David Yakobovitch
You are listening to the HumAIn podcast. HumAIn is your first look at the startups and industry titans that are leading and disrupting artificial intelligence, data science, future of work and developer education. I am your host, David Yakobovitch, and you’re listening to HumAIn. If you like this episode, remember to subscribe and leave a review. Now onto the show.
David Yakobovitch
Welcome everyone back to the HumAIn Podcast. Today I have a very special guest, many of you know that as I’ve been growing in the data science and AI field, I have a special place with video games and my love as being a geek, a fun fact that many people may not know about me is that back in high school I would nerd out Two MMORPGs, like RuneScape and a lot of these other cool games and so our guest today is the founder and CEO of AI Dungeon, AI Dungeon franchise Nick Walton¹. For those of you in the industry, you’ve probably by now heard of AI Dungeon and how it’s taken the industry by storm as one of the paramount examples of the success in the GPT-2 movement. Nick, thanks so much for joining us on HumAIn.
Nick Walton
Happy to be here.
David Yakobovitch
I’m really excited again not just as a nerd and a geek and a video game enthusiast, but seeing what the industry’s going about six months ago on HumAIn and I had the former CEO of Epic Games we got to speak about the unreal engine, the unity engine, the evolution of games, both from GBC and these conferences but now fast forward into 2020 so much it’s changing the gaming industry is now a hybrid of makers of creators who can code and artists, and we’re seeing all this merge together and what I’d love to hear for our listeners is tell us about #AIDungeon and your inspiration from where this came about into the movement that it is.
Nick Walton
So a good place to start is at the beginning of last year, I started playing Dungeons & Dragons with my family, just a small group and I hadn’t actually done that before and what was really interesting to me with AI Dungeon compared to other video games is just that freedom because you had a Dungeon master that could say like, whatever action you try to do, he could think through what would be the consequence, how that would affect the story and so even though the #DungeonMaster would think of a world and possible quest things you might do, he could adapt to any action you did and that was something that just doesn’t exist in any video game so then in March of last year, I was at a hackathon and GPT-2 had just come out and so I was playing around with it and they had only released the smallest model.
So it wasn’t as good as like their largest model, but it was pretty fun and I was able to make like a decent little like AI Dungeon master with it and, I didn’t win anything at that hackathon, but I thought it was cool enough that over the next couple months I continued to work on it and worked on how to deploy it and I made a little web app, which I called add Dungeon one or now air does in classic and it got played by probably something like 10,000 people so it was like a decent number, but the model was much smaller so it was less cerent.
And also in order to save on compute, I just had it generate a couple like pre-generated actions you could choose from so it wasn’t free for him, like and too so then fast forward into fall, I was working on it time neglecting my classes to a degree and GPT-2 released the largest model and I also found a dataset of text adventures so I trained the largest model on text adventures, and that’s where it got like much better and so released AI engine two in December and it was initially just released as a Python code and a Google CoLab notebook, which is just a way to run.
AI machine learning models and Google provides a free GPU for that and it exploded much more than we thought, which is crazy because although, although we didn’t have to pay for GPU compute, which was our biggest concern, some quirks with the Google CoLab and downloading our model made that really expensive and so within a few days, we’d have racked up like 30, 40,000 in bandwidth costs and so that was scary and so we shut that down, but then pretty fast, community members had implemented a peer to peer torrent sharing network so they could still play the game like within 12 hours so that was one of the first demonstrations to us that we’re onto something cool and so from there we’ve gone and we’ve made like a mobile and web app version that you can play and now we’re, we just passed 750,000 registered users and so it’s been growing pretty fast.
David Yakobovitch
It’s so fascinating what you just shared, Nick not just about the cool factor in the growth, but that it started similar to one of the big movies last year on Netflix Bander Camp or Bandersnatch, which is all about choose your own adventure with these two scenarios and now it’s dozens and hundreds, and it’s maybe infinite number of scenarios, which is really inspiring that you were able to take a product from an MVP or a very simple product, and then start to see how you would scale it up over time.
I wanted to dive back to that MVP stage because a lot of engineers and #datascientists and founders always get stuck on how did we even get to MVP? How do we test and iterate a #product? And a you mentioned you got to participate in a hackathon and that’s where this inspiration got tested and you saw some results.
I have participated in hackathons, usually in a group setting. But prior to recording today’s episode, we talked about that you are solo competitor at this hackathon. You didn’t have a group so what was going through your mind when you were at this hackathon and why didn’t you go about it solo?
Nick Walton
I haven’t even ever really gone to hackathons to win. I honestly don’t really care. The thing I love about hackathons is, it’s this place where you can do like build things completely and explorative way like you’re not, you don’t have any pressure or time crunch. It’s like let’s just explore possible things we can build but the nice thing is if it doesn’t work out, you can just say like I tried that for a day it didn’t work out and I don’t have to keep going with it and you don’t really feel bad about not continuing this project.
And so it’s a great place to be able to explore different ideas, but keep it like bounds on how much time they take and just get to the point to see whether there’s something there or not. I went there and I had like there were people I was there with who are like working on their own fun things but partly because I didn’t care about winning, I was just like play around with different things to see what would be fun and so that’s why I did it that way.
David Yakobovitch
It’s part of having fun and myself I used to be involved as a North America, regional manager for angel hack so we used to run hackathons and Detroit, San Francisco, Seattle, New York City and so many ideas would always come out of the hackathons I want to know that as someone myself who is both mentored, judged, moderated, I love seeing the feedback that participants get out of a hackathon and how that evolves into pivoting or giving validation so what was some of that feedback or validation you’ve received from the hackathon that convinced you or led you down the path to continue with AI Dungeon?
Nick Walton
Just like the people I was working around with I would like share the stories and the adventures and people were just busting out laughing so that was a good sign of like, how fun this could be and then one of the issues with like in terms of the competition is I didn’t have like a web playable version by the end of the hackathon cause that took quite a bit more work, especially on the machine learning side and so that was one of the hard things is like I was showing it off.
But it’s just this text thing and it’s hard for people to like, get what’s going on versus if I had been able to like share a link that people could play, people would have been like gotten a much better idea of what it is but about seeing how much fun people had playing it just around me, sparked the inclination to be like I should take this to the next step, make a web app and I spent like a hundred hours over the next, like several weeks doing that and then it went from there.
David Yakobovitch
I see this recurring theme about cool and fun and even as Marie Kondo, one of these well-known things sway masters on Netflix now says, does it spark joy? If it sparks something that you’re having a great time, the great effects are going to happen.
It’s going to compound, it’s going to scale and build and it did scale and when you deploy the model at scale, you started small, you used #GoogleCoLab and like any developer who thinks Google CoLab is for free, it is to a certain capacity, but then there’s always those we got you scenarios and so you mentioned you experienced over 30 to $40,000 of bandwidth costs. What happened there? Did any sponsors come in to absorb this to Google say, well partner how did you manage those costs?
Nick Walton
The professor whose lab I was working with as soon as it was clear, it was going to be over like a thousand he’s like I’d be happy to, like we can cover it out of the lab budget. That was when he thought it would be like a few thousand every time it hit a new milestone, I was like professor it’s at 5,000, are you still sure you want to do this?
And then we got to like 6,000. I’m like are you sure you still want to do this? he’s like, it’s still fine 15,000, he’s like maybe let’s look into ways we can reduce that 20,000 is like shut it off so we ended up with quite a bit there thankfully we were able to ask the Google Cloud billing people just to forgive some of that, because it was this just university project that went viral and they forgave a lot of it and then we’re planning on paying back the rest to the professor from like the startup but it was crazy and part of it was a few quirks of Google CoLab.
Like we didn’t realize that Google CoLab servers, a lot of them were in like Asia and Europe and our model was hosted in the U.S. so we were getting like international egress bandwidth fees and so that made it a lot harder than otherwise would have been and it was ironic that like we did Google CoLab because we are afraid of spending all the money on GPU compute but now actually our GPU compute infrastructure is much cheaper than the cost of downloading all those models for the initial Google CoLab version.
David Yakobovitch
It’s incredible to see how once it’s set up, once these models are set up, then the cost rapidly reduces because as a practitioner do as scientists myself, it’s all about the model, how much time and compute goes in there but once you got that steady state, then you’re just feeding in a few new data points it’s not that much compute, which is great and it’s really great to hear that most of that’s been resolved with the costs.
Thank you, Google for being generous with the compute recovery there and so now as you’re continuing to scale, you have multiple products both online where people could download and now even on Android and other operating systems, what is that looking like now? Because you are a #startupfounder, scaling AI Dungeon too, what are some of the goals and what does the product look like?
Nick Walton
I want to make clear that since we released the co-lab version, a team started to come together and a guy volunteered to build up the mobile apps.
I started working with my brother and he’s been really awesome. He’s got a lot of tech and startup experience and so now we have like a team together that does the mobile and the web and the model serving infrastructure on the backend and we’re looking at growing that team, but what we really want to do is explore all the awesome directions.
This can go because the awesome thing about Dungeon that players have been hungry for and that’s one of the things we’re hitting on is that player Freedom and like, so like Bandersnatch is a good example where like you have choices, which is completely new for like a TV show or movie, but at the same time, they’re so limited so I haven’t personally watched it, but my brother did it and his main issue was like sometimes all the choices just suck. It was like, kill this person and it’s like, I don’t want to do any of those and so with AI Dungeon, you literally have an infinite set of possibilities because anything you can express in text, you can do and that’s a completely new idea for a game.
And there’s so many interesting things you can do with that, but there are also technical challenges and we have a really strong team to solve those, but we need to explore and figure out how to resolve those technical challenges and how to make, for example, right now, air Dudgeon is just story so if you think in terms of D&D, D&D has two halves, there’s like a story half and there’s a mechanics half so mechanics, you have levels, you have health, you have damage and dice and everything and then you have the story that they’re the Dungeon Master is feeding off the player’s inputs so being able to merge those two in a video game format would be really powerful and that’s one of the main things we’re working on.
David Yakobovitch
I can only imagine just speaking from my experience with video games there’s so many possibilities whether you partner with big franchises or developers with anything from, could be like a Kingdom Hearts to an EA Games to what sports players say to the gamut could just run wild. I can see so many possibilities. Are you thinking that this Dungeon theme for games is the ideal, vertical or scenario to start focusing on?
Nick Walton
I guess, what do you mean by the Dungeon? Seemed like the more D&D thing or that’s right.
David Yakobovitch
Because as someone who’s attended the game developer conference, there’s so many emerging fields of video games. Everything from live streaming and Battle Royale to MMRPGS and so I’m just my wheels are running right now thinking. I could see so much implementation here, even if I play a game and normally I choose pre-populated text to respond to my team, whether they’re always evolving with technology like this it just seems really cool to me.
Nick Walton
Actually that’s something that we definitely in the long term, we’re thinking much broader than just like fantasy RPG type genre, and already you can do really anything, which is one of the cool things about #AIDungeon it has this vast knowledge from the vanilla GPT-2, which was trained off 40 gigabytes of text data. So we’ve already seen sports news sites saying I played this sports adventure on AI Dungeon and like, these are the cool things that happen or you can do all kinds of fan fiction types adventures.
Or you can do, just normal you’re a normal person so it’s really powerful and its ability to transfer theme so it’s not just like Skyrim type scene it could also be The Sims it could also be like FIFA and we’re definitely interested in exploring that broad set for the initial game, the fantasy theme has been really powerful because it taps into that getting closer to that D&D field that people are hungry for, but we definitely have like larger long-term vision.
David Yakobovitch
That makes a ton of sense even as you just mentioned with all these industries, whether it could be like Skyrim or Sims or FIFA we’re seeing the GPT-2 technology emerging across the whole industry news publications, like the wall street journal and the New York times have started dabbling with creating content or auto generating content.
Traditionally we’ve seen it with there’s the stock market and at the end of the day, a stock closes and do we really need a human to draft the whole article on it or can not be automated, but now it’s also the text generation so how interesting would it be? If two people play a sports match and then a very creative story is generated from that so there’s a lot that’s going to happen there in content and then media content could be with games or it could be anywhere in the industry.
Just in the past month I attended an art gallery in New York city and they had this exhibit and the exhibit was themed the view from nowhere and I thought it was so interesting because the entire exhibit was based on AI and most of it was automation and so forth, but they had an exhibit on GPT-2, this artist Juan Cortez had basically a computer robot arm a FANUC robot arm, and the FANUC robot arm was connected to a pencil with a Quill.
And that basically was going down and up and drawing these hieroglyphic characters based on training on, as you mentioned, the 40 gigabyte GPT two dataset so it’s incredible to see the expansion across verticals and speaking across verticals you’re now at the forefront of this industry, especially in gaming and where entertainment’s going. I wanted to hear your thoughts on whether some trends you’re going to see that you’d like to see in the gaming industry.
Nick Walton
There’s two things that this AI generated content makes really powerful: one is this player freedom where you could potentially, rather than having this preset list of possible options you can make it much more expansive the other thing is much more dynamic and interesting content because you think the games like Oblivion, Skyrim, they’re really cool and they’ve got quite interesting worlds, but they’re parts of it that break down and its uniqueness, for example, like you talked to one guard and then you talk to another guard in a different city and he says the same things and so like that starts to break down this huge expansive world, because then it starts to feel that some of these are actually the same and scare them, does a pretty good job of making doing enough handwritten stuff to make the world not feel that way, but that takes a lot of manpower to do that and so with AI generated content, you can have less developers and less like creators and maybe the creators are creating more of the longterm and the overarching themes and then the AI is filling in all these details and helping create this super expansive world and so that’s a big thing.
You’ll see also just the power to create unique content depending on the player like individualized so the player can do a things that are much more interesting to them then just following the same storyline that everyone else has and you could see potential frameworks where you’re learning based on what the player does like what things they like in a game.
David Yakobovitch
What this makes me think of is personalized gaming hasn’t been around for a long time of course, we’ve seen some of the games out there like Minecraft, especially around education where you can build these worlds and they’re very custom but even beyond that, I know in the last year Fortnite came out with their creator content pack and asset pack so you could build these custom worlds to do Battle Royale, but even then it’s still very simplistic we think it’s custom, but you have 20 or a hundred Sprite assets these 3D objects and you build your world but it excites me to think whether this could be unlimited, infinite, and it’s always evolving for every game.
I can only imagine, like in the future Fortnite doesn’t just let you create your own assets to sell these different objects, but they’re invented by the AI and the ones that are good become popular and stick around for the game that’s pretty cool and perhaps we can see this both with text assets and digital assets as well so doing a lot of this work, doing a little segue, it’s all around in the AI space, NLP, natural language processing, natural language generation, natural language, understanding this whole natural language field is where the work is happening and you’re seeing all this scaling going as you mentioned, Nick over 750,000 players at this moment are having fun with AI Dungeon too. What is your thoughts around other languages? And if you’ve been exploring, going into languages beyond English.
Nick Walton
There’s a good amount of research into how to do that in the near term, probably the closest thing we would do is just route things through #GoogleTranslate or some similarity service cause to, to make a model that is natively in some language is a lot of work and we’re not quite there yet, but we’ve already gotten a lot of requests from people who are like, I want to play this in Spanish or I want to play this in Russian and we just don’t have the ability to do that yet, but there’s a lot of potential in that area and in terms of AI generated content for games, NLP is going to be one of the first ones, just because so there are a couple of things that are powerful about NLP.
One is you’ve a lower dimension space versus like images and so it’s a little bit easier also the data you like we have so much text data on the internet and you can use that to do really interesting things and the powerful thing about GPT-2 is, so the way it works is it says it’s basically an advanced text predictor given the last text, what’s the most likely next word and then it just keeps generating words so the cool thing about that is in order to predict the next word really well, you have to essentially build some model about the world, for example, if you stab someone they might bleed.
Or if you bake a cake for someone, they might be happy and so GPT-2 has really learned this model about the world, and sometimes it’s not right like sometimes it does crazy things and you’re like, what? But it’s surprisingly good and so that’s really powerful as you, you have this framework that can learn from unlabeled text data and create all of this, create this model and use it and so that’s probably the most powerful thing about Judy Too.
David Yakobovitch
The scenarios you just gave stabbing, and bleeding and baking a cake. I remember those in, even in Brainscape so it’s all these games come full circle with these scenarios and going back to games that I’ve played in my experience the more personalized that is just the more I’m attached, the more I love it, I feel it’s individualized so that’s very exciting and I can’t wait to see what’s on the horizon as well for what you’re doing with AI Dungeon too, in the whole.
Nick Walton
And I just want to add along with that, one of the cool things about this as well is, you can make surprisingly life-like and Dynamic NPCs. For example, one of the cool things I’ve seen people do in an AI Dungeon is you can go and talk to NPCs and tell them that they exist in a video game and they’re not real and they’ll like get depressed and sad and be like what? And like freak out or like in one of my games, I searched for this magic tome and I learned this new life magic and I brought a tree to life.
And he was like, hello, I’m tree man and I’m like, are you happy to be alive? And he was like, yes, I’m so happy to be alive thank you for bringing me to life and like, I just became best friends and I had this like emotional attachment to him and that’s like something really cool. That’s much more possible than in Skyrim when you have something like the main NPCs. You can create a little bit more of that, but you can do this with every NPCs and that creates a lot of really interesting individual emotional connections.
David Yakobovitch
As someone who comes from the educational arena, I even wonder what expansion this could have for education like improv comedy acting or speaking and gaining confidence it sounds like there’s so many scenarios that can be generated now of course we have the GPT-2.
I wonder how long it will be till #GPT3 or even other updates come around. Do you have any predictions on where that technology’s going and particularly around? Some of the things you’ve hinted throughout our conversation is the emergence of transfer learning so what’s your thoughts on that technology?
Nick Walton
We’re already seeing really fast development in terms of general purpose language models so we have things like Albert that is a much more efficient and better than previous versions there’s other potential things like Google has one called T-five. That’s supposedly Seven-X as efficient as GPT-2 there’s other ones that have a much larger context that they can take into account and so we’ve over the next year or two, we’re going to see really rapid development of these kinds of models and that’s really going to enable us to do a lot more, even than we have already done with ad engine.
David Yakobovitch
One thing I wanted to dive into beyond what you’ve done with AI Dungeon is this is not your first rodeo you’ve been going about research and technology for quite a few years and I had a chance to read that you’ve been working on CubeSats at Brigham Young University BYU and got to actually work on building these mini satellites up Into Space.
I actually, a few months ago talked about that the future of technology is not just where we can be on our terminal like a computer, a phone but also having the access of data especially on the edge and in space to enable these transactions and to enable games like this I wanted to hear your input from the work you’ve done with CubeSats to till today AI Dungeon too has compute and if I’m not on the internet connection, how powerful is that game? Do you think there’s I’m asking a lot of the questions here but whether you think about the emergence of #CubeSats, and #5G and the edge as it comes to playing these dynamic games, like you’re building.
Nick Walton
I did a little bit with Q-tips. I didn’t do a ton so one of my brothers so I’ve I have two older brothers, the oldest one is the one working with me on endangering.
The second one is he’s completely in that space industry and so he’s the one that so I helped him with this process a little, but he was the main driver where he got a grant from NASA to build a cube, sat at our university and assembled this team and so has done a lot on that front and so he’s much more experienced on that I helped for a little bit and then eventually I transitioned to like more of the robotics and machine learning side of things but in terms of like doing things, so one of the issues with AI and Dungeon is it actually has the highest minimum required GPU spec of any game we know of it.
So your GPU has to have 11 gigabytes of GPU RAM. So that puts it out of the range of all but the most high end gaming GPUs and so that’s why from the outset we’ve gone rather than playing it locally, we host the GPU’s on servers and then have people play there and in the future, that’ll continue to be the case because there’s just so much more you can do on server side deep learning customized GPU’s then you can locally and also right now, just the tooling to package machine learning models into games is just not that great right now and so there’s a huge barrier on that front as well.
David Yakobovitch
Speaking of the tooling for machine learning models into games, do you think any of the platforms are going to accelerate that whether that’s unity whether that’s unreal or even another platform?
Nick Walton
it’ll be interesting to see it makes sense for them to incorporate that more and that’s something that will enable really interesting things with games and so I would hope they do that, but I’m not sure what’s on their roadmap.
David Yakobovitch
I’m looking forward we’ve had now this dynamic conversation on AI Dungeon and AI Dungeon Classic and AI Dungeon 2 and where the product has evolved from your story into now, this startup that is accelerating with growth forward thinking, looking at the next three, six months, what are some new ideas that you think you’re going to be working on or where you’d like to take the product forward?
Nick Walton
So we have several things that we have good prototypes for, and that we’re working on implementing so things like multiplayer where it’s a turn-based so we’ve got a lot of interesting ways you can modify the game than some of the next steps text to speech could be really interesting so we have samples where it’s like a pretty good, like British storyteller, like voice that really adds to that you could do interesting things like adaptive music.
So like, imagine you have a very simple sentiment classifier that tries to detect what the mood of your, the current place in the adventure is so it’s like mysterious things are happening so like place a mysterious music or like triumphant cause you defeated the dragon so that’s a really fun direction, but long-term the biggest thing that we’re working on is, how can we incorporate state into the game? How can you have locations or inventories or quests that merges in a way that works well and adds this discreteness, because you really need the discreteness to feel a strong sense of progression, because without any of that, you don’t feel the same level of progression as you do when you have levels and like discrete things you can track and watch improving.
David Yakobovitch
And that makes sense we can think of all this exciting technology like synthetic voices and adaptive music and all the glamor that we want to move to in technology, but you’re right. You gotta get the fundamentals going and now that you have this traction, it sounds pretty cool that your is focusing on state and of course, when we think of AI and data science, state is talked about a lot but when it comes to video games, I don’t know if most developers think of state, they just think of levels so that’s an update that I’m looking forward to and especially everyone in the gaming industry and everyone listening to the HumAIn Podcast.
Nick, we appreciate your time here and thanks for joining us on HumAIn.
Nick Walton
Happy to be here.
David Yakobovitch
Thank you for listening to this episode of the HumAIn Podcast. What do you think? Did the show measure up to your thoughts on artificial intelligence, data science, future of work and developer education? Listeners, I want to hear from you so that I can offer you the most relevant trend setting and educational content on the market.
You can reach me directly at humainpodcast.com/contact. Remember to share this episode with a friend, subscribe and leave a review on your preferred podcasting app and tune into more episodes of HumAIn.
Works Cited
Companies Cited