Podcast

Dhaval:
Welcome to the podcast, Nick. Tell us a little bit about your product.

Nick:
Yeah, our main product is AI Dungeon. So it’s an AI powered role play adventure game where people can jump into variety of different AI experiences and make choices that result in a fun and interesting story. So, traditional text adventures where there’s a limited set of options you can really select and you go down a path that a developer pre-imagined with AI dungeon every time the story is unique written by an AI. And so you have freedom to create all kinds of different stories that the developers us would never have imagined possible.

Dhaval:
That’s amazing so. Is this like a video game? I’ve never played Dungeon before. So what is this for people who have never played Dungeon? Is this a video game? Is this like a regular board game? Tell me a little bit about that.

Nick:
Yeah, it’s like a video game like a classic text adventure game where the game shows some story of where you’re at and you type actions and then there’s a result. It’s a little bit more towards the end of a creative sandbox than it is a really structured game with lots of mechanics. So it’s more on the story roleplay side.

Dhaval:
Got it.Wonderful. So how do you serve your customers? Is this a mobile app? Is this a web app?

Nick:
Yeah, We are on mobile and on the web both.

Dhaval:
Wow Okay. And is there a specific segment of your customers, of video game players or other players that you address with your game?

Nick:
Yeah, I think for ours, we’re especially targeting kind of role play gamers. So gamers who are interested in role-playing games like DnD and tabletop ones, as well as our traditional game RPGs, like Skyrim and things like that. Where you get to kind of decide who you are and what direction, and what actions you want to take in a more open environment

Dhaval:
wornderful Tell me a little bit about your own journey. Are you a game developer or are you a techie who got interested in games? Or are you a business person who decided to try a new idea? Tell us little bit about your personal background.

Nick:
Yeah, definitely the techie category. So, I come from a machine learning background. Before I ended up doing this, I spent a couple summers at self-driving car companies. I was going to go work at Aurora. And this side project that I’d been working on AI dungeon, just kind of took off on the internet and I realized there’s something cool here that people are excited about. So, I pivoted from where I was going to be a founder. So I have learned lots about kind of game industry and the business side since starting Latitude. But going in, I just had the kind of tech machine learning AI background.

Dhaval:
And Is Latitude already generating revenue? Is that a revenue positive organization?

Nick:
Yeah, we just reached profitability pretty recently.

Dhaval:
Wow. Congratulations. That must be a big deal. So if you could share any information on your revenue or your number of users, or any capital that you may have raised that will help us understand the context of your product.

Nick:
Yeah. We’ve raised over 4 million. Since we’ve reached profitability, we haven’t had to raise in the last little while. We’ve had millions of users download AI dungeon and we have a pretty active excited player base.

Dhaval:
Wow. Millions of users profitability. And when was the last time you raised?

Nick: End of 2020.

Dhaval:
Wow, more than two years ago. So you are a fairly successful case study here in terms of building an AI product. I would love to dig in a little bit here in terms of what was your journey like in terms of building the product? So you said you are a techie, you have background in machine learning. Did you find some video game players, video game creators? Did you have a passion in this space yourself? How did you go about creating a business model and a product canvas for your capabilities?

Nick:
Yeah, it’s kind of funny because I think I very much came at it from like hacker creator who got kind of surprised. And so I feel like it’s taken me a while to catch up on some of these other things around understanding our product, what it is kind of game industry space. So the way it really got started was just at a hackathon. I was just playing around with the smallest version of GPT-2 had just come out. So it was a hundred million parameters, a thousand times smaller than the language models of today. But even then I could see how this kind of AI’s ability to do dynamic storytelling was really fun and interesting. So I got super obsessed with it. Worked on it over the next like nine months. I would have my friends play and I’d just like watch them and see kind of what their experience was, what was challenging and what made it more interesting. And so put it on the internet, not expecting that it would really launch off and be a product. And, so I was kind of surprised of scramble after that to figure out how do we make a product out of this so that people can continue playing. Because when it first launched, it was on Google Colab. And there was so much traffic. I knew that they could not let it be there for very long. So, I wanted to get an actual app and product place where people could continue to play it and not rely on Google Colab’s GPU generosity for more than a little bit.

Dhaval:
Wow. Colab, for people who are not aware of is like a script writing tool. Is that right? It’s like one of those web editors in which you can write code. Is that true?

Nick:
Yeah, the thing that was cool about it is Google generously lets people use it and run on their GPUs for free. And they have paid plans as well. So it was a way that. People can run machine learning models and try them out themselves, and so that’s how AI Dungeon was first distributed was just as code that would run in these Google Colab notebooks.

Dhaval:
Wow. So what was the journey like for you when you finalized the product and you are ready to launch it? Did you partner up with other business co-founders or are you the sole founder of this product? Tell us a little bit about your partnership.

Nick:
Yeah, once the game really started kicking off and I realized there was something interesting here. I pulled in my older brother who’s worked at a bunch of tech startups to co-found a startup around this with me. So for the first couple months, it was just us with a couple people advising or helping out here and there. And then we raised our first round of funding and got kind of initial team together. And then we started going forward from there. So that’s kind of what it looks like.

Dhaval:
Got It. So when you started this, you built this on top of GPT-2, which was a thousand times smaller than what we have now. And when you say a hundred million parameters I’m sorry, 10 million, did you say 10 million?

Nick:
Yeah. It was the smallest version of GPT-2. The very first version, the version that took off was a billion parameters. So it was about a hundred times smaller.

Dhaval:
Got it.Now, did you create a lot of custom capabilities on top of the foundational model or was that built in and you just used it in a clever way?

Nick:
Yeah, we did fine tune the model on second person kind of story adventures so that we would have an understanding of what that format looks like. Especially because there’s not a huge amount of that kind of text format on the internet. And so I don’t think the base models had a great understanding of what that was, especially earlier on when they weren’t as smart as they are today.

Dhaval:Got i t

Nick:
There were also a lot of things around how you manage what you feed into the model, what it comes out, how you alter the output so that it would feel more like a chat experience where the AI is telling you a story rather than stopping mid-sentence on the way

Dhaval:
That was also a big effort in your engineering, in your output engineering for the prompts. Is that what you’re saying?

Nick: Yeah. In the early days. Getting that piece right.

Dhaval:
Yeah.Where did you spend most energy in the early days? Fine-tuning the model, getting the data, training the model with the fine-tuned expectations. What was that like? What was the energy expenditure in the early days like?

Nick:
Yeah. So Pre-launch it was a lot on the dataset. Changing the data set, like kind of getting it into the right point and fine tuning it so it worked well. After launch, a lot of people nowadays use language model APIs, but back then there was not language model APIs. We had to spin up clusters of hundreds of GPUs serving GPT-2 and people weren’t really doing that yet. And so we spent a lot of time just managing those, trying to keep them up, trying to request more and more GPU quota from Amazon AWS because they don’t like it when you want a ton of GPUs really quick. And a lot of the early days was just managing and adapting to the huge amount of traction we had early on.

Dhaval:
Interesting. So, would you say that falls in the category of machine learning operations, orchestrating all the different pieces together to run an app together that has a foundation of ML? Am I getting it right? Was it ML ops or something else?

Nick:
Yeah, I think you could call it ML ops.

Dhaval:
Okay. And then your strength is in building ML models or is that in making the business case out of the ML models? I believe it’s a former, but I just wanted to unpack that a little bit.

Nick:
Yeah. I think we’ve even kind of changed how we’ve thought about it over time. So we used to consider ourselves more of like an AI startup. And I think in a lot of ways we are, especially compared to some, but at the same time we don’t really think of ourselves as an AI startup anymore because I think that’s it’s easy to think of like the hammer you have relative to the nail. Really what we care about is helping users have a rich, like meaningful role play experience where they can jump into the shoes of a character, be someone they’re not in normal life, and go on you know exciting fun adventures. And we use AI to enable that much more effectively. But AI is not the goal. It’s just one of the big tools we use. Like If there’s lots of other work we do to enable that kind of fundamental human desire that aren’t AI. And so i think now like we certainly have a lot of things we’ve learned about how to leverage AI models to do that, but I think we’re especially focused on that user desire and need and fulfilling that and using AI where appropriate.

Dhaval:
Yeah. Interesting distinction. You are not a AI company, you are a game company that uses AI. Interesting distinction.

Nick: Yeah.

Dhaval: What was the journey like for you in terms of experiencing all the growth? What changes did you make in the fundamental technology once you reached profitability or once you reached a certain level of confidence in your product?

Nick:
Yeah. I think one of the hardest things was, has always been the unit economics. So, anyone who has worked with these large language models, especially especially GPT-3 size models, knows that they’re very, very expensive and especially for a consumer use case and a game where people don’t expect and want to kind of get limited and say like, you can’t play anymore because you don’t have enough. I think adapting to that and user expectations was very, very challenging and figuring out how to make the unit economics work so we could provide a unit consumer experience at a price that was within consumer ranges because you can’t do 50 or 100 dollars a month for consumer subscription because that’s out of the ballpark of what the majority of consumers can afford or imagine paying

Dhaval:
And where did you acquire that? How did you come to a place that was a happy place for you in that area? A lot of testing, a lot of iteration or something else .

Nick:
Yeah. A combo of playing around with kind of the structure and our monetization and pricing structure over time, as well as just working on the how we deploy the models. So for our free models, we deploy them ourselves on GPUs because that’s how we can get the cheapest cost possible so that we can support free to play, which I don’t know how many like yeah it’s hard to do free to play with kind of the cost that go into these models. And then with, I think the other thing is just testing and evaluating different AI providers and just finding ones that match. You know needs and what we can do for our users. So I think those are kind of the combo of those things. We’ve just iterated on as well as there’s some other optimizations in terms of like and how and when you call the models, right. Like if you do one call with multiple generations versus one call with only one generation, like you can cash generations and there’s some things there you can do to get some gains as well.

Dhaval:
So it has been a lot of optimization in the operation and the unit economics. What advice do you have for game creators or for product creators who want to use the new technology to build a product? Like, any advice for new and up and coming entrepreneurs using AI.

Nick:
Yeah. So, one of the things and I mentioned this earlier, one of the things I see and we had ourselves for quite a while was I think a lot of products and things you see are somewhat like people love trying them out and they play them for a few days, or they like play around with it a little, but then there’s like, there’s not staying power. It’s not really a product that someone will use long term as much as it is kind of like a fun toy they play around with and set aside. And I think the generative AI tech lends itself very well to like, oh, look what you can do. And it’s kind of fun. But I think the challenge in the space for people who want to build big businesses will be what is the like fundamental human value that you’re solving here, Right? like One of them mental models I like to use is like, what’s the like caveman equivalent of what you’re doing, right? Like for us it’s role playing where for thousands of years people have played in a way where they step into the shoes of another character and kids do that kinda like, it’s an instinctual human thing. I think if people can understand that really well, like a good example is chatbots. Right like I think that there’s a lot of AI companies doing chatbots, but I think the problem is a lot of the ways they do chatbots is as a toy and as a novelty, not kind of a fundamental human need, right? And so I think companies who will be successful with those won’t approach chatbots as kind of that fun to chat with this a thousand. You know, these a thousand different interesting chatbots, but rather it would be It will try and match the fundamental human experience of like building a relationship with another being and like the progression in the relationship, how you start more guarded or unfamiliar and you develop greater kind of like emotional connection over time and feel this closeness. And I think that will happen with only one or a couple characters rather than this wide swath where you try one, set it down, try another, set it down. So anyways, that’s one thing I’ve thought a lot about in this space.

Dhaval:
Yeah. So going back to the value proposition or any product creator who’s interested in creating an AI product using some of the fundamental foundational models, it’s not the fact that you are doing something cool or something novel that’s going to make you successful. It’s a fact that you saw a real pain point of your users. That’s gonna make you successful.

Nick: Yep yep exactly,

Dhaval:
Yeah. And I love how you have have this analogy of tapping into the instinctual user needs. That’s even like more powerful than just a simple pain point, right? So if you can tap into those instinctual needs as a product creator, you increase the retention, the stickiness, the relationship with your user needs.

Nick:
think the pain point model works well when you’re building a tool. Right. Like I think pain points are like humans have been using tools for thousands of years, right? Like fire and weapons and all these things. And so I think it works well there I think it isn’t always as applicable for consumer things that aren’t a tool but are are more of an experience or fill some other need besides helping you accomplish something else you already want to do.

Dhaval:
Yeah. Very interesting distinction there. Thank you. Thank you for sharing that. Wrapping it up, I’m wondering what do you think is the future of your product? Where do you see it going?

Nick:
yeah so , two things that we are really excited about in the long-term, enabling with AI in the gaming space. One is creating and enabling people to create games that are dynamic and alive in a way that they haven’t been before, right. So past games, everyone goes on the same quest and everyone is kind of the same hero you know you have three options you can choose, right. So there is kind of projection of choice of choice But your actual choice is fairly limited. And so we imagine games of the future where you have this like living dynamic nature where you might have individual quests that are just for you. You might have conversations with characters that you build deep, meaningful relationships, right. And so I think there’s this whole class of games that we’ve imagined as humans, but we’ve never been able to make before. And we’d love to see those come to reality. The second thing that we’re really excited about is the ability of AI to democratize game creation. So you think about what the digital phone did for creators who wanted to make video or picture content right?. They spawned entire platforms platforms like Instagram and and just this wealth creation that before could never happen because it required kind of a professional Company level. And I think we’re going to see the same thing with games where games have traditionally been somewhat hard to make. And I think AI will lower the bar for that. And so we’ll see a lot more. Everyday people be able to have a cool idea and be the kind of creative director with an AI team to bring it to life.

Dhaval:
Wow. I love that. I love that analogy that you just made. Thank you so much for coming out to our show Nick. And I learned a lot just from our brief conversation. Looking forward to have you back on the show once you have a few more learning lessons to share with us. Any last thoughts to share with us?

Nick:
I think there’s a lot of kind of like fear and worry with some people in kind of the AI space. And I think technology, new technologies are always hard. The way I see the future is not that AI really replaces people’s jobs, but rather, I think the best example is with chess Right. So like for a long time, humans were the best at chess, and then a computer finally beat the best human at chess. But now the best kind of competitors in chess are not human or AI alone. It’s human AI teams that work together. That are more successful than human or AI by themselves. So that’s kind of how I see the future of AI and technology is not that AI is like replacing us as humans but rather that it’s augmenting us so that we as a team can do much more than either can alone, because I think human will always have huge advantages in how we can think and our flexibility and our ability with ambiguity that machines won’t. And I think machines can continue to kind of automate the less ambiguous You know sides of creation, enabling us to do more interesting things.

Dhaval: Thank you. Thank you so much, Nick.