Latest / Convo AI World / From Real-Time Voice to AI Avatars: Live Play Mobile on the Next Era of Gaming

Transcript

From Real-Time Voice to AI Avatars: Live Play Mobile on the Next Era of Gaming · Convo AI World

• Geremie Camara (00:00) We've come through close to 100,000 hours of streaming live entertainment. I consider convo AI from Agora the heartbeat. The frontier of the LLMs is the brain. They're helping the humans. And then the humans are helping the Ais get better and it's this nice recursive loop. Whether it's the entertainment value, what the host is saying, the voice fidelity or the avatar fidelity, we're racing towards the Turing test. But the initial feedback from a data standpoint and candidly from an audience standpoint is better than expected. I think we're going to find out a lot quicker then we expect to find out. John Moore (00:48) Welcome to the Convo AI World Podcast, where we talk with the builders creating the next generation of real-time, voice-first and AI-powered experiences. Today we're here with Geremie Camara, one of the founders of Live Play Mobile, creating at the intersection of live gaming, AI avatars, and entertainment. The team has already built one of the most interesting things we've seen in conversational AI live hosts engagement. Now they're exploring how they can take that further and where that innovation leads. So we'll take a look today at what it takes to make an AI character feel alive, how the AI hosts and human hosts play together. And really excited to talk about all things convo AI today with you. So for starters, would love to have you introduce yourself and kick it off. Geremie Camara (01:46) Thanks, John. Super excited to have you in the studio again. It's been a long and fruitful partnership between Live Play and Agora. We actually go back, as companies, we go back now, I think almost seven years. This is my third, video-based entertainment hybrid startup. Actually worked with Agora even before Live Play, but we started Live Play back in 2019. We went live with our first 24-hour live game, live bingo game, peak COVID, so June of 2020. We've been live basically 24/7 since then. So we've done close to or We've come through close to 100,000 hours of streaming live entertainment. through the Agora stack, which has been pretty amazing and fruitful for us. And yeah, my background is mostly from the entertainment side. I spent a long time at Paramount Pictures and sort of every kind of job from marketing to distribution. I was at Sony Television for a while, right here down the street, working with some amazing game show IP, 'Wheel of Fortune' and 'Jeopardy!' and 'Who Wants to Be a Milli a Millionaire' and a lot of their game properties as well. So I've always had this sort of hybrid background between making games and making broadcasts and back in 2019, we basically stuck those together within Live Play and essentially we build mobile games and web-based games in a bunch of different categories, but our special sauce is streaming live hosts into those games to entertain the audience and make them feel seen and special and connected. And so that's what we do really well and we do it all over the world and we have hosts all over the world. And Agora has been the core video partner for us since day one and continue to innovate and we continue to innovate on top of that stack. So it's really good to share some of that with you guys. John Moore (3:37) Yeah, we we've often talked offline about what makes your games different, the community built initially. So what makes your games different with the live hosts from your perspective and why does that give Live Play Mobile games a leg up? Geremie Camara (03:54) Yeah, I mean, we know games are amazing in that they they elicit certain emotions really predictably and reliably. And you see this in games that or I first saw this in games that came from TV IP. So for example, at Sony, we did a lot with Wheel of Fortune. And you take that evergreen billion-dollar mechanic that Merv Griffin came up with, spin wheel, solve a puzzle, which is basically hangman with a wheel. And it that little mechanic has created a 40-year dynasty on television, but the emotions when watching that are that are elicited are very different than when we would build Wheel of Fortune games. And when someone's actually playing the game, it's a different set of emotions than watching the show. It's progression to mastery and a sense of accomplishment. So games have always been amazing at that. What we came on top of that with, in addition to those emotions, is the ability to make people feel seen and recognized and celebrated. And when we launched during COVID, we were shocked at how many people wrote in and said, 'Look, I haven't had a kind word said to me in weeks'. Or 'I've been playing your game because I've been in chemo and no one can visit me and your hosts are so kind to me'. And so I think that's you know, hearing your name and feeling connected to someone who's predictably or reliably kind is special. So that's what we try to create are those special moments for our players. John Moore (05:26) What keeps those players coming back? What have you learned through building with live hosts in the picture? Is it anything that you can put a finger on? Geremie Camara (05:35) Yeah, I think depending on the game that we put the host into. So we used to say they come for the bingo but stay for the connection. And I think it is that human connection, because in the bingo example, there's thousands of bingo games on the App Store. the people who stay in our game, they connect with one or many of our hosts, and the host also connect with those players and it's just this friendship that evolves over a long period of time in the best case. And so I do think it's this genuine ability, authentic ability to make people feel seen and appreciated, cheer them on when they get lucky, console them when they don't, you know, all the human things that we all want. John Moore (06:17) Yeah, before we get too technical, so we've talked about human hosts. Can you describe in a really simple way what you're looking at and what you've built so far with the AI Avatar hosts as well? Geremie Camara (06:31) Yeah. So we know on the human side, from a general hosting standpoint, we know what grade is, right? We have a hundred thousand hours, we're very data driven, we're very creative in our pedigree. You know, again, I came from game shows with some of the best hosts literally on the planet, with Alex and Pat Sajak and just pure masters of hosting a show. So we start out knowing what grade is, and then we can distill those lessons into the AI avatar-generated experiences. And obviously they're not human. So you know, you're not even going for exactly the same thing. But there are a lot of places where the Venn diagram overlaps emotionally. And there are things that humans can do that at the moment the AI host can't do and vice versa. There are things that we can set the AI host to go do and learn, but it's very hard for the human to go do. So it's this unbelievable privilege to have both running in shows. John Moore (07:32) Yeah, it's really cool when we've talked again where you've separated some of the experiences into two pretty distinct camps where there's the human engagement and the AI avatar engagement and their goals are to do different things. I wanted to come back into the origin of why live hosts work for you. And I know you've talked briefly about your background and we've talked extensively about it, but what was the origin idea of bringing a live host into a mobile game or just games in general in a way that hadn't been done before? Geremie Camara (08:13) Yeah, so when I was at Sony, there was an app that went supernova viral called HQ Trivia. And I was working on a bunch of the biggest game shows on the planet at the time. And within a few months that live-hosted trivia show, it was basically 'Jeopardy!' with a live host on a mobile phone went from zero to a couple million concurrent viewers twice a day. And I remember there was a call on Christmas Eve, the first year that thing launched, with a bunch of the showrunners and executives saying, 'This little mobile app is getting ratings that most new syndicated shows would dream of, right?'. And it was virtually overnight. So that piqued my interest. Again, working with some of the best talent on the planet, you see how a human host can pace an entertainment experience with an audience so that we know that works from the game show world, we knew it would work on mobile from the HQ Trivia example and then you know the only thing that HQ didn't have and I wish they had because I missed that show was a business model that worked that made it sustainable. So my co-founder, Stuart, he really comes from that casino gaming world. Which has, by the way, a lot of live hosts, right? If you go to a casino, there's a dealer at most tables, right? Kind of helping you enjoy the experience. And so we really took my broadcasting background and his casino background and started building things that could sustain themselves economically, but have a very entertaining and human experience to them. And so I think the combination of those two things works really well. John Moore (10:05) Yeah, continuing on the host side, do you find that they're they're there for primarily entertainment? Are they coaching people through the game? What are some of the main goals or I know you've instructed and taught hosts how to be great hosts, so what are some of the main things that you say, hey this is the element of becoming a great host? Geremie Camara (10:27) Yeah. That's a long conversation on how to be a great host. Their primary, like any good game show host on television, and if you know, in until you watch 'Jeopardy!' or Wheel or any of these big shows get taped. What gets on TV is such a small tiny segment. And if you watch when Alex was doing it, how much it takes to keep the contestants happy, to run the clue board, to, you know, in between takes he's talking to the audience. There's a lot that goes on that doesn't make it into those twenty-two minutes of television every night. And our hosts are trained to do that, and I think the outcome is the same, which is they're trained to direct everyone's focus on the best parts of the experience. And so that's welcoming new players, it's teaching people when they recognize them that they're new how to get started, it's making the people who've been with us for five years feel special, conveying the best part of, you know, the schedule for the day. So it's part educational, it's part human connection, it's part entertainment, it's part retention because people come back day after day after day to see the same host because it's part of their routine. And there's different skill sets in each of those categories and then there's also just being charismatic and persuasive and you know, endearing and we teach them all the humans and the AIs by the way how to how to do all. John Moore (11:59) Clearly engaging games. I know I've played and watched a few. Where would you say the origin story with Agora being part of the picture and really thinking about real time started and how did it come to fruition that real time's an important part of this infrastructure? Geremie Camara (12:21) It started with a startup I was part of before Live Play called Ripcord TV and that was on the back of HQ Trivia. I spun out of Sony and we did I think a dozen live game shows. Ran into the same pros and cons that HQ did. High engagement business model was a challenge. And I had met one of the founders of Agora through that first company. And I wasn't the CEO of that startup, but I watched how well the company was treated by Agora so that when we started Live Play, it was at the top of our list of vet. And at the time and still now, you know not to be technical, but glass-to-glass subsecond latency was paramount for us because these games are extremely fast paced and people are winning in real-time and you really need a real-time experience with a live host. So it's always been at the top of the choice matrix of real-time glass-to-glass ultra low latency video streaming that is available worldwide that also has all these other features on it and allows us to continue to evolve. John Moore (13:33) Yeah, uniquely and I mean important for your use case, but just the global reach of what you're doing as well and where you're headed with Live Play Mobile. All right. Now I'm fast forwarding. Jumping into the AI conversation, the avatars, there's a lot of news about AI avatars, voice AI, etc., when did it first hit your radar and when did you think this could be potentially a part of what we do with our live hosted games? Geremie Camara (14:09) It was always a curiosity for us. For years, it's been a curiosity. In terms of when it crystallized, I would say it was little by little and all at once. And it was probably back half of last year. Everything had to fall into place at the same time, which it just did. So we needed and when we trained the host, by the way, as most people who are into communication or study communication know, the visual is not the only thing that matters, right? The tonality of voice and voice quality and what is said and you know the avatar component, the video manifestation, in that tech stack is just one requirement. But you needed, you know, the LLMs and the Frontier to get to a place where in real time you could make an avatar sound human in terms of what it sang and when it sang it and what it's responding to. You needed the voice models to get to a place that sound human or human enough. And you needed the avatar tech to get to that place, and you needed places like Agora to be building the heartbeat. And to me, it is kind of like an anatomy, right? Where I consider convo AI from Agora the heartbeat. It keeps the whole thing moving. The frontier of the LLMs is the brain. All the stuff that we're building around the memory and the harnesses and the structure of it all and the recursive learning and the ML and all of that is, you know, is also in there as well. The voice models are obviously massively important. And so all these things came into alignment around well, first of all, November last year, I think, you guys were really starting to roll out convo and we did some voice-only experiments that reminded me a lot. So one of the the inflection points when I was at Sony is we got a dev kit, this is years and years ago, a dev kit of one of the first Echo devices when it was still strapped to a board that you couldn't take out of the office. And I remember building a 'Jeopardy!' game and we used Alex Trebek's voice. And I remember showing that both to customers or watchers of 'Jeopardy!', but also to the 25 producers and staff that had been on that show for 20 years. And really I think the authentic nature of a voice coming to life was probably the thing that I was waiting for most. And so all of that happened around November, and then on the video side it was really March with some of the with the partner that we're working with on the video side. So yeah that was the evolution. And you know and what we keep telling people and telling ourselves actually is it's the worst it's ever going to be today. Which is so exciting because it's getting really, really good and it's only going to exponentially get better. John Moore (17:06) Absolutely. I wanted to comment on one of the things you said at the beginning of that statement where Agora is the heartbeat, the pulse, the LLM, all the information that you have on your side is is the brain and it allows for this really cohesive, sturdy infrastructure for convo AI while allowing you to keep your secret sauce and all the mechanics of how these hosts work, what increases gameplay or the metrics you're trying to hit and I thought that was a really cool statement that that you said. You also mentioned back in November it all started to fall into place all at once. What were some of the main changes in the technology that made a production ready implementation of convo AI avatars feel possible now? Geremie Camara (18:03) Well, to Agora's credit, you guys saw the big, Agora saw the big picture. And I think a lot of folks who are trying to piece together this human-like avatar experience. And I was in that camp six months ago or a year ago. You don't really know how to get started or what the best stack is to get started. We've built so much internally since that time and we're rolling so much of our own stuff now. But when you're just getting started, it's like a Mr. Potato Head that you guys created, right? Where it like comes with the heartbeat and you guys integrated partners on the video side, you've integrated partners on the voice side, you've integrated partners on the LLM side. And so it allows a customer to not have to go up the learning curve completely alone. Everyone's going to go up that learning curve and probably start to customize depending on where they see creating value and what's proprietary to them but yeah, it was just a very amazing, well-thought-through template. And the heartbeat is not, it's not nothing. It needs to keep the patient alive 24/7 hours a day without fail, right? So that's why it is the heartbeat and I call it the heartbeat because if that stops ticking, it takes down everything else. And it's no small feat to keep that heartbeat running without incident. You know, we need these hosts to be avatars or human, to be up 24 hours a day, 7 days a week, 365. That is a very robust infrastructure. And so yeah, you wanna pin your experience on the strongest heart you possibly can. John Moore (20:05) What would you say are some of the main capabilities that are core to human hosts that AI can't do? And then what would you say are some of the capabilities of AI avatar hosts that humans might not do at scale? And what are the two camps looking like today? Geremie Camara (20:26) First of all, we don't have a prediction on what is going to resonate over the long run. So we're trying and experimenting in a ton of ways. I mean look, humans are humans, right? I mean, people are incredibly worried about AI and the level of its intelligence and what you know that AI is already smarter than most, if not all, humans and but you know, humans have free will and they have desires and they have ambitions and they have lives outside of the LLMs and so the avatar hosts can't connect with our players outside of the show. They don't live a life, they don't have Instagram, they don't have things that are rootable. So yeah, they're not human, we're not trying to make the AI hosts human. What the AI hosts can do though is that they can learn insanely fast. They can hosts for 24 hours a day without ruining their voice or getting sick or you know what I mean? They can experiment in real time, they can speak any language. They can show up in a a bingo game in Mexico in 5 minutes, whereas I can't get a host necessarily out to Mexico to be on camera in five minutes. Yeah, I think on the learning side there's so much that the AI Avatar hosts can help us with. And by the way, They're helping the humans. And then the humans are helping the AIs get better and and it's this nice recursive loop. for sure. You know, if I want to look at ten ways to open a show and I'm just noodling, and I have to pull a host for three hours and play with that in a studio, that's fun. A lot of hosts don't want to stay another three hours after they've done their three hours on air live. I can experiment with an AI and get a sense of what's going to work, you know, basically infinitely. So we see ourselves as this pyramid of content that we can supply our own games but also any company's games or any live experience where at the bottom of the pyramid, yeah, a big portion over time of that content will probably be some version of AI. And then in the top 20% of that pyramid, we're going to have the best human host on the planet with the best tradecraft and the best way to long term connect with other humans. And then at the very tip of that, because we're in Hollywood and we come from that background, we'll also be able to provide live events with AAA celebrities. And so you know I think if you're looking at like unlimited scale of content to keep 20 languages around the world engage and then have more special entertainment on top of that and then really special on top of that. I think it's a pretty good offering and all three of those layers help the other layers learn. John Moore (23:37) Awesome. Stepping away from AI just for a moment because I like this part of what you do so much is the magic moment you've talked about that you create in games for players, whether it's AI or or human. So can you describe what you previously talked about as the magic moments created within games for players in your audiences? Geremie Camara (24:02) It always starts with someone hearing their name. You know, it's that classic win friends and influence people. No one ever gets bored of the sound of their own name. And there's this, you know, reticular activating system in your brain where if you're walking through a crowded casino and you hear your name, you immediately tune in, right? It has such deep limbic resonance to it. So we're really good at that. We're really good at making people feel uniquely seen. A lot of that is having the tradecraft to know when to address people one-on-one versus when to move the whole arc of the show or the experience forward. And then, you know, this is going to date me, but my favorite show as a kid was this show 'Cheers'. And the theme song to that show was 'A Place Where Everyone Knows Your Name'. And this is where the human piece, yeah, players show up every day and they get to interact with someone that knows their name, knows them, knows their history, knows that their birthday was last week, knows they're going through a hard time. You know, we're not therapists, but there is a level of I get to connect. It's no different than, you know, the your favorite bartender or your favorite barista or someone that just feels consistently connected to you. There's a lot of other magic moments that we create, but it's simply human. It's simply just remembering someone and focusing on them. John Moore (25:25) With AI becoming part of your strategy, do you see that expanding what can be remembered about players either through the avatar host or giving human hosts more tips on it's someone's birthday, their friends in the chat, how are you thinking about increasing that level of personalization and engagement in both camps? Geremie Camara (25:49) Yeah, I mean look, I think every company is wrestling with that. I think at the far end, we all know how creepy it feels when you're talking about buying a new car to your spouse and then all of a sudden your feed blows up with ads for that car or whatever, right? So there's that creepy side, which is not what we're meant to do. Now we're getting into AI production. So how does AI, not the avatar, but AI writ large, help us do a better broadcast, right? And every entertainment company is going to be, every company period is going to be using if they're smart, using AI to make whatever their product is more satisfying, more limbically resonant. I kind of think of it like having a really good spouse. Who walks, you know, if I go into a party and you have a really good partner by your side that's like, 'hey, do you remember it was his birthday?' or if you forget a name, kind of that like really good just sidekick, a good spouse or partner that can nudge you to be the best version of yourself. I think AI can be a good sidekick to our hosts and be that little nudge of, 'hey, it was John's birthday yesterday, don't forget'. John Moore (27:03) Being that this is the convo AI podcast, wanted to dive slightly into the the technical requirements that that you see and start to talk about really at a high level what has to happen when a player engages with an AI host at a minimum requirement. What are you putting together on your side and what does that engagement have to be like to be passable from a player's perspective? Geremie Camara (27:36) It it has to do the vast majority of what the humans do which means it needs to recognize every player in the game, it needs to be able to uniquely address every player in the game, it needs to know the game well enough to teach people how to play in real time. It needs to in real time read all of the game KPIs so it can actually move the audience through the beats of a game or rounds or however the game is structured. Yeah, I mean it it really has to do almost everything that a human has to do outside of, you know, posting on Instagram after the fact and having a life, although the best AI hosts that we have have insanely complex backstories. You can chat with them like you could chat with one of our humans, and they're going to have a pretty robust story that is their backstory, right? So we're creating these AI hosts to be as human-like as humanly possible. There's a ton of tech underneath all of that. It's not just the frontier models, right? I mean there's millions of data points per hour that come in that are live in the game, that the sort of logic of the game and of the broadcast needs to follow. We are, you know, we're obviously training the avatars from a performance standpoint, just like we would train a human. Everything from voice tonality and language patterns and body language and all of that's going into the AI as much as it goes into the human training. The AIs can chat in real time with the player chat. They can learn minute over minute and hour over hour based on the recursive data that's coming in, what's working and what's not working, or topics to stay on or stay off. So yeah, it's not unlike training a human. They need to do everything and probably more over time than the humans are doing. John Moore (29:48) You had a really cool story about when you were experimenting, and I think this is November or even earlier when you were experimenting with text-to-speech about when you saw some of that expressivity at least initially match what would be a minimum viable speech pattern. I don't know if you want to share that story about one of your hosts when you were doing voice cloning or if there were anything else that came up on the text-to-speech side that one you needed to have initially or you're wanting to have in the future. Geremie Camara (30:26) The voice is actually as important to us as any other part of the stack and most people overlook it. I know through all of the communications training I've had how much voice tonality matters to any communication. And so that piece took a long time to get to a passable place. And it's still very fluid, right? There's new models coming in every day. You have some of the big frontier models getting into the space. We cloned one of our original hosts' voice, who's been with us for six years, back in November, and the fidelity of that was astounding to us. It was one of those moments where you if you closed your eyes, you'd be fooled that it wasn't her speaking to you in real time. That said, at a production level with 24 hours a day times X amount of shows, we are looking at experimenting still. We're kind of hacking some of it because you know, at high latency, you can get unbelievably high fidelity and expressivity, and you can dictate all of that if you want to wait, you know, seconds, 8, 9, 10 seconds. When you get to our level of real time, up till now the models are less expressive. And so we've had to write a lot of custom voice manipulation to be both low latency and high expressiveness. I imagine the models will get better and better and better and lower and lower latency, and that'll come out of the box. But at the moment, you don't want to listen to someone for an hour if they're super monotone. So yeah, I mean we've just done a ton of custom work on that side. It's really important to us. The models are catching up, but we're always going to be experimenting and you know Whether it's the entertainment value, what the host is saying, the voice fidelity or the avatar fidelity, we're racing towards the Turing test. of this category. We want to get to a place where you cannot tell emotionally that it's not, that it's an avatar. And we're lucky with you and some of the partners we're working with that we're all chasing the same goal. John Moore (32:45) Definitely. What was harder than expected when trying to bring that real-time expressivity together, one, making it real-time and also keeping it hyper expressive and as engaging as you might want it to be. Geremie Camara (33:04) Yeah, it was just literally the models at the time didn't have that built-in out of the box. And so you had to do a lot of additional development that was very proprietary to us and our type of pacing and our type of expressivity. I mean we're not you know, there's a lot of the palette of creativity and human actors that you know, we're not doing Shakespeare, we're not like there's a lot of stuff. We just need people to be authentic and high energy and kind. I mean there's a limited aperture of what we need the models to do and so we can be really specific and really fine-tuned. You know, we're not building we're not boiling the ocean. We don't need it to on the one hand be sad and yeah we our hosts are trained to be a certain way and so we just need to mimic that the high you know, with a lot of custom development. John Moore (33:55) Agora's the Heartbeat, part of the initial goal is and Agora's mission vision around conversational AI is allowing you to experiment with different models, build something that is keeping up with the fast-paced changes in the AI ecosystem in general. Have you been able to take advantage of those and what's been helpful through being able to test different models and test different implementations that in production for different games? Geremie Camara (34:28) It's this nice sort of co-development model where you guys have an aperture on the world that is much wider. Convo AI is supporting use cases that are out of our scope. And so you're going to get requirements from many different sources around the world to experiment in different areas. And so to the extent that we benefit from that and we keep in touch and I notice you know, something that I can turn on and play with. It's going to help us innovate and help us stay fresh that you guys are doing more than we are doing. And on the flip side, in 2026 Q3, our ability to quickly go deep and customize things with agentic development is also incredibly exciting. So we can kind of rely on the heartbeat to stay strong, number one, and keep everything alive. And you guys are always adding new things to the mix that we can quickly would literally the flip of a switch turn on and test. And then also it's modular enough that where we want to go really deep and build very proprietary subsets of the stack where you guys, you know, for example, in the beginning, all of my deep dive around expressivity wasn't an Agora supported feature, but it was also easy for me to go off and dive deep and bolt that back into the heartbeat. Right. So it didn't matter that you hadn't come up with a client yet that needed that. We were able to take that little vector and go really deep and then add that back into the heartbeat. So it right now, it's been this good balance of a stack that's evolving, but that also allows flexibility for us to be really specialized where we know it drives. John Moore (36:22) It's a great point. I know we always talk about at Agora, we've invested the past 10+ years in building this infrastructure that supports hyper real-time and invested heavily in highly-forked WebRTC implementation that will work anywhere around the world. And it's a really good point, especially for some of our medium technical co-founders now that can take that infrastructure, combine it with agentic development, as well as supporting software development teams to build production-ready implementations, but take that and run with this very solid infrastructure that we've invested a lot of time in. So I think it's just awesome to see how you've taken that from idea to full concept to initial implementation and worked with your teams to then take those features in production honestly quicker than we've seen most, but that pace of play, so to speak, pace of iteration just continues to to increase. All right. Now I wanted to talk a little bit about the experience. I know we've said this is just November when we really started to put anything into testing. We worked together on defining an MVP and ultimately got that minimum viable product into production with you. So at a high level and we can go deeper, but what are some of the initial things that shocked you that you didn't expect when you took the ideation to concept to production with the AI avatar hosts? Geremie Camara (38:07) Well, we didn't expect the rate of innovation to look like a hockey stick this early. Back in November, when we initially did the MVP it was voice-only, right? And so we were super impressed with that. It was around March where we started to figure out that we could add the video on top of that. April or May, we had avatars in production with some of our games. We will probably be in a position to start sharing results soon. But the initial results, we have a lot of games that are not hosted, and so we know what that baseline is. Here's the experience, there's no host. We have a lot of games that are hosted 24 hours a day, and that's always a different level of KPIs. Candidly, we thought adding the avatars at this stage would slightly increase from no host to towards the hosted experience. It closed the distance a lot more than we had originally expected. Now, until we have avatars at scale, both 24 hours a day, different types of avatars, different types of products, we're not going to get our hopes up too much, But the initial feedback from a data standpoint and candidly from an audience standpoint is better than we expected. and we also know the shape of the curves at the frontier for everything in the stack. And so it's not hard for us to extrapolate out, you know, model updates on each of these areas of the company or of the experience. And they're not additive, right? They're multiplicative. So I think we have a unique maybe not unique, but a special window into what the future looks like and it's definitely something with that we are chasing and wanting to lead and innovate in. Like I said, we have a 100,000 hours of what is good already from a human standpoint, so we have a ton of a training material and benchmark KPIs and other things to make the models or make our not models but make our entire stack learn the fastest that we can make it learn. John Moore (40:29) That fits pretty well into that pyramid of the bottom chunk being AI hosted experiences, human hosts and those celebrities. Are the players, are the people participating in the games, are they interacting the way you'd expect? Is there something that you didn't expect when launching this? Geremie Camara (40:46) I think depending on the game and the avatar that we're using, so at the minute we're not trying to pretend that an AI is a human. In fact, I don't imagine us ever doing that. So the AIs that we have are discernibly planned AIs. We're not trying to say this is a human. There is experiments that we want to run around, you know, will players want to interact with true clones of the people they love? Knowing that this is, you know, we have a host Haley who everyone loves. Would players wanna, you know, but they only get her for four hours a day because she's a human and she has a life and things to do. Would they play with a hologram Haley if human Haley and hologram Haley could sync up every day and inform each other of what's going on with the players? I don't know. Interesting. We know characters have worked forever. Pixar or Disney quality, any of us who have kids and take them to a park and watch them lose their mind to go hug a stuffed Mickey mouse. I think again, there's so many ways this can manifest and we're just going to experiment with everything we possibly can and learn and double down on what works. But yeah, so you asked specifically, is there anything unexpected? Because we're not trying to trick people into it, even with the avatars that are discernibly not human, what we didn't expect is people will question it. Is this a human? So even when we're trying to back off the fidelity, it takes a minute for because they're so, the way they're hosting is so lifelike. John Moore (42:30) Yeah, what do you think that's related to? Is it the Geremie Camara (42:34) I think there's a if you come into the game, even if the avatar is discernibly not a human, but it's human enough and it remembers you and it says your name, there's a 'What is this?' kind of question. So yeah, there's a lot of chat. It's like, hey, 'Are you real?' kind of as opposed to when we do real characters like a teddy bear or they obviously know that's not, but even then we've seen people get attached just like you would get attached to a Mickey Mouse or whatever. If it's personal to you, then it's no less gratifying in some ways if Mickey Mouse compliments you in a way that feels good versus the human. So it but it's all it's all for us to learn. John Moore (43:21) What would you describe as I guess keeping the AI avatar on brand for you? What are some of the like core mechanics of the Live Play experience? And then how does the convo AI engine help with that? How do you use other technologies to help keep it on brand? Maybe put guardrails into the interactions? Just a wide-ranging question there. Geremie Camara (43:44) There's so many layers to the stack that keep that, all those guardrails in place and the experience to a certain degree. We're not letting the frontier go crazy and hallucinate and even though there's a million token context window that we're feeding with every utterance. All of that core knowledge about the game, about the players, about what to say and how to say it, and how many times to say it, and how to filter out things that are not to say, and knowledge of the game and knowledge of our third rails that we'll never cross. That's all internal. So convo AI is again it's a good, it keeps everything on track. But there's a ton of proprietary scaffolding that is on our side, and we're not passing that to any model actually, right? So at the end of the day, what the host is doing, how they're doing it, why they're doing it at a certain moment in time. What they know about the game, what they know about the players, what they know about how to communicate well, how to do calls to action, how to design behavior in the audience, that's all us. And that's obviously a proprietary harnessing and memory and recursive learning and a stack on our side. And then what convo AI is amazing at doing is keeping the heartbeat running because these things need to happen in a very succinct, coordinated set of steps. And there's a lot around just like a human heart, you don't focus on it beating. You shouldn't need to focus on making your heart beat. It just needs to beat. And everything else is downstream of that. We get to then instead of focusing on the sequence of the utterance, turning that text into a voice, and then turning that voice into an avatar and then repeating that over and over and over again with every utterance for 24 hours a day. That's kind of on autopilot. And we get to then dive off in each of those areas and customize. John Moore (46:00) The guardrails are definitely something we talk a lot about at Agora. I've gone to a lot of meetups where that's a huge part of the conversation. What advice would you give to other developers that are trying to balance that creativity and control of what that can do? And I know that might look pretty specific in the social gaming space, but just in general as you build, how do you think about guardrails and that ? Geremie Camara (46:29) I mean I think everyone now is talking about what is going to the frontier and what's proprietary to your company and you know, whether you're a bank or a movie studio or a healthcare company, everyone's going to have different core competencies, different domains of expertise, things that are incredibly sensitive, both from a competition standpoint and also from a regulatory standpoint, PII and protecting. So every stack is going to vary. But you wanna have an ability to benefit from bleeding edge of the frontier interchangeably if you're smart, and also have this core value proposition that you protect on your side. And yeah, that's just a matter of having smart architects that and experimenting, but really having smart architects that are building for performance but also building for protection both of your customer and your company. So I mean, you guys know you can go on any CNBC or podcast of anyone talking about AI, and that is the conversation, right? Do you hand over everything about your company to the frontier or any model for that matter? Or are you building and harnessing the right things on your side and using the best parts of the frontier when it's applicable? And we're trying to you know, invent the plane like everyone else as we're flying it. Every month, you know, you're having to upgrade and change and build new things. John Moore (48:02) I know a lot of people are looking at you and Live Play Mobile specifically because it has been so much conversation, ideation around this and live plays. Now got this entertainment, real-time avatar into production, which is a, you're in a really cool spot. You talk about cutting edge and being at the frontier and you're there. Are there any just coming back to the advice piece? Are there any common mistakes or things you might avoid when trying to get something into production again? or what would you say is something a dev team should avoid when experimenting with conversational AI? Geremie Camara (48:46) I mean not to plug you guys, but you know, your know the area that you want to invest in. And you know, we didn't have to invent the wheel around the heartbeat, which let us spend a lot more time building the parts of the stack that make it feel incredibly human. And we're lucky that we have tens of thousands, if not more, hours of knowing what that is, number one, we're also lucky in that we're in a product category, both in social gaming and in casino, where the dog has to eat the dog food, and this is a commercial enterprise, and the closed network of data that we have is incredibly valuable because it's high-value commerce. So the ability for the model to, or our stack, to perceive what's happening in the game, in real time, plan what the avatar is going to do, execute it, you know, create a performance in real time that is like human-like, and then observe what happened. And then perceive the KPI and go through that recursive loop against data that is commerce data, engagement and commerce, people actually buying the eating the dog food, so to speak, and have that all be in a closed recursive loop of learning is super valuable. So we get to focus on that versus recreating the wheel on the heartbeat, or you know, we're not building video avatar models. We're not building voice models. We want to use the bleeding edge in each of the categories to make human experiences work and delight our customers and make them feel just as connected as they are with the humans. So that's we just get to focus on what we do well. and I would say to any company, like just in the beginning at least, who knows what'll happen with all of this stuff, but in the beginning, try to put a stack together where you can continue to focus on what you do incredibly well and what's defensible to your product and yeah, lean on others where don't roll the whole thing. That's probably what I would tell. John Moore (51:06) Great advice. Tempting right. Geremie Camara (51:08) Tempting with it with agentic coding to try to roll the whole thing, but you're not going to at this minute you're not going to catch up on certain segments of it to where the frontier is, right? I mean none of us would go start a frontier model company today, right? John Moore (51:23) Yes. Even with I know just with my familiarity with Agora, this is ten years of infra infrastructure, SDKs, APIs, all designed to create the real time experience that you now have. And so that takes a while, right? Geremie Camara (51:39) And then there's also the part that we're not talking about with Agora, which is really the you know, the long-term secret sauce, the sub-second WebRTC distribution. At the end of the day, whether we produce a broadcast in a studio with humans, or we produce a broadcast in the cloud with an AI avatar, you gotta pump that through pipes that get to a million phones with in 600 milliseconds, which is also what you guys do, right? And so we're in that video, real-time video world either way. John Moore (52:16) It is cool to see how that infrastructure has gone from specifically human-to-human to human human-to-agent or avatar worker. And even now something that we haven't really talked about, but I'd love to take some time to walk through it because it is innovative and it is happening with you all. But one avatar or AI agent to many humans or hundreds or thousands of humans all in in subsecond latency. Stepping into the future, I guess just a look at what you're doing that's pretty innovative today as well and something you really have designed in many ways is the one-to-many avatar broadcasting solution. Is that something that just worked by nature of your game environment or how did that come to fruition from from your perspective? Geremie Camara (53:11) I think one-to-many works in a lot of use cases. It was the HQ Trivia use case that you know, I'm playing amongst thousands or tens of thousands of people, yet I have a chance to be special at any given moment. So I think that special amongst the crowd is the dynamic that drives this use case of one to many. It's kind of like if you go to a Taylor Swift concert with 60,000 people and she singles you out in the audience, there's that's a like a once in a billion year experience, right? That's what we, a version of that is what one-to-many does. However, we're going to play in every space. It could be one-to-many for 90% of the experience, and then it could go one-to-one which I think is a super special. The tech is now getting to the place where you can seamlessly cut between those two things, which we're experimenting with as well, and we're also experimenting with straight up one-to-one as well. And as the frontier and as the cost, token costs come down, right now it wouldn't be cost effective to do a million or even 10,000-person show all one-to-one. It just wouldn't be economically feasible. But that doesn't mean there couldn't be a lot of one-to-one side experiences or other types of applications of AI avatars. And at the minute, we're focused on making our content, all the stuff we publish and build, better. But probably in early 2027, we'll productize what we're doing and our expertise and we'll start to help other companies. And that could be in any category. It could be in live shopping, it could be in fitness, it could be in other game types. So yeah, I think we wanna be facile and innovative in from very, very limbically resonant one-to-one experiences all the way up to a one-to-a-million live event. John Moore (55:11) Very cool. We've talked a lot about your implementation so far. I want to step fully into the future. Let's look what's to come for Live Play Mobile, especially in conversational and voice AI. So where do you see AI hosts and avatars going over the next six months, two years, five years? Just a prediction on that and how that affects how you're approaching implementations. Geremie Camara (55:38) I think the voice models and the video, the avatar models are call it 80% of the Turing test at the minute. I don't know when we cross that line. It probably feels somewhere in 2027 where I could feasibly show you two things and you won't be able to tell the human from the from the avatar. Maybe it's early 2028, I don't know, but it's definitely knowing what I know of the companies and their models, it's close, close enough to be really doubling down. And then I think where do you take, you know, where do you take a lot of the experiences? We're just going to on a broadcast side, there are so many things that you ask what humans can't do, it's not so much necessarily what the human host can't do, just technically, you know, to build a set in real time and I think there's so much you can do in the cloud with an avatar and other elements of a show mini games that you can create in real time that the avatar can react to that I'd have to go hire a set designer to build for a month and then bring into my studio and train the human on how to deal with this Plinko board. It's just, you know, with agentic development on the front end side, plus the avatar, plus multiple avatars having conversations together. There's just so much experimentation that we can do to make these experiences feel really, really more and more special that only the biggest call it game shows, 'Price is Right', 'Jeopardy!' could and even then they're not iterating on their sets and what their humans can do that fast. John Moore (57:21) Brings in a good point with multiple avatars, also human host, avatar co-host. You think of so many shows that have the co-host. I don't know if you've thought about that at all. Geremie Camara (57:32) Well, I have. We're doing some TV production at the moment and we're talking to a bunch of other people about it and you know we are working with very famous classical TV hosts, late night and social influencers and none of these folks would have the time to come into our studio and host for four hours a day. I imagine a world where you get to the Turing tests, maybe there are celebrities that would be totally fine to have their likeness, their avatar part of experiences that, you know, because they know it's a really great experience. And so, you know, we're going to create our own IP, we're going to augment our human hosts. But sky's the limit, you're right. You could have a, you know, an AI famous influencer with a human having a talk show. And that AI, that human influencer who's wants to be in Spain on a boat might be perfectly fine having his avatar do it, and vice versa, right? And so there it's just too many possibilities to chase down at the moment, but they again, I think as we hit that Turing test moment, a lot of these things that seem kind of silly today might not be so silly a year from now. John Moore (58:50) Is there with all the possibilities, is there anything that you're most excited to build next with those configuration options? Geremie Camara (59:01) I'm excited to see if this is more of an experiment, we've built it. I'm excited to see how an audience would react to a human that they love saying, hey, I'm done for the night, but here's my clone. When I come back tomorrow, my clone is going to tell me everything you guys talked about. There are things like that are just pique my interest. They could flop and people could hate them and that's all fine. That's all totally fine. But there's so much to go build and test that yeah, there's no shortage of ideas that it's just picking the right ones or iterating as fast as we can. John Moore (59:40) Yeah, and when you're building and testing those things that might work and might not, you know it's not the quality of the expressivity, you know it's not the infrastructure that's the problem, and you're truly testing the concept at that point. Geremie Camara (59:52) Yeah, at this point we are testing all of it, but I think all of these things are going to come down to more and more how human can you make it outside of what it looks like? I mean the avatars already, they're already at a place where it's very hard for most people to tell, especially on a mobile phone. If it's blown up on an IMAX, yeah, you could tell. But it's getting close. And the voice tech is getting really close. John Moore (1:00:20) Very cool. All right, Geremie, with everything that we've talked about today, just in short sentence or phrase, the future of gaming is what to you? Geremie Camara (1:00:32) I don't know, but I think we're going to find out a lot quicker than we expect to find out. Because the tools allow us to create basically unlimited amount of experiments. I do think you'll see if it hasn't happened already, unicorn companies that have a solo founder. So what, you know, a year ago it would take me with a team of five for five months to create, you'll see a person or two do that in five days. So I think whatever the future is, or the futures are, I don't think it'll be static, obviously. I think the rate that we figure that out is just increasing like none of us could even have ever, ever contemplated. John Moore (1:01:12) Great answer. Well, thank you so much for being here, Geremie, this has been a great look at how live entertainment and gaming are being shaped and shaping conversational AI. it's really just great to talk about your implementation, the infrastructure that you use, and where you see it going. So thank you to everyone listening to another episode of the Convo AI World Podcast by Agora and many more to come. I know we have a lot more to talk about. Geremie Camara (1:01:44) Amazing. Glad to be here.