Kylan Gibbs: How Inworld is Bringing Games to Life with AI NPCs

26:20And like, that's a couple of years out. All these big companies are challenging, right? So if you look at any massive tech company, it means they've had some form of product success. Yes. In Google's case, it was, you know, largely search and YouTube and ads businesses. And I think that the thing that it also taught me and inform also my thoughts on like the viability of startups in the current ecosystem is when you have companies that large with a very clear growth trajectory and they're, you know, they're doing their quarterly reports, their interest is to maintain a current growth trend. Right.

26:50So you basically hit, you know, as a startup, you're trying to really have that like hockey stick curve of like growth. when you get to the size of Google, just maintaining 20%, you know, quarter on quarter growth is like immense. Like if you can do that, you're like, that's incredible for any given product area. And I realized that dramatic shifts introduced maybe even if it's like a 50 % chance of an upside, but a 50 % chance of like dropping by 50 % performance, they'll never take the risk because like, it's not worth it when you're dealing with billions of dollars in revenue. And I was like, what that that frankly, like was one of the things that we think about as we started in worlds, like Google can't just like step out of its current lane to go and move into these other areas because it's far too risky.

27:38And I think that's why we have startups and why we have this ecosystem, because these big companies are working super hard just to maintain their current businesses. And that's especially true for, I think, large consumer businesses where, you know, you make one small change. For example, let's say you changed out Google search to be completely generative and half the world for some reason didn't like that. All of a sudden you've just destroyed Google's core business. And I think that's one thing. And then I also realized like lawyers are really powerful. And, you know, we, there's like SVP level meetings at Google where like a lawyer, like SVP is like, this seems like a great idea.

28:15And the lawyer is like, well, there's actually some risk here. And then, you know, the conversation gets shut down for another couple of years. And so I was like, well, a startup would never have to go through that, that, you know, that process and review. And so it feels like a, a big opportunity. And I was like generative AI, especially at that point, I was like, I don't see this being deployed by Google, who yet innovated the entire technology itself. And so it was like, all right, we've got some blue sky here. It feels like within the past year, big companies have now gone all in on AI. Do you feel like they're actually moving quickly now, or is this all just catch up?

28:53There's a few fundamental innovations that we've seen around sort of like the image generation. But a lot of the products and features I see being launched now were possible four years ago. I mean, maybe three years ago. If you were really good at using GPT. The problem is it would have been super expensive. There was no one that had any expertise in it. Was it really viable? Probably not, right? But in terms of the technical capability, I think we are not using the state of the art today very much. We're still kind of using a lot of technologies that we're available for. um i definitely have i definitely have seen a huge shift i mean even in the experience with inworld you know we went from having conversations early on trying to convince people that ai like we were spending months just trying to convince large studios and companies that ai was even powerful to powerful enough to like to to drive the use cases that we were talking about and like we you know we did they were like i don't i just don't really think that ai is there like you know maybe in a couple of years, it'll be there.

29:55And then boom, like, you know, a year later, everybody has their new AI strategy. They're thinking about what their product and service looks like in this new age of AI. And everybody's kind of like rushing to that front. And that I, I mean, I've never personally experienced anything like that. Like the dramatic shift in general sentiment that kind of seemed to happen almost overnight. So big companies are slow and it's by design. They need to protect what's already working. And that's why startups exist. And I totally agree with that. And I think that's a great way to look at it. And InWorld exists to do something that no one was really doing.

30:33You gave a little bit of the story early on about how you founded it, but maybe just like at the highest level, tell us what InWorld is. What is the product? What is the service you're providing to the world? Yeah. So InWorld is, I would argue, the developer platform for AI and games and largely focusing on the interactive portion of games. So looking at how you can actually introduce characters like AI NPCs is the tagline. So we've had games for 30 years now where NPCs are kind of this static object that you go up to and you have some dialogue lines that you can send into, but it's really just like almost a non-living sort of static part of the game that kind of you know, how many helps give you information, but isn't actually part of the core gameplay.

31:20And that's why we've kind of, you know, there's a lot of memes that have even popped up right around NPCs. And I think it's become kind of part of the colloquial folklore. So the first thing we do is, you know, just implement that ability to take a character within a game and bring it to life. And that's really cool. You know, being able to have a conversation with, you know, your favorite characters in the game. But we very quickly realized that it's not enough just to have this platform to actually make characters talk. And so the platform itself now is really able to take the existing form of games that we have today and implement sort of this layer of AI on top of that.

31:58So that's taking a player going through a world and now understanding, okay, you know, what is this player's current mission? What are they supposed to be doing? And then using the sort of characters and world around them to actually dynamically adapt. So that means, the characters are able to speak to the user in a sort of very interesting, engaging way, but also to facilitate that player's progression along through the game. And that's done through the dialogue. It's done through them actually engaging with actions with the player. It's done through understanding the current environment and giving suggestions or hints.

32:30It's also allowing, for example, the way that the player speaks to a character to actually change things about the game world or to generate an emergent narrative. and you know i i used to think about this sort of from the lens of you know you're taking skyrim or you're taking assassin's creed you're making the characters talk and increasingly i'm realizing that this is actually implementing like a new form of gameplay most of the triple a studios we're working with aren't just interested in you know just taking ip and bringing it to life but actually thinking about like what does the next generation of games look like that was never possible before without these mechanics.

33:07So for example, some of them are looking at now these types of games where the core gameplay is actually that social interaction. So for example, you know, you're going along and you know, you encounter two characters and your mission now is to convince these two characters, like let's say it's a king and a princess, and you need to get them to sort of, you know, unite. So unite the two kingdoms. But in order to do that, you need to go and talk to like, you know, their three trusted advisors on each side. And you need to find out, you know, What are they interested in? What are their biases? What would convince them that you're a good person?

33:40You need to literally befriend them. You need to have like, you maybe need to manipulate or trick them. You've got to collect information from them. And the actual core gameplay, it's like the quote I've heard is like, we're not taking Legend of Zelda and making NaviTalk. We're taking Legend of Zelda and actually making the characters into the water temple. So the way that you actually interact with them, the way that you puzzle, like, you know, you solve social puzzles with them is the part of the core gameplay. And I think that will actually dramatically change games overall, because we're so used to the types of game mechanics that we've had in the past, which are very, as we were kind of talking about in the beginning, right, this kind of like jump and shoot dynamic.

34:18And now we are actually able to, you know, go through these like complex social simulations in order to progress in a game. And that, you know, completely, I think, changes the actual core of the gameplay. the other thing that i've really seen is you know by having these characters that are both aware of the world that are aware of you within the world and are responding to you in real time the it completely changes the capacity for role-playing and i think that's one of the things that i've been really fascinated by is how how core role-playing is to i think the the actual value that people derive from games and how characters can help for that immersion by making you actually feel like wow this is a living world that i'm interacting with it's not just a game that I'm stepping into and, you know, having much controls.

35:02It's like, this actually feels like a real place. And I feel like the consequence, like my, any actions that I have, anything that I say, anything that I do has real consequences for these characters within here. And that means that you, you, you feel a deeper emotional engagement with the world in terms of like the relationship with the characters in the world. And you have a greater sense of like your, your role that you're playing with in that world. And I think the way that I summarize this now is like the three things that I think we're kind of transforming is the way that role playing works within games, how we develop the relationship within games or virtual worlds from being this sort of external thing to actually feeling like a living place that has consequences.

35:44And kind of to that earlier point, it also changes the way that we think about progression. Like, how do you create those mechanics to unlock player progression within a game by actually having it like a social dynamic? And so that's the kind of outcome. But just going back to the actual core and world platform, like what we actually have now is this studio where you're designing the character brains and you actually go in and it's, you know, you're like mad scientist mode, actually designing the character brains, describing who they are, setting their emotional profile, how they speak. Then you're designing sort of world and narrative around that.

36:15And then what you do is you basically take that brain and you take your, you know, your game clients. So like, let's say it's Unity, Unreal, a web application or whatever it is. And you're literally kind of like, I think of like an avatar sort of moment, right? You're like plugging in the character brain into the inanimate avatar and then it's, you know, coming to life. But with that sort of world awareness and then you're actually driving that. So we have like the engine that's running, you know, 30 machine learning models of parallel to kind of simulate that real behavior while taking into account that world context.

36:46And so that sort of studio integrations and runtime is kind of how we think about the end to end platform. I can imagine how you do sort of the language of the characters and the speaking. You know, I imagine there's I'm going to overly simplify it, but you're using a language model and some sort of voice synthesis. But like, how do you create the movements and sort of like the actions that the characters take? That that seems like whole new territory. I can't really think of any comps. Yeah, it's very hard. So one of the challenges is like generative animation has yet to kind of really progress.

37:23It's, you know, there's a lot of papers that have come out, it's still really wonky. The way that it's largely done right now, and I think humans are actually somewhat similar to this is, there's sort of a mapping. So you basically have an animation set. So for example, you have a bunch of just some gestures, animations, actions that are implemented on the client side, you can kind of think about this. is like your canned options for different things that you can do in any given moment. And then we have that sort of on the client side and on the server side or on the model side, what we're actually doing is generating a bunch of meta information.

37:57So that's like what the character is saying. We also have like a thing where we're actually generating like a text description of the gestures and actions and animations that the character has. And then we do a mapping between those. So it's kind of like you're taking what is the character meant to be doing in terms of the context, what they're saying as well as the context. And then you're taking the sort of set number of options that you have, and then you're basically selecting the most appropriate ones for the given moment. And that creates this sort of generative capacity to have those interactions while still not requiring what ideally would be sort of like fully generative animations and actions at any given moment.

38:33That's largely how it works. And then we also have this system that we call goals and actions, which is kind of a logic that sits on top of the conversation and dialogue. And that's actually what controls what a character does in every moment. So the problem with LLMs is like, it's super cool, but it's really open-ended. And within a game, you probably want to have more constraints on what the character can do and how they respond at certain moments. And so this goals and action system basically can take in any arbitrary context trigger from coming from the world or any intent expressed by the user, and then directly tie that to a specific type of response or action from the character.

39:10And that gives you this ability to sort of have like embed heuristics basically within the world that allow you to align sort of that specific narrative or character behavior with the actual technical implementation. So yeah, it's a very weird system. And you realize how much, you know, as humans, we don't just generate arbitrary random behavior. We follow very specific heuristics that based on the context that we're in, who we're talking to, you know, what we just did actually result in very different actions, given even like similar prompts. How do you make it so that an NPC or a series of NPCs help drive the game to the appropriate conclusion while also at the same time, it sounds like the benefit of the technology is that these things are fully interactive now.

40:00You can have open-ended conversations. how do you ensure that they actually do what you know the story requires them to do people often think about it from like the one-to-one character level um i always think it's important especially in the game context to go like up one one i guess one level so you have kind of think about like a node like a graph where you have sort of the the narrative parts so you kind of have your introductory scene you know you've maybe got some like additional scenes that follow that and then there's like a decision point where depending on the certain things that i've done And as a player, there's like five different narrative branches that I can follow.

40:35And you can follow down those different points. And then depending on all of that, they might all converge back to like one or two different outcomes. And so you can literally think about it as like this graph representing the storyline of the entire kind of game or narrative. And then think about the character as kind of like a node within each of those, which within each of those kind of like narrative nodes. And so you're kind of like embedding the character. and so you kind of think about let's say it's like you know you're in a a detective story and i i come up to you and say hey michael you know your first node is introducing you to the world letting you know what the crime just happened um and giving you some sort of you know prompts in terms of what might have happened and what the things you want to investigate and you can then kind of talk to the characters and that sort of introductory node to learn more about the world and everything then once you're like okay i want to start the investigation you click over to the next scene which is like actually now starting the investigation.

41:28And the character's goals will change in that second part from being introducing you to the world and giving you context to now actually, let's say, helping you discover clues from the world, helping you interview other people, possible witnesses in that scenario. And then let's say you make a decision to upset a police officer and that leads you down sort of this one narrative branch. and so you kind of like you can kind of see that like the we're kind of maintaining state of this entire storyline and where the player is in that and the decisions that they've made and then the character's job we're actually using a bunch of the controls in the back end to prompt them to help you kind of progress on the specific storyline and even if you really try and deviate and players of course do this where they you know they try and break the system but ultimately we will push them back to kind of that core storyline and and this is super important and probably one of the most challenging parts of using generative AI in games is like, it's cool just to be able to chat but like as a you know, a game designer and narrative designer, you have a very specific opinion for how you want the player to be progressing through this world.

42:34And the dialogue and the generative interactions are really just there to kind of make it feel super open. But ultimately, the player is still following the storyline that the designer has set out. And how you add in all those controls to make sure that you're not breaking outside of that core storyline or outside of the world overall, right? So like if you ask a question about Joe Biden and you're in a fantasy world, like they probably should not have any idea who Joe Biden is. And so kind of keeping all those constraints on the kind of core storyline and narrative and information of that world is like one of the primary challenges that we're constantly solving with partners.

43:09Like setting guardrails for it to make sure they're not breaking the world. Is that what you mean? That's right. Yeah. So basically you can kind of think like three levels i think of as like character world and story so for the character making sure that they're not breaking outside of their given personality their assumed behaviors their identity um for the world so let's say it's star wars you know making sure that let's say you're pre when luke skywalker became darth vader imagine if like all of a sudden you're like oh you're going to become darth vader said something like that right i think it's going to break the world and so making sure that the world sort of maintains its coherence and also you know star wars shouldn't start blending with dune for example despite the fact there's like some similarities right between them uh and then on the last point around story like you know if i've gone through three main steps of the experience it should not then all of a sudden you know jump back to some prior um you know point in the story i need to make sure that i'm keeping that narrative progression and that storyline maintains coherent so got it got it tell us what it means to be chief product officer for in world like what what do you think about every day it sounds like there's multiple sides to the product.

44:13There's sort of the developer side, I guess. Then there's probably kind of the actual game side or the character side. I don't know. Tell us what it's like. What do you think about every day? I mean, I kind of balance between, so especially now we're kind of moving from a state of what I would call like internally motivated development to partner motivated development. In the beginning, it was a lot more of like thinking about what does human interaction look like? You know, how do we make characters interact like humans? How do we help people design characters that are, you know, related to humans?

44:44Because that was kind of our best benchmark. And before we actually had very specific requests from partners, our goal was just to create realistic interactions. Now, what I do a lot of is, you know, interacting with top tier studios, creatives, and understanding what is their vision for their games for their future part of media, and then thinking through, okay, how do we actually support that? And it's kind of like, Like, you know, people always go to the characters and the dialogue and everything. But usually I'll start with something like, OK, this is the game loop that you want to implement.

45:14These are the kind of general technical capabilities that you need to be in there that you need like the different architectures for. And then breaking that down into different parts. And so, for example, I mentioned that scenario before of like the king and princess. So I'd say, OK, we need, you know, a relationship system that's going to manage the relationships that like as you progress in relationship with the character, it will trigger something. I need to be able to create multi-participant conversations. So, you know, multiple characters need to be able to interact with one another. We need to be able to have some form of, you know, triggering system so that when you kind of unlock these different parts of characters, it triggers a change in the scenes and the kind of general.

45:49And so basically you kind of taking that customer requirement and then breaking it down to the different parts. And then given sort of like more and more technical background, I tend to try and go deep on certain parts. I've got also a great team, of course, who does this. but you know thinking through how we design each of those different components putting together requirements and then working with the engineering team you know at this point we have we're still all very involved you know we're not a tiny company anymore but we're still at the sides where like everybody's involved I think in everything and so a lot of the time is as well working through like really nitty-gritty problems for example like we're trying to solve hallucination has a problem right now, which is an unsolved problem in general of AI.

46:32And so I'll also be able to, you know, going deep in terms of, okay, how do we set up the data labeling task for that? How do we set up everything else? And yeah, I mean, it's, I always kind of think about the, you know, product as a general domain, as the mapping between problems and solutions. But for every single domain, for every single context, for every single team, that looks very different. But I think it's still like the best way to think about what I do today. It happens to be that at this point, though, a lot of that is also done through like organizational work. So whether that's like, you know, helping manage our internal team or helping manage our partners to either define the problem or build the solution.

47:10That's, I guess, where I spend most of my time now is probably like help helping work with people to accomplish those different parts. You mentioned hallucinations a couple of times. How do you actually go about doing that? How do you prevent hallucination within these models? There's like deeply deep, more deeply technical solutions. So for example, through fine tuning and RLHF, you can, which is reinforcement learning using human feedback, which is basically where you train the model using positive and negative signals from humans. You can take examples of hallucination and not hallucination given similar prompts and the model itself will learn to hallucinate less.

47:47Given a more tight domain, it's harder when you're trying to make it very general. So that's like kind of on the deeper technical side. And that's also done through things like data labeling and support on that side. that's like probably the main thing I just say, like, you know, model improvements to reduce hallucination. RLHF is seemingly one of the most promising areas there. And there's a lot of interesting new techniques that also allow you to do that faster and more efficiently. And then the other kind of core ways that we do this is one is like fourth wall is so fourth wall is kind of the feature that we define, which is effectively being able to detect when something is outside of a given universe or a certain conversation.

48:28And then we basically tell the model, okay, you should not, like, let's say I asked the question, you know, who is Joe Biden? And I'm in the Star Wars universe. We'll basically send a signal to the model to let it know that you should not answer this question, like, or at least express that you don't know who this is. That's like the fourth wall. And then we have another set of features that we call cognitive control. So fourth wall is kind of like thinking about it as like putting a ring fence around the given information or IP. And then cognitive control is I if I'm giving grounding information, so I'm actually uploading knowledge and facts about the world or character.

49:05Cognitive control is effectively forcing the grounding of the current conversation to those facts and making sure that it doesn't talk about anything else. So if you're going to be really strict, it's like I will only refer to the facts and information that I'm provided with, and we'll never go outside of that. And so those three main ways that we're those are kind of three main ways that we're getting around it is one is, you know, core model innovations, the fourth wall, which is kind of like the ring fencing of the of the context. And then the last one being the actual cognitive control, which is large, like people are calling it now like retrieval augmented generation.

49:36But it's that's kind of like, honestly, I think a couple steps behind because now it's kind of like, okay, we have we can we can use knowledge to augment the generation. But we also need to make sure as like a forced constraint to make sure that only that knowledge is used. And that's how we're solving it now. So for fourth wall, do you literally have to define everything you don't want the character to know about? Like, do you need to say you're unaware of Joe Biden? Like, how do you create that fence? Yeah, so it's this few different like layers. So what we have is like for like every customer, we have this like generic fourth wall, which basically blocks like real world information.

50:15and that's just trained by saying okay this is real world this is fantastical information and kind of generally saying like anything that is about the modern you know 2023 earth is kind of out of balance so that's like the broader ring fence and then with customers we build custom fork wall filters so for example if we were working with a you know a lord of the rings domain we could train it on all lord of the rings lore and then the model would actually detect is this lord of the rings related or is this not and if you ask about you know darth vader that's outside Lord of the Rings world. So that will trigger and you basically get it.

50:45So that's kind of how you can do it. So you kind of have that generic version and then you can build it very specific for a world. We've never gone this way, but like the system is very adaptable. You could even build it for a specific character and saying, you know, if you're talking to Buzz Lightyear, there are certain things that Buzz Lightyear will do and there's certain things he will never do or never say, you know, you could constrain that as well. And so it's kind of allowing for that general constraint over all of that. One of the things I imagine must be hard for building a product like this is because of the long development cycles of games, you don't actually see this product make it out into the real world sometimes for, I'm guessing, several years.

51:24Is that true? And how does that impact your product development? Yeah, it's a challenge mainly from the angle of scalability. So we have a lot of partners at this point that are doing play tests. we've also got a lot of partners that are much faster to launch so if they're not building a AAA game they might have a one to six month development cycle so we do have experiences that have launched within world powering them and we get a lot of data from that but certainly when we look at the AAA studios and large stuff like their the development cycles are very long and so a lot of it is doing that sort of internal work with the team trusting their development and design teams to understand like what will make a good game.

52:05And then, you know, basically doing lots of things like play tests. We also have, you know, our internal in worlds, like first party products and tools that people can use for testing. And so kind of trying to triangulate those two things. So we're, you know, working with the developers and designers from the AAA studios to understand, like, what are they working towards? And then we're using the examples that we do have live, which you know there's there's a few dozen um like smaller experiences now that are live and we can actually see the telemetry from those and understand okay you know for example when someone tries to bring up a not safer work topic within this domain like how do you respond right there's like very common issues that we see every day with a live experiences that help us inform sort of like how to improve the product overall but ultimately like our innovation is driven by those those larger partners um we also are now looking at there's sort of like for sure i think most of our business will be driven by those large studios in the long run.

52:59But we're working with a lot of other groups to see like that are much faster to market. So for example, streamers are a really interesting domain where we're starting to build products and tools specifically for because one thing that we've identified is, you know, there's like a I mean, one one estimate that I saw was there's like around a billion people who at some point every week watch a game stream, which is like a lot of people. And at the same time, streamers overall as a group is growing. and the interesting thing about AI of Howard games is they're much funner to watch and to stream because you don't know what's going to be said and every single playthrough is completely different.

53:33And so we've been building a lot of the early experiences on that so that streamers can take advantage of them, play them, and that also gives us distribution. That's like one area that we're looking at. A lot of the other ones, mobile games are much faster to watch. Sorry, play them in advance, you mean? So we have a few games. There's In World Origins, which is like a technical demo that we produced. There's a few other games that have already launched. And then we are building like some they'll be launching in a couple of months, but actually specific like products and tools that are very much built for streamers themselves.

54:07All this to say that we're kind of identifying certain segments of like the gaming market that are faster to launch and easier to get stuff out and get some early signal before we have those major AAA games coming out. Yeah, that makes sense. You're trying to find paths to kind of get feedback more quickly, given that the AAA studios probably just take a long time. That's got to be really challenging for a product team. I feel like, you know, I feel like all the products I've ever worked on, you can basically build something and launch it, you know, within a week or whatever, get feedback, change it up the next week.

54:38That's got to be really challenging. How do you think about AI in sort of non-character aspects of gaming? does in-world aspire to apply artificial intelligence to other elements maybe kind of you know the world around characters what's the opportunity like there yeah i mean admittedly i think we're kind of getting to a point with characters now where they're pretty good like in most cases it serves a large majority of use cases very well and so a lot of our efforts right now are actually going beyond characters um and so we we kind of generally like call two parts of our product, like the character engine, which is actually driving those character interactions, which is largely what I've been talking about.

55:20And then this contextual mesh, which is basically organizing all of that world information, the storyline, and actually grounding the conversations in that. And so we've been looking at tools, for example, to actually generate narrative. So that at design, there's kind of two parts of it, the runtime and then design time. So how can you actually use AI to not only drive the characters at runtime, but actually to generate parts of that narrative that are playing in the game. And so you're actually giving tools to narrative designers, to game developers, they can build their storyline, they may be able to also design things like dialogue trees if they want to use them for offline use.

56:00And they can even do things like auto-generate a lot of the character information and the world information. And so that also gives us, to your point, on the long development cycles, if we're serving more of that design time as well, it gives us a lot more early feedback because the end user in that context is the actual developer themselves. And so definitely, I think we kind of see this progression going from characters to a lot of the context around characters. And then we're even looking into things like the animations and gestures and assets that you were discussing before. And I think, frankly, we maybe just didn't anticipate how fast it was all going to move because we were like, okay, characters will probably get us all the way.

56:38And now we're thinking about what is a more fully-fledged, end-to-end AI game engine look like. And of course, characters, I think, still sit at the center of that because they're kind of like what you interact with in the world. But organizing all of that information and the quests and the story progression, you know, generating, like one of the things we're even looking at, like, could you, for example, in a game like Skyrim, actually just generate arbitrary quests that actually then organize all of the characters on the back end of that. So you could actually say, I want to go on a mission, you know, where I'm fighting with a wizard going against a dragon.

57:09and you just like kind of spawn all the game information and context around that. So you can actually have this kind of customized quest that you've just, you know, you're about to embark on. And I think that's where I see a lot of it going. And of course you want to respect the creative control of the developers, but definitely there's a lot of opportunity there. What's interesting about that is you could imagine that the games become playable for much longer, right? Like Skyrim could go on forever if it could just generate quests kind of infinitely on demand. which obviously like you feel like the game studios would love um just from like a sort of retention and kind of playability standpoint yeah yeah we actually did a gta mod um or in the summer which funnily enough got well a developer working with us did the gta mod built on top of in world rockstar um ended up having it taken down um but it uh it was really interesting it was basically a full mod of gta where you have this quest where as part of the quest you've got to interact with these characters and actually convince them.

58:08And there was kind of a social simulation stuff that I mentioned to you before. And people loved it. Like it was really, really well received. And I think looking at things like Starfield, Baldur's Gate that are, you know, these kind of primo of the new releases, every single one of them would benefit from the ability to basically once you finish the game or even the middle of the game to being able to extend that with arbitrary quests while still maintaining within that core game world. and I think we're also seeing a this sort of merging between I think narrative games and UGC where the ability for example for audiences and players to actually influence what is being built within the game itself and having an influence over the content of the game seems like a lot of interest and I think this is going to dramatically increase that replayability and also increase the kind of specification personalization of the experience to each audience or each player.

59:01I'm super excited about it. So, yeah. You guys have been thinking about the metaverse for a while. Obviously you mentioned that a Facebook shift to meta was a big moment for you. What do you think about vision pro and sort of what's coming within the next couple of years? Do you feel like there is going to be an actual shift now that people have been waiting for? So I see it on two angles. One is I, I think there is certainly a new form factor that is emerging with VR and AR. My biggest concern in terms of like general market uptake is like device adoption. So, you know, vision pro is 3 ,500 meta quest pro is, is cheaper now, but I am, I am skeptical of the, basically the price point that it's at, given that it's sort of still a a fairly small amount of experiences that people can play on it and then i also question whether or not people want to actually have these things on their heads like i don't when we started in world our main test platform was actually the oculus and i spent like four hours a day with that on my face and it did not feel good i can tell you by the end of it um so i i definitely think that the next generation of ai and gaming will involve um some form of ar vr i don't know if it will be adopted in sort of like the metaverse style of way where you step in this world.

1:00:22We, for example, just kind of as a point where I actually see a lot of promise is we did a project with Niantic where we basically created this talking character that you can talk to on your smartphone. You can have this AR-AI experience where you're talking to this character in AR. And there's a few groups surrounding that project that are also thinking about how do you actually integrate like like role-playing games into the real world so that you for example could be like walking around the streets of san francisco and you know whether you're wearing some ar glasses or you're just using your phone you can go up and talk to characters and they give you some quests you've got to physically move through world like like pokemon go crossed with you know like dnd um i see some really interesting potential there and i i think that there will be some experiences coming out over the next years that combine a lot of those form factors and i do think that gaming integrated into like everyday life uh is an interesting consideration and i for that reason see that like ar probably has more promise than a vr in the in the near term but i certainly also think that like you're gonna have the best like immersive experiences always in vr like we've done tests um with some of our like top end demos just talking to characters in vr and it's like it's crazy right you're now face to face with the character you're looking around and they're actually speaking to you and you know you and what's interesting is also we've noticed that when people play any um in world power games in vr they're much nicer to the characters because it feels like i think you have this more of this personal connection you're like i'm not gonna like you know i'm not gonna say anything provocative because it's like a real thing standing in front of me all right and so i'd start i certainly see all of that coming together i guess i just have some skepticism of how much how quickly the market will actually adopt vr just given the current state of devices.

1:02:12But I bet over the next five to, you know, five to 10 years, that will be the next kind of frontier for gaming for sure. Are there other areas where in-world tech and products can be applied outside of gaming? I mean, there's two big ones. One is what we call brands and the other is education. So brands is, you know, you can think about a ad that you click on on Instagram or Google. And instead of, you know, getting a blank page or whatever, a landing marketing landing page, you actually have an attractive character that you're now talking to. So for example, let's say, you know, the new Marvel movie is about to come out.

1:02:47You see an ad for the Spider-Man, you know, you click on it. You're actually now having a conversation with Spider-Man. And like the, I assume that the engagement there is gonna be much higher than it would be on a static page. And so I think the ability to take brands and actually inject them into through ads or online or even taking them, for example, and like there's a lot of, for example, Roblox games that are like Nike and Gucci world and whatever. And so how you, for example, also help brands enter the surfaces where markets, like where consumers already are, which is largely within things like Roblox, Fortnite, existing games.

1:03:19And I think there's some interesting ties between like core gaming business models there and also how we can inject this interactivity into that and then combine that with, you know, brand voices and marketing and all of that. So I think that's one place that we've already been working with. We've worked with some large partners on. The other is around education. So of course, the traditional way to think about this is, you know, basically taking a character, making it speak like a tutor or a teacher or whatever this is. We've got a few actually experiences that are powered by us now that are already doing this.

1:03:55So whether that's language learning, social skills, people are also doing like sales training, you know, manager training. You know, for example, you have to like let someone go. How do you do that? You know, do you want to simulate that experience? So like training and simulation is like definitely a big area. And one thing that's also been really interesting is also to think about those, that convergence with gaming and how people will likely build games in the future that are actually educational experiences. Because if kids are now spending most of their time on games, trying to pull them out into a traditional classroom setting or even just sitting in front of a tutor is probably really boring.

1:04:27But if you can make take a pick a Roblox game and make that educational and you learn about, you know like let's say roman history and everything else as you're playing through that experience that feels like a really interesting domain so yeah brands and education are the two other kind of large domains that we we see a lot of traction that's really really cool uh chat gbt we talked about was sort of this big inflection point for in world were you expecting that did i mean you've been so close to all this stuff for so long i feel like you probably had to know this moment was coming or did you not was it a surprise we no we there's a few interesting things So actually the guys, like my co-founder's previous company, they initially built it right before Siri launched.

1:05:08I say this kind of as an interesting parallel because they were building this conversation AI platform based on, you know, the previous generation using like dialogue trees and everything. And then Siri came out. And like the first reaction that you have is like, oh, no, this thing is going to take over the market. Like we're done. But actually what it does is it wakes up the market, right? to realize like, oh, this is a real problem and businesses should actually be taking this seriously. And ChatGPT was very similar. Like I knew for sure, like, I mean, just seeing the potential of what I had seen within DeepMind and even what we were already building in the world, I was like, I am shocked that people are not more excited about this.

1:05:44Like this is crazy, right? Like we can talk to AI the same way we can to humans and it can do other things that humans can't do. Like why are people not, you know, more interested in this? And so, yeah, I mean, first, actually, there was like mid journey and stable diffusion and then chat GPT. And frankly, from within world, I had in the earliest moments, all those launches, I was like, Oh, no, oh, no, oh, no, because it was you kind of assume that the market is a set size, and it's cannibalizing attention. But you realize at the end of the day, actually, what it's doing is like spreading out that attention across the entire global population.

1:06:16And what we saw happen was, you know, ChatGPT followed a lot of kind of like the coming open AI hype. But the form factor also proved itself and was really beneficial to us because clearly you can have these generative models that can just generate arbitrary text. But to put that in a chat format and to interact with it as a character clearly was what resonated with people. And so we kind of had this overnight shift when chat GPT came out from companies coming to us and saying like, oh, I don't know if interactive characters are actually that interesting to be like, wow, I just use chat GPT. How do I put this in my game?

1:06:53I get it. Yeah, they get it. Yeah. And and that was huge. And so I've realized now at this point that, you know, every one of these other product launches that comes out is actually a huge boon for us because at some point it hits a game studio and they're like, oh, we're going to hack around and try and build this ourselves. and then they run into a million and one issues and they're like, okay, we're going to go to whatever the best solution is, which at this point I believe is in world. Um, and so it's been really good actually to see how these other companies are pushing the market to kind of wake up to, I think what the potential is.

1:07:25What is the next shoe to drop? Like what will be the next big cultural moment for AI? Something I've been thinking a lot about lately is like this, the general significance of characters in our lives. So, you know, whether you're going into even like religious texts where we have characters that have shaped, you know, the history of humanity by telling us certain lessons of ways that we live to now, you know, in the 20th century, we had huge media moments, right? Around Disney and like every, all these Hollywood movies. And I think people actually identify a lot with them and it teaches them a lot about their lives.

1:08:00And we have celebrities and we have people that we come to love. I think having an AI character that is like a or a multiple AI characters perhaps who are actually like culturally significant so Chachi Wiki I actually think is kind of one of these if you kind of think about in a weird abstract way but actually having AI that are known to us as these like important entities so maybe it's going to be I don't know some pundit on some sports show or it's going to be something like uh you know a newscaster that's AI powered or it's going to be the next influencer and they're all going to be the AI influencers.

1:08:33I think having personified AI that have cultural significance to us that we're interacting with or watching or viewing on a daily basis, I think is going to kind of like shake the social fabric of things. I don't want to say if it's good or bad, but I see that being the big thing that will happen over the next year to two years. And I think it's already happening, frankly, I think if you kind of because of these chatbots as like the early form of that. And I think also then eventually that will blend with media and gaming and everything else. Like I think a lot of the beloved characters we have come out of movies and games and then thinking about how that gets integrated back into our regular life.

1:09:11Maybe you're, you know, your favorite game character is now you're following on Instagram and, you know, autonomously makes their updates and everything like it. I think that that sort of, yeah, representation of AI as characters in the cultural zeitgeist is the big thing that I expect to happen pretty quickly. Where can people find out more about InWorld if they're building a game or if they just generally are curious? InWorld.ai is our website. You can learn everything about us there. We've got our studio where you can actually just play around with the characters and actually try and create them.

1:09:45You can play a bunch of our demos and games that are powered by InWorld there. Awesome. And are you hiring? We are definitely hiring. Yes. We probably every role that you can consider we are hiring for. You know, we're looking for a lot of things from developer relations to business development, you know, deeply technical product managers, great AI engineers, you know, working on the research or engineering side. Basically, everything that you can imagine in a company we're looking for and, you know, really looking at people as well who are really passionate about the space and kind of the future of media and gaming.

1:10:22because I think that ultimately, you know, what we're doing, I think is pretty, pretty self-explanatorily cool. And it's, it's awesome always to find people who kind of resonate with that. And the team is in San Francisco or are you distributed? We have most of our team in Mountain View. So we've got probably, I think 40 % or so of our team in Mountain View. We've got an office up in Vancouver, a few people kind of spread across the West coast, some on the East coast. And then we've got a smaller group of folks over in Europe as well. And so it's, you know, centralized to a degree, but yeah, with a really, really good, like remote presence as well.

1:11:02So everyone is, everyone is welcome. Awesome. Kyle, this has been fascinating. Thank you so much for giving us so much of your time. This was a great conversation. Thank you so much, Michael. thanks for listening to generative now if you like what you heard please rate and review the show on apple podcast and spotify that really does help and if you want to learn more you can follow me at mignano on all the socials or you can follow lightspeed at lightspeed vp on all those same socials generative now is produced by lightspeed in partnership with pod people i am michael mignano we will be back next week with another fascinating conversation see you then Bye.