S07, E16: AI and video games (with Julian Togelius)

Mobile game developers no longer [music] need to pay up to 30% in major app store fees. With Xsolla Web Shop, you can create a direct storefront, cut fees down to as low as 5%,
[music]
and keep players engaged with bundles, rewards, and analytics. Start today at xsolla.com. That's x [music] s o l l a dot com, or use the link in the episode show notes.
The problem is that the distinction needs to be drawn between [music] the competence of the economists and the correctness of their analysis.
Hello and welcome to the Mobile Dev Memo podcast. I'm your host Eric Seufert, and I'm joined today by Julian Togelius. Julian, welcome to the podcast.
Hey, thanks Eric. Good to be here.
It's great to have you. So, I know you through having the great pleasure of of working alongside you at Unity, where we

are both on the AI Council, but you you also do a number of other things, a very sort of prestigious CV. So, I'll let you kind of introduce yourself to the audience before we dive into the conversation.
Awesome. Yeah, I'm a professor at NYU, and I also have a habit of coming up with startups. So, basically, my main research is artificial intelligence for games and games for artificial intelligence. So, how do you make games better with AI, and how do you use games to test AI developed through AI? And when I say games, I mean anything that people play, not so much like traditional board games, but more like, you know, Starcraft, Minecraft, any craft, Super Mario Bros, Doom, etc. And poker, and so on. And then, I've founded a game testing company called model AI, m o d l dot a i, and recently, I've joined as head of AI another startup called An oft one, where we building air for finance and do I know anything about finance? No, but you

know, it's just another game isn't it? And by way of background I'm from Sweden originally but I have been living in a bunch of places Switzerland, England and so on an academic career tends to take you to places some more interesting than ours.
Yeah, so I actually had Christopher Homewood on the podcast in October 2024. So he's the the CEO of of model. So it's great to have you. It was tough to kind of come up with the questions because I think there's just so many interesting directions that we could take this conversation. I think just given our sort of, you know, mutual interest in games that's probably the most logical place to start. So maybe just as a as a big picture kind of question to set the tone for the conversation. What what makes games a particularly rich area of research for AI? Why why start there? Why why focus there?
Games are fun and that seems like a sort of inane thing to say, "Ah, you're just working with games because it's fun, not because it's fun stuff." But I think it's kind of it.

Something very deep about games being fun. What does fun mean? I mean fun is a very under defined concept and we all can use it from time to time and basically it means that something is interesting. It means that something engages our minds in a very particular thing. There is a famous theory which comes in different guises. In game design it's popularized by Ralph Koster who is a designer of a bunch of different well-known games. Star Wars online is maybe the most famous and it comes from developmental psychology like the Piaget and Vygotsky and so on where you look at like the zones of proximal development but any course in machine learning is on my former post doc advisor Jurgen Schmidhuber. And it basically converges on that things are fun. Games are another type of fun because we learn as we go about them. They push our limits. They put us in a zone of proximal development. They force us to learn what's coming next. And when the sign game is a extremely well-designed machine for

pedagogics, essentially, for teaching it itself. It's not all there is to game. There are many other things to game. There's games There's also a thing that like most games, almost all of them, are an abstraction of some interesting real-life activity. So, you know, chess is about battles, and Tetris is about, well, stacking things, you know, SimCity is about building cities, Super Mario Bros. is about playing on a playground, essentially. I was thinking about this the other day when my son was, who is a big Super Mario Bros. fan, was like jumping into like a cardboard sheet on a playground. He was just like, "Oh, I'm going into the pipe. Let's see where I come out." He's like, "Oh, yeah, right. This really is a playground. This is a platform game, or this playground is a platform game." So, there are many good reasons why games are so interesting to study and to study AI with. And this has been recognized for a very long time. But for a very long time, the only games that were like

kind of kosher to work with AI were like boring stuff like chess and checkers and Go. I mean, I'm sorry. Apologies if you love chess and Go and so on. You know, it's great, but it's not my interest, and and basically so much There's so much more to games. So, you saw after AlphaGo, which is now 10 years old, people saying that, "Okay, now we're done with games. We beat the games. Now we can go on to more interesting stuff." And I'm like, "Wait. Wait. We're just getting started." And we're still just getting started because there's so many different games, and there's so many different challenges, and so many different things. Like basically, chess and checkers and Go and stuff, they're all like, you know, zero-sum, discrete, perfect information games with the relatively few turns, relatively low branching factor, etc. And really not very indicative at all about all these fascinating video games that are out there. Well, the other thing is like it's not just about

playing games and it's not just about playing games well to win them. It's also about playing games in the style of a particular player, generating game content, generating complete games, understanding players, and so on.
So, Raph Koster is someone that I look up to a lot. So, I actually saw him speak once at an event in Helsinki, but his book, A Theory of Fun, is I think kind of a must-read for anybody that works in gaming. And it's interesting cuz I mean is is So, you said that you think Star Wars is the most famous of his games. I think it's Ultima Online.
Oh, yeah.
But what's interesting about those, you know, these MMORPGs is it that that concept is a fundamental deviation from like a platform game, like Mario, right? Like so So, you think about like computer like interactive, you know, video games, you know, there there were maybe sets of paths that you could take, but there was an endpoint, right? And you would beat it, right? That was the you'd beat the game, right?
Mhm.
And when you got to MMORPGs, there's no

beating it, right? I mean, this is there's like this living society that persists. And you know, the incentive structure was different. A lot of it was like social, right? There you know, different ways of like there's no essentially like a score, right? Like there's maybe different ways you could consider like a score in terms of like the stats of your character, right? Or the number the amount of money you had or the side like in Ultima Online, you could have a house, right? And like you could have a castle. And the thing is like maybe that's a way of keeping score in some sense or comparing yourself against other people, but there was no sort of singular you know, kind of quantitative benchmark.
But but isn't this the case? Isn't this the case for like many other games as well, like you know, so my son who turns four in like two weeks. So my son and I play Super Mario Wonder and we beat him and it's great. It's great Nintendo design. We played this together and it's like they they're so good at designing for like you know, asymmetric co-op basically and we beat the game, you know, we kicked Bowser out of the

castle. But that's just really the beginning of it. Like even though it is in a sense as traditional a game as it gets and there's so many things you can set yourself as many goals to set, like you know, and sometimes my son wants to replay various levels with no particular reason. Just like you know, we're inventing new things of like you know, carrying each other around and so on. There's so many kind of like there's so much game invention that goes on inside the game playing with
No, yeah, yes, that's my point. That's kind of where I was getting at. Sorry, I was I ended up getting caught in a whirlpool there. But so you know, Ralph Koster, I mean with with MMORPGs and that's part of what he talks about in the theory of fun. It's like game is not some one defined thing. Especially like like interactive video games, right? Like they're not there's not one template for it. Like and and really what you're trying to do is you're trying to materialize that that the primitive is fun. That's the whole theory of fun. Like how can we define that some way where we could apply it in all these different settings that may or may not look like what you would consider to be a game. And certain certainly like in many MMORPGs, I've

spent a lot of times like being mad,
[laughter]
right? Like it's just frustrating and and irritated. And so like how does that factor in? So so like now you think about it like okay, well now it's it's this this much more abstract idea, right? Like you could teach a an AI like there's probably like an AI agent that could like beat the original Super Mario Brothers and like you speed run it and beat it very quickly, right? Like that's not really a great use case of AI, right? I think like the more sort of applicable and like helpful and valuable use case is probably summoning that that primitive of fun and allowing it to sort of explore that. Is that is that maybe a like a a topic of research or how how do you think about that?
Oh, yeah. Oh, yeah, definitely. So, Super Mario Bros. is kind of funny because it's like it's a ubiquitous kind of model organism in game AI research, partly because of like Notch, the original creator of Minecraft, Markus. He before Minecraft he did this did a Java-based version of a Super Mario Bros. clone Mario Bros. and me and one of my students this back in my post-doc

days turned this into an AI competition and that's basically been proliferating all over like this and it was back in 2009 and it's also funny like, you know, could you make an agent that just wins and plays perfectly at Super Mario Bros.? Well, that depends, you know. It's like the I can turn on academic mode. Well, it depends on what kind of inputs you give the agent. Do you give it a simulator? If you have a simulator, yeah, obviously an AI star agent can do extremely well and so on. And if you don't have that, it depends on is it like level 16 before or level seen or not seen before and so on. Unseen levels, perfect play on unseen levels is perhaps surprisingly we're not there at all. We're not even up there. We don't even have an agent that can reliably win unseen levels in the absence of a simulator. Perhaps surprisingly. And this is a game that's been studied to death. This is like an insane amount of it. But a lot of work with this that is the model organism. I'm I'm using model organism in the same way as you talk about fruit flies and genetics or like

this little snail with 608 neurons whatever in neuroscience. And it's basically people study Super Mario Bros. because it's easy to work with and very understandable. And there's a lot of work in level generation and generating levels that bring out various types of challenge and that would incentivize various types of fun, that would allow for as many different play styles as possible and so on. A lot of the research is not really about playing the game. It's about measuring interesting aspects of a player's play at generating new content, generating new variations, and using AI for that.
And you said something really interesting about So, it's on on your website, too, like about What is the connection that you have to finance? Well, it's the stock market is a game, right? Essentially, like it's you know, it There's obviously like real-life consequences there. But they talked to me about that because do you think that is like a motivate like is that is that a fundamental motivation for you know, people that just do Like

people like high-frequency trading, right? Like Jane Street or Jump Trading or HRT. I would have to imagine that a lot of people go to those firms because it's just an easy way to say you're the best, right? Like a lot of these people did Math IO. They're probably a lot of the people at the competitive firms, the competing firms, are people they know that competed against in Math IO. And you know, if you implement an algorithm or you just have a better year than somebody else, you beat them. Like I have to imagine that's part of the motivation for some people to do it other than just the amount of
Yeah, definitely. Lots of people go into quantitative finance and other types of finance because they just like the competition. They treat it as a game. But then again, on another level, the agents that you use for trading, because there's a lot of agent trading going on, even before I don't know if you guys know what I'm talking about, but agent trading is little little snippets of code that execute buy and sell and hold actions all the time. I mean, that's how the kind of like cyberspace Wall Street kind of thing works, right? And they are pretty game playing agents. They are

basically uh little agents acting in reinforcement learning environments, and they're learning to get better. The issue here, the problem here, is that the game is changed all the time. So, basically, you have the non-stationarity of the markets. The market conditions are different today than they were yesterday, and the and the policy you had yesterday that made you money may actually be counterproductive today. And how are you going to adapt to that? You can do it in many different ways. And one way you can look at it is that like it's not the local thing, it's like the whole kind of thing, everything you can observe in the world of and all the actions you could take. You could short Bitcoin, buy Volvo, and you know, keep your something else like options on something weird, a South African gold or whatever. But, the action space and observation space for this problem is like humongous, it's enormous. You can't play that game. You don't have anything near the amount of data you need to learn a policy for that game. Where do you go from there? That was what everybody does is they go the other way basically. They try to find very

limited little subgames, so they try to find the subgames of like, you know, this kind of option today based on these kind of inputs. We can learn this and then we keep retraining it and that's why the financial industry tends to work with very small models. So, people who come from machine learning and go into finance, especially the people who go from deep learning and go into finance, they tend to get a little bit shocked like, "What are you doing? Why are you still having linear models and logistical regression and decision trees and stuff like this?" And like the finance people will look at you like, "It works, you know." We and we can retrain it extremely fast. So, what's the connection to games here? The connection to games is in generality. I said it before that, you know, there's actually not an agent that can play all of Super Mario Bros. levels, you know, proficiently in the absence of a good simulator. And this is surprising and it's and if you take it one step further and you want to have an agent that can play any 2D platformer game in the style of Super Mario Bros.

We definitely don't have that. We're not even close. I mean, [snorts] the the problem of the general game playing is not like it's not only not solved, it's actually we barely have any progress on it. Some people will take issue with this statement and I will say that I actually know better. But, it is surprisingly hard and general game playing is like general finance. And that's why people focus on playing one game like Go or Super Mario Bros. Or like, you know, trading one type of option and one type of horizon or something like this.
Yeah.
Because they can learn something.
Let's get real. Half the time, you're not sure if your campaigns work. You just kind of hope they do. Attribution numbers look good, but what if those conversions would have happened anyway? Incremental helps you find out for real. No tracking people, no guessing games. Just clear answers on which campaigns actually [music] move the needle. If you want to stop guessing and start knowing, go to Incremental. Because good

marketing isn't about getting clicks, it's about getting results. Check them out at incremental.com. That's i n [music] c r m n t a l dot com or click the link in the show description.
I want to talk about like generative tools for game production, right? So, you've done a lot of research there. And is the gaming industry fighting an uphill battle in terms of consumer sentiment around the use of AI in development? So, right? Like I I have the Crimson Desert experience in my mind, right? There was this and I mean this is I think this is kind of like a Redditor, you know, a backlash. I don't know if this is like a general public backlash. But, you know, these on Reddit people were upset that these obvious like placeholder graphics were made by AI were shipped in the game and how does the gaming industry position the use of AI as a benefit for consumers? How do they make the case that look, you're better off if it costs us less, we can create more with the same budget, you're going to get more content, you're going to get more games as a result of these tools? How do they

how do they make that case? Or is it is it just too late? I mean, we've just missed the boat on that.
No, I I find it super fascinating. Most game developers are just staying away from it. We run this game AI summer school every year since 2018. This year is going to be a June 15th to 19th in Leiden. And we always we we think it's it's a lot of fun doing this. It's me and a colleague of mine, Georgios. And we invite people doing AI from within the games industry in a bunch of different roles and a bunch of different companies. And weirdly enough, it's actually become harder to get the speakers we really want. I mean, we have a great lineup this year because we worked a lot on it. But a lot of people said no because they don't want to be associated with AI. Like, maybe not them personally, but their companies said basically, "Look, we cannot go out. We, this game studio, cannot go out and do anything at the conference. What if our players find out?" Which I think is super fascinating.

I don't know where this is going. If this is if we're going to get into even more polarized environment, everything is polarized these these days, right? And the AI debate is like in a polarized and very very unproductive way. There's a certain kind of boosters that are like, you know, "AI is gigantic. Everything is the future of everything." And you know, "Please outsource your brain to Claude as soon as you can." And then there's like, you know, this kind of like Luddites that are like, you know, "I won't touch generative AI with a pole." And it's these are very unproductive positions, both of them, right? It's like there are real issues. And I think the games industry as a whole is just staying away from it. They're just trying not to mention it. I don't know. Do you see Do you see Do you see it differently or
No, I I do see it that way, which I think is is uh unfortunate.
Sad. Yeah. I'd like to think about like, you know, there is AI in game dev, and there are things that are like, yeah. I mean, no one wanted GTA VI to take how

many years to develop? Like, you know, some kind of crazy stuff. Clearly, a little bit of Clearly, there are like, you know, things that could be made more efficient, right? But then there's also other things that AI can make possible in terms of new game experiences, new things that we couldn't build otherwise. Like AI of various kinds inside the game, from non-player characters to environment generation to like player adaptation and a lot of fascinating stuff. And I fear, because that's one of my main interests really, and I fear that that's is going to that's going to get caught up basically. People are going to sort of not want to experiment with this because it looks bad.
Yeah, or just that it gets leaked that they used it and there's some kind of moral panic about it, which is so bizarre to me because I mean, look, if you want indies to make commercially successful games that reach your phone or reach your, you know,

PlayStation or whatever, I mean, what better pathway to that than these AI tools? I mean, like it's such a net good. I I don't I don't even really understand what the perceived downside is from the gamer's perspective. Like other than they think that they're just getting like a kind of algorithmically generated experience versus something that was like, you know, hand-crafted and artisanal. Which which but I don't think anybody is intending to use AI for that purpose, right? And like and that was the weird thing to me about the Crimson Desert episode. It was It was like textures, right? Like is that an issue like that the textures were made by AI? Does that really matter?
No, that's it. It's It's kind of silly. I don't know. I tend to think of it as like what people feel. There's a lot of people's reactions to AI in society that's kind of proxies. People are People dislike generative AI in general and they fear that it's like taking over things that they were good at doing, things they enjoy doing. And I I mean, I can relate. I think I'm a good writer. I

like writing. I do not like that the models are suddenly writing. I I this is my personal opinion, but people see this and they kind of perceive that's okay, this is nothing I can do anything about and then they get angry at like they just end up being about that take the like basically that soak up the electric power or they get angry at textures generated in some game, you know, it's like it's a proxy error, I think.
Right.
It's also true that I don't like reading something. If someone sends me some AI generated piece of text, I get slightly angry at them. Like, you know, why why why do I have to read this if you didn't write it? And it's a bit like I'm playing this level in this game I paid money for and you know, I expect the kind of intentionality where like if I run into like this kind of data and and there is not something for me there. There was no one who thought about this, like, you know, part of playing a game is that you are interacting with a system that someone built for for people

like you, right?
So I I get offended if someone sends me obvious AI output like ChatGPT. I I get offended. That's offensive. But at the same at the same time, I also get offended if someone sends me sloppy, [snorts] grammatically incorrect text that is riddled with spelling errors or or punctuation mistakes. I expect now I expect perfection. You've got all the tools at your disposal to send a something that's unclear, it's difficult to parse out any meaning from. You could have sent this to ChatGPT and had it critique this for clarity, for readability. Like you should have. I expect that now. I expect perfection. So I don't don't give me something that it just output, but give me something that it checked.
As we say in gamer speak, get good.
Right, get good. And like you have you have no excuse, right? You have no excuse for that.
Right.
And so like I I I I you know, I I take your point. It's like, hey, look, I paid, you know, 70 bucks for this game and you're sending me there's AI slop visible in the textures, but it I mean, in that case it was like someone found

it. And the thing is going back to this like kind of idea of the theory of fun and and MMORPGs, like well, at some point you've got to expect that that's the only way that some truly open world game is going to be possible because otherwise, you're you're going to 20-year dev cycle and it's still it's still going to be a constrained thing, right? And actually, that's the other that was the other thread I wanted to pull on here, which was GTA 6, right? I read an interesting, you know, thought on on Twitter the other day and someone said like GTA 6 is the last pre-AI open world game. And from now on, these games, you know, will have like much shorter dev cycles, much lower costs. I mean, it's a a billion-dollar game. GTA 6 is a billion-dollar game. That's how much it cost to make. But they they've recovered that in pre-sales. They recovered that in in pre-sales, right? So, but nonetheless, I mean, like that it's it's like it's I I think that's probably true. And my kind of reaction was, "Great. That means I don't have to wait more than a decade, right, for the next GTA 6. Uh that's fantastic. That's fantastic news."
I think that's true as well, you know, as my kind of I want better games

sooner. I mean, sure, there are like too many games already and there will always be too many games, but you know, yeah. Can we get better games? That's great, you know, I like that as a game consumer. But what I'm really interested in is like as a game researcher, an AI researcher is can we put run-time AI in here to sort of because I've had this like idea for a long time about like imagine some GTA X, you know, and you're basically you're in some place and you basically decide that you want to go 4 hours in that direction, you know, just like randomly choose it and then it would basically create a new world or like new regions. There's a city there that no one has ever been to before and it's created and it's basically has new architecture, new quest lines, new people, new sort of scenery, whatever. A little bit No Man's Sky like, but it's kind of it fits together. It fits together with like, you know, story lines from the rest of the world and maybe some of these challenges in there

are like what the game thinks that you may get enjoy doing next, you know, with like elaborations on the kind of quest you like and so on. Will people like it or will people hate it? Because on the one end it's like, you know, generative AI and game design on steroids. There's literally no human has been involved in creating this new city because it was created right now for you, you know? On the other hand, there's this long tradition of PC. Roguelikes and so on. I love roguelikes. And a significant stream of my research has been like, how can we make this better? How can we make, you know, more personalized or more controllable or more kind of like better in some version or some variant content generation for like online creation, co-creation in games. I think that's going to be a big thing and I don't know what people will think about it. I mean, I'm looking forward to it.
Well, I think that's actually that's that's up my next question really well. I think what do you what do you think the optimal AI-enabled content generation pipeline looks like? Is there a is there a human in the loop for that

reason that you can't just let this, you know, procedural train completely run away because you know, that that might I don't know. Maybe that induces some kind of psychosis or something or you know, like but then people aren't feeling like they're they're cuz it but also it's like the the loss of connectivity with the other experience that the experience that other people are having, right? Like okay, if I have this totally unique experience, like who do I create right with about it, right? Is it So is there is is that the optimal pipeline does include a human in the loop in some way? And these are just kind of like these are productivity tools. It's not generate It's not pure play totally expansive infinite generation tools. They're they're productivity tools. They make us faster. They make one person can do the work of 10 people. Or is it really No, we're going to have these experiences at some point and that actually is the optimal outcome?
I think optimal outcome. I mean, we all know that games um started off I I there's like, you know, for people who are not gamers, which is not us, you know, we're gamers. People are not gamers tend to look at games and like, oh, it's this thing, you know, this is what games look like. But, you know, games has always been like this

super multi-faceted and diverse um thing and it's getting even more like game whatever this is, you know. You look at like what kind of games were released last year and you find like everything. I mean, even if you look only look at games that actually got attention like what's this game where you're like basically unpacking someone's life? It's called unpacking or something and then, you know, you find like walking simulators and RPGs and, you know, grand strategy 4X games and, you know, and and and like I don't know how many kind of like bizarre kind of mobile game sort of genres come up there. So, I think the future will hold all of this and much more. There will definitely be games that are more or less like they are made like they are done today, mostly with a little bit of like you probably use cloud or cursor or something like this to write your code and probably some kind of generative AI tools to a certain but it's mostly like pattern on game development as today. And then, you know, there will be other games where like the design work of the game design

the game designers who I guess most game designers are perfectionists, right? They want you to have the illusion of your own agency where they actually have designed all of this in and ahead of time for you. But, there will also be a room for another type of game designer that designs the experience machine more like that and basically they set up like tool chains and so on for like things that to change as you are playing the game. Now, the question the other question you have, I mean, for me like personally I I I I think there's so much interesting to do research-wise in like, you know, this kind of more generative tool chain and this and and what the design look like when you sort of jump up an abstraction layer and you kind of never as a designer you never touch the action level or the action rooms for driving or something. You design the rooms for how this is going to be automatically designed. Some of it in during design time, but not sure if yeah, at runtime. Because that's another thing about like truly open-ended games. I mean, memory

is expensive. And the world is enormous. And that's always been the case and that's that's always going to be the case. There's always going to be more world to model than we can fit into memory in some way. That was true for like before for like any game. That was true for No Man's Sky. It's true for like, you know, GTA, whatever. This type of game will have to have some kind of runtime generation going on. The other question you had about like what is what about the social aspects of this? It's interesting because it's going to be really different, right? I used to be the kind of person who I'm like It's funny. When I was working at IT University Copenhagen, I was sitting right next to T.L. Taylor who's like a world-famous online so MMO anthropologist. She basically started that whole kind of line of research. And she was going on about how games were always social and I was like, "No. I I basically I pull the curtains and then I sort of disconnect my Xbox from Wi-Fi and I sit there and play all them

alone. The The outside world, I don't communicate with anyone. This is I'm an asocial gamer." But then she was like, "Oh, do you never read strategy guides?" Well, then I was just like, "Shut up."
[laughter]
So So basically, I think that in that case it mean might be about like inviting other people into the world that the game has created for you or that you have co-created with the game.
To your point, it's just up to the designer to make that happen, to figure out what that framework is. Maybe it's sharing, you know, generated worlds or maybe there's just some sort of new framework for making it social in that same way. It's like, "Well, I'm playing it and I'm in my own generated environment. No one has ever going to see this the exact same way I saw it, but I there's still some other touch point that I have with other people to discuss or to experience, to share the experience. Maybe it's just like some kind of system of trade or something. I don't know, it could be anything, but but that's that's the job now. That's the job is figuring out the theory of fun, right? How do you connect that to

this theory of fun? That that's the true primitive in like the actual format that it takes or the just the visual output is is almost like secondary to that and that's that just is what the job becomes. But I think like, you know, I was talking about this yesterday. I was I gave a talk at this event and I was having lunch with the CEO and this the chief product officer of this company. And one of the guys his his son is like 17, right? So he's thinking about colleges and I was saying like, well, I don't I don't envy you. You know, my my kids are very young, right? So I'm not to think about this for a while. But I don't know what I would tell my son to study now and probably it would be just humanities. Like I I genuinely don't think I could recommend going to study computer science, which is unfortunate because I love I love coding. I used to love I found it really relaxing. Like I felt like it was So one thing is that I've always just found really interesting and maybe this is just it's peculiar to me, but I can't listen to music that has lyrics if I'm writing. It's distracting.
Same. Lyrics in languages I know. Sure. I love finding like really aggressive hip hop in languages I don't understand.
Yeah.

[laughter]
I like how you emphasized the the aggressive hip hop.
Yeah. Yeah. Yeah. It should be like it should sound like like like like they're threatening you. You're like, I don't understand a word they're saying, but this is awesome.
But but when when I'm coding I can't. And it's because like code coding is just rules based, right? It's rules based and things work or they don't. But what it isn't is our architecture design. And now coding is essentially just architecture design, system design. And that I feel the same way about. I can't have lyrics in the music if I'm actually thinking through the prompt that I'm going to send to Codex or whatever you want to call it code whatever because it's distracting and there's got to be just something less than there in like cognitive load and whereas I could just sit and code turn out code all day because it's just like kind of procedural and and I understand the rules. So if the computer scientist of the future or you know the the systems architect or or the game designer of the future really is just mediating systems design only

then what should they be educated in? I think and I think you're just talking about like critical thinking and I think you're probably just talking about the humanities like you are thing you know like philosophy like a literature like how do you how can you read something and truly parse meaning from it like that that to me feels like the skill set that you need even if you are working at a tech company building products.
Yeah my first degree is in my bachelor's is technically majoring in philosophy. I did end up taking a ton of other things but you know
[clears throat]
I wrote my bachelor's thesis in philosophy and I was so into And yeah this is like still the core of what I am and this is like has shaped my thinking so much so if my son or daughter when the time comes said they want to go study philosophy I'm like yeah great daddy. I mean do take a lot of classes you know philosophy needs like side dishes to work but it's like it's definitely a luxury. The thing about coding I mean the cognitive load I think cognitive load is not the one one dimensional thing. It's like I think

there's so many different ways you can do I mean when you draw I mean drawing or painting is is is is like coding to some extent right? And I I come from a family of like visual artists so both my parents and a bunch of other people in in my in my family were painters so something something around that. And it's so fascinating to see that you know this is like kind of they often need music to do this. It's like you know it gets you in that right kind of you kind of need to occupy some part of your brain while you're doing the the rest of the other brain. Now, I think the same for coding. I also think coding is like the reason why LLMs are so good at coding and reinforcement learning with verifiable rewards have worked so well before that is that coding is a game and that's also why coding is so fun. I mean, you are playing a game as you're writing this code and it's like a game in a kind of like classic single player games with

with with clear rewards. Like you're playing it's like playing Super Mario Bros or playing Minesweeper or something like this. And the corollary of this the one corollary of this is that it's not clear how well coding is game really transfers to other things that are not like coding in the same way as like you know, there's no observable transfer basically in between Super Mario Bros and you know, chess or StarCraft or something like that. Which I think gives an a a perspective on what foundation models know and don't know.
Plenty of mobile growth strategies [music] are still built on broken journeys and incomplete data and it's costing brands big time. Branch fixes [music] that. It's your AI-powered MMP that unifies your links, measurement, and user experience [music] across every channel to successfully acquire, retain, and engage your app users without wasting spend. Learn more at branch.io. [music] And while you're there, check out

Branch's AI search and discovery report covering insights from more than 300 marketing, growth, and digital leaders on how they're responding to the [music] rise of AI search.
Talking about the promise of world models. So, fundamentally, what do people hope world models accomplish that language models can? Like when Yann LeCun says look, we're going to hit a wall with LLMs, world models are like the next big leg of progress. Like what what is he talking about? it. and how and and then relate this to to gaming and why is gaming important in that discussion?
World models is a very broad term. It's basically in a very technical sense, it's in reinforcement learning is a transition model, basically. If you are in this state of the world and this action is taken, and then what's the next state of the world going to be? But then, you know, what exactly that means and how you represent that it varies a lot. What people tend to think of is something like Genie and so forth. And so forth. Genie. Essentially a steerable video model. You know, you can code up something. You

you can prompt up like, "Hey, I want to play a racing game set in Greenwich Village, New York." And you get something that looks You know, it looks nice and it plays a little bit like a racing game. It kind of sucks, but um and it stops to decoheres after a bit and you know, even within the one the one minute you're allowed to play like, you know, you see that the physics starts behaving and then and like the the sort of the the other cars start looking wonky and wonky and stuff like this. Anyway, what's fascinating is that it works at all. It's really interesting and I mean, I I know both you and I and many others have racked our brains among like, "What does this mean?" And it's not really clear because like when Google sort of went with the on the big sort of publicity about Genie, you know, stocks and game companies tanked. And this was kind of weird. And it's just like some kind of strange overreaction because that's not what it is. It's not a replacement for game creation. It's not a replacement for game engines. It's like it it has like lots of disadvantages and few of the

advantages of game engines. It's like really hard to edit things, you know, you have to edit it with a prompt, which is like really inexact. You don't have the exact editing you have in a game engine. It's slow. It's hard to predict, you know, it's like the it's not deterministic fundamentally and so on. So, basically it has almost none nothing of what came the silence value of our game engines. But clearly it is very good at something. You know, what is the something? What How can we fit this into the sort of content workflows we can think of or how can we think of new content workflows and new types of experiences? It may be that this is I mean, hopefully I had this Sora video creator for a bit to try to make a social network. It was fun for a bit and then it was too limited. You could very well imagine that something something like Sora about interactive experiences, game creation, social networks where you basically create them in games for each other. It's unclear if we have enough and many

people have tried that sort of thing with kind of not well models but more like you know, some kind of more traditional canvas game engine like thing. I think it's fundamentally good idea. Someone is going to succeed with this. Maybe based on their own game and game engine maybe about based on something else but it's going to be someone is going to make something big with this. The other thing is that it's a cool prototype thing. Like you know, you can just prompt your way into a weird world you can interact with. It's broken as hell but then you can basically get ideas that you can remodel in a game engine or basically have a model look at like play through what you created and create a game engine version of it that you can then actually build a game and out of. And probably other things. On a very technical level, you could imagine using this as forward models if you train if you tune the neural game engine on a particular game and then create a very low dimensional version of it that runs

really fast because an issue with current game engines is that they don't simulate fast enough so you can't really run planning in them. Maybe you can use neural game engines for that. It's one possibility. But I don't know. It's very fascinating. I don't know where it's going.
Yeah, I think my sense is like the prototyping piece is probably like in and look I mean I like with a lot of this stuff you don't want to get anchored to the current capabilities either of the model or just the carrying capacity of the hardware because I mean this stuff is changes you know too rapidly but my sense is like the prototyping piece makes a lot of sense right like I mean if I want to just render a full environment and even if it's a minute of gameplay but just have a like and then you okay well does this work and then if I want to like scrap all that and then try something that's like fundamentally different like like you imagine this like I always kind of return this idea of like in mobile you know you want to test theme tone and like environment right and setting right and so like well all three of those things in combination is a pretty heavy lift if I needed it to like radically alter all three of those

things like imagine like photorealistic noir you know crime game or like cartoony Viking battle game like all three of those if it's a totally new game right all those assets changing everything I mean that's you know if if I needed to do all that like now if I could just render those with prompts and see okay which of these do I feel like you know best captures what I'm trying to do that then okay at least then I have like some I could I could just in a day I could test 50 variations of those and just see which one of these best fits my vision and maybe even test them as as ads right and then see who clicks on what that feels like a use case
Ads by the way is the kind of content that is time limited enough that they kind of limited coherence and limited kind of like interactivity of your game engines would actually be really good for us if you just managed to basically the problem here is the size of the training network
Right
But that's that's going to be and someone will figure that one out soon

It could be cool like you could you you could get a lot of like really personalized so hyper personalized ads that are generated and generated game world for you that you interact with and that tries to sell you something.
Yeah, no, I mean I'm sure we're not too far away from that.
[laughter]
Julian, this was fantastic. How can people consume more of your wisdom? How how can they stay in touch with you?
They just type my last name, T O G E L I U S into their favorite social network or their favorite search engine or their favorite animal and it's very useful. It's one of the great things I was the best thing I was born with is like an almost unique last name. There's only 10 of us in the world. So, Togelius, that's me on Twitter/X and everything else.
Oh, wow. And actually, one thing I want to highlight and I'll link it in the show notes is like you wrote a really great post, I don't know, a couple months ago about how, you know, you had a a productive, successful career in AI

research without being an expert in math, right? Without having, you know, done
No, no, no, no, no. No, without knowing any math. That's this what it was.
Okay, without knowing any, you know, I didn't want to feel like I was exaggerating, but I I actually really I really appreciated that post because I think it a lot of this can be very inaccessible. And I think a lot of people, wrongly so, get turned off from AI completely because they'll take a look at a paper and they're like, I have no idea what that is saying. Like it's just it's like alien language. And and I think the reality is like I I understand why, you know, the default is to just go to the to the math, but like the reality is there's not I don't think and and you know, peo- people would probably argue with this, but I don't think certainly to read these papers, there's like some math you need to know and you you could teach that yourself to yourself. Like now now maybe to do the most cutting-edge research in like quantization, maybe it's different, but like I think to to read a lot of this work and to understand it, there's a it's a lot of reusable. It's a lot of the math is like it's it's once you get familiar with it, you it's reused all

the time. You recognize, oh, okay, that's a soft max. I get what that is.
Yeah, yeah. A lot of the math in the research papers are completely gratuitous. It doesn't need to be there. You could just cut it out and and focus on the ideas. The real problem is that a lot of research papers are badly written because a lot of people don't know how to write.
Right. Right.
And that's we should just do you know, we should demand that they get good at this, but Yeah, but it's perfectly fine to just ask, you know, them for two sort of explain it like like you're like 5-year-old or golden retriever or something.
[laughter]
Johan, thank you so much. I hope you enjoy your weekend. Thank you for slotting me in on a Friday afternoon.
Hey, thank you so much, Eric, and it was great. Enjoy your weekend, too.