Why Stripe Might Acquire PayPal, Agentic Shopping Course Change, ChatGPT’s Audio Language Barrier

19:46Stephanie Palazzolo:How did you come across this issue of how effective AI audio models not just are, but are in different languages?

19:55Ann Gehan:Definitely. So, you know, in recent months, I've been writing a lot about the kind of rise of audio AI models and how there's really been the shift towards AI that's able to kind of listen and talk to humans the way that we do with each other. And especially, you know, with a recent story from last week about OpenAI's device efforts, which are going to be audio first, meaning that people will be talking to the device and it will be talking back to them. You know, you would imagine that these companies want these sorts of devices to be used by people around the world, meaning that they have to, you know, understand and speak all sorts of different dialects and languages.

20:33Ann Gehan:And then, you know, the more I talked with researchers, the more that they told me, you know, there's already this gap between Western and kind of non-Western languages with text-based models. But that gap is even bigger with audio models.

20:48Stephanie Palazzolo:And what are the core reasons here as to why the audio models are not as good in non-English languages?

20:57Ann Gehan:Yeah, I mean, honestly, it really just boils down to the fact that there just isn't that much training data there, especially audio data. So in order to train these models, you know, in the best way possible, you basically need data of people of different ages, gender, speaking about a variety of topics. So everything from, you know, customer support to medicine. And you need that in, you know, all sorts of languages. And so that's, you know, that type of data doesn't really occur naturally. And so companies would kind of have to go out of their way to collect that sort of data, which is obviously very difficult for them to do.

21:36Stephanie Palazzolo:You focused on OpenAI in your column. I'm sure there are other companies that are working on the exact same thing and have the same issues. But I mean, if we just stick with OpenAI for the moment, tell us about their expansion plans outside of North America, how important those international markets are for them. Yeah.

21:55Ann Gehan:So as you mentioned, the column kind of focused on OpenAI, but we can imagine that this is going to be a big deal for any sort of AI company that wants to have a global reach. So companies like Google and Meta, for instance. But with OpenAI specifically, we saw Sam Altman say in the last month that the company has 100 million weekly active users of ChatGPT in India alone. And so that's more than a tenth of all the people that are using ChatGPT on a weekly basis around the world. That's located in India. And, you know, many transactions there are being handled over the phone or by people speaking out loud to each other, which means that in order to kind of meet that kind of natural way that people there do their work, OpenAI is going to have to meet that with their models and ChatDBT.

22:46Ann Gehan:And it's going to have to teach ChatDBT how to speak out loud. And so expanding to these regions are very important for OpenAI because as my colleague Shree and I reported in a story last week, in recent months, basically since the release of GBT5, OpenAI has seen a slowdown in chat-GBT usage. And so there's only so much that they can do in the U.S., so they're really going to need to expand to other areas around the world in order to keep that chat-GBT growth going.

23:18Stephanie Palazzolo:So basically what I hear you saying is that it's not just important for the growth element of the story. I mean, it's also important to be able to accommodate different behaviors and the way different people use AI because not everyone is texting anywhere. And it sounds like people are using voice a lot more in Asia and other continents.

23:41Ann Gehan:Yeah, I mean, there definitely is this element of kind of like meeting the users where they are and kind of matching. Yeah, just these different cultural preferences around the world and the ways that people want to work and kind of communicate with each other, whether that's through, you know, text or speaking out loud.

23:59Stephanie Palazzolo:Now you're starting to sound like a product person, Stephanie. You've really been studying up on the user journey, as they call it. I like it. Okay, before we let you go, you talked a little bit in the column about the startups that are trying to tackle this issue and the opportunity that stands there. Who should we be watching and who are you talking to?

24:23Ann Gehan:So I'd imagine that, you know, all these data labeling firms are starting to expand into audio data. But, you know, one startup that's kind of early and up and coming that I focus on for the piece is called Poseidon AI. So this company basically has an app where any sort of user around the world can upload audio files of them, basically reading out loud different transcripts on different topics like customer service or law. But this is very difficult for Poseidon and other startups trying to do this to do very well, because you can imagine that they need special technology to make sure that people are, you know, actually following the script correctly, not going off script or speaking a different language than what they're supposed to.

25:10Ann Gehan:So that requires these startups to build their own kind of proprietary technology in order to make sure that the audio data is of the highest quality possible. Great.

25:22Stephanie Palazzolo:Well, Stephanie, I want to thank you for coming on. That is Stephanie Palazzolo, our AI reporter here at The Information. As advertising and AI chatbots starts to heat up, so too are investments in startups that are offering products to ride that wave. Koa is one of those companies. The startup raised$20.5 million this week in a Series A funding round led by Theory Ventures. I want to bring on Koa co-founder Nick Baird and Theory Ventures GP Tomasz Tunguz to help us unpack the deal and what's coming next. Nick and Tomas, welcome to TI TV. It's great to have you both here. It's great to be here.

25:58Stephanie Palazzolo:Thanks for hosting us.

26:00Tomasz Tunguz:Tomas, you brought a friend with you this time. Normally, it's just you. More the merrier. I'm thrilled to be part... We're thrilled at theory to be partnering with Nick. And as you know, I was a product manager on the AdSense team. And the world of AI is absolutely taking over and a big part of that monetization and the future of it will be driven by Koa.

26:21Stephanie Palazzolo:Well, okay. So Tomas, I'm going to come back to you, but I'm going to go to Nick first. Let's talk to him for a while here. So Nick, tell us what Koa is and how old is the company even?

26:31Sheel Mohnot:Yeah, the company is very young. We've moved quickly in this space. I had my dog in the background. And so we're a little over a year old. The product's been live for 13 months. And what we do basically is we help the application layer of AI monetize via sponsorships very simply. So we install SDKs, and from those SDKs, the applications that use AI, that use generative interfaces, that use kind of anything that's interactive and dynamic, can serve basically a native app format that works for them, that can help them sustain their growth, their revenue growth, and, you know, cover some of those inference costs.

27:05Stephanie Palazzolo:Okay, so let's break it down a little more simply here. So, I mean, on one hand, you have the chatbots, you know, that we're all used to. You also have some of the publishers, I guess, you know, the thinking about the the ride hailing company, the Uber Eats type of platforms. Right. And there's advertising there. And then you have the brands. So just help me understand where you fit in there, what the actual product is and who you're selling to. Yeah.

27:32Sheel Mohnot:So we're we're a marketplace model similar to, you know, Uber being the intermediary between the driver and the rider. right? And so we are the intermediary between the inventory, the apps that have space to advertise on and the advertisers themselves. Okay.

27:45Stephanie Palazzolo:And so what is the actual product that you're selling to both parties? Yeah.

27:51Sheel Mohnot:So basically for the advertisers, similar to how they might show up in a banner ad on Google AdSense or a search ad on Google's AdWords or, you know, really actually very similar to what ChattoBT is doing, right? Advertisers are able to come in and buy slots. And what we're selling is basically an HTML snippet, if you will, a part of the publisher or the AI application's real estate that we're carving out as a separate entity to mark this as a sponsored post or a sponsored piece of content. And so the advertiser gets to make sure that they show up alongside the model response.

28:26Stephanie Palazzolo:So in a world, Nick, where OpenAI is pursuing their own ads, why is this not something that they could just do themselves and create these products themselves?

28:37Sheel Mohnot:Yeah, I mean, it's a great question. And I think the easy answer to that is that look at the mobile marketplace, right? So in mobile, you know, we love having all these apps that are free and they're powered by ads. And there is no player that owns more than 50 % of the mobile ad market. And yet AppLovin exists as a company that's worked$225 billion on 40 % of the mobile market share, right? So it is possible that everyone is able to do this. And by the way, that's beating Google's AdMob product. And that's a huge company in its own right. So there is just so much inventory, so much opportunity here.

29:06Sheel Mohnot:I mean, we're about to see the entire internet change from something that's more static and consumable to something that's dynamic and interactive and agentic. And so how do we allow sponsorships to exist in there? How do we allow those publishers and people creating great experiences to make money and sustain the business? And, you know, I think that ChattoBT will do their own thing. Maybe they branch off into off-platform similar with Google. And there's plenty of space for all of us to play. There's a lot of publishers out there and a lot of internet experiences that exist.

29:35Stephanie Palazzolo:Tomas, walk us through your thesis here. Why Coa among all the other companies that are in the space?

29:41Tomasz Tunguz:Yeah, great question. I mean, you look at the search ads markets, about$250 billion in size. You look at the social media ads markets,$265. Last year, it was the first time social surpassed search. And both of those and most online advertising is predicated on the idea that there's some data that you can use to target individuals. And the incredible part about AI, as we all know, is the amount of information that we share and the length of the queries that are going into AI. And so we think that the amount of context, the amount of knowledge about a particular user and their wants and desires is strongest and not by a small amount within the world of AI.

30:21Tomasz Tunguz:So there needs to be these advertising systems. And, you know, we haven't really invested in an advertising company in the last 10 years because of some of the dominance of the existing platforms. But AI changes all that. So we were fortunate enough to meet Nick and the team, Mike and Herrick, early on in the business and track their growth. And as Nick was saying, the hard part about these businesses, they're marketplaces, just like Uber. You need drivers and you need riders. In the world of online advertising, you need publishers, people building AI applications, and then advertisers who want to appear on those sites.

30:54Tomasz Tunguz:And building that marketplace, see something between the liquidity dynamics, is not an easy thing. And it was clear to us as we tracked the progress of Koa how well they were executing that liquidity building. And ultimately, like any marketplace, that liquidity is the competitive advantage.

31:10Stephanie Palazzolo:So, Tomas, you started your career early on before you were investing at Google working on the AdSense product. And I wonder as you think about how this marketplace will develop in the AI context, there's going to be a lot that's similar. What do you think that's going to be different about how this is built out and all the ad tech around it? Great question.

31:32Tomasz Tunguz:I mean, I think conceptually it's similar. You have systems that work for advertisers, systems that work for publishers, application creators, and then a marketplace in between. So buyers and sellers in the marketplace. I think that conceptually is the same. The major change and the reason why we think this market opportunity is even bigger is the quality of the data. So the thing that we learned, Google discovered that it was great. It was best business model on the internet because they knew exactly what you were searching for right at the moment that you were searching for it. And then if you're on Instagram, like, you know, many people think of the ads on Instagram as part of the product.

32:11Tomasz Tunguz:The ads are so well targeted that it's actually, you know, I'll go buy that sweater from Quince or whatever it is. The great part about AI is now you'll have tremendous context. And so the monetization capability of a publisher should be significantly greater than search or social. Google makes about$120 per user per year. I think it's very easy to see an AI system making$250,$300 per user per year in a way of subsidizing inference costs and driving real business model innovation.

32:41Stephanie Palazzolo:Because of the fact that you have more information about the person, you can target things better, and ultimately it's a higher conversion likely to a click-through or a sale.

32:53Tomasz Tunguz:Exactly. So at Google, when we were building these machine learning models, the more data we had about who you were, what your age was or age range was, what your interests were, the much better the monetization. And we got orders of magnitude increases in overall ad performance as a result of some of these signals. And so if we look at the... I mean, just think about like a Google search query. I think average is three and a half words. for a Google search query, and they built a$4 trillion company on three and a half ads targeted on three and a half words. What can you do with a conversation that measures in the thousands of words?

33:31Stephanie Palazzolo:Nick, tell us who are the publishers that you're working with right now?

33:35Sheel Mohnot:Yeah, so early adopters are similar companies you might expect, like core chat apps. Example would be like Liner. Liner is a search tool for students, similar to kind of like a perplexity, kind of aimed at graduate level students. A company called DeepAI, for example, which is, you know, the founder of DeepAI has figured out how to be really good at SEO. And so it's a very similar experience to ChatGPT. But if you Google AI chat, you'll actually get them organically showing up before ChatGPT does. And so those are just a couple of examples of kind of early adopters. And then, you know, the direction that we're going, and I think what Tom is explaining very articulately is that this is not just kind of like the chat bot era and these clones and stuff that are kind of popping up, but there's new vertical use cases, there's AI pediatricians, there's AI math tutors, things like that.

34:24Sheel Mohnot:But there's also now being created dynamic experiences within traditional publishers and people who are kind of already distributed. And so a great example that I like to point to is, for example, Quizlet, right? Quizlet has about 50 million students that use the platform every day, and they're using it to quiz themselves. But now Quizlet has launched a conversational interface, an AI tutor, if you will. And this is very similar to somebody like Duolingo. Duolingo has their mass program. And that is like an AI surface area that now is a new surface area that's getting tons of engagement that these companies don't know how to monetize effectively because like Tomas says, we actually, they don't know how to take those intent signals and actually make them effective for the user.

35:06Stephanie Palazzolo:And so ultimately the goal for you is to work with some of those larger publishers eventually.

35:11Sheel Mohnot:That's correct. When you have those conversations with them, what concerns, questions do they have for you?

35:19Stephanie Palazzolo:Are they even receptive to this? I mean, I'm just imagining the abyss of uncertainty around, we're just trying to figure this out. I mean, you know, take us inside those conversations.

35:30Sheel Mohnot:And yeah, so early, early publishers that we work with on the chat app side are basically just they're just feeling the pain. The reason why we started this company is because so many of these businesses were just struggling to cover the inference costs and actually deliver a great experience to the user. So that's kind of one bucket of customers. And then if you look kind of upmarket at these other interfaces, and again, like we think that the entire Internet is moving towards dynamic interfaces, towards agentic interfaces. Right. So eventually everyone will move in this direction and we have to create personalized, you know, user experiences, but also personalized monetization services.

36:02Sheel Mohnot:And so usually what the companies are sort of worried about when we talk to them is that it's just very early. We're very early in the market. Right. People are still learning how these interactive interfaces are going to make sense, how we're going to be able to engage with users. But the main question that people have, and I think, again, the reason why we exist is that they want they test these AI features. They get great engagement. They get, you know, people love them. They get a bunch of messages saying like, we love this thing. Can we distribute it more? And they're only testing on, you know, 0.1 % of their traffic or 1 % of their traffic.

36:33Sheel Mohnot:But they're losing too much money because the inference cost is much higher than hosting fees, right? And so they actually cannot scale those products unless they have a sustainable monetization method to do it. That's why we started the company.

36:43Stephanie Palazzolo:And last question for you here. I just want to make sure I understand the inference costs here in the context that you're talking about it. So explain that for me a little bit. You have these publishers, they're running their own chatbots, and they're saying that we can't even handle the inference load that it requires to run ads, let alone the core chatbot. Is that what it is?

37:04Sheel Mohnot:That's right. So if a mobile app today, they're spending maybe a few dollars per user per year on kind of like hosting costs and everything that comes with having an app exist. And now you put in this conversational interface and you're actually, you know, it costs you two, three cents per session per user, right? And so if the user is there every single day, that becomes significantly more expensive than what it costs to have a free user on the same application in the past, right? And so if they want to deliver these, you know, interactive personalized experiences, it's simply a more expensive thing to do.

37:36Sheel Mohnot:And so they can't do it for free forever at scale.

37:40Stephanie Palazzolo:Great. Well, Nick and Tomas, I want to thank you for coming on. That is Nick Baird from Koa and Tomas Tunguz from Theory Ventures here on TITV. That does it for today's show. A reminder, we are on this stream Monday through Friday at 10 a.m. Pacific, 1 p.m. Eastern. I want to thank you all for tuning in. We really do appreciate your viewership. Make sure to subscribe to the information on YouTube and follow us on X, Instagram, TikTok, and check us out wherever you get your podcasts. I'm already excited for our next show tomorrow. Have a great rest of your Wednesday. Bye-bye for now.

38:17Thank you.