Latest / The Information Bottleneck / Why Video Games Still Don’t Use Modern AI | Julian Togelius (NYU)

Transcript

Why Video Games Still Don’t Use Modern AI | Julian Togelius (NYU) · The Information Bottleneck

• Ravid Shwartz-Ziv: Hi everyone and welcome back to the Information Button Egg podcast. And today we have a great guest, Julian Togelius. He's a professor at NYU and he's mostly working on AI for games. So hey Julian, thank you for joining us. Julian Togelius: Hi Rid. Hi Alan. great to be here. Ravid Shwartz-Ziv: So I I want to start from the beginning and I I want to understand like why we should care actually about AI and games and maybe like games in AI. Julian Togelius: Yeah, I'm not gonna lie. When I started doing this, back in my PhD days, which started in 2004, I was like, hey, video games are fun. I wanna work with video games. but I was also Right, exactly, Ravid Shwartz-Ziv: This is the dream of every child, right? Julian Togelius: exactly. I was dreaming about this as kid, wow, Ravid Shwartz-Ziv: Yeah. Julian Togelius: like video games that play themselves, all that kind of stuff. And And it was also because I had done my masters in I was kind of focusing on evolutionary robotics, so evolving neural networks to control robots, but the robots were slow and you know, wires got tangled and batteries depleted and all kinds of really boring stuff and like I didn't want to work with that. I don't really care about the physical world, so I so hey, you can do this in video games. So and Of course I got a bunch of pushback. Back then, people to some extent thought that chess and go were like real AI problems, but video games like no. You know, I I I went out and showed some of the stuff I had done and people were like, Okay, this is fun, but where's your real research? No, but this is my real research. my god. This guy is not gonna have a career. and no but like seriously though, Every video game, every game is some kind of abstraction of the real world, and some kind of abstraction of situations we may encounter, and that's partly why we consider them interesting. The other part of why we consider them interesting is that we learn to play them, we learn to get good at them as we play them. A well-designed game is a good, kind of nice, smooth learning curve. game design is pedagogy to a very large extent. Now you may ask what in what in what sense are these video games like the real world? Tetris is the it's like it's like extremely non-representational, right? But the Tetris is of the on the most abstract side of video games, still. It is like literally like packing packing boxes, you know, things like this. It it it is like We have a not very big car and when we need to bring the whole family somewhere with all bags and everything, it's like it's a Tetris problem and my wife is much better than I am at it. and my solution is that next time we're gonna buy a bigger car. This is America after all. but then you have like also other things, like StarCraft is a military campaign, but it's also about running a logistics for an organization because that's what it is. You need to get things there at the same time and there's these role-playing games. About managing emotions, understanding complex narratives, etc. It's all like abstractions of the real world, and they're so diverse. People have this idea about video games, like you know, there's Super Mario Bros. and Space Invaders and you know maybe Grand Theft Auto, but actually they're like insanely diverse. I mean, did you know there are dating games where you're literally dating pigeons? And and there are other games where you're controlling the Prussian army. And there's this game's about everything. So in a sense the wide variety of even just video games is like much, much, much, much wider test bed of intelligence than the kind of benchmarks you usually use. Ravid Shwartz-Ziv: I don't know to if I want to solve how I organize I don't know boxes in in Amazon, I can design something like that. It may be like that I want simulation, fine, okay, I can't do it. But I don't need a game for that, right? I don't need like human interaction, I don't need to design like entire world for that. So Julian Togelius: Hmm. But but you want that because that's what you learn from. Because there's the other part. Games have this fantastic feedback all the time. That's what makes them so so so good, you know. That's makes you want what what makes you want to play them. That's what's fun is, right? and that's also the reward signals you want to train a controller or an AI as the kids weirdly say these days. Allen Roush: So so that gets me to a to a question I wanted to ask you, actually. You know, AI has historically referred to all sorts of things, particularly the behavior of the game's enemies, for example. with certain games, especially from the old days, having been coded up as very simple decision trees, leading to really great outcomes like the original Fear, Halo 1, and probably several other examples that would s I I guess my question for you is what is AI? I take I personally take an take an expanded universe type selection where I think that this kind of programming by video game developers is effectively just a different paradigm, but basically the same thing as computationalist neural networks today and in that they were dealing with like creating the appearance of intelligence in a nonlinear system. But like do you agree with such an expansionist view of AI or do you want to like draw much more stuff in the sand? Julian Togelius: Totally, totally. No, no, no, no, no, no. I think AI is like I have a very kind of like broad definition. I think that's the only kind of intellectually honest thing to do. And I I've stopped getting angry at like kids who were born yesterday and say that AI, that's LLMs. AI was invented in 2022, you know, November twenty two, twenty two, we invented AI and like, you know, first I was like, my God, these people, and then I'm like, then I saw myself as old man yelling at clouds and and and and then I sort of like okay this is the the language they use. This doesn't this doesn't change the fact that that's wrong. You know, AI is a wide variety of things and We've had superhuman AI since the nineteen fifties. We've had like superhuman intelligence at like solving particular kinds of theorems and we've had superhuman chess play since nineteen ninety-seven without a single neural network involved and no nothing pre-trained on the bunch of language and and and stuff like that. So when it comes to what you're saying, like game AI. So when I first went out and talked to game companies in twenty two thousand five two thousand six you know and and told them that hey I've done all this work I've been evolving neural networks that train that drive racing cars. Is this interesting to you? And you know I I was mostly roundly dismissed by a bunch of people not There there were a couple of really receptive people. For example, at Microsoft Research that actually were working with Game Studios, but but but there were also a bunch of people that were like, what are you even talking about? go home, you don't you don't know us, you don't understand what the actual problems in game development are. And to some extent they were right. First step is to actually realize that things like Finite state machines for agent control, very elaborate derivatives of A-star search for navigation, behavior trees. These are genuine inventions, they are really quite sophisticated, and they are there for a reason. And the reason is that designing a video game is so much about creating the illusion of life and the illusion of intelligence to some extent. But then all you know. GPT 6 is also the illusion of intelligence in a different way. This is real AI, but a different kind of AI. And even 10 years ago, if you went to game developers conference, where the industry congregates. there would be basically, there was such a huge rift between the game industry and the resear game in a in AI research that people were like basically this is what we care about and this kind of machine learning stuff and your network stuff, even in the games industry would like They thought that that was bullshit. We can't design with it. We can't use it for anything. And there's no point in even talking to these people. They don't know what it what are doing. They don't know games. And to be fair, there's a lot of people in in AI still to these days going into thinking that, I'm gonna revolutionize games, and then they haven't done the basic homework of understanding why game AI is what it is today. Allen Roush: So so f r related to this, I think that we can, you know, look into video games, even some new, many old, that had excellent AI, also excellent other features, like for example, fully destructible environments, Red Faction, Gorilla, I believe, was the one from 2009, the finals from today. and for modern AI, actually, with the exception of Arc Raiders, I can't name a single video game that's using like neural network connectionalists. based AI systems. And I guess my question from all that is why has AI like the of the type that we are doing, and well before LLMs, I mean like machine learning, why has that had so little impact on video games and their their style of AI? Julian Togelius: I'm gonna give you a separate answers that are mostly compatible with each other. The first one is that you don't know. I don't mostly know either. I may know a little bit more than you but you don't but but we don't know because the games industry is highly secretive. Like coming from AI where people like everybody except OpenAI is very open about AI. and you basically put out your code on GitHub, you put your models on Hugging Face and you tell the world what you did the day before you did it. This is not at all how the game industry operates. They traditionally have much more like other parts of the entertainment industry, and they keep quiet about this. So the way to figure out what's actually used in games is to go to game developers conference, hang out with people, go to the after parties, get really drunk, and then they come up and and tell you like in confidence what they did. This is this is how I find out to some extent. and yes, there is more modern AI. actually used in in games than they want to let through. This is for multiple reasons. They don't want to give away the business secrets also because AI has there's a there's a a huge sort of Ravid Shwartz-Ziv: Yeah. Julian Togelius: ongoing culture war slash civil war within the games industry about the role of AI in games. Where a lot of the games industry, a lot of people in the games industry in particular, artists and designers are very, very negative towards AI because it thinks this out to take their job and steal their content. And I see their point. They're right in some way, unfortunately. But so that even for the more innocuous users of AI, people may not want to talk about it. They get a lot of pushbacks from even from players. But to take an example that actually is public, people did have talked about this. EA is now shipping the game that is not called FIFA, it's called EAFC, their big football game, which is one of the most played games in the world, if I'm right. Ravid Shwartz-Ziv: Wait, there is no more FIFA? Julian Togelius: Well yeah, because FIFA, the organization, did not want to they they had some crazy contractual deals to continue working with EA. So EA just chose not to not to keep working with them and EA came out of top because they still have the main football game and FIFA does not have a credible football game. which s makes sense because EA is Respectable Games Company and FIFA is a deeply corrupt organization. so so EAFC, Allen Roush: Yeah. Julian Togelius: the main football game, the goalkeeper is now controlled by a reinforcement trained neural network. I also happen to know that there exist, for example, fighting games with RL train neural network control characters and so on. You you may have to dig to find out about this, and they don't publish to the extent they even tell you, they don't publish all the details. you're very unlikely to find a neuros paper about it, but it exists. that was one of the answers. It's true. Now there are other answers as well. one is like the very, very trivial answer. for as long as I've been in this game, people come and tell you that we don't have the compute to do any kind of relevant AI. Ravid Shwartz-Ziv: Are you talking about generative AI, to generate the game itself or the decisions? Julian Togelius: Okay. Now we I think we're mainly talking about the things that happen at runtime in the game. Ravid Shwartz-Ziv: Mm-hmm. Julian Togelius: Because Consumers don't have the hardware that is needed to run these things. And anyway, this is a game. The GPU is not supposed to run a network. It's supposed to do you know you remember when GPU meant graphics processing unit? So you basically what's gonna run your neural network if your GPU is doing graphics? it sounds silly, but that's it. And they're like, can't you and and if you want to have a big model on a server somewhere, can't you call that? No, because the economic model of games are that people want to pay upfront for a game. They absolutely do not want to pay per hour of gameplay. You're not gonna get players willing to do that. So the economic model of games is that any kind of real compute needs to happen on on on device. So now that's another answer and that's real. Now the third answer is the deep answer, the kind of real, the the hard the truth about this. And that is that games are designed to not need AI. Why is that? Because Game design is like all other artistic expression. it's built on earlier design decisions. You have a lot of design decisions are kind of frozen in time. in video games you can think of genres like you know, MOBAs were built on RTS and RTS came out of like other strategy game designs and so on. The kind of import design conventions at the history of the whole game genre. And when game genres were invented, you can we can like somewhat arbitrarily say that, you know, the turn-based video game RPG was kind of div invented with Final Fantasy, which is 1982 or something, and like action RPGs invented by Legend of Zelda, 1987, maybe. I mean might be slightly off on these things, you know. first person shooters were invented with Doom, 1991 3? and etc etc. and at all of these cases we did not have the modern AI algorithms we have now back then, and we certainly did not have the hardware to run it on. So for example, why do we have dialogue trees in role-playing games? That's a really, really stupid, boring way of doing dialogue. Well because everybody expects it. Because players expect it, because they've been expecting it since Phantasy Star, since Ultimate 4 and so on. like all these classic games and that's how you design things. And that's how designers know how to design these games. you can Ravid Shwartz-Ziv: way to to dis to show dialogue? Julian Togelius: These th these the ask anyone in AI and they're like, obviously you use a large language model for this. But we don't know how to design game with conversational non-player characters or even if they don't even if it's like turn based we don't even know how to design fantasy star or final fantasy 7 or something with LLMs in you can do this thing there are people who do this experiment you can mod Skyrim for example and put an LLM in there and and instead of like the existing kind of dialogue trees and scripts and the result is disastrous it just destroys the game Because it's Ravid Shwartz-Ziv: In what sense? Julian Togelius: you try asking them things and you're gonna ask them the wrong things, they're gonna answer the wrong things, they're gonna hallucinate stuff, you're gonna quests are gonna go completely off rails because and you and you don't know there's nothing there to push you towards the right quest objective. And what if you ask someone in Skyrim what's the price of a hamburger? Does Skyrim have hamburgers? what would the price be? you know, how is that gonna advance your your mission? Like like we don't know how to design for this. So it's a completely new thing. So a lot of the reason we don't have modern AI in games is that we don't know how to design for it. And there are people who have really, really gone out of the way to really try to do this. Black and white is a great example. Ravid Shwartz-Ziv: and white? Allen Roush: I played this. This is a famous ex keep going, keep going. Julian Togelius: Yeah. Yeah. No no no. Tell me, tell me, Alan. What what's what's your take on that? Allen Roush: Well, beyond being a buggy mess both games, right? it's they're great god simulators, right? I'll let you Julian Togelius: Yeah. Allen Roush: continue, but they're buggy. man. Julian Togelius: Yeah, yeah, yeah, yeah. isn't amazing that This actually works to some extent. It's a god simulator, thing populus or something like this, but you also have this giant creature and you can tell this giant creature to some extent what to do. You have so you're kind of like reshaping the landscape and taking care of the villagers and stuff like this. And you can you can it tries to imitate you and you can also reward it. Or you can punish it for what it's doing. don't eat the villagers, and help the villagers, expand their land or something like this. And it's so funny. but there hasn't really been many I've barely seen any real examples of people trying to replicate what black and white did because it's hard. Sure, they're computational problems, but most of all they're design problems. And these design problems is not gonna be solved by by AI bros barging into the game space saying, I know how to do this because I don't. And it's not gonna be Allen Roush: Well well w Julian Togelius: solved by like big big big big AI companies because it's too risky. So it's like we're waiting for like some really cool indie game designers who that get enough to get enough compute and actually do this. Allen Roush: So so related to that, I think most would agree, both inside and outside the game industry, that we're at probably one of, if not the lowest points, since the nineteen eighty-three crash. You know, with with whole game consoles looking like they might not continue to exist, the entire concept of couch co op outside of Nintendo is dead, which is crazy, right? most manufacturers anyway. it seems, in my subjective opinion, that the people who work in the game industry, the coders, the designers, artists, are enormously talented, especially relative to the absolute pennies they're paid to work, right? And it it scares me when I see how hyper competent even mediocre games are in comparison to really so-called good software and enterprises made by 400K whatever average people who don't care at all about the software and are phoning it in. versus, median bottom twenty-five percent developers, ultra passionate. So I guess, you know, given that the game industry is is powered by ultra passionate, ultra creative people who are willing to be paid almost nothing, and that it was already henceforth even before AI dealing with serious quality problems. Like what do you think the future is in a way that tries to make people as happy as possible? Julian Togelius: It's a little bit bleak at this moment, So so so part of it is like a massive hangover from the pandemic, when the pandemic happened, y game companies were kind of raking it in and kind of like expanding massively. Which makes sense. You know, this is what I did. Like, like that the first lockdown happened, I was playing Assassin's Creed, Odyssey and I was like you know, pretending that I was traversing ancient Greece and there was sunshine and I was fighting the Cyclops and stuff like this. It's awesome. and and then the world restarted and game spending fell back to more to more normal levels and a lot of companies just like crashed. then there's a couple of other things. One is like all the investment goes to AI and everybody wants to invest in AI even though they have no idea what they're doing and then they're like funding the seventy second company that does the same thing as a bunch of other companies. because I don't know. insert quip about the efficient market here or something. then there's another serious thing though. For a long time the game industry has been has been very hit driven. It's like it's been a tech industry that worked like an entertainment industry in the sense that you try to go for the next hit and so on. So so it's been a very precarious existence. So like companies that actually survived long term as game studios and had multiple titles and like kept making money like EA, Ubisoft and so on were kind of the exception. A lot of game this is a lot of churn, a lot of game studios that that were created and then They did a great game and then they didn't nail the follow-up and then they had no money and they crashed. So now some companies have managed to break out of this cycle. notably, you know, Minecraft is a game that people just keep playing for years upon years upon years, Fortnite as well. and they and they keep making the money and they keep the players there and that's not necessarily a a good thing. In a sense you can argue it is a very bad thing for innovation, it's a very bad thing for the game industry, or for most of the game industry. I I try to not put a value judgment on it. It's just like it's it's it's it is what it is. Allen Roush: I I just want to quickly observe that I believe if you check on Twitch right now, I feel like it's the same games being played as ten years ago, and that the average lifespan of the game the good games are d well like people still play a lot of Team Fortress 2. That came out in 2007, right? And I think there's even more extreme examples of this everywhere on Twitch, for example. Ravid Shwartz-Ziv: But why wait, what so w why is that? Because for sure like AI but like also like other like technological like things that related to games are like were becoming like much better, right? So like why why people still using or like still playing like these old games? Julian Togelius: The the for a long time this was driven by You you had better computers and you have above all better graphics cards. So you could new games would look better. But now, the advances aren't so manageable. Plus, people don't necessarily have better graphics cards because all the graphics card get expensive because you use for AI. Plus, you can just update like every game is to some extent a live service, and if you basically you can upscale your textures and ship new textures and then you're kind of like I'm fine with that. So so there are many reasons why you kind of don't have the reason to upgrade anymore. Now, the real reason to upgrade to a new game, to switch a new game, would be like real design innovation. But the same factors that basically make people want to because everybody wants to have a Fortnite or a Minecraft or something like this, because people just do Team Fortress do, people just keep playing. These the same factors make people very con developers, large developers particularly very conservative in terms of design. Now if you look at the indie under vegetation, game design is very much alive and well. There's a lot of fascinating kind of fascinating kind of like innovation happening, but doesn't reach most players really. So yeah. Allen Roush: So Julian Togelius: Well y you could imagine, someone you can imagine talent scouts listening to this and saying that, wait, we have a massive passion tax on game development in the sense that in the sense that people people get paid like a pittance compared to AI companies and they're super talented. Why don't we just like go and raid the game companies and invest hundreds of millions of dollars and turn them into AI companies? Allen Roush: No, I I I fully believe. Yeah. Julian Togelius: The problem is that you have the aforementioned civil war slash culture war, whatever, inside game companies, where a large amount of l of game developers passionately hate AI. Allen Roush: Well, okay, I now now that I wanna get to this too, because I actually think that there are s there are very few industries where I believe it's okay to actually blame the customer and say the customer is wrong. I think the game industry is the number one place where the customer is straight up wrong. And I'll give several examples. One place, Julian Togelius: Yeah. Allen Roush: gamers are flipping out at the idea that they're 60 or even 70 dollars a pop games, which it's been the same price since like 2000 or earlier or something, by the way. are now eighty dollars. Julian Togelius: Yeah. Allen Roush: my gosh, right? And then they're also, I won't wade into the death of CD ROM controversy, but I think that's a little bit Luddite too. I guess you've got this, right? You've got game devs are already treated terribly, you know, c crunch time and all of these other things, impossible demands. Julian Togelius: Yeah. Allen Roush: You've got the weird connections that gamers have to the rise of right wing politics. there there are pretty credible connections of Gamergate, the original Gamergate against so-called what what do you call it, game journalism, whatever nonsense circle like 2014, directly leading to Trump being elected and and the rise of, you know, modern right-wing populism. Julian Togelius: We we we we had there there was a conference that I was a little bit involved with. I never actually went to this one, but we had like some overlap with it. that was taken over by one of its founders who then at the next iteration wanted to invite as a keynote speaker, so it's a games conference, Steve Bannon. And and and and of Allen Roush: Yeah. yeah, it's exactly. Julian Togelius: course everybody's just like what the f And but the bizarre thing is it was kind of lactopical because Steve Bannon used to run an online games company. Allen Roush: Yeah. But but I guess my point is the the the stereotype of gamers being basically man children chuds is true. And so like can we just like blame Ravid Shwartz-Ziv: Mm-hmm. Julian Togelius: But which but which inlets? Allen Roush: the like customer here? I I claim yes, like screw them. And you know Julian Togelius: But but but it's very immature, right? But which gamers? Because this is a funny thing, Yeah. Yeah, you you yeah, you ask Allen Roush: Mul multiplayer gamers in particular. Julian Togelius: your aunt and she would say that she's not a gamer, but she would like Seventy-eight percent she would be wrong. You'll you will look at her phone and she will be playing a whole bunch of games. She will be playing Candy Crush, she'll be playing some other kind of like casual games that are actually really fairly deep, you know. And actually like like the vast majority of the adult population of rich countries are gamers in terms of they regularly play video games. They don't identify as such. Ravid Shwartz-Ziv: But you you don't think there is differences between a game that I play in my phone, I don't know, Candy Crush or even simpler games to these super Julian Togelius: I I I I I take offence with you calling Candy Crush simple. Ravid Shwartz-Ziv: Okay, okay. I I I say okay, let's say let's go you're right. I I Julian Togelius: Well, Ravid Shwartz-Ziv: Let's say that like games that that are not like involve like high v super high quality graphics with that that that you need a lot of GPUs to to run them. Julian Togelius: I I I I I am just watching you paint yourself into a corner here. Ravid Shwartz-Ziv: Ha ha Julian Togelius: Because Allen Roush: He means games not aimed at the traditional gamer audience. Julian Togelius: Right, right, Ravid Shwartz-Ziv: Yeah, yeah. You Julian Togelius: right. Ravid Shwartz-Ziv: think so you think like AI I I I will I will try to to re- re question it. So like you think like these games are kind of like the same, there are no differences and do you think like AI can help like these types of games much more? Julian Togelius: Yeah, okay. Then there's a real real real real real very interesting question here. So games are very, very, many, very, many different things. And the games that get the mind share and like, you know, like everybody team we play Team Fortress 2. It's like it's not true. I don't play Team Fortress 2. And only a few of my PhD students do. I'm both a gamer that plays like, you know, I I I was I played a whole bunch of Elden Ring and you know and stuff like this and but I'm also like I'm also a kind of casual gamer, you know. I there's like no no kind of opposition here at all. AI can definitely help here. I actually did work when in in my role at Model AI, which I co-founded, we did work with King on level generation for Candy Crush. I don't know how many levels Candy Crush has now. At some point it was 12,000. And you may like be like, ha ha, who needs 12,000 levels? And the answer is Candy Crush needs many more than that because they have an issue in that they're really they're really cool. More player demographic because there definitely is such a thing as hardcore countercrush players. are kind of like, you know, they have a problem with serving them with good new level content as as fast as it is being cleared or consumed. so there's like there's a thing, level generation. There's a lot of like player adaptation and player matching going on. Again, all of this is not necessarily public, but But the King actually has a very strong AI team as they as I know them personally. Electronic Arts has a very, very strong AI team. And a bunch of like sort of mobile developers certainly have that and do that. The main thing, the the pa the the point where they usually start is in understanding their data. because you have you have all this data coming in and you want to use it for something. Like, you know, the lowest hanging fruit is like how can we optimize to make people pay? how do we serve content that make paying users pay more? Whale optimization, essentially. but then beyond that, you can go on to like, you know, how how do we do like active learning for A B testing, for example? and then you can go into actually generating some of the content, generating some of the other kind of NPC parts and so on. Then you can go into generating many many more aspects of the game. there is a lot of AICs in game development, as you all know. You these days you can just ask Claude to basically make me a game. What game do you want? I want a game about the X and it should look like Y. I did in my game design class right now. I the first week I had them design minimalist games totally by hand. By hand I mean like they had to write the code and and no no AI allowed like and and no gra no graphics allowed like or like basically just basic geometric shapes. I I'm basically telling the opposite. You're no longer allowed to, you're gonna do a maximalist game with as much game as you can cram into five minutes of playtime, as many features as you can, and you're not allowed to touch the coder assets. this works really well in terms of volume, you or in terms of quantity at least. You can get a lot of for free. So so they but smart mobile game studios, the ones that do casual games, they do this. There's a lot of like AI assisted game development. but as we all know, like if you just ask it to make me a new casual mobile game, it should be good. It's not gonna be good. It's it's gonna be extremely derivative. So you so the tri Yeah, don't make any mistakes. Allen Roush: Yeah. Ravid Shwartz-Ziv: Don't make any mistakes. Julian Togelius: It should be better than Candy Crush. Allen Roush: I have a few questions on that that I want to bring up that I I I want your take on. So there are two new phenomenon I've seen as a result of the the new power of AI in video games. One of them is taking often I call them kit games, so like Rim World, Skyrim, games where the real shtick is to go get like three hundred or four hundred mods and turn it into your experience and fiddle with all the settings. bait for for folks like me. Unfortunately. and then what now you can do is you can just Gen AI create whatever mod you imagined. And so I actually did this from RimWorld and created an actually random version of Randy Random, the storyteller, and I got top of the RimWorld subreddit and everything for multiple days. So now I'm seeing I'm also seeing way more mods on the Steam Workshop for every game that are clearly like claw-generated but still good. So that's one phenomenon. Another phenomenon. Julian Togelius: Yeah. Wait, wait, wait, wait, wait, wait, wait. So the so the mod is cloud generated, but not not game essentially. Allen Roush: Yeah, correct. The mods and even up to and including reverse engineering games and decrypting and all of that, so being able to modify parts that the developer didn't originally think. But there's another phenomenon I'm seeing simultaneously. So Escape from Tarkov is one of these many multiplayer games that's plagued by hackers, and if the hackers would just F off, the game would be amazing, but it's terrible because of them. And they've made a single player version called Escape from It's like single player escape from Tarkov. There was a bunch of hooplaw where they had to name change recently. But I have seen one of the most advanced AI developments in this single player variant mod where somebody has created a mod for it called Sane, which is like AI designed to replicate human behavior, up to and including like chatting a human to you, like in multiplayer. They even have a version where you could simulate the hackers if you wanted 10 like 1% of the AI bots to spawn in and hack as like a joke. But that really does imitate human-style behavior in a way that I have never seen in any game, minus a little bit like Fear. And so I guess what what I'm wondering then is like, do you think that we'll see people using AI to like bootleg mod games to be their heart's content, to be like single-player versions of multiplayer Julian Togelius: yeah. Allen Roush: games overran by hackers, or even simulating multiplayer game multiplayers in dead games? games where if they had player bases they would be really excellent but just aren't because we're all playing crap new games or something. Julian Togelius: That's that's just it's it's it's just so poetic, It's just me in a dead MMO and all the ghosts of players that were once there. Allen Roush: Yeah, exactly. Julian Togelius: It's it's it makes sense. I think yes, I think all of that will happen. And and and I think it's kind of cool in an from an like an art perspective, like it's this performance art part of me. I am not that player. Most People are not those players. most people want, like, you know, a well-designed game experience. I I have a family now. I don't have time for this. I still want to play games so much, but like I I basically almost exclusively play console games and sometimes mobile games because just like I don't have time for all the mess. I don't want to hack around with the shit. I hack enough otherwise. so I kind of honestly this might be a very strong argument for kind of more locked-down console games. Like Nintendo is gonna win on this and Sony to some extent. Because yeah, you can keep your play experience controlled, you can stay keep the hackers out and so on. Maybe we'll see a renaissance more for like this very kind of like more tightly controlled single-player games. but this being said. in in theory it's a super cool thing where people going wild with modding multiplayer games to their hardest light. I don't see what the business model is in any of that. maybe I'm just like but but whatever, it's sounds fun. It's like yeah. I I I think this this this also means that there's gonna be like probably for things like Fortnite, there's gonna s at some point gonna be some kind of arms race with like detecting people who use it in AI assisted illicit ways in some way Ravid Shwartz-Ziv: I I want to talk a bit about the the other side, about AI that actually play games, so you have a recent work about game pain game planning, right? it's a strong test for AGI Julian Togelius: Yeah, it's it's kinda funny. Because Ravid Shwartz-Ziv: Tell us a bit about it and then I w I will I so I have a few questions, but let let's start with like what what does it do? Julian Togelius: I'm not enti entirely sure which work you referred to. Was it this little opinion paper or was it like one of our benchmark papers or like Ravid Shwartz-Ziv: So so actually both, the opinion so both the the opinion paper about yeah, yeah, yeah, exactly. Julian Togelius: What can what can AGI learn from game AI? yeah. And and and and and yeah so there we basically looked at like okay what about what what does game I mean this is kind of re re Telling the story I told, twenty minutes ago or something. games are like extremely multifacet. There's so many different types of games and video games and board games and card games that that engage different kinds of the mind. And a good game is designed to be a good learning machine and therefore you may want to try train it. And it's like it's engaging because because it tests your s your your some aspect of your cognitive skills. I have like in one of my fundamental philosophical beliefs is that there isn't a single thing such as intelligence. There is a large variety of different abilities that may or may not correlate to various extent in different ways. and and and basically game design you can see this as an as an empirical cognitive science where people design new games to explore new niches of human cognitive ability. Like a new good game design tests the mind in a different way than existing game designs. As long as it is an activity that people want to do and people people want to do activities that test the mind in some way. So you can see like you know if had a fully fledged game design and you found all the game designs were good, you would kind of have an inverse mold or like a mold of human cognitive ability, right? So so this means that if you really want to test AGI, which is like absolutely bullshit term, that people who use it should be fined. And mm thrown out on the street and then you should pour like cat urine on them. and it's it's it's it's how bad that term is and how much it has destroyed. But say you wanted to basically test AGI, well then An ATI should be able to play all games that humans find interesting. This is not about learning how to play them from scratch. It's about or this is not about being able to play them out of the box, but being able to learn to play them really fast. because that's what kind of what that's what humans do. Ravid Shwartz-Ziv: so you basically you argue that games are like still a useful test of general intelligence. So what would an AI need t to do in I don't know completely new game before you would consider it as as as general what does it mean now to have a general system to to play games, right? Like what now the I'm putting an AI agent in a totally new game. So what you consider as like general behaviour? Julian Togelius: to play it in the time taken for a single human to do it, using the same not by like, you know, brute forcing it to put back having ten thousand games in parallel. It should be able to p have the same amount of gameplay time as a human would do it and and and learn to play it as good as a human would and and f and and finish it. And it should have done this without having played it before or even being trained on anything that's really close to it. Because when people see that, you know, GPT-6 can step step its way through many games by pausing, taking taking actions and pausing, taking actions. And of course, we have the real-time issue here that it's too slow and you need to pause all the time, but even if it was hundred times faster, The issue is that these games, it has almost certainly seen them before. It's like, you know, when people were like the Voyager agent back in 2024, maybe, or three, I can't remember. that could play Minecraft by having by basically writing code to do things. It It did so because I mean it was cool. It was very good work. But it was b Minecraft is literally one of the world's most famous games. And it has you go out you go out in the world like an insane amount of like of stuff written about Minecraft and YouTube predictors and whatever. So basically you need to do is a new game you have never seen before and in the same time as a human, learn to play it well. And it needs to be not one game, but all of the games. Or like the top thousand on App Store and Steam and something like this. Allen Roush: So so what do you make of the GPT six Astra results of it playing games on Twitch? I just saw it like complete Fallout three and it used a lot of vats, right? The the the pseudo turn based mode. But in most games so far that they've been able to throw at it, including Rimworld, like or and and I think probably now I wouldn't be surprised if they did it to net hack. and in Rimworld there's like a f it it's you know goes on forever, but you can have the optional ending by getting off the planet and so it was able to do that quite quickly. So I'm seeing I claim now most games that aren't gated by tokens are really expensive, I don't get them fast enough to handle MOBAs or whatever, right? It seems like most games are now meeting that criterion. Would you Julian Togelius: not the cost but how how fast it is. Plus Allen Roush: Yeah, speed, yes. Julian Togelius: yeah, speed. Plus that it should not have seen it before. So we need we we need to bring up games that we know that GPT six had not had not seen during training. So we need to make a good game that we know it has not been trained on, so we l basically need to air gap We make a we take a game studio, put them on a South Pole, air gap them and make them develop a game and then and Allen Roush: Ha ha ha ha. Julian Togelius: then we try it. You know. But it shouldn't be one, it should be like like, you know, all of them that are good for humans. Ravid Shwartz-Ziv: What about reward hacking? so I'm assuming that these agents, these models will find some hacks inside these games. do you think this is something that like it's good, it's bad, like do you think this is we can prevent it? Julian Togelius: Well well, in a sense that's from the perspective of a game designer, that just shows that you had an issue with your game design and that's the that's where automated game testing comes in. That's basically trying to play them and finding a reward hack. Well then you found an issue with the game and you found a bug that needs to fixing. So so it's Ravid Shwartz-Ziv: But this is like this is like with humans, right? you basically give humans to to test it. Julian Togelius: Yeah, that's what we do now. But the game company Ravid Shwartz-Ziv: Yeah. Julian Togelius: I founded seven years ago now, that which is like Model AI, which is about automatic game testing. We tried a lot of things. and it to some in many ways we're too early. The company still exists and we have some new stuff happen going, and it's I dare say it more or less works for at least some Cancer games these days. but it the basic thought is this, like you know, try to play a game and if you do it in a way which is not what a designer intended, well then you found something there's an error in the game that we should fix. there's lots of other bugs you can find as well. this is a real use of AI to try to automate away the most boring part of game development, the part that people don't want to do. which in striking difference to the kind of AI automation that makes half of the game industry hate AI. Ravid Shwartz-Ziv: what do you think about like let's say more collaborative system or cooperative like AI human like I don't know that now I'm trying with like I don't know swarm of agents to I don't know conquer some castle in in a game or things like that. do you think this is something that we'll see more and more? You know, like today like we talked about it that we that we we can we can use AI, right? to to have the game more dynamic and to have it like less strict, but do you think we will see it also with other like other players and that communicate with each other? Julian Togelius: Yeah, I think this is a very fertile space for game design in a sense that it basically means that you can do a lot of game design around what would it mean to play w together with AI agents. design games that are meant to be played with AI agents. I'm not much of a game designer. I'm mostly an AI researcher, but I'm an AI researcher with some game design sensibility. And one thing that strikes me with like the thing that we're all doing these days, like, you know, outsourcing tasks to AI agents and making them do things, it's kind of a game like interaction on his own right, right? You're kind of like you're telling Claude, like, hey, Go and do this for me, and then you don't know if it's gonna do it well and what's gonna come out of it, but then you go back and check on it, and like, this is what happened. That's kind of a game already, right? And and I think there's a lot of this is one of the best angles for someone who wants to design the next AI-powered game, I think, to to work with, like to work with designing games. based on human agent collaboration. This is this is another one of like I'm not sure if I answered your question, but I said something I wanted to say. Allen Roush: Do you do do you think we'll have Ravid Shwartz-Ziv: Yeah. Allen Roush: a harder time verifying if somebody playing a game is is AI? And this not just in hacker sense, but I mean across the whole internet that seems to be one of the particular fearful scenarios from the labs who are trying to talk about pausing development as the internet being overran. By by by just bots and AI systems. I also want to point out that's totally a plot device in a lot of cyberpunk and especially Cyberpunk twenty seventy-seven. And I I really do just tell people in general we're living in a cyberpunk novel. It's just figuring out which one. Julian Togelius: Yeah. It's amazing. It's like, and all the ways in which the internet is just actively getting worse for for for all the a for all the agents. Like, you know, there's like there's a broader thing I've been meaning to write at some point. There's all these things. I write a bunch of things, but then I want to write many more things. There's all kind of like things where like, you know, I wrote this piece compli called Complementary Intelligence recently, which is like some of my kind of philosophical view on on intelligence. But I wanted to in in the history of AI, how it's like all taken different directions and this is one different one direction. But I've been thinking of writing a more strongly worded thing, basically like saying something like, We invented the wrong AI. Come on, it's just like you know, just imitate first like imitating all the human text and then teaching it to do the things humans do. It's just like the wrong thing to do. It's like we should not have done this. We should have done other kinds of AI instead. and for people to think that intelligence is Intelligence, then they have a lot of growing up to do. But basically it's not. It's we decided to build this. thing that is just as much as possible like us, but we could have been doing there's so many other much more interesting directions in AAI that we could have taken and we have Ps the past to do things that are genuinely complementary to us and are good at completely different things. But anyway, here we are. Ravid Shwartz-Ziv: do you have specific annual directions? Julian Togelius: here here's the weird things. The stuff that is very cool and different tends not to be called AI. Like a lot of computer science was about why was the initial invention of the digital computer not called superintelligence? Because something that humans did and didn't really want to do and thought was boring, like carrying out lots of actual computation by hand, like adding and subtracting and multiplying and like solving like you know according to formulas, s solving like second degree polynomials and stuff like this. This is boring and it's totally algorithmically solvable. Then we made machines that we follow these algorithms. We should have called that super intelligent, shouldn't why didn't we? Because literally because we didn't have the word at the moment. But that is what it was. And that's like the best kind of AI. It's doing something we don't want to do, right? And and letting us focus on stuff that we want to do. and And a lot of computer science inventions that we don't think of as as as AI is really it's kind of the real AI, like databases and boring shit like that. it should suppose AI is supposed to be boring, shouldn't it? Or AI could be like all the stuff that comes out the artificial life tradition, all the kind of like little sort of, you know, life forms on cellular automata and things like this. These are not they can't talk to you, they can't write code for you and they can't like do your taxes but they are like genuinely very different and very interesting in that sense in terms of like little artificial life forms and so on and Yeah. this is this is in a sense what's really interesting. The nice thing with video games is that there is still a s a space for this of designing around like, you know, how do you interact with really different things. Whereas AI has become here's this extremely long long laundry list of things that humans do to and make money out of it. Now make make the machine do exactly them exactly like humans want, like like humans did, so the humans can't make money. It's like the amount th th the extent to which we just went the wrong way in AI is like staggering. Ravid Shwartz-Ziv: What do you think about personal assistance? do you think this is something that the right direction. because like from one end like it replaced a lot of like boring, time consuming things that you're doing, but it's also like right, like it can be I don't know, some replacement of like human interactions and Julian Togelius: It is there's is always a very fine line to walk here. But but which which side you wanna fall on this. and Yes, it's about helping you without substituting you. Very tricky, complicated, but I think there's a lot of scope for that. This is also, by the way, why I think like, you know, everybody will ask what what's the biggest news in AI from last week. And or like if if you ask people, they would say JEV, you know, because and a JEV is cool because it's genuinely a different thing than it's not an LM. It's an interesting thing. We're all trying to grapple what what we can build with it, but And like I don't know, if you also were like running experiments all weekend with trying to do j make Jeb do do fun things. but I think the real big thing was that series AI finally came out. this is like so extremely uncool to say. This is why I love saying it. This this is the real frontier. And because sure, like the series AI app is behind on raw Large model capabilities is behind the frontiers, maybe a year behind the frontier, whatever. But in terms of system integration and actual computer use, not the pretend role-playing computer use where you look at pixels and and and move a move a mouse cursor, but like the the way it should be, where it's actually calling calling the system via API calls. It's light years ahead of everything. You can just ask your iPhone and it does stuff for you. It's great. That is personal assistant in a good way, I think. And most of the time it does this locally, running and running locally. So it actually makes me happy. It's also like, you know, that there's such enormous financial firepower behind doing something that's the w where where you are the consumer because you are paying for Apple's overpriced product rather than the product. Allen Roush: I I wanna quickly talk about Jeb for even just a minute because Julian Togelius: Yeah. Allen Roush: Jeb seems fascinating and I love the memes of people asking all the millennium questions and you know, Jeb having ninety nine percent confidence P is not equal to NP, for example. But but but like how is this different from structured generation again in terms of use cases? Because I could trivially make any language model into a classifier with structured outputs. Julian Togelius: Yeah, but it's like Allen Roush: Sorry, what? Julian Togelius: But it will be slow. Allen Roush: Slower. It's still very fast. It just pre-fills my input and then it's like one or like a very small number of decode tokens. Julian Togelius: Yes. Yeah. Allen Roush: Yeah, I did. Not everybody else did, but I sure did, and Raveed can vouch. Julian Togelius: Right. Yeah, okay. My mental model is that sure it has a text encoder coming in somewhere, and then on the other side it's something much more like a classical neural network with a number of outputs. And you can probably take an existing anonymas what you said and basically have it output a few tokens, interpret them as like these outputs and you tr find training tuned to do that. And that would probably be a way of doing the same thing and maybe it would be as good. But probably just doing it natively, whatever architecture JEV is, is probably better and more efficient. What's cool is that, when someone comes up with a version of this that runs locally, which of course will happen, then it opens up like a whole category of very, very fast things, you know, that it's like that the fact that it is like it is The way I think of it is yeah, we've all been able to train neural networks for specific use cases for a long time. That's that's great, but that then you need to train it, you need data, you to do the training and so on, but then it's super fast. Here is something super fast, but it's like text conditional instead, right? That's kind of like that that's kind of novelty from a conceptual standpoint. you're specifying your fast classifier instead of training it. And Ravid Shwartz-Ziv: I I I have yeah, I have a different question. So I I want to talk a bit about code versus like I don't know like other modalities like maybe more general approach. So in games Julian Togelius: yeah, because that's the was the other part we didn't get to about that position paper. what can AGI learn from game AI? That Ravid Shwartz-Ziv: Yeah. Julian Togelius: coding is just one particular game and it's like an extremely good game. That's why it's so fun. And and when there's insanely much data about it, and like each programming language is kind of one level, and current LLMs are extremely fine, extremely sort of overtrained on this particular game, that's where they're so good at it. Ravid Shwartz-Ziv: So wait, so do you think that every task that we can actually transfer to be a game is actually a good task to for for LM? Julian Togelius: Depends on spam, I guess. any task that we can transfer to be a game is yes, if we can make a game out of the task, it is it means that it's relatively easy to learn it in that sense, yes. Ravid Shwartz-Ziv: AI models are really good in code, right? So they can use like existing game engine, right? That like that you can actually like generate games or like videos or that and whatever through code, right? This is like kind of they're good on in that. But they are not so good in yeah, now just generate a game with animation and graphics and whatever without code. First of all, do you think there is something fundamental there or and and and second do you think we can there there is a way to to solve it? Or like do do d does it important to solve it? Or like basically you think like every game that now we want to design from from from from this moment like we basically need to to to do it through code. Julian Togelius: But don't don't only do that. If you look at Unity AI, default AI assistant in the world's most widely used game engine. disclaimer, I'm on the Unity AI board, so you know, it's just like I have an opinion here. But basically it's it has a strong LLM behind it, but it also has a whole bunch of other models. There's an orchestration where it does like a whole bunch of like you know standard diffusion models and like you know there might even be transformer-based general models but models that are specifically make and 2D and 3D textures for example and in and objects So and those are these days pretty well specifiable. and there's a whole kind of ecosystem around them. They are kind of in a sense they are they get less attention because most of them are so small that they run efficiently on local devices. So there's like in some sense less money in them. And it's true that if you ask If you ask Claude Fable to make you a Grand Theft Auto 3D game, then it's gonna make you one with these ugly fing sort of you know polygons that he sort of just like maths out. and I actually haven't tried. I'm gonna do that next. I'm gonna try to tell it to just download a bunch of state of the artificial models and use those to sort of generate all your all your assets as you make the game. Probably doable. I bet it knows how to do it. So so it's not only code. You can definitely do all the there are often models for all the other modalities as well. Ravid Shwartz-Ziv: And do you think th they're they're good enough or as the same as coding models? Julian Togelius: 2D, yes. 3D almost. you still need a little bit of like it's mostly the feedback and like on on the 3D part. The 3D models itself are actually good at generating things, but when things get wrong, it's I haven't really yet seen a harness that's really good at giving the feedback that it needs to sort of fix the errors that come out of the 3D models. it's not, I mean, think it's solvable with a really good harness and a really good kind of like way doing that. But yeah. an interesting part, and this is something that all the research people in in the games industry are thinking about right now, is to what extent a video world models are going to play a role in this and which role to play. I don't know if you know that like last year when genie the genie three like the project genie was like officially released it it it had been known to the world for quite a while before then, but when it when it was officially released, the stocks of like most games companies tanked. They basically lost like, you know, double digit percentages that day. Which is an insanely weird reaction. But like but then if you really look into what it is, it's like you cannot build a game on a on a video world model as it is right now. And even if they got way better at what they're doing and way more efficient and so on, you have the basic authorability problem. How would you author with that with video world model? But what is going to happen probably is that you're going to see way more of these generative models on the presentation side of games. So we may see future game engines essentially handling all the logic and then basically producing a bunch of structured output that basically renders as like white box, like like a bunch of like geometric figures that are have semantic codes to them, and then that's turned into beautiful graphics by something genie-like. Ravid Shwartz-Ziv: But you don't think we will see end to end video world models, models that we will Julian Togelius: no, we would see them, but they will not be as good. I mean it's like obviously Genie kind of is, but they are not I have as hard time to find the real niche where they would have an advantage over a combination of like of existing game engines with really really strong graphics models on top like DLSS 10 or whatever. yeah. And and much after that, most of that has to do with how you author them. Either you're a human or you're a language model or most likely you're both in some combination. Like it's easier to author in a structured format. Ravid Shwartz-Ziv: I have one more question. so you're you're like you and your PhD student, you so you said it like you're you really love ga games and you played games even before, but is it true for your PhD students also? Or like what Julian Togelius: Yes, Ravid Shwartz-Ziv: you actually need in order to be a researcher in in in a lab similar to to yours? Julian Togelius: Well. So I've I've I've had a lot of reason to think about this because I'm actually leaving New York University at the end of the year and moving to Chalmers University in Sweden because we've been wanting to live in Sweden for a bit with try to see if the kids like it there and learn the language, the culture, stuff like this, and and you know, the overcast guys, the grain the rain, the the the the kind of the boring food, all this stuff, you know. but Ravid Shwartz-Ziv: Fine. Julian Togelius: but Anyway, so I'm I'm recruiting a whole bunch of like I was given a deal where I mostly don't have to worry about funding, which is also always very nice. so I'm recruiting a whole bunch of people to my to my lab. So I have PhD and postdoc openings. search for my name, find my PhD and postdoc openings. And I've been thinking a lot about this. And I'm not constrained to only working with games, but I find find that games are just an inexhaustible supply of good problems and good opportunities. so yeah, I like people to play games. I like people to think about games. It's great if they actually design games as well, because that makes you think about games in a in a different manner. And it's great if people have somewhat esoteric tastes. but beyond that, I mean you need I really like when people have a bunch of different interests that you don't typically see together. That's what's but what's what's makes me sit up and look at like, this is cool, this is interesting. These are a couple of new things like you sort of deeply into these completely different fields then we'll see what com comes out of that. Yeah. That's what I thought. Ravid Shwartz-Ziv: Sounds great, yeah. Yeah. okay, Julian, I think we are out of time. Do you have anything else that you want to add? Your favorite game. you didn't say your favorite game. Julian Togelius: Yeah, no no no. I just like what is my favorite game right now? There was a time when Our son was just born that I was sitting up at night with him and I was finishing Elden Ring. and and Elden Ring is so good. And at the same time I was playing Battle of Polytopia on which is kind of like a Fisher Price mini civilization thing on my phone all the time. And then and then I realized that these two were the favorite games of Elon Musk, so I had to stop playing them because I don't want to become like him. and but otherwise if you look at I see lunch has arrived outside very nice. otherwise if you look at like favorite games of all time, like you know, there are games that are such extreme masterpieces that you know you can you how could not mention Zelda, Breath of the Wild and Tears of the Kingdom, like their objective is some of the greatest games ever made. But you know, and and and I've been deep into the Mass Effect series, I really like. procedural content generation and I been into a bunch of roguelikes. played a bunch of Spelunky. Hades is a fantastic combination of generation and storytelling in Hades too. But if you go back to the game series that probably taught them the most is probably Civilization. and I feel like I learned half of all my English through playing Civilization as a kid. Allen Roush: I I I just want to quickly quip that Civilization 5, in my opinion, is probably the best of at least some modern Civ games. That's the only Civ game where the Gandhi nuclear war bug thing is actually real and not an urban legend. They put it in just to be funny in there. Julian Togelius: That's awesome. That's awesome. It's just a such a great thing. yeah, no, no. Civilization is awesome. yeah. Ravid Shwartz-Ziv: Anything else that you want to add? Julian Togelius: no, I think this has been a great, very fun discussion. let's talk more. and hey, we're still in New York, so let's have coffee. Ravid Shwartz-Ziv: Yeah, for sure. Thank you so much, Julian. It was a real Julian Togelius: Awesome. Ravid Shwartz-Ziv: pleasure. Julian Togelius: Real pleasure. Right. See you later. Allen Roush: yeah, real pleasure.