't have communities as a strong. That is true.
And yeah, it's probably a bad thing. I don't know. It's bad thing, good thing, hard to say, but it is true. My claim is that it is what people claim to want. So they will get it.
Someone will figure out a way to give it. And in some ways, it's also one of those things where like, you know, people wanted faster horses or somebody invented the car. They wouldn't have thought that they can rent cars, but like somebody made the car. In the morning, I was listening to something and they were talking about how the difference
between a brand and a commodity is that a commodity has no differentiation apart from the fact that like, you know, it has a certain price point and ease of access. Right. How do you think about as more and more AI content becomes a thing? There is like anyway, social media has commoditized a lot of cultural artifacts.
Let me say that. Right. Like, let's say I watched reality shows, Netflix releases a love is blind every other week. Like I'm not even joking, right? And how do you feel about it now that like AI also comes into picture and accelerates
already like a snowball rolling down the hill? How do I think about like, how will that change things? There is a commoditization. So there is like some amount of value that things have because of the time it takes to build, let's say comics and stuff, right?
Like Superman, Marvel comics, all of these, they have been around for like 80 to 100 years. And that's also why they have the capital that they have. I don't think I get the question fully that. As in, I'm just wondering with the increased pace of production, how do you see that changing how people relate with media?
So like part of it is you're making things and part of it is also you have, like you said, you've consumed like a very particular kind of media with a certain language all your life and you do deeply connect with it. Right. So like, how do you think about that?
I think one thing that I am fairly sure of, which is what my thesis of my startup is, is that more people will become creators. That's baseline. Like that is, I don't think there is any doubt there also. It's not a risk, but which also means there'll be more, more shitty content also, which is
also true because everyone has not. Barrier is lower. That is great because I think that the selection that happens from a loads of shitty content to that one content that like a few content that stays, they often are really, really good and they often are far more bigger IPs in a business term and higher shelf life.
So we would start seeing more of those bigger IPs happening. That is one that for sure, because there'll be more content. And second is, I think decentralization will increase, which also means that bigger IPs may not get that big anymore. There may not be as one piece like big IP that often because you will have 10, one piece
like IPs for each niche of the, each sub cohort of the sub cohort of the world. Right. So things can get that big, but there will be people will be more satisfied because they are getting their niche more often. That is what I envision how things will fall out.
For example, I think I'm a big fan like Shah Rukh Khan says, right? I'm last of the stars. He has an interview where he says I am last of the stars. I really believe he is because decentralization as it happens in media content, anything, right?
And it's increasing so much that it is hard to have one big Shah Rukh Khan anymore. It is, we are going to have a lot of Anbir Kapoor and these things and with their niche audiences looking at following them, but it is very hard to have one big star as this keeps happening. Okay.
So I don't follow Bollywood at all. Right. So I don't know the scale difference apart from the fact that like Shah Rukh Khan has had a very, let's say rich life in terms of like how long he has been a creator and how many films then he has produced as well.
Like it's been a long time, but like all of these other guys are younger, right? They haven't had the amount of time also. What's the difference? Right. So I think the difference is the pool that one person can bring.
You are saying no more Rajnikanth, no more Amitabh Bachchan, that kind of a thing. Yeah, it is that hard because people have access to more content now because people have access to varying content, varying platforms. The distribution has become so decentralized. Right.
Let's forget Shah Rukh Khan generation before that, right? There was only one way to connect with your actor. There was only one way to get the content that you like. You go to a movie theater and you will see that or Aravarnand or Amitabh Bachchan or Shamil Kapoor, you get them, you connect with them.
Right. And people had emotional connection with those people. They made their emotional connection with those people. Right. So they built a following around it and then they grew.
So 50% of the content following came from them and around 50%. I'm coming up with a random number right now, but there was X and Y content that following that came from them and some from the story. And when there is a great story and a great past following person, you would see a super duper.
You get a show there. Yeah, you get a show there. The show they did not work when it was launched, but it didn't. I didn't know that. It was a, it didn't.
It was relaunched is when it was. But anyway, coming back to this is what happened then. Now in the, in the era of a Shah Rukh Khan, Sarwan Khan, Amer Khan, this went a little dis-decentralized, not as much, not as much, but little decentralized. But what also happened is that there was more money flown into the market because content
became bigger. There were more directors, more money. India became slightly richer and there were more people who would spend on content, discretionary income increase of people and all of that. So they saw a huge wave and we saw like the golden video body, what was I would call it?
There was really, really amazing content out there. People would watch and all of that. But the moment OTTs have come in now, last 10 years, last 10 years, right? Now consumers have more options and they will not compromise. This is what is out there.
So my entertainment has to happen by going to a movie theater and now I have more options. I can watch a movie on Netflix. I can watch a YouTube video on YouTube. I can go and sit on Instagram and watch Reels. I can do all of this.
So now the bar for me connecting with a content piece has increased. Now I don't connect with Ranbir Kapoor as much. I probably connect more with my story, my audience, my creator. Like I probably connect with Kumar Varun. He makes quizzing, right?
If given a choice, will I watch an average Bollywood movie on, on going to a theater versus watching a one hour quizzing with Kumar Varun? I'll choose a Kumar Varun because I don't want to watch an average Bollywood. So the bar has increased a lot. So now because that I have so many options and my bar has increased so much,
those actors which used to get X following has reduced and X has reduced massively. Because now I don't connect with that actor as much. I probably want a better story. I probably connect with the creator like Kumar Varun also. I probably want to create with creator Ranbir Kapoor also.
But now Ranbir Kapoor can't become 10 X because I have five, 10 options. Interesting. I wonder that's also probably why like a lot of stand-up comics have become popular because they are doing this like very niche kind of comedy and they're talking to a very specific, like each one is talking to a very specific crowd.
But then there is enough of that. I think it was Zig Ziglar who said 10,000 true fans are all that you need. And like I wonder like if that's what you are also trying to say that like the high like the ceiling has been also reduced. Correct. Yeah.
So I do believe that this happened in the West far before like in India even today a Shah Rukh Khan movie makes more money than many other movies because of the following of Shah Rukh Khan. In the West that skew has reduced already like a Tom Cruise movie does not bring in as much audience easily. Like Tom Cruise has to also make a great movie. The West saw this slightly earlier. The starification or the star power reduced slightly earlier because they already had so many options.
Yeah. They already had. India started seeing this recently and that's when it's going to happen. Interesting. Like now that I'm also thinking about like other things maybe earlier like non Hindi films would not gain any traction in the audience.
Correct. So many people have told me like let's say about that Malayalam movie Great Indian Kitchen. Yeah. Like you must watch it. You must watch it right.
But like it's a Malayalam film. I cannot think like even 15 years ago somebody telling me to watch a Malayalam film. Yeah. That's what because distribution was so limited like control control like cable TV was the one distribution of old movies and you don't control that. And theaters was the distribution for new movies which also you don't control.
So if the power has gone more to the consumer they will choose what they like. Interesting. That is something that I will supercharge it. Right. If I can make a really good story and put it out there I'll become popular.
Will I become Shah Rukh Khan? No. But I can get somewhere. Right. I'm saying no one can become Shah Rukh Khan.
Interesting. Or maybe like people whose sole job is to create content to create parasocial relationships. Like let's say people like Joe Rogan or Andrew Tatars. Like I can only think of these like you know people like very extremist right now. But like their whole stick is that they want to create a parasocial relationship with their viewers or consumers.
And maybe like they become that kind of big maybe not that like not Shah Rukh Khan big but yeah. It's very hard to become a Shah Rukh Khan big in a in a in a new world now. Very very hard. Very very unlike. And with I'm claiming this also a new claim that even with this no one can become a Narendra Modi big also after a few years.
The moment politicians start understanding that this is the way to market to your audience like P.J.P. nailed that marketing better than anyone else. Right. They nailed it. So and they nailed it with the face of Narendra Modi. But the moment people realize that I get to choose my politician again that will get decentralized and that should happen.
It is the right way to do. Okay. My claim. No no no it's interesting because like that's also making me think about in a very squeezed time frame something similar happened with K-pop. So BTS and Blackpink were the first wave of internationalization of K-pop that happened during covid time.
Right. So 2020 BTS and Blackpink like they shoot up the sky and people are like you know everybody around the world is their fan. More people listen to K-pop but there isn't any more BTS or like you wouldn't find one person that like oh everybody is just a fan of this band. Exactly. A Taylor Swift is another example.
Will we have another Taylor Swift? I don't think so. Very very hard. It's going to be insane. Very hard.
Or you can be that good. Maybe you have to be way with the bar to be good has to be that good to be that big. But like don't you also feel like some of it is just like these are Black Swan events. Being a Reshmi Kanthar, being a Shah Rukh Khan or being a Taylor Swift is a Black Swan event. Like it's very hard to get there.
It was already so hard. Now it's going to get 10x harder or 100x harder number but it is going to get harder. Only what is going through this if you follow enough interviews is going through this identity crisis that actors are not getting in enough audience. Great directors are not getting in enough audience consistently. Then what will get enough audience.
So my claim is that good stories. Good stories but then it is still not be that big right. If you make a non-massive story which is really good, it will get a non-massive audience. It will get that audience. But you have to make money off retention and not off acquisition.
You have to get that sticky audience again and again and again and again. Which is what all apps have already, app companies have already understood this. That it is the retention that matters. You can solve acquisition in different ways but it is the retention that matters. All you need to look at that.
I think but even saying something like oh Shah Rukh Khan brings in a lot of people or Reshmi Kanthar brings in a lot of people. That is also a retention thing. Correct. But the problem is that. Who is retaining is different.
The relationship that is being retained is the focus is different. Here there is a person and what you are saying is now it is going to be a little more ephemeral than that. Yeah, it is going to be more content driven. It is going to be much harder to have it person driven. It will be content driven.
I like action movies and I want my Nisha action movies every Friday. Someone has to make action movie for me every Friday. Otherwise I will not retain. And this is something we notice in our users also. It is very interesting that we make romance daily soaps practically.
That is our core content. Romance daily soaps. The moment we try to shift our acquisition to something even slightly different from the romance daily soaps. We see a massive drop in our retention or whatever. Average revenue per main user.
We see a massive drop because if my content is all romance and my acquisition content is not romance. Then they will not move around shows and if they don't move around shows, I'll make less money. How do you not pigeonhole yourself in that case? Because like after some point people are also going to get bored. That point will come.
Correct. Which is why the emotional journey of my user is this right usually. That okay, I saw this ad on Insta. I like this. I clicked on it and now I want to read this.
I want to know what happens at the end of the story. That is the emotion that the user is. They went through that story. Then they are going to be like, okay, I want the next story and most often, more often than not, they want a similar story. They don't want to explore.
Most people don't want to explore very early. After three, four, five stories, they're going to be like, okay, now can I think about more this platform has built my trust. They're giving me this quality bar and then they're going to explore. I mean, my company is not there to get to a fixed show. So right now, the pigeonhole is the goal for us.
But at some point, yes, like Netflix, very distinctly, there is a Netflix data also, right? That the long tail is what retains users. The acquisition is what gets users. Like it's the sacred games or reality shows. Love is blind.
Love is blind. Love is blind. Those are acquisition shows. They're going to spend tons of money in those acquisition shows. But if you don't get things like love is blind every week, every month, or whatever that you like.
You will not retain on Netflix. Netflix ensures that that tiny cohort of love is blind user also has enough content. And me, the the the Masi Hindi film movie, I mean, love is blind is also not tiny cohort. Okay, that is also Masi. I don't know how big the cohort is.
I'm saying there is a each cohort. Netflix is is able to address to really well. That is insane amount of content that they are able to do. Insane. It's very hard.
It's amazing that like, you know, alongside creating stuff like love is blind, which let's be honest, like it's not very finished content. Right. I love it, but like, you know, it's not finished content. They are also able to sort of invest in content like our cane, which I would say is on the
other end of like, you know, front edge of, okay, what can content actually be the ceiling bar. It's interesting that like, you know, they are able to do both of those things. Yeah. Yesterday we met someone who's used to run hotstar, right?
And which is like a platform not Netflix big, but recently big, right? The core job of the content person is to understand the emotion this content creates in the user and a person has many emotions like and different moods, they have different emotions, different things, but a person also has a default that this is what the emotion that a person mostly likes.
Right. So you want to acquire on that kind of a content and then you want to retain by giving them slightly similar, but different and keep nudging and rotating them in the same type of content. Right. For example, when you think about this, this is not totally counterintuitive light on a
Saturday. Like I'm emotionally most susceptible to start a TV show, but on a Monday night or a Tuesday night, I will not start. I won't. You cannot.
Please give me something that I know what is going to happen. Netflix has nailed that thing down really well that went to show more and they do think that deeply. Yeah. I'm sure.
I'm sure. They do think that deep. I'm 100% sure. Like I remember when I was watching a lot of K-dramas and I was watching these romance reality shows, Netflix started pushing like these content, which are Asian reality shows on romance to
me. And I was like, my response was, they know me so well. It is amazing. The music is by Akshay Ramu Halli of BTRPT music. Editing is by Beatnik.