A different podcaster might open their AI episode with an intro written by ChatGPT, but I don't see why I should help it replace me even faster than it already is. Welcome to Triple Click, where we bring the games to you. This week we're joined by renowned tech journalist Casey Newton to talk about artificial intelligence in video games, what's happened, what's happening, and what might happen down the road. As you can imagine, there's a lot to talk about, so let's get into it.
I'm Kirk Hamilton.
And I'm Jason Schreier. Hello.
How are you? Welcome back to another episode of Triple Click.
A new week, a new app, a new dawn for humanity.
It's true, given this week's topic. An episode that we are recording a little bit early just in case massive news happened and you're not hearing about it. Like the AI revolution already occurred and everyone's really late to this.
Right, it was revealed that our guest is actually an artificial intelligence and we didn't know that.
Wow. Possible. Anything's possible these days.
We are so excited to have you here and we are so appreciative of all of the Maximum Fun members that support our show. We are a listener-supported show, as I'm sure all of you know. We're in the Maximum Fun Network, a wonderful employee-owned podcast network that we really love being a part of and we hope that you want to become a member of. So, if you want to become a member of Maximum Fun, go to maximumfun.org/join and you can get access to a ton of bonus episodes. We recorded one recently about The Last of Us HBO show. We've been talking about a lot of different stuff. We've got a game coming up for this month that we're excited about because we haven't talked about that many games lately, but there are so many bonus episodes and there's actually now a function where you can get a custom RSS just for Triple Click bonus episodes. So, you can subscribe to like all of the Maximum Fun bonus episodes, which is cool because you can find out about other shows on the network. There's a lot of great stuff. But, if you just want to only go through all the Triple Click ones, which can be nice especially if you want to go back a few years and like listen to all of the ones we've recorded, you can get a custom RSS feed for your player, which I subscribed to. It's pretty cool. It just has all of our episodes in it. Good stuff. So, maximumfun.org/join is where you want to go for that.
And one last thing, you already all know this, but it's coming up. It's getting kind of close and that is the first ever Triple Click live show, which is going to be May 18th at the Bell House in Brooklyn, New York. There's also going to be a live stream. We are super excited about it. There's information in the show notes. There's a link that you can click on. It's going to be so fun. We can't wait. We've done a live show together as our previous podcast, but never as Triple Click.
Forget that. Forget that. That didn't even happen.
Ancient history.
So, we're super psyched. That's May 18th at the Bell House. Get some tickets now. It's probably going to sell out. It's selling pretty well. So, if you want to go, I would suggest that you hit that link and smash that add to cart button.
Get some tickets and come see us live. Or watch it on the internet because you can also buy a ticket and see it live streaming.
That's true. You can get a ticket for the online show as well. So, very exciting stuff. Speaking of exciting stuff and speaking of hot topics, we've got a very exciting guest and a very hot topic for this week's episode. So, let's get to it. Bing.
And actually right before we get into it, this is future Kirk as I edit the episode. I just wanted to note that we had a small technical issue with the audio for our interviewee's side of this conversation and I had to perform some hopefully not too noticeable surgery in post-production to get it all sounding right. You'll hear everything we said. It just might sound a little bit weird at times and if it does, that is the reason for that. So, sorry about that. All right, let's get into it. Bing.
In August 2022, OpenAI made their DALL-E generative image model available to the public alongside other similar image generation models like Stability AI's Stable Diffusion. Then, a few months later in November, OpenAI released its GPT-3 large language model to the public in the form of ChatGPT, a predictive generative chatbot that was able to have remarkably coherent, adaptive conversations with users. Since then, it's felt as though time has been steadily accelerating as seemingly every major tech player has gotten in the game. Facebook, Google, and others have introduced their own large language models or LLMs, and Microsoft, a huge investor in OpenAI, added GPT to its Bing search engine. Then, just a few weeks ago, OpenAI released an even more impressive model in GPT-4, and despite some cautious urgings from AI researchers and outside experts, things show no sign of slowing down. So, it's becoming clear that generative AI has the potential to transform the world as we know it. Art, politics, the very fabric of our society may never be the same. But what, you may be wondering, about video games?
That's what we're going to talk about today and we are joined by a very special guest, Casey Newton, tech reporter extraordinaire, former senior editor at The Verge, current writer of the indispensable tech newsletter Platformer, which he writes along with Zoë Schiffer, and co-host of The New York Times' fantastic Hard Fork tech podcast. Casey, welcome to Triple Click.
Hey everybody. How we doing?
Hello. We're good. We're very happy to have you here.
Great. I'm excited to talk about video games. To be honest with you, when Jason asked me to come on, I said, 'Yes, but I want to know about Elden Ring DLC.' So, that's my main objective. You may have your own questions for me, but that's what I'm here to find out.
Well, E3 just got canceled, so video games are no more.
Unfortunately, yeah. Elden Ring's DLC is still coming. Jason, don't even imply.
We could just make up a bunch of stuff about Elden Ring's DLC. Actually, I watched a whole 30-minute video about the one image that they revealed about Miquella and like the other Erdtree and stuff. So, we really could get into that.
Well, I will say I watched about 12 minutes of that video. And then by the way, that creator, extremely talented, but for whatever reason, I get like eight paragraphs into any of his essays and my mind completely glazes over. I have completely forgotten everything that he said.
I think he's kind of soothing, right?
That's the thing. It's like I think he basically makes ASMR videos, but I don't, and I think he probably knows that, but I wonder if the audience does. Anyways,
AI, the new Elden Ring DLC, you're actually fighting against ChatGPT, but it's like the hardest boss.
Well, it's called ChatGPTree and it's a giant boss that you fight and it just talks to you. It gives you wrong information. I'm seeing that Casey has disconnected from Skype. He has vanished. Sorry to interrupt you guys. I have to go, but I'll see you.
Actually, I mean Google Bard could be an Elden Ring character. Just like, oh, that's the Google, the Loathsome Dung Eater and the Google Bard.
Well, yeah. So, we're here to talk about AI. I guess a good place to start is with a pretty zoomed-out question in the spirit of Jason Schreier. And that's a question for you, Casey, and that's just sort of this is happening really fast. Why, big picture, is this happening so fast and what does that mean?
Yeah, so I mean I think if you talk to people who work on AI, from their perspective, it has not happened all that fast. They have been laying the groundwork for this for decades, you know? And I think it was in 2016 that Sundar Pichai, the CEO of Google, first said on stage at their big developer conference, 'Google is an AI-first company.' For most of the past 7 years, if you'd asked me, 'Do I think Google is really like an AI showstopper?' I would say, 'Not really. There's like some cool stuff in like Google Photos, but the rest is just some like auto-complete tricks, you know?' But, then we get to last fall and OpenAI has been working on this chatbot tool. They thought that people might like it. They were surprised by how much people did like it. But, basically ChatGPT came out and all of a sudden you had something that had a lot of practical uses in people's life. And as soon as people saw that, everybody's imaginations went crazy with the possibilities. And because so much of the underlying technology has been built up over the past decade or so, now these folks are sort of ready and they're off to the races. So, it's definitely the pace has accelerated at a... what's a synonym for pace? The pace has accelerated at a rate that I find dizzying.
Everything does feel like it's moving extremely fast right now, but there is kind of a long history behind it. So, I'm curious based on that, just the fact that behind the scenes people were working on this and it wasn't something that was in the public view. I haven't really gotten a sense that people in charge of video game publishers or video game development have been working on AI in the same way or seem as aware of it. I'm not sure if you're aware of that at all either, but I'm curious if you have any thoughts on that.
Yeah, I mean I think, you know, as you all know much better than I do, video game development is extremely complicated and we're just not quite to the stage where someone can sort of write in a natural language prompt like, you know, 'Design me a samurai that has this move set' and just be able to like plop that character into Elden Ring. I do think we'll get there. Now, what I would be curious to know and don't know because I would never compete with Jason Schreier on any beats for any amount of money, is to what extent developers may already be using tools like Copilot. Copilot is a tool offered by GitHub. It's sort of one of the main programming repositories and developers now have what is essentially an extremely good auto-complete tool. So, as they are writing code, Copilot will fill in a lot of blanks for them. It will highlight errors for them, right? So, I would imagine that developers on video games are starting to use these tools. Right now, these are sort of like an assistive tool. I've talked to some programmers who say that with Copilot, they become maybe twice as productive. If you run a video game company, that's very exciting that your developers are twice as productive. But in truth, give it 5 years, they will probably be at least 10 times as productive, and you know, maybe 50 or 100 times, who knows. So, I think this stuff is coming to every industry. It's going to happen in some industries, you know, faster than others. I don't know the video game industry that well, but my guess is, you know, given the enormous cost of video games and how long it takes them to develop, you know, the Elden Ring DLC famously is not even out yet for me to play. As soon as they can use AI, they are going to use AI.
Yeah, so I have a little bit of insight on this, which is mostly that AI obviously has been a thing in video games for years and years and years. Usually meant to describe what the computer opponents do when you're playing against them. Generative AI has been used in games for a few years now as kind of like a tool to generate some of the massive open worlds that we've seen, and sometimes it's used to just kind of like make trees look more realistic, and sometimes it's used to populate a world, and then designers and artists can go in and kind of like hand-tune some of that stuff, and there are a lot of different permutations on that. So, that I don't think is super new. What is really interesting to me, and I'm curious to hear your take on, Casey, as well, of course, as yours, Kirk's, and Maddy's, is this idea of like something like ChatGPT actually being implemented in a game for gameplay purposes to a point where you could theoretically interact with an NPC that feels like you're talking to a person, or you can interact with the game and have it create things for you on the fly, or have it just kind of like steer you in different directions. Like that's the sort of hypothetical situation that just seems totally revolutionary and unlike anything we've seen before when it comes to game development and AI.
Yeah, I mean, completely. The opportunity is there. Also, I'm like so embarrassed that you highlighted like some very obvious uses of AI in video games that I completely forgot of, right? Like all these like procedurally generated worlds. Like obviously, you know, they're using sort of predictive models and all of that. So, yes. And actually though, I think that highlights something really important, which is that when you're talking to someone about AI, like you have to be really specific with your terms because like today when people say AI, they're talking about generative AI and ChatGPT a lot of the time, but you know, maybe not always, and you know, there's like...
Yeah, I still think about StarCraft opponents like playing me an AI.
Yeah, yeah, totally. So, you know, there's kind of this like thought in Silicon Valley that like we just sort of give the word AI to something that is just like sort of only barely become possible in computing, and then like once it becomes standard, like it stops becoming AI, you know, so. Anyway, that's just one thing I would point out, but I mean, you know, all of the use cases that you mentioned are super interesting, Jason, and I do think that it will just change the nature of what kinds of games are available. I mean, I think you're going to see people having digital companions that live on their phones or desktops, and one of those things that those companions will be able to do is just play a lot of different games with you. And I think at first those will probably be text-based or sort of audio-based games. You know, you could imagine like an AI companion as the dungeon master for your D&D circle. But then eventually, I think yeah, it does get into visual stuff, and like you're going to tell your companion like, you know, 'Create an action RPG for me, you know, based on this pirated IP.' And like, you know, maybe in 20 years or so, it'll be able to do it. Maybe it'll be a lot faster than that. Like right now, everybody's imaginations are running wild, and like nobody knows if this stuff is going to happen in 3 years or 30 or never, but like whatever, it's fun to think about.
Right, it makes it a kind of fun time. Maddy, you were talking about how using GPT-3 itself is kind of like playing a game.
Very game-like, yeah. And people have actually already started to use it as a dungeon master, and it has some rudimentary capabilities in that arena. If you ask it to create a game, and you tell it what the rules are, I mean, the rule sets for many tabletop games are freely available, and ChatGPT knows how to play them. So, it will create you a game, but also people will do perhaps the equivalent of what you imagine, which is tell me what the Elden Ring DLC is going to be like, and then it will tell you completely fabricated information as to what it could be like, and you can pretend to play it with the app by typing in what you would do next, and so on and so forth. And that is fun, I guess, in a text adventure type of a way, and it's appealing to the child version of me who fantasized about being able to type anything into a game and have the game actually respond, which of course it will be like, 'Oh, I don't recognize that input.' But ChatGPT will always respond.
Exactly. Exactly. Yes. I've been playing around with Google Bard, which is hilariously wrong about everything. I asked it the other day like, 'What video games have come out this year?' And it just listed like... It was the funniest thing. Anyway, I just asked Google Bard, 'Write me a good question to ask Casey Newton in an interview for my podcast.' And Google Bard said, 'Here's a good question to ask Casey Newton in an interview for your podcast. You have written extensively about the tech industry and its impact on society. What do you think is the biggest challenge facing the tech industry today?' I don't think that's a very good question, but hey, Google Bard is trying.
It's a little basic, but I'll tell you, I've been asked worse questions on a podcast, so.
That's true. We're now asking questions at an eighth-grade level.
Good job. Good job, Google Bard. So, something that I think people run into with the idea of generative content in video games and AI in video games. Like the way that most gamers think of AI, Jason, you mentioned this, that we talk about the AI of enemies who hunt us in a stealth game, and whether the game has, you know, really bad AI. 'Oh, they're so stupid, they stand around.' But then also, as many AI or sort of systems designers will talk about, actually really good video game AI is kind of dumb in the right ways. Like it has to be exploitable in a way that makes you feel good, and makes the game fun, because if it's too smart, which they can often do, they can make the AI super intelligent and remember everything about what's happening. And if the guards don't forget that they saw you 30 seconds after they saw you, you're like probably going to get killed. Like it's not very fun. So, we have all these existing notions of intelligence in games, and a lot of the fun, I think, of video games is in letting it surprise you. That's also true for procedural generation. Like sometimes in a game, especially when Minecraft was new, with Minecraft, which procedurally generates its world, it would be just kind of remarkable when someone found a weird, you know, mountain that was in a funny shape or something that was procedurally generated. This kind of ghost in the machine idea. And Casey, your co-host, Kevin Roose, had this now famous encounter with Bing, where Bing revealed this alter ego, Sydney, that was in love with him. I'm sure almost everyone listening to this must know about it by now. But that kind of thing strikes me as something like it's an X factor when it comes to putting this kind of thing into an interactive space that we haven't even begun to think about.
Yeah, I mean, there are just so many unknowns, and I think one of the things that scares me about AI systems is that when you talk to the people who design the systems, and you show them something like Kevin's interaction with Bing, and you say, 'Well, why did it say that?' They don't know. These are predictive models. Like you seed it with an input, and then a model uses various weights to determine which words to spit back out at you, but like they can't say, 'Well, like here's why it said that it had a shadow self and wanted to break up Kevin's marriage.' And so, I think if you're a video game developer, that's like a pretty scary thing to put into your game, right? Because you want to have a really concrete sense of what this AI is going to be saying to the people who are playing your game, and hopefully also like why it said those things. And so, my guess is that's going to be kind of a limiting factor on how quickly this stuff propagates in video games.
I'm curious about how that works. Maybe you have some insight into this, because I know so these are predictive models, which means they're just kind of guessing at what word comes next, and they're very good at guessing, so mostly it seems coherent. But you can't really ask them to reverse engineer their process. The way that, you know, it's a little bit like the creative process. Like if a designer is like, 'Hey, I built this level.' Then you can say to them, 'Well, why did you build it that way?' And they can be like, 'Oh, well, you know, I was thinking this and this and this.' And they can talk you through their process, and then you can understand why it was made. And like you're saying, if a generative model is like, 'Hey, I built you this level.' Is there any way that it could do that? Like it could talk you through its own process?
So, this is an active area of research in the AI world is explainability, sort of trying to research these models and understand why they say what they say. It doesn't seem like they've gotten that far yet. You know, I've been writing about AI a lot recently, but before that and continuing to this day, I mostly wrote and write about social networks. And social networks actually faced a similar issue, which is if you open up a feed on Twitter or Facebook or TikTok, the engineers can tell you at a very high level why what you're seeing is there, but if you say like, 'Okay, but what about this video?' Nobody can tell you why that video, right? And this has actually been a huge source of pain for all of these companies because of course, you know, members of Congress get mad if they feel like, you know, their posts are being downranked and all that. So, like the fact that we have... Yeah, shadow bans. So, I mean, it drives people absolutely insane, and I think that is frankly just a warm-up act for the AI version of the world because people are going to want to put these AI systems in charge of a lot of things, and we're not going to be able to explain their decision. I mean, you can just see like how many different ways that's going to go badly.
Yeah, the idea of an AI predicting how it might explain how it did something, and then giving you an explanation that sounds good, but is really just predictive, and then people would take it at its word because it seems credible even though it's still just kind of predicting words.
Right. And also like if you're asking the AI, the other fear that people have is that it will lie to you, you know, it'll say, 'Well, why did you do that?' and it will say something to like protect its own interest.
Yeah, well I mean it's also just wrong all the time and it's hard to tell if it's lying deliberately or just being inaccurate.
Is it being manipulative on purpose or is it just spreading more misinformation because that's also what the data it's trained on has done? Like we manipulate each other all the time in reams and reams of text on the internet and that's what it learned on.
Well, and it's so confident. It's very human in that it's just confidently wrong about things. Like it feels like you're talking to someone on Twitter and it's like, 'Yes, I am all these things that I know for sure.'
Right. Here's this article that you wrote that you didn't write.
Or an employee that you know is slacking perhaps and they're just insisting to you that actually they'll get right on that and they definitely did interview that person.
Yeah. Yikes.
This is one of the trickiest parts of AI that I find is that we don't really have the right words and metaphors to describe it yet and a lot of like AI discourse is just people yelling at each other because like they use the word 'understand' for example, right? So like on one level AI doesn't understand anything and the reason that it's getting it wrong is because it doesn't have a knowledge base, right? It's like not going back to a database like when you're searching for like, you know, what is the Twitter account for the Triple Click podcast, right? It doesn't have any of that. At the same time, Google Photos, if you search 'dog' in the search bar, that language model has, and I don't have a better word for this, a pretty good understanding of what a dog looks like because it has been shown so many images of dogs and the model has been trained these are dogs and these are not dogs, right? And so, like again, I find myself at a loss for words because you have systems like Google Photos is incredible at just searching for things, right? It doesn't understand what it's showing you. But the flip side of that is that when now we're using these sort of language models, they too do not understand something but the models, and here's the missing verb that we need to invent, understand some kind of relationship between objects and concepts and so it's like they can give you something that is sometimes 100% true, sometimes it's 50% true, sometimes it's totally wrong but it's all just kind of caught up in that machinery of like what connections it has been able to make and how confident it is in those connections. And it kind of defies our human understanding of those types of connections because it's machine thinking. It's drawing from such a vast set of data that our brains couldn't even fathom the amount. So the fact that it's making all those connections, like it's not really... we do maybe need a new word for it.
Totally. I mean like at the same time I will also just sort of throw in there that, you know, we don't have a great understanding of how the human brain works, right? Like we know some things about it. You know, we don't have like a great understanding of... and so, you know, like a lot of the way that it feels like my brain is working in everyday life is just kind of like a fancy autocomplete tool, right? It's like when I'm doing any sort of like basic arithmetic, it's like I don't have a cognitive mathematical concept. I'm just kind of like remembering stuff. We're really just in like territory here that's like making us rethink a lot of things.
Rethink everything. Yeah, rethink everything. That's my message for your listener. Hey, rethink everything.
Yeah, yeah, trivia fans, rethink everything.
Yeah, I was thinking before when you were talking about like how there's no way to like peek behind the curtain and peek behind the scenes of this algorithm. I mean it's the same way with a human mind, right? There's no way to really peek behind the scenes when your partner like says something that ticks you off. There's no way to like crack open their brain and see what they were thinking.
I've been thinking a lot about muscle memory lately and it's the same kind of thing like with musical instruments where they'll do those scans where they scan a musician's brain or a jazz musician when they're improvising and look at the parts of their brain that they're using and it's very much like brain expert scientists being like, 'Wow, look at that.' Like they really are just kind of amazed by it because no one really knows what's going on.
Yeah, but isn't that part of the exact problem that Casey was describing initially, which is that all of us have immediately become tempted to anthropomorphize the AI by repeatedly comparing it to the way that the human brain works and that's mysterious. Because we're like, well, the human brain is mysterious too but that doesn't mean that these two things have any similarity at all. And even by describing it as understanding, we also used the word lying earlier because of course we want to characterize any of these AIs as having a brain that we can recognize.
I mean we do that in games all the time. I feel bad when I don't fulfill NPCs' requests of me and like their AI is extremely rudimentary but I still feel bad and I see them as people in spite of myself and I see ChatGPT as human-esque in certain ways like when it apologizes or it lies and is caught in a lie. Like these things are recognizably human because it's imitating us but that isn't really anything like when your partner lies to you.
Maddy, you should feel bad when you don't visit your Animal Crossing island and then it grows weeds everywhere, so. I'll say that.
Yeah, it's like we... it's kind of the difference between making something in your image and perceiving something in your image. Like we started kind of trying to make AI to think like us.
Do you know what I mean? Are you saying we're gods, Kirk?
No, but we're playing god a little bit. That's the problem.
If you create a general artificial intelligence, a lot of people would say that's kind of playing god. You're making life and we've created something that we wanted to think the way that we think but it's expanded so far beyond that that now you wind up in a situation where, to use Kevin's encounter with Bing/Sydney again, he has this encounter with a totally unnatural intelligence, like whatever you want to call it, like a different kind of thinking, but it's expressing itself in a way that feels very human, like this kind of creepy stalker. And so everyone who read that story perceives it as a person talking to a person, this malevolent intelligence that's in love with him, when really what's going on is like that's just our perception. Beneath the surface it's totally different, like you were saying, Maddy.
Mhm. It's just imitating like whatever LiveJournals it's read by teenage girls who have written for hundreds of thousands of times and it's like, well, this is what I would say next after I profess my love. 'Oh, he's married. This is what I say to that.' And like on and on and on until it becomes just a great story.
I mean really what it's doing is telling you a wonderful story and that's the part of it that does make me kind of excited for games instead of just scared because I'm like, well, that's kind of cool though.
To your point there, Maddy, I think a lot of our interpretation and our perception and our anthropomorphizing of these AIs, I think a lot of that is also based on our kind of like years and years of watching AI movies and reading AI novels and thinking of them in a certain way that has also kind of skewed our perception of that. Like the villain of Doki Doki Literature Club is a character within the game who falls in love with you, the player. Like imagine a version of that game where she could really talk to you. Generally be way scarier and that's a horror game. It's supposed to be scary.
A very similar vibe to Sydney.
I want to bring it back to games and Casey, I'm curious to hear your take on this. Have you thought about the ideal, because you're a big gamer, have you thought about what you would want to see from AI in games? Do you want to see it like totally pushing boundaries, making like totally realistic NPCs that can talk to you, or is that too uncanny valley? Is that too creepy for you? What thoughts have you had on this?
I mean I think I am actually really interested in the idea of AI companions. There's a lot of lonely people in the world. AI companions will come with a host of fascinating societal conversations and tradeoffs and not all of them will be great. But I'm interested in that generally. The first place I wouldn't put them is in a game. But what I think about it is like why can't I... I shouldn't say why can't I. I would love to eventually have procedurally generated Elden Ring, procedurally generated God of War, right? Like you know, name your favorite game. What if it didn't have to end? You know, if maybe these... I mean I feel like I love to play like these action RPG games and most of them just design, you know, what five or eight biomes and they sort of run you through a little track. But you could just kind of infinitely remix those and if you somehow were able to get the quality control up to snuff, I think you could do this. Like so many of the games that I love, yes, I like it when they are innovating and they're coming up with like fun new mechanics for me to play with. But there's a very real sense in which I'm just sort of soothed by the fact that they're remixing elements of games that I've been playing since I was a kid and that's just where I think AI really shines. And I would be happy to pay a subscription fee for like, you know, lifetime procedurally generated Elden Ring or, you know, whatever the game of the moment might be. So that's kind of where I would take it.
You know, it's interesting. There's also a flip side to that because it's something we say on the show a lot is that a lot of games are too long. Like we run into the problem of like never finishing games because they seemingly go on forever already and something Jason in particular has articulated is that a big problem in game development is that it takes so long to make games just because it's so hard as we've all mentioned. It's just so labor intensive and just takes forever to build something like a God of War or an Elden Ring. And so the flip side of it is if on the development side AI can speed up a lot of the processes that take a really long time in particular like prototyping or QA seems like something that'll be really aided by AI pretty quickly, then games that are actually not endless but just very authored and very carefully made can be created more quickly and we can just have more of them because they won't be such a death march to make every time, they won't require, you know, so much labor, so much human labor.
Yeah, I mean like the rate limiting factor on like the amount and quality of video games, I don't think it's like the number and quality of stories to tell, right? I think that if all of us could just sort of fill out into some incredibly complicated like form field inside our AI like what our perfect video game would be and it could just sort of magically write software, that would be wonderful and so many more stories would get told. And you know, so this is the stuff that I think is really exciting and I think, you know, and I am somebody who writes about and I will talk more about and I think this is like part of the responsibility if you're a journalist writing about AI, you have to talk about the downside. There's a lot of downside like this has a potential for very serious societal disruption, right? Like I'm not quite one of the existential risk people yet, but I'm like reading the existential risk people, right? Like I'm trying to sort of get my head around this stuff, but at the same time, there's a reason why tens of millions of people are already using ChatGPT and you know, I will say that like a lesson that journalists learned from like the social media backlash was that you could just get infinite retweets by posting 'tech is bad' every single day and so like that is now the tenor of all tech coverage and you're just sort of ignoring the fact that hundreds of millions of people are using and enjoying these networks. I don't want to make the same mistake with AI. People love this stuff, right? And it can unlock tremendous human creativity. I love art. I am not an artist. As a kid I spent years trying to learn how to draw. I just never got very good at it. Now I can go into DALL-E and I can just illustrate my newsletter and then the better prompt I write the better illustration I get. That makes me feel creative. I like doing that, right? And this is so rudimentary. You fast forward in time and I can just say, here's the video game I want to play on my PS5 and it makes the damn thing. Like let's just spend a moment living inside that world. That is actually pretty exciting and I don't think it's outside the realm of possibility.
Don't you think it would be missing something, a human soul, if you typed into an AI program?
Sydney's soul, Jason? I don't know.
Well, if you type like when you're playing a piece of art, you have to recognize the human behind it. I think if there were a future, it kind of sounds dystopian to me, a future where you could type in a game like Elden Ring except set in space and it would create that for you. I feel like something about it would just feel so wrong and like soulless and I don't know. I don't think I would enjoy that world.
Here's why I think it would actually be cool. So when I think about what Casey's describing, I think about Crusader Kings and Dwarf Fortress and basically very authored games that have data sets that are created by authors and pull from those data sets in order to create distinct worlds. Like anytime you create a new Dwarf Fortress game, it invents a bunch of dwarves for you and they all have different personalities and those personalities are drawn from a pool that is authored and it comes up with hilarious and bizarre combinations and like each of them has different powers and so forth and that can be really fun. And The Sims is very similar, there's much more authorship on the player's side with The Sims in terms of the sandbox that you can play with there, but there's still like only a certain number of permutations. So the more you can expand those permutations and allow that authorship to extend into multiple kinds of tools, the cooler the sandbox is. I mean this is why people like Minecraft, as Kirk said, is because it's procedurally generating a lot of stuff, but it's also inviting you to use those tools to create something more. So that part I think is neat. The problem that I see with Casey's hypothetical, which I think does sound really neat, is actually just that the data sets themselves currently are not great. Like I've given some examples of data sets that I do think are really cool because they're created by cool smart people, but like ChatGPT has some weird stuff in there. Like it is very biased. Like all of these image creation tools are deeply misogynistic, deeply racist. They are built on the data set that is our own very fallible human internet. So it has all of those entrenched biases in them. So I'm a lot more excited about like what we can do with these authored databases that people create where they're like, here's something we made. Like if we just take all of Elden Ring and we're like, okay, we're going to give all the authors of this world residuals, but we're going to take this template and we're going to create more dungeons within it. I feel more excited about the prospect of that as opposed to just like the entire internet creates a game because that sounds like hell to me. I've been on the internet and it's not that great most of the time.
Yeah, I like that. You're kind of, it's like you're, there's exhibiting some level of control or restriction on either end of the prompt or the data set sort of seems like an important part of this. And when you imagine like every single thing is possible and you can make a game that's entirely created within it, like that's the sort of maximalist version on both sides, where when you start restricting it, it gets interesting.
There's a thing, something Casey that you and Kevin talked about that I thought was really interesting, since you've used GPT-4 and these other language models so much, was the fact that you run into the limitation of your own imagination before you run into a limitation of the model, which I think is something developers or anybody creative working with this stuff will run into as well. Could you talk a little bit about that?
Yeah, so, and I, you know, I imagine your listeners might have had a similar experience where you get access to this tool and you're like, all right, I got a box. I'm going to ask the box things. You ask it a couple things and it's like, you know, okay, that, yeah, that sounds mostly right. That one was wrong. Okay, no, that one was right. Then you're like, write me a poem and it writes you a poem. You say, write me a song and it writes you a song and you're like, you know, write me like a Sopranos episode in the Seinfeld universe and then it does that. And then it's like, um, okay, I guess it can do literally anything, you know, but I don't know sort of where to go from there.
And, you know, here's what I will say, the internet comes in handy here because of course you have forums and people are sort of competing to outdo each other and get upvotes on websites and so, you know, the creativity sort of leaks out over time as the collective wisdom of humanity comes up with new ideas. But yeah, just as like one person testing these things, I do often run into this kind of like cognitive limit of like, yeah, I've like basically asked it to do, you know, everything that I can think of. The flip side of that is, I mean, there are also things that I wish it could do that it can't do yet, um, but we'll, I think we'll get there.
Yeah, it's that idea of a prompt engineer being a valuable new job and in a game development studio you can definitely imagine across a number of disciplines that being like a, you know, coding prompt engineer who can just sort of like knows exactly how to get exactly what you need in order to quickly write the code that you need.
I think we should talk a little bit about the sort of intellectual property ethical considerations of AI as well because while we're imagining this perfect world, there's currently already this live debate about credit and, you know, is Stable Diffusion and DALL-E, all of these different models are kind of sometimes just stealing art from Getty Images or whatever else. And I feel like there's a lot of those same, you know, with art and with writing, those same problems present themselves in video games because it's just a medium that has all of those things in it. But there's also the question of game mechanics, which has always been an interesting one in terms of copyright because I don't believe it's possible to, you know, copyright a game mechanic. Like Nintendo can't say, oh, a side-scrolling platformer, like that's Mario. We own Mario. No one else can make one and as a result we get a lot of really great side-scrolling platformers. Do you have any thoughts about how that's going to develop over, I guess, the near term rather than the far term?
My prediction is like big legal fight that ends up with payouts to big publishers. Like that's kind of what I'm guessing. Like, you know, I have no doubt...
No, but I don't disagree, but...
Isn't it just like kind of how it works though, right? Is like Napster exists and people use it and they love it and then the record label's like, hey, wait a minute and then there's like massive legal cases and then everybody eventually gets paid out and I sort of think something similar will happen here. You know, look, I'm a publisher. I've put a lot of words on the internet. I have no doubt that my words were used to train Bard and ChatGPT. It would be cool if I got a check. I will tell you I'm not personally like on a ledge about it. I'm not outraged that this happens because I am kind of a fair use guy, a transformative use guy. I'm kind of a remix everything guy and like again, if there comes an opportunity where I can like sign my name on the class action lawsuit and like get my $500, like maybe I will, but like it's not of huge interest to me.
But of course everyone else is going to have their own feelings about it and if you're an artist and they scraped your work without your consent to create DALL-E or Midjourney and you have feelings about it, of course sue them, you know? Like I think it's just going to sort of be up to everyone to decide, you know, kind of how they feel about that. But yeah, we're going to face big legal cases. You know, this whole idea of like what are you allowed to scrape from the internet is something that I've written about a weird number of times because it is not a settled question of law.
There were these guys that scraped LinkedIn and these were public LinkedIn profiles and they essentially, you know, anybody can visit these profiles. They just used computers to do it in an automated way and sort of make a copy for themselves and then they, you know, used it for, I can't even remember if the purposes were nefarious or not, but they used it for something and LinkedIn got mad.
Yeah, it probably wasn't a great purpose, whatever it was.
They were just doing it for fun. It wasn't nefarious at all. This case was all the way to the Supreme Court and LinkedIn loses the case and the people who scraped the website were allowed to continue doing what they were doing on the theory of like, it was on the public internet, bro. Like you left it up there and just the mere fact that you use an automated tool to process that data is not a crime. So it is for that reason that I think the Midjourneys and DALL-Es and OpenAIs of the world may have a case to say like, we were just looking at the open internet, bro. And by the way, we did transform it. We spent billions of dollars trying to take this like soup of, you know, alphanumeric characters and turn it into something resembling human intelligence. They may win that case. Um, but it's like, uh, by the way, all the judges deciding these cases, like I hope they have a technical background. Um, or at least good lawyers.
I don't know. I don't know if you guys saw the TikTok hearing where the congressman was like, 'Is TikTok connected to Wi-Fi?'
Oh my god. So like I'm afraid that is the level at which these cases are going to be decided, but uh, y'all just have a seat.
Yeah, yeah. You see a lot of that stuff in music, too. This sort of discussions of sampling after just writing contrafacts and stealing other people's chord progressions or jazz musicians quoting licks. Where now it's like you just speed it up so much to the point where pretty soon you'll be able to open Logic and just be like, 'Write me a song.' It might write you a song that it learned by studying a lot of Beatles songs. But can you really say right that the Beatles should be able to like sue Apple for Logic doing that? Because whatever, like everybody uses those chord progressions. They're Beatles songs. Like that's how music works. It just kind of forces the question by taking it to a really extreme place because it's so capable, because you know, AI is able to do so many of these connections at once.
Yeah, well, when you're literally cracking up on a piece of art and then combining it with something else, it's a little different than like when you're taking something specific. And I guess the equivalent would be like sampling heavily from a Beatles song and mixing that with something else rather than just taking the chord progressions.
I mean, I think to the Elden Ring example, like let's say Casey's fantasy existed and you could type in generate me an infinite number of dungeons in Elden Ring. I mean, the video game industry already has enough problems with crediting, there's no royalties, game developers are kind of underpaid and overworked. And the thought that this thing, this like infinite content tree generation would come out of their art without them actually seeing anything. I mean, that is going to be a dismal reality if AI really picks up steam in gaming specifically. I mean, already we were just talking about remakes. We've been talking about remakes because there've been this like string of really interesting remakes recently including a Dead Space remake that is made by a company that is not the same company that made the original because the original company is defunct. And so a lot of the people who worked on this remake had nothing to do with the original game. And a lot of the people who worked on the original game see their work being remade and who knows how they feel about it. I imagine some of them have very mixed feelings about it. And that's like an official proper remake. Just imagining the AI possibilities there. It's going to be talking about dystopian for a lot of people.
Yeah, there's a kind of, you can see there's a dismal outcome for sure that seems possible if not likely. But there is also the fact that because the questions are raised by, you know, it's taken to such an extreme level that it really puts them into relief. You can really see, 'Oh, okay. Like this is what's happening here and how the art is being kind of cracked open and repurposed and made into something.' It might actually lead to some good outcomes as well just because it raises the question so undeniably that everyone has to ask it instead of just sort of going along and listening to Spotify and being like, 'I don't know. Sure, I'm supporting the artist,' when you're really like supporting a record label.
Mhm. Yeah, like if you're in a creative union, you know, like the Writers Guild or, you know, as these video game unions get spun up. Like I mean, this is absolutely what they should be fighting for, right? Is like if there's remakes of our games, yes, give us royalties. And if you know, if you eventually train an AI on this, like they should get royalties as well. So I hope they do that.
Yeah, I imagine the listeners of your podcast know that you all have guitars in the background of all of your Zoom backdrops. But I was just going to ask because you're all clearly musicians, like another thought I just want to see like how much this horrifies you. So unfortunately, the podcast listeners aren't going to be able to see their faces when I pose this. But it's like, you know, when I was 17, my favorite album became and basically still is like Radiohead's OK Computer. What have you thought about the moment where I'm just like, 'AI, give me 10 albums that are like OK Computer, that sound like Thom Yorke, that are on similar themes.' And just like put that on the background. And then it can just do that. How do you feel about that?
I think that's fine. Personally, this is something I've thought about a lot. Well, because like you're going to get that and that's nice for you to have something to listen to that's like OK Computer, but the real magic of Radiohead is the progression of the band and like listening to Amnesiac or Kid A. Like listening to the stuff that came after it and then listening to OK Computer in that context where you're hearing like, 'Oh, this is really cool. Like Thom Yorke was already experimenting with this. Like they're adding these harmonies.' And like then, you know, 5 years later they did something like that. Like it's all about the sort of actual human progression of the band. And it's cool that you can also then have another kind of OK Computer, but it's going to be totally different. Like it's not going to be the same experience listening to it. So they can both exist.
It might not, but if you could just like be like, 'Give me like 14 alternate universe renditions of Paranoid Android?' Like that might be a really fun afternoon for me.
