Podcast

Dhaval: This founder and his team built a platform to create life-like characters for immersive experiences like games or books. In this show, Kylan Gibbs shares his thoughts on generative AI and how to manage products where you create immersive experiences. He also discusses his career journey. Kylen is a former product manager at DeepMind. He was also a consultant at And a co-founder at Flow X. Now he’s a C P O and co-founder at In World AI.

Hey Kylan, welcome to the show. Tell us about where you are in your product journey, a little bit about your product as well.

Kylan: Awesome. Yeah. Thank you so much for having me. Super excited to be here. So At InWorld, we’re building a creative suite of tools that allow people to build these AI characters and then integrate them into immersive experiences, games, entertainment, enterprise experiences as well. And we started just over a year and a half ago. And basically where we’re at now is we have this studio where people can come in and actually craft their characters. We got integrations with things like Unity, unreal, where they can bring them into games, as well as Node for actually bringing these into web experiences. We’ve also then got sort of this arcade and, basically, which is a way to actually share these characters directly to the web. And then a suite of experiences that we’re gonna be releasing this year that are self-produced. So we’ve got, for example, in world origins to sort of show off the power of the product. And this is all to sort of say that this is kind of crafting that end-to-end user journey of being able to build these characters and then bring them into worlds. And basically where this has all been going is kind of setting the foundations of actually being able to not just create the characters, but, build them into experiences and deploy them scalably. And I think, compared to a lot of the things you’re seeing in generative AI right now, like it really is production ready. And we’ve been focused on that. And so we’ve already seen a lot of developers starting to churn out games and experiences now that are integrating in world characters. So it’s super exciting to kind of see that. So, of course still early on, I’m excited to see where it goes, but already seeing hundreds and thousands of users using it, plus, actually seeing live experiences is pretty magical.

Dhaval: Wow. There’s a lot there. So let me just quickly ask you a follow-up question on how you support game creators. Is that right? Or do you support all kinds of creators, like writers, novelists, or just focus on gaming experiences at this?

Kylan: We’re supporting really any type of creator. So ultimately, we have users who are, for example, very, well-known science fiction authors who are using their characters to iterate on experiences and potentially write new books with, we have people actually building like AAA games and these types of experiences where you’re building multiple characters and integrating them into worlds. We have entertainment companies who, you imagine integrating these into parks or like live experiences where you’re actually interacting with. And then we also have enterprises who are using these for things like, brand representation, corporate training. So really we’re open to any types of creators. Of course, we’re really focused on narrative oriented, immersive experiences, and our product is best built for that. So, you know games, narrative, entertainment, you can think about the adaptation of movies and IP to these experiences. And that’s really what we’re targeting. But we’re really open to any types of creators, and we’re finding new ones every day.

Dhaval: Yeah, I picked up a keyword there. Narrative oriented, immersive experiences that could be representing multiple customer segments. What is your product journey like? How do you go about defining your product capabilities with such a broad range, range that you know you could be?

Kylan: I think that, so abstracting out of our specific use case for a second, like I think when you’re building a developer tool, you always have two customers. You have to think, keep in mind, one is your creators, your developers, and the other is your end users. And so for us, our creator journey is really people coming in. They have ideas either their building an existing experience or they’re ideating on a new one and they’re using the studio and our characters to basically iterate on that and then integrate it through things like Unity and Unreal to actually bring those to users. And so when we think about success in that, it’s basically: is this person able to create the character that they love and like, and kinda ultimately represents the vision that they have and fulfills the purpose? And then are they able to integrate that and deploy that successfully. Then you have the actual end users, which are the people actually interacting with the characters that the developers built. And that is really like, is the interaction enjoyable? Is the person you know staying around to talk with this character? Are they finding out what they need to progress through this experience? You can imagine throughout a game you have someone that’s a guide or shopkeeper and they have to fulfill a particular role. Are they doing that successfully? And so you really have to balance the two of those. I think that’s true for most developer tools, but it’s kind of unique here because, ultimately there’s a key part here, which is like, it’s the generated content. So it’s almost like we are allowing developers to create characters that are fulfilling the wants that they have for the users in the end? And so it’s always a little bit of an art and a science.

Dhaval: Yeah, there is a lot of art there, especially when it comes. The narrative, the experience, right? The character creation and then narrative and the experience. What are some of the ways you create these characters that actually immerse themselves and are conducive to the experience? Is there a difference in product creation? How do you actually adapt to the narrative or the experience that the creator is trying to have?

Kylan: Yeah, I guess there’s two points, which is like design time and run time here. So at design time, you could take for example, Arah LA a large language model and generate sort of responses that are aligned with a particular character. But getting them to do that reliably and stick to a story and actually fulfill goals and actions is very difficult. So we actually allow users to come in and they can specify, for example, scenes for their story, and then the characters will actually stick to basically, the kind of motivations and goals that they have for that specific scene. Then we allow them to specify, what is this character supposed to accomplish in this? How are they supposed to speak? What types of things in the world might they be reacting to? And all of that sort of is going towards controlling and biasing the character behavior in a specific moment within a specific story. So they’re fulfilling what they’re supposed to. And that’s all kind of the design side. When you’re actually interacting with the character. Then we introduce, for example, emotions. So the characters actually react with emotionality. We have voices that also integrate that emotionality, so the characters can actually, you know, you could hear when you upset a character, for example, and you can react to that. We then control gestures and animations. The characters can actually react to you, or you can ask them to perform a particular action and they can actually act on that. And so, that ability to actually have, and we’ll be releasing this soon, and we have a new system where you can actually, for example, give a character a goal and a series of actions that they can use to pursue that goal. And they’ll autonomously pursue the goal until they’ve accomplished it, which is like a pretty magical thing if you think about the ability to actually create this living thing that’s kind of, pursuing goals and motivations. But of course all of that comes back to the ability to actually drive this story or narrative forward, whether it’s something like Assassins Creed or Far Cry or one of these games. Or you can think about even in an enterprise experience where you’re trying to usher a user forward through sort of a brand experience. All of it is really the key point is that the characters are filling a specific purpose in some broader experience. Yeah. And so we’ve got a huge amount of controls that are available to enable.

Dhaval: That’s very interesting. You mentioned that you are able to add emotionality to a character. Wow. That blows my mind. Is there anything more you can share about how you go about doing that? I’m personally interested in that because a few years ago I created a startup to inject emotionality in writing. And that time there were no elements or any of that available, and it was pretty challenging, but I would love to hear how it’s done in 20 23, 20 24.

Kylan: Yeah, so right now our system is really composed , so when you as a user are interacting with a character, you’re taking all of that input. So whether it’s audio, visual sort of event side, so you imagine a you unreal, you have these events that happen and all of that get fed through basically a series of, you can think of ’em heuristics models or models that are representing different parts of function. So for example, you have a goal-oriented module, you have emotions, all these kinds of things. And so emotions are literally kind of, you can think about as an engine. Is taking in the character’s current state, taking in all those inputs from the user, updating the current emotional state, and then that feeds out emotion as basically one input to the overall behavior of the character. So then they’ll respond to you incorporating that emotion into their conversation or their visual representation, all these kinds of things. And so literally the emotion engine is effectively this module that takes in inputs from different. You know, updates basically this emotion, which is represented as a continuous state, and then outpays, a discreet state, which is like, happy, sad, frustrated and distressed, these kinds of things.

Dhaval: Okay. We are gonna switch up the gears a little bit. I think you have, this is not your first AI startup. I think you have created many AI startups in the past. Tell us about your journey leading up to this involved experience. And what were the learning lessons that you had to share with aspiring product creators? What are some of the, so first, set the context and then tell us your learning lessons.

Kylan: Yeah, so my background was pretty, atypical. I didn’t set my trajectory to become an AI product leader. It all really started with like, I was actually starting to go into politics, and then realized that nor to aggregate people’s interests and insights and actually understand what they really needed. You couldn’t do that manually. You need to do something like data science or machine learning. And so I actually started using it, for that purpose as basically working with the city to understand, you know, population interests. And then, my first steps into the startup world where after I finished my masters, I worked with a few friends to basically create a company where we were applying machine vision to CCTV, which is like security cameras. And basically extracting out information, and the idea there was to be able to feed that then into traffic systems. You could actually optimize traffic systems in real time in cities in the uk. We were working with Singapore and the UK at the time. And that was a super interesting first learning experience because of course it was like, going through the, straight out of university doing the startup life was very different than it was this time around. But it taught me a lot about, I think like that, the importance of. I think one, accepting the technical complexity of what you’re working with, which I think can be very Dramatically underestimated, and really being realistic with yourself about what you can do and how much you can accomplish, without a team. And then also thinking a bit more scalably, for example, working with the public sector is a big challenge, so maybe it’s not the best place to start out a machine learning product and these kinds of, And then I left that and went to consulting. So I worked at Banning Company and helped set up machine learning practice there and worked as a data scientist cross consultant, which was very interesting and got exposure to a lot of industries. I think the key thing was at that point, which is several years ago now. Getting in uptake from a lot of these large organizations, especially around innovative things like machine learning at that time especially, was very challenging. And also the way that consulting firms or large corporate work sometimes is amenable to building really hardcore tech. Whereas, you’re basically focused on churning out client projects versus actually building something foundational. And so you have to put the time into that. Then. I was lucky enough to meet someone who was moving over to DeepMind while I was at Bain and transitioned over to DeepMind. While at DeepMind, I first started working, helping set up the research strategy and doing, you know, basically we were doing a meta-analysis of research across, you know, the industry to inform like research directions. And then around that time, sort of G P T three came out and a lot of these sort of large language model efforts came out. And I was working then as the product lead for conversational AI and generative models, as they pertain to sort of your language tasks and these kinds of things at DeepMind. And so I was working across Google, to basically integrate DeepMind tech into Google products. So it’s super interesting, you know, basically learning about, even within Google, where are the challenges of applying generative models and like, you know, this more advanced AI, which is a lot more challenges than people would probably anticipate from the outset. And then that also informed, how I thought about, okay, the benefits of being in a large company versus being in a startup or working in product in a small startup where you have a lot more capacity for experimentation when you launch something, it doesn’t have to work for a billion users from day one. You have that capacity to work with a smaller group of users to refine a product and then actually be able to scale that up. So while I was there, I met Mike, who is our current co-founder and c T o, who is leading AI and ML at Cloud conversational ai. And then Ilia, who previously founded Dialogue Flow or API at ai, which became dialogue flow when acquired by Google. And so that’s how Inworld was born. Yeah. And then I think there’s of course, this has been sort of a rocket ship, so there’s been a lot of journey learning along last year, so I can pause there, but I’m keen to share as well. But a bit about what we’ve learned in the last year or so.

Dhaval: Wow. There is. Five careers in that one career. Thank you for sharing that. What do you recommend, what are some of the things that you recommend avoiding new brand new pm just thinking of becoming a product person or a product creator. What are the things they should avoid doing? Absolutely. If they want to be thriving in the aI space.

Kylan: Yeah, I would say, to boil it down, I would say like, don’t philosophize. What I see a lot of people doing, I think because they’re mimicking what, so when people have been successful towards the end of their career, they start to, philosophies and they go on Twitter and they go on, post blog posts about all their philosophies and learnings. And so what I see people doing at the beginning of their career, Is basically starting out with that versus actually building something that validates what they’ve actually done and proves that they’re valuable. So don’t start by trying to teach people lessons before you’ve actually learned anything. But that also translates, I think, into actually how you build products. And I think it’s increasingly where. The differences between even software development as it existed 15, 20 years ago, and what we’re doing now with a lot more AI and ML is that things need to be a lot more iterative because outcomes aren’t as deterministic. So you don’t know, for example, if a feature is going to be successful. One, because the actual development of a new machine learning model doesn’t directly translate into something specific. It’s not like you’re adding a button to a page that you know has some outcome for a user. It may or may not improve, increase, performance and quality in some other ways. And so what I frankly learned the hard way was, you shouldn’t set out, for example, like long-term visions or roadmaps. This is like a very classic thing that I see being told to a lot of early PMs is like, build out your roadmap. Build out your five year roadmap, break it down to one year, and then six. I think that the process is a lot less sexy, but is ultimately far more important. And I see so many PMs, like across their career base, like, oh no, I don’t wanna be involved in any organizational stuff. Like I don’t want to be involved in the Jira tickets or these kinds of things. But I think a really good PM. Is basically able to take the vision that they have in their head and then turn that into basically a machine or a process that naturally pushes people towards the direction that you’re working in, but is able to adapt to things like the market pressures, the reactions, the user feedback. Because if you basically plant a point on the map that you’re working towards, you’re gonna be so biased by every time that you perturb ba from that point that I think you’re not gonna end up in the optimal state. Whereas what you really want to do, and this is coming from the AI ML world, in AI and ml, you have this idea of gradient descent, which is basically you’re doing an optimization problem where you’re trying to find like the lowest point in a space, I think of good product building as basically inverse gradient descent. So you’re trying to always find the highest point. And if you say, I’m gonna, getting to the top of that hill is my goal. Hands down, that’s all I’m gonna do. You might be missing the fact that there’s like a mountain on the other side of that. And so you should build yourself, you should build your product, you should build the organization and the process around your product all to be able to constantly take in that feedback and shift dramatically. So for example, I’ve now set up our organization to be more focused on smaller iterative cycles. So, breaking down, you have like larger term, longer term efforts, which you know, are important and thematic. And then you break those down into two week increments. So it’s like standard, more standard sprint planning. But the challenges that With A N M L products, it’s a bit more experimentation. So it’s not like you’re necessarily saying we’re gonna have this launch and have this impact in two weeks. I’m gonna run this experiment within the next two weeks because I think it might have this outcome, and if it’s successful, we’ll build on that. And so it’s really like taking that long-term vision that you have, setting the metrics, setting the sort of general directions that you have, setting your feedback channels from. And then being able to constantly turn that into experiments that you wanna run, both with developing hardcore tech. If you’re doing a deep tech startup or, turning that into, user research or user experiments. You’ve been doing something that’s more front done based.

Dhaval: Very interesting analogy there with gradient descent. I often use the same analogy, but I use a different term for it. Global maxima, find your global maxima. Yeah, I love that. Let me ask you this question. Where do you. This is going, this whole, generator AI especially. And then if you wanna narrow down to your specific space, I would love to hear that.

Kylan: Yeah, so I, I generally think where we’re moving towards is, In the same way that you had basic automation of things like databases and things over the last 20, 30, 40 years. I think we’re moving to a place where a lot of like hardcore knowledge tasks are going to be automated. And ultimately, this means things like you imagine as a PM creating a P R D, you can probably generate a pretty good PRD of chat GPT, barely better than most starting PMs to be honest. And, you could probably, you can create, there’s like slide generators now for people who are in consulting. There’s a lot of these things. So knowledge tasks are becoming more automated and I think it puts a lot more pressure on ideas and creativity as like the actual gold of what actually humans contribute to this whole process. And so what I see happening over the next 10 years is, A lot of the things that people held as basically their like unique value are going to be eroded as what happens like with any technology innovation. But what that means is, so let’s say that was only 1% of the population, now those skills are accessible to 80% of the population. So the overall value created by everyone goes up dramatically. But it takes away the feeling, the niche, sense of value that small group once had. And so what I think needs to happen is for people who are very, you, Highly technical knowledge oriented that feel like this is impacting their work. The question is not “oh no, how do I react to these things being taken over?” It’s like, how do I also now become an expert in taking advantage of these tools? So, for example, in our case, so you have. Amazing game developers, narrative designers, writers who are creating video games. And it’s not the fact that creating these AI characters who are able to basically generate dialogue is gonna take away from that. It now means that a single writer could create a hundred characters and create all the dialogue for them in the time that it would’ve taken them before to create one character. And what this really means, I think, is you’re just gonna have an explosion of content overall. So both in terms of games and things that are directly impacted by us, but of course there’s also the image generation and text generation, and I see music generation happening now, and so. I think we’re about to enter an era for probably the next five years of just like a content explosion where there’s just gonna be so much of everything and a lot of it’s gonna be complete noise. And so I think the key thing for like those people who want to continue to be leaders and product leaders in this space is not what can you create, but. How can you filter out the noise? So how can you ultimately define what good is? Because it’s gonna be a lot easier to create something. And so I think that like over the last, while we’ve seen the benefits of people who are just able to build some hacky prototype. Building a hacky prototype is gonna get really easy. So now you’ve gotta figure out like, what are you actually creating that is completely unique and like applying some creative sense, not just necessarily some technical or general sense, or applying some framework because guess what? All these generative models have now taken in that framework and know it, so you don’t really have anything special. So you have to come up with something actually new now. And I think that’s where it’s all going is like a content explosion that biases us to really thinking about what the content is that we want to create, not just the fact that we can create content.

Dhaval: Yeah. I love that. Thank you so much Kylen. Thank you for making time for. Really appreciate it. Looking forward to having you back on the show once you have a few more lessons to share with us.

Kylan: For sure. Thank you so much. Yeah.