Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
S7 E9: Up Late - Naturally Unintelligent
58 Plays5 months ago
We take a DEEEEEEP dive into artifical intelligence in the video game industry, from the AI that drives enemy and NPC behaviour through to the mechanics of your avatar moving through a virtual landscape. We also cover the introduction of deep learning and generative AI and how that has and will continue to impact both the way we play games, and the way games are made. Finally, we cover the eponymous ChatGPT and its Large Language Model brethren - the tool that put AI into family dinner table discussions into a couple of years ago, and how it's revolutionalising the way we interact with video games.
Beers reviewed this episode:
- Akasha Brewing - Whisky Barrel Aged Imperial Stout - 10.2%
- Dainton Beer - Apple Brandy Barrel Aged Barley Wine - 14.1%
- Hawkers Beer - Gee These Woods Sure Are Dark - Imperial Stout - 8.0%
Recommended
Transcript
Introduction and Hosts
00:00:18
Speaker
with their listeners. Welcome back to another epic episode of our podcast, where we blend the realms of pop culture and frosty brews. I'm Pete. You guide through the virtual landscapes like a pixelated Indiana Jones. Howdy, folks. Tom here, your cinematic sherpa navigating the peaks and valleys of television and movies, just like Frodo on his quest to Mount Doom. And good day for me too, mates. Down here, your craft beer trailblazer, seeking out the holy grail of beers.
AI in Video Games
00:00:48
Speaker
It journey as legend dares Captain Jack Sparrow's escapades on the Black Pearl. Today, we are diving at first into the evolving world of artificial intelligence in video games, from non-player characters with minds of their own, automated game mechanics, the unpacking health technologies reshaping our gaming experiences,
00:01:06
Speaker
So strap in as we dissect the pixels and algorithms that bring our favorite virtual universes to life. Exploring artificial intelligence's impact on gameplay and storytelling. Plus, we have each got a unique Australian craft beer to sample and rate. I will be cracking open a brew that promises the flavor of a thousand tattle in some sets, as bold as Thor's hammer. I have a beer that is complex as the blood twists in Westworld. Brewed to challenge even the most decent palate. And I will be tasting a beer that is smoother than Neu's moves in The Matrix, guaranteed to make you question reality and savour every sip. So grab your controllers and your favourite pint glass, because this episode of Pixels and Pints is about to level up.
00:01:50
Speaker
So that's our intro. I don't know how we want to introduce the emphasis. Oh, fuck, we're recording. Right. It's a thing we do. Yeah. So our introduction was scripted by chat GPT and then voiced by a completely different artificial intelligence that does text to text to speech. And I thought it was a good, end well, actually it was Tom's and then Dan's idea independently and similarly. um But we do. To do some janky intros using AI. I think we've succeeded. so freeze Free software too. Yeah. Well, I mean, chat GPT is completely free. And then, um yeah, the text to speech, it says it's free. And then you read after using the website, I read its terms of service and the amount of data it has now pulled off my computer and phone and probably every other device on my network. Pretty fucking scary.
00:02:43
Speaker
400 and 450 vendors pulled out or off your phone. Oh, those are cookies. I declined. Anyway, Jesus. Probably not say you used to being pulled off with your phone, but we usually I'm doing the pulling off. All right. Before we jump into our topic of this evening, which obviously is artificial intelligence in video games. Uh, do we want to kick off with some berries? Yep. um Can I address the elephant in the room? Yeah. Okay. ah Peter, you've shaved your beard differently and it's throwing me out. Yeah, it's it threw everyone off. I had a beard shaving disaster and it ended up completely gone. I've been holding my breath for two weeks while I look a little bit like Jabba the Hutt with all those neck rolls.
00:03:29
Speaker
No, it's the cheeks. You got the Tony Stark like cheeks appearing. That's what's throwing me out. Not quite as straightly manicured. It's not. It's not. It's not
Personal Stories and Craft Beer
00:03:38
Speaker
the usual manicured Pete we're used to. Anyway, that's fine. I'll be able to deal with it. a bitt Maybe. There's something completely off topic. Seems tim Tom's thrown a spanner in the works this early in the episodes. You see those like face stencils that people hold do to get their ah beard line straight. It's like a plastic stencil that they put onto their face and then they shave around the stencil. so save me head straight lines every time i forgot the old old shaky hand too many tonight i flight I'm shaving with the Craig David stencil. Yeah, exactly and strip I find that the epitome of wankiness. You're not wrong. Oh, fuck. Anyway, time already what are you drinking, Dan?
00:04:20
Speaker
um I have, I don't even know if this is available anymore. I've had it in my courtroom. I was gifted this by one of the guys from Akasha, probably 18 plus months ago. This is one of their whiskey-aged, ah whiskey barrel-aged imperial stouts from their barrel room, the Akasha barrel room over in Black Art? Black Art in Norton Street. It's a 10.2% Imperial stay out in a 750 ml bottle. I'm pretty sure it's 750. Yeah, 750. So I will be consuming most of this on my own tonight. So we'll see how this ends up. And I'll give my rating of that later on. OK.
00:05:03
Speaker
Cool. Excellent. Uh, these, what have you got? I am drinking Dayton beer's, uh, apple brandy barrel aged barley wine. So I've actually got two tonight. I don't know if I'll survive. We'll, we'll see. Uh, so this is, uh, age six months on apple brandy barrels. Uh, and it is 14.1%. And it's fucking delicious so I will say yeah I think it's probably gonna do to withhold scores until somewhere in the middle unless we. Forget like we have before sometimes um but first off the bat it's nice and caramel and loads of apple which is exactly what I was hoping would be.
00:05:42
Speaker
I've got its sister wine, so there's a, sorry, it's sister beer, there's another Bali wine with the same base that is aged 12 months on whiskey barrels that is also 14%. I may, if I'm feeling particularly adventurous, I may crack that later on tonight too. Excellent. What do you got, Tom? Fuck, I'm the sensible one tonight for some ungodly nine reasons. Oh yeah, not the 400 beers you had last night. That's not the reason. Shut up. Shut up, you fucking dog. I'm drinking the hawkers. You can't see this because the can and the writing is so dark. It's just, it's lovely and purple, but it's a G these woods sure are dark and it's a black forest stout. It comes in at a nice round 8%. Um, so look with you to goobers drinking, you know, giant beers. I think I might just have to treat myself to another one, but it's poured lovely, lovely and dark. Um, yes, I,
00:06:36
Speaker
Look, I will admit, I saw this in the bottle shop today when I went to buy beers and I got it purely because I know Dan would probably want to try this because you love a black forest out and hawkers naturally would make a good one. As soon as you soon as you said that, I'm like, oh, resist the urge to get onto Moondog site to see if K Cole's been made again. There should be alerts. You should set up like an RSS feed or something it pops up. should That is Very good. Very close to the cake hole. It's probably got more cherry though. Okay. In a good way. Yeah. That's good. That's why I'm going to happily enjoy that. Ryan's in your cake hole. It's sweeter. Actually slightly more bitter. Okay. Like the the bitter chocolate and then the sweet cherry on top. The cherry is really at the forefront though. Yeah. It's good. It's good. Very good. Yeah. Nice.
00:07:25
Speaker
What a surprise. Hawkers make another good beer. In other news, water still wet. I've got an 8% double, I want to say double um IPA. It might be a double dry hopped IPA sitting in the fridge waiting for me, but ah yeah, no. I want something a bit heavier. Should we kick off?
AI Mechanics and Concepts
00:07:48
Speaker
All right, so this is going to be a big, lead captain a big topic. We're going to try not to go too deep. um And ah this will be a lot easier to consume this content on our YouTube channel because there'll be visual aids, but we'll, we'll do everything we can to make it understandable or at least.
00:08:06
Speaker
simple if you're listening audio only. So ah we really wanted to talk about AI and video games. I think the reason that started this was we I put up a short, like a 60 second short on YouTube about AI and video games. And it's to this day is still the most popular short we've ever put up. So clearly there's interest out in the market on it as a topic. So we thought we'd just pander to our fans and ah possible sponsors. And that's why we got here today. So we'll start with real high level. Possible sponsors. I thought you'd like that, mate. I made that just for you. Luna, you, Dr. Squatch, send me some soap. to me some so I'm a dirty, dirty boy. I am a dirty boy. so um So I thought we'd start with what is ah artificial intelligence. I think everyone kind of has a concept of that. It's simulation of human intelligence processes by computer systems.
00:08:54
Speaker
um And then, you know, I don't know if we want to get into the theoretical debate about intelligence versus creativity. Dan and I have thrashed this out ad nauseam in chat before. Um, the reality is. I just stoked the fire on that one. I was going to throw another call. Um, you know, the AI at this point is not creative. It's, it's capable of doing some cool shit that looks like it's being creative, like generative AI, which we'll talk about later on, but there's no true original thought.
00:09:25
Speaker
um There's no original work, it it creates derivatives of data sets that it's gathered. um And that's because it's all based off prompts, right? No, that's because the way that the actual algorithms work, it uses pattern data, it uses patterns in data to infer results or infer an answer to a problem. um Right. And so if the data is not If the pattern's not there in the data, it can't just invent it. There's no, you know, there's a level of creativity in some of these algorithms where it will explore, you know, a 5% random variable. But the other thing to bear in mind with all this, and but we'll talk about a little bit later, is there is no such thing as randomness in computer systems. Everything's binary. It's either on or it's off. And there's no way you can you can fake randomness through seeds and through fuzzy logic and stuff that we're not going to talk about.
00:10:15
Speaker
But it's important so to keep in the back of your head. Computers cannot be truly random. And that's really important when you get into some of these algorithms, if you were to do your own research, because we ain't covering that shit tonight. It's too heavy. um What I will say, though, is in psychology, and I thought this was interesting then after our conversation, there is there is already a difference in and in terms of a textbook definitional difference between intelligence and creativity. um You know there's a lot of lot of research that says you need a high level of intelligence intelligence to be creative but you can be intelligent without being creative. So really intelligence is about you know being able to work through problems.
00:10:57
Speaker
um understand patterns in data and and derive results or likely results to a problem, whereas creativity is really about generating you know new stuff that no one's ever seen before. So they don't necessarily have to cross over for this to be true artificial intelligence. So, yeah, I think my, my argument on that side of it was, wasn't so much about intelligence. Yes, it's smart. There's no question that it's, it's smart. More about ah the sentient side of things. yeah It's, yeah it's, it's, it's not sentience. It's not spending its spare time contemplating. Electric sheep. And yeah, electric sheep. Yeah, exactly. Electric Boogaloo.
00:11:37
Speaker
Different reference, but we'll go with that too. Could be the episode title we don't know yet. It's early days. um Look, I just wanted to, before we get much deeper, I just wanted to say... Electric shape to electric bodily.
00:11:49
Speaker
I have to give credit to a guy named Tommy Thompson, ah who has an awesome name. He runs and a YouTube channel called AI and Games. um if anyone's If anyone's interest is piqued throughout this episode, I'd strongly encourage you to go over and check him out because his breakdowns will treat each one of those single topics we'll talk about very briefly tonight. He's got like 30 minute explorations into video games that use those algorithms. So, good wealth of knowledge. so Going back to AI I mean this stuff's been around since 1956 so it's not a new concept and even machine learning has been around since 59 so um this is old school technology it's really just been in the last.
00:12:31
Speaker
last couple of years, the processes and data storage devices have caught up in terms of density to be able to deal with stuff in real time that we're seeing the proliferation of new AI. um And in fact, chat chat JBT is kind of like saying, you know, the iPhone was the first PDA bullshit. There were loads of PDAs out there before. It's just that most of society didn't really interact with PDAs until, you know, one liked their compact iPad with their shitty little stylus that got lost every day. um until the iPhone dropped and it was just user-friendly and easy to wrap your brain around in terms of the output. So when we're talking about AI in video games, you kind of have to split it into two major categories. There's a thing called symbolic AI and and we'll talk about that in a minute and you'll see that it's what's been driving non-player characters in video games for decades.
00:13:22
Speaker
Symbolic AI is is basically um a series of states or it basically breaks down a problem into a series of states and how those states can be changed by our actions. And then it allows the system to solve a problem by figuring out what the the final state is, what the goal state is, and the series of actions will get it there as quickly as possible. um So that's that's kind of symbolic AI at a 10,000 foot view, which we will talk more about and will make more sense. like Okay, so we'll get to that because I've got an example, a very good example of that in video games. Go for it. Yeah, interrupt me. about it Yeah, yeah, interrupt me. otherwise telling me me fucking talking So there was a there was a game for an FPS that came out in, I'd have it on my screen, but it obviously it doesn't have the 2005, October 2005. It was called F.E.A.R. and it was an FPS survival horror. FPS had some um some thriller aspects to it, but ultimately it was um
00:14:22
Speaker
an FPS where you had sort of bullet time abilities and things like that. And the AI, I say that loosely in the ah in the enemies, in the non-player enemies was fantastic for the time. And I actually, but all the FPS I've played since, I still don't think in a tactical shooter, they've reached this level of complexity or not complexity. ah believability in the enemy AIs in this. And so I've pulled up, I just literally just typed in what what is the AI in the fear game go and it said the T. Gope. Goal oriented action. That's exactly it. Yeah. So the team had just one AI programmer and so the design, and the designed
00:15:07
Speaker
A plan based system that lets the ai soldiers think think for themselves instead of telling me exactly how to behave in every situation he gives the soldiers a set of goals. And a set of possible actions and then let's them figure that figure it out for themselves and not only goes to the there the. The programming for the soldiers but it also. Blends into the level design because they had they had multiple angles to get to where you were you could never really back yourself into a corner in fear you always had an alternate route to get. Some of the some of them to get behind you ah in in the game so.
00:15:47
Speaker
It wasn't, they might've had one per programmer writing the goal orientated AI for those soldiers, but then they had another team designing the levels, allowing those soldiers to act much more naturally in a flanking way yeah than than um then I reckon almost any other non-player soldiers in an FPS since. Yeah. um So, ah goal or interaction planning is is a ah type of symbolic AI that that is on my list of things to talk about and you've kind of pretty much covered it, but it's also a perfect example of
00:16:22
Speaker
Of simple like i am general so simple like i think of it as ah it uses knowledge to solve problems as opposed to machine learning which uses data to solve problems. So when you talk about knowledge you know every i bought in fear has a world state and a blackboard so the blackboard basically has a bunch of variables that it knows can i see enemy equals true or false do i have a weapon true or false you know um is cover nearby true or false so it's got a series of. of data points that it and or knowledge points that it knows. It has a series of goals programmingming programmed into it. So if if enemy you know if enemy is visible equals true, then the goal is shoot enemy, until enemy equals a alive you know enemy alive equals false. So it's got some very basic goals. And then you've got a series of predefined actions. So can see enemy, enemy alive, shoot gun until not alive anymore. And that's its action sequence.
00:17:17
Speaker
And then you basically every frame or two that i runs back through its full action tree and it's set of goals and it just prioritizes got several other goals like you know don't die is a secondary goal to to kill player. And so if the end if if it can't see the player it'll next prioritize go find cover or advance towards enemy. And so it uses knowledge and a series of of ah predetermined actions and goals to get you know to drive its behavior. So it's a really good example of symbolic AI. And we're going to part machine learning, but um really the the the introduction to the topic is there are two major branches of AI. There's symbolic AI and there's machine learning. And so I was going to talk about, we'll talk about symbolic AI in a little bit more detail.
00:18:04
Speaker
I like the way Tom Thompson and I haven't stolen it, but Thomas Thompson talks about, or Tommy Thompson talks about, you know, AI that interacts with the player directly, AI that creates content, and then AI that models player behavior. And then it talks about, you know, game production as opposed to game interaction. But we're going to just kind of go top to tail. So in terms of, I'm going to make sure I don't lose my place.
Machine Learning in Gaming
00:18:33
Speaker
I was going to talk about so so symbolic AI as a concept. Machine learning, again, existed since 1959. As opposed to symbolic AI, it really talks about data. so It's a collection of algorithms that learn a solution by looking at existing data about a given problem. They can do it supervised, ah meaning you tell it, this is a picture of a cat and this is not a picture of a cat. and If you feed it that information enough times, it'll be able to figure out what's a cat and what's not a cat.
00:18:59
Speaker
um You've got unsupervised learning where you basically just feed it a shit ton of data and let it look for patterns in the data. And it won't know that something's a cat, but it might be able to group together all the cats not knowing what it's called. And then you have um reinforcement knowledge training where it basically just uses trial and error. It's got a known goal. It wants to get within 0.001% accuracy and it will keep just fucking around with, you know, different ways to solve the problem until it will hit its target or get closer. And it's not about solving a problem. It's about getting less bad at solving the problem over time. And so that's really what machine learning is. And it does learn over time.
00:19:42
Speaker
Tom, you look like you're going to fall asleep in this episode, mate. but No, no, I was just I was I was really curious before I didn't even want interrupt your your flow. But when you guys were just talking about the fear AI and that um symbolic AI, how how is that AI? Because the one thing I was thinking about was like, not in a in a first person shooter, I think it's a bit more A to B kind of thing. But then you think of like an open world game, like I've just been playing Assassin's Creed 3 and you go to the homestead and the AI are doing their little things around the homestead, but they do it at certain times. That symbolic AI? yeah Or is that like, no but and but a different different kind of strain of it Yeah, yes so there's about and like, i'm I'm not going to go through all of the different algorithms yeah in symbolic AI because and like, again, this has been around since like the 80s, our first AI in games. There's shit loads of ways to do things. I'm going to go through five or six as a as examples in symbolic AI. um Right. And a lot of them are either derivatives of
00:20:43
Speaker
or people mix them together. So you might have a goal or interaction planning, but you might have and and that itself is a finite state machine, which we'll talk about. But then the game might play over navigation measures. So that's a ah completely different set of ai algorithms that interact with the first. So let me let me jump into that in a sec. I'm really just kind of trying to branch out AI between symbolic and machine learning. So you got machine learning. It learns from a bunch of data. We then created a subset of machine learning called ah deep learning or neural networking, which people have heard banding around.
00:21:20
Speaker
Really, it's it's a way of storing data that models the human brain or what very little we know about the human brain. We know we've got a bunch of neurons. We know that each neuron is connected to a shit ton of other neurons around it. We know that if you were to unbundle the human brain's neural network, it had reached like ah twice between here and the moon. Like it is an incredible amount of neural fiber between our our our neurons. And what's cool is we've also discovered that things like memory the data isn't stored in the neural in the neuron itself it's actually stored in how many connections to other neurons around it that are related. Is how strong the memory is so the brains kind of fascinating but we know fuck all about it so yeah in terms of machine learning.
00:22:08
Speaker
What we did was build non-linear data sets where a piece of data can be connected to five hundred other pieces of data simultaneously it's not traditional if i was be then saying it's, when i got a property it's red and you might have an apple is another new ron and it's also red so it's now connected to the color and you might have a red boat and and you can create these really dense non-linear data structures and throw, Data like machine learning data at that and it's capable of doing a shitload more than a traditional machine learning model can.
00:22:42
Speaker
But we don't really understand enough about the human brain to go any further than that. But some cool examples. um So Google's DeepMind was a early example of of machine or neural networking, and they taught it how to play StarCraft 2 by watching professional players over and over and over. And it got to a point where the AI was capable, the neural network was capable of outplaying a lot of those professional players. And I'm kind of taking us down a couple of layers in machine learning, because we'll get to the bottom in a second. Generative AI is a type of neural network that uses, um that instead of solving problems through inference, it generates content. And we all know what generative AI looks like at the moment. It produces character dialogue, art assets, chat GBT is a good example. And then you've got large language models, which is a type of generative AI. And that's kind of why we stop where we'll stop in terms of depth.
00:23:38
Speaker
So, examples of symbolic AI. Yeah, I'm trying to keep it light. um And look, there's not. actually The doors are locked, by the way. You can't leave this lecture. yeah There's not that much more content. So, symbolic AI, um a lot of symbolic AI is used to drive NPC and player behavior in games. So, i I mentioned navigation meshes. So, a navigation mesh is basically a way for a computer game to figure out what you can walk on and what you can't walk on, or what the NPCs can walk on and what they can't walk on, and the most the easiest optimal path to get from point A to point B over the over the mesh. And then you break that in Skyrim by walking your horse up a mountain. Yeah, but the the the inability to walk through a wall is a good example of a navigation mesh. So an AI bot knows if he wants to get to you and there's a mountain in the way, he has to walk around the mountain. And navigational mesh is the way that the AI does that calculation.
00:24:35
Speaker
If you think about ah Quake 3 Arena, which was actually one of the first games to implement it, um you've got all those those kind of galley ways floating in space, and so the AIs needed to needed to know how to get from one side of the Quake 3 Arena to the other without falling off the map continuously. And so they had to create a navigation map so we could figure out how to get there. Death Stranding is a really obvious example. You know you can slow down and you if you walk on an uneven surface on an angle, you're your um packages start to tilt and you'll fall over.
00:25:07
Speaker
This is a nice easy example. Finite state machines, and I'm throwing some jargon up, I will put them up on the screen in the YouTube video. A finite state machine, good examples do. It understands what actions the player could execute and the conditions that drive what action remains active. so it It's particularly useful for systems where multiple tasks need to run at the same time. So examples might be um you could run forward, you could turn left, you could shoot, you could do both at the same time. um There are only a finite number of states in the system and the computer knows what those states are. And when you pair it with something like a goal, you end up with a goal or interaction plan and that's actually where that came from.
00:25:52
Speaker
So a GOPE is a finite state machine. There are only and a certain number of states that it can you know that the system can exist in at any given time. Let me know if I'm not explaining anything very well, because I could jump into more detail. So when you were talking about that before and you were talking about it, um it's ah it's got that checklist of things it has to go through, or it goes through every every now and then. Every every cycle, every CPU cycle. So they they have to be obviously not obviously that's why I'm asking the question. They would be binary choices. So simple, simple things like you were saying, see the player true or false. Keep shooting till player is.
00:26:38
Speaker
ah alive Unless another priority takes over and this is where you get the complexity in those systems because it might have several, several goals and the goal might be kill the player unless its health is low and then the the priority of the goal of surviving and running off for cover might override the goal to kill the player. And so those complex interactions is how you have, you know, you see bots in Call of Duty all the time, attack the player until you start shooting its mates or you shoot it a couple of times and then instead of attacking you, it'll run away. So those are where you get the the interaction of and that's and that's what makes, um you you know, these complex systems seem like human behavior because you've got a lot of variables interacting with each other that can reprioritize their goals.
00:27:24
Speaker
I suppose that's where it would cut in like, because you do play certain games where it's when you want to die, it's almost impossible for the bots to kill you. When you're not trying to die, they obviously output more damage. I imagine that's where that kind of factors in. Probably. I don't know. um Okay. but I think I know exactly what you're talking about Tom. Sometimes you'll just leave, uh, you'll, you'll have screwed something up. So you're like, Oh, just kill me so I can restart. And it seems to take three times as long for them to kill them to to kill you because you're not actively participating in their, um, in, in their, in their goal structure. So they probably go, okay, we need to ease up because it's a bit of a noob here and he's not doing
00:28:14
Speaker
they're not doing what they should be. I've seen videos on that where it's like these are the types of games where it's, it's, it is part of their thing where it's like it'll actively try and kill you harder when you're playing better. But when you're playing worse, the difficulty, like it's not, it's not a set difficulty level. Like, you know, Doom's a perfect example. It's, you know, you you set the difficulty level and that's the whole way through. But this is more like, ah I'm just gonna use it again. the Your Assassin's Creed, where it's it's, you don't set a difficulty level. You don't go playing hard, medium, or easy. yeah you You just go in and it um it adapts as you play. But that, that I imagine that's probably that where that factors in, is that what you were talking about, that that kind of blackboard bit. Yeah, it doesn't. And it because it's not modelling the player behaviour, so it might seem like it's interacting or it's responding to you.
00:29:03
Speaker
by learning your behavior, but it's not, it's more that you've, you've changed the variables. So the minute you go in suicidal, it has to reprioritize all of its, value all of its actions and all the variables change on the blackboard because you suddenly stopped behaving the way that you were behaving. And so it has to reorient itself. So the, the, sorry, Dan. No, I mean, it was just more of a thought. I guess there's something probably built into those sort of games too, that would be, um, Uh, like you said, say ah I've lost my train of thought now. Um, it would.
00:29:43
Speaker
No, it'll come to me. It will save you some. There's a lot of information to unpack here, so I scrambled my brain already. So let's talk about finite state machines. So finite state machines are the granddaddy of symbolic AI. And the easiest way to think about it is Pac-Man. So every ghost on the board can either chase the player, roam freely, or evade the player. Those are the finite states that those ghosts can be in. And it uses, and the artificial intelligence algorithm is really just about responding to the player and making a decision on, am I chasing the player? Am I running away from the player because he's got the magic pill or am I flu freely roaming because he's not near near or close enough to me to attack him? And that's a finite state machine. And it just basically makes decisions on those those actions that it can take based on the state that it's in. If it's in a roaming state, the action it can take is run away or attack.
00:30:37
Speaker
And so it's it's not that complicated, um but a lot of games use it like Last of Us, Half-Life, which we talked about, Doom. all All are good examples of finite state machines. The clickers can either be randomly roaming on its patrol path or it can be responding to a player that has made a noise or it's you know it's it's detected somehow. um Behavior trees are another good example of symbolic AI that drives NPC behavior. so um but Behavior trees are pretty much used in all AAA titles. It's completely pervasive over the last 10 years. It basically identifies a sequence of actions relevant to the current context. and this is where it gets This is the improvement on the finite state machine as it understands context, um but it ensures it can rapidly shift based on in-game events.
00:31:25
Speaker
so There's a set of constraints that define the sequences of possible actions, and it regularly reevaluates the tree when an action finishes. so You can literally think of a branching decision tree to say, well, um am I hurt? Then if the answer is yes, then I will go find a medkit unless somebody attacks me. um am i not If I'm not hurt, then is the player nearby? um Then I'll attack. Is he close enough for me to use a melee weapon? Then I'll stab him, otherwise I'll have to shoot him. And it's literally just a tree and all it does is constantly reevaluate or it might go back to the top and say, right, I've just finished shooting at the guy. I didn't kill him. I'll go back to the top. Am I too damaged? Then I'll run off and find a med kit and it goes back through the decision tree again.
00:32:09
Speaker
and these iter really you can You can kind of overrun like a perfect game for that. It's like Dragon Age where you got to set what your AI partners did. yeah So they ah would only use a ah health pack if you, depending on what level of percentage of health they had left. And what you're doing is one above yeah yeah and what you're doing there is changing the the constraints of the decision tree. the constraint for MI damage is 10% health instead of 50% health. yeah um And like if you think about Halo, so it's good examples, Bioshock Infinite, Division, Halo 2, they're really just, it you feel like it's and a bot that thinks like a kind of a dodgy computer version of a human, but in reality, it's just going through a decision tree 4,000 times a second if you've got a full four gigahertz CPU. And it makes it seem like it has
00:33:00
Speaker
intelligence and this is how we've faked intelligence, literally artificial intelligence. and We talked about goal-oriented action planning and the last one in terms of NPC and player behavior. Sorry, Dan. No, just in that goal-oriented action um is the. the alien in Alien Isolation. ah just it's It's another one that's perfectly ah perfectly balanced. and From what I understand, I found an article on it too, but it seemed too long to read in the 30 seconds while that intro was playing, that the
00:33:36
Speaker
the They developed two separate alien brains for it. Did someone send that to the chat the other day? I did, yeah. yeah yeah that one One knows where the player is all the time yeah and the other side reacts to things that the player does, but also gets drip fed things from the knowing brain to the unknowing brain. So if you can- That's really interesting. and stay too long in one spot, it goes, the knowing brain goes, they're not moving, so here's where they are. and that's been It felt in some stages, you need to be cautious in alien isolation, but if you're too cautious and too slow, you always get caught, because it's like the alien is sniffing you out. and it's it's
00:34:22
Speaker
it's It's patrol circle feels like it gets tighter and tighter and tighter on you. And if you don't move, there's no stopping until it will go away. yeah But if you, if you stop for too long, it'll just come straight back and kill you because the knowing brain's gone, nah, go kill them. This, this is their exact location. This is where they hide and go get them. I'll give you another good example that you'll, you'll be able to relate to shadow of Mordor uses goal or interaction planning. And so all of the NPCs that you interact with, all of the the the enemies that you interact with have a series of prioritized goals and that drives their behavior in the video game.
00:35:00
Speaker
But isn't that isn't Shadow Mortal a bit different like because of the the memory aspect of it how the. the all it No so so in fact all that means is the Blackboard doesn't get wiped off every time you reset the match. Ah right. So some variables remain Blackboard. Some some variables will remain cross game and remember certain variables about the player. Yeah, that that was the whole big thing that they they touted was the whole fact that the the enemies don't forget you and, you know, you kill a guy with fire once he's now either scared of fire or he's immune to fire. i Yeah. um I'm hoping you come back, Tom, because I've lost video for you. Oh, no. Not again. And and look I can still hear you perfectly if that helps. So ah Tom is having a problem.
00:35:50
Speaker
ah So Tom is always having a problem. I just take a quick, a quick break just to reset and I think. Pass the health check, which is more than I can say for my actual body.
00:36:02
Speaker
So um look, I know that was pretty heavy, but that's probably the least that's probably the most abstract stuff we're going to cover in this episode. The rest is a lot easier to kind of kind of directly engage brain with. But in 30 minutes, we've basically described every artificial intelligence driven NPC in video games in the last 40 years. because they all use a blend of those behavior of those algorithms or they use just those algorithms it pretty much covers every npc that you'll encounter in almost any video game we'll talk about a couple of exceptions to the machine learning section in a second but just about everything.
00:36:40
Speaker
So, I mean, the only other thing I added for symbolic AI being used for NPC and player interaction in games is Bach systems, which is exactly what it sounds like.
AI in Game Development
00:36:54
Speaker
When you're in the middle of Call of Duty and you narrowly escape the bomb and someone jumps over the, like an NPC jumps over the mic and Bach something at you, that's the Bach system. So NPC audio comms. And they just based on triggered events, uh, got a wall users. So every time you get in the boat and he starts and, uh, Mamiya starts telling your story. That's an example of a box system. So. Lead with that example next time. Yeah. Sorry. Will you play a lot of duty? No, I don't. I'm glad to call it duty since black, black ops one.
00:37:29
Speaker
Fair enough. Original zombies was where it was. Yeah. Fair enough. Yeah. And that's it in terms of- Play a lot of Call of Duty. I am offended. I will leave this podcast. I am offended. That's okay. You can now be replaced by the guy from the narrator voice from Star Wars Clone Wars. Admiral Yularen. That's it for in-game stuff, and then Symbolica is used in a couple of other places outside of actual video games, or outside of the the gameplay, so quality assurance. um In terms of bug testing,
00:38:03
Speaker
yeah I just find it funny, you're about to talk quality assurance and Tom's talked about Assassin's Creed several times in this episode already. Well, I'm going to use it tour don't you are i'm going to lose start unity i'm going to use Division as the example for QA, which was riddled with bugs when it came out, but yeah more and more. Dev houses are using symbolic AI to kind of automate um ah iterative testing of bugs because you have to go in and interact with the game content a million times and change one or two variables to figure out where where the bug is. And actual human bug testers are bored shitless and it's a lot of money that you have to spend on a shitload of testers to find on the bugs. That's why they just charge people for easy early access these days and let me let the player base do it for as well. like
00:38:49
Speaker
I know crowdsourcing. thought it was interesting that in Division 2 they did a video like ah it was not a tech demo so much as a video about their process for QA and they talk about that they've written some AI algorithms. That specifically tweak to a curiosity setting so it'll do stuff in the mission. That isn't strictly speaking what you're supposed to be doing in the mission. So you know players don't go from point A to point B on a map and kill the thing and move on to the next map. People like us especially you and I Tom looking for fucking feathers or whatever's buried in the corner will search every square into that map and interact with everything that you're capable of interacting with and so.
00:39:38
Speaker
A lot of basic symbolic ai qa testing platforms won't do anything more than just that goal oriented action planning we're talking before. you know Go to the thing kill well you know diffuse the bomb move on to the next mission and so they built curiosity into these algorithms now where they will go and explore they'll do non-linear stuff they'll do shit that isn't necessarily. playing the game as much as look at the pretty lights. Oh my God, look at the screenshot I took of Tatooine Sunset. You know, those sorts of things players actually do. um That's now baked into QA platforms too.
00:40:13
Speaker
and that's all like They'll still never be able to break a game like a speedrunner though. Very true, very true. um And then the last, before we move into machine learning, it's actually a good segue into ML and that is, um there's a couple of algorithms now that combined simple like AI with machine learning to drive um ah like an even better outcome in then and then that the main topic is an AI opponent so I'm particularly in racing game so the artificial intelligence. Cars in Gran Turismo 7 and Forza use similar but different systems that were trained on actual humans playing the game.
00:40:57
Speaker
And so, they have symbolic AI in the sense that they their goal is to beat you. And and in GT7 specifically, their goal is not to crash into you at the same time. That's a lower priority or it might be a higher priority because it's got to adhere to the rules of the actual licensing in the game. um But the behavior and the racing line it takes and the response to what's going on around it. If you break heavily in front of it, it may veer around you or it may break behind you. That's actually trained behavior using machine learning algorithms that has like read and interpreted real world player footage. Right. And so it's kind of kind of combining symbolic and machine learning. The last example I had is the the behavior of factions in Age of Empires 4.
00:41:43
Speaker
Also trained on watching footage of human players play other players in PvP matches. And so these- That's fascinating. These AI bots are getting smarter because we're training them to be smarter. I think that's really cool. They're clearly not watching me play because it'd be just dying every 10 seconds. yeah Yeah, we're there to train the algorithm not to be unbeatable. From the opposite end of the ball curve. As one of my mates used to watch me play games, he was just like, I can't Sam watching you play because you see a high thing and you just jump off with no regard for your leg bones. Yeah. And I'm just like, whatever. Like it's.
00:42:21
Speaker
but's <unk> Not a real person. I can jump off that if I want. If I land and I survive, that's fine. If I don't, I'll sort of restart the checkpoint. He's like, no, no, it's too chaotic. I can't watch it. See, that's the curiosity baked into the QA engine that we were talking about. Exactly. ah That's a perfect example of that kind of testing. Yeah, will I land this jump? My favorite thing to do in video games, will I land this jump? Every new game I play, that's one of the first things I do. I climb the highest thing I can and decide if I can, I find out if I can land that jump. Yeah. Nine percent of the time, I can't land that jump, but I find out the hard way and it's worth it. We're about halfway through. So we're going to talk about machine learning. Do we want to talk about these beers? Yep. So I'm ready to roll. I almost finished mine because it's delicious.
00:43:10
Speaker
Out of five, what deliciousness ranking does it get, Tom? Oh, it's definitely getting a four and a half. Um, this is, yeah, it's just, it's liquid black forest cake, like a little bit extra cherry, but not in a bad way. Like it's, it's quite well balanced still. Um, um, real dark, real dark chocolate bitterness coming through. Um, but and that's counteracting the sweetness of the cherry. It's a bit like. I will say it's a bit maraschino cherry, like that but extra sweet cherry-ness. That's why it works, yeah. Like if you're looking for the like natural cherry flavor, it's not that. It's kind of what I was trying to get to, yeah. um Yeah, it's fantastic. It doesn't taste like 8% either. It is, for a little hungover boy, it is going down smooth as silk.
00:44:05
Speaker
And look, if I had another, I would crack it right now and I would keep topping up that glass. I wish this was the never ending pack of Tim Tams right now that I could just continually keep it. But i can't it's empty. You're dead to me. Where'd you get it from, Tom? The Union Hotel is where I visit. In the bottle shop. Yep. In the bottle shop. I had a lovely red IPA in the bar and then I went around the bottle shop and got a nice little mix six. Nice. 15% off, don't you know? Um, see if it's available at one of the online shops somewhere. Oh, I'm sure it will. It's, it was one of their newer Hawker releases. And so Hawkers are pretty good at getting to most, most major bottle shops in in the area. So yeah. Yeah. if Easily four and a half stars. Nice. Easily. Yeah. Then.
00:44:48
Speaker
Oh, I wasn't on mute. Uh, I was, I was trying to put it on mute while I clicky clacky away. Fucking MX cherry reds. yeah I was just about to, I got the ah the b double E written for beer cartel to see if it was.
00:45:07
Speaker
yeah um ah yeah a Very good. um it's It hides its alcohol really, really well. yeah um I don't think there's as much woody barrel character there as what I would have liked. There's a little bit of whiskey character to it. But I mean, it's been sitting in my courtroom at work for probably 18 months or more. So um maybe some of that, that, um, barrel characters sort of just, uh, blended into the overall character. It's a bit sweeter than I standard, like any sort of beer. Um, but it all balances really well. It's got a, um, I don't know if you guys have ever eaten the old gold rum and raisin, uh, chocolate one of my favorite chocolate blocks.
00:45:52
Speaker
that It's pretty much what it is. it's it's It's got a little bit more sweetness to it than what that does in a little less of the sort of cocoa bitterness to it. um But it's definitely got that that rum and raisin character to it through the dark crystal malts and things like that. It's really enjoyable. I've re-poured myself another glass and I'm sitting at probably half the bottle now. So I will have to drink the rest of it while I go and cook cook some cook some food later on and finish it off. But I'm quite happily sitting on it. It's a solid four. That's sweetness without an extra little bit of
00:46:32
Speaker
bitterness either from the barrels, the dark malts or just overall bitterness to the beard just overpowers a little bit for me. um But the overall flavor profiles, really, really enjoyable. Yeah, cool. um And mine ah is kicking ass and taking names. It's it's It's got that really lovely apple cideriness to it um but it's got like a, you can really taste the complexity and flavors from it being barrel aged as opposed to it just being like an apple brandy.
00:47:06
Speaker
um I don't taste any alcohol and yet I know there's plenty in it because I'm feeling the effects even if I can't taste it. So, I'm going to give that a 4.75. That's pretty darn close to a perfect beer for me. um Yeah, nice. I'm actually surprised. Actually, I got to say, I really can't stand Barley once. I really get over them very quickly. Yeah. I'm absolutely not a fan. but It is, it is available on beer cartel as well. So we'll be placing an order. I also just looked at the, uh, the back of the can. I didn't read it at all the first time. It was yeah the most generic blurb, but, uh, they used oats in the malt bill and I really think that adds to it. Yeah.
00:47:49
Speaker
That's, yeah. Yeah. The rest of it was just waffle. Something about grabbing a spoon. Like it's a cake. It's like, yeah, yeah, we get it. Um, but yeah, tell me the ingredients. Don't need to make the joke. It's right there. Yeah. Yeah. Um, yeah. High score for a 14. Sorry. I'm still stuck on Pete's 475. It's 14%. It's fucking spectacular though. It really is. It's, it it look, you know, it's not the best beer I've ever tried. That would be the strong ale, but that we, that we've talked about to death. But for a- Key for Pete's score, high alcohol. Yeah. I think that might, might be on some sort of, yeah.
00:48:25
Speaker
Get as close to whisky as possible and Peter's happy. yeah ah um No, but its it is bloody good. It's not overly sweet, but it's sweet enough to kind of take the edge off the booze and balance out the booze. um You get a distinct flavour and smell of both the apple and the brandy, which I'm impressed that it's layered that way and that you can separate those those aromas and those flavours. So, yeah, no, that's why it gets a good score. Um, Pete, maybe you could start subscribing to the, uh, the BIL way of living the Bali wine is life.
00:49:00
Speaker
Yo, I don't know why it surprises me. it's it's interesting Sub cult in, in, in the beer world, but yeah, I can see you joining it. Yeah. Um, do you want me to launch straight into this or do you want to take a proper break before I do? Cause we're at kind of the halfway point in terms of the content. I'm going to take this opportunity and pee. Yeah, cool. All right. We'll be right back. ah that then ten then it then then and and and and hey and then then then then and then then and and and and hey hit the third and
00:49:32
Speaker
and and then then and then and hey You guys ready to talk about some machine learning? Do it. Do it. So I personally have a lot more like work professional experience in this space and I find these concepts a lot easier to wrap my brain around rather than kind of abstract algorithm concepts. So ah procedural generation is the most obvious form of machine learning and we've been using this for ever. um it you know it's a It's a set of algorithms that um applies a set of rules and and ah intelligent processes based on your parameters set around quality. um It includes pseudo randomness, which is why I specifically said before, there is no such thing as true randomness in computer systems. um But things like No Man's Sky, we all know that every single planet is procedurally generated in real time or ah upon load. um The cityscape in the Spider-Man game was a bit of a surprise to me.
00:50:30
Speaker
And yet the more I thought about it, the less surprising it really is. Does it mean the activity or the cityscape? The cityscape itself, the actual block structure, there are specific craft, handcrafted buildings ah that are points of interest, but the generic buildings in between are all procedurally generated. So my game doesn't look like your guy's games. No, I'm pretty sure they baked the seed, so a seed is not a concept I wrote down, but but a seed is kind of the the parameter you feed a random number generator and you'll always get the same outcome based on a set seed.
00:51:10
Speaker
So that's how random number generators work. But you'd have to stop and really look at that too. For Dan and I, if we're playing Spider-Man, you know, versus you, like we're all going to know. Do you have two blue buildings? Yeah. When you're swinging past at full speed, you would never notice it. Yeah. And I've got to be honest, I don't know if they bake the seed in and they used in order to save In order to stop the developers from having to pre draw the map. They could have just fed it a fixed seed for everybody's game and so everybody city scapes identical it just happens to be procedurally generated on the fly for you to interact with. or It's different per player i don't know i didn't do that research i just know that no man's sorry i can't know it's going on spider man was procedurally generated in terms of the city scape.
00:52:02
Speaker
I think that's super interesting. Certainly the player behavior is. Yeah, definitely. Yeah. um and And like the sorry the NPC behavior, like the the dudes just walking down the street and you think about watchdogs that was all procedurally generated behavior as well. All those those citizens that you interact with. um So yeah that's kind of the first layer of machine learning and one that's real you know people ah it's very accessible in terms of of people interacting in those kinds of systems. um And then we get out of gameplay and more into the actual generation of the game so real time rendering.
00:52:37
Speaker
um is used a lot in games now. So ah it's real-time rendering is where artificial intelligence helps to draw graphics on the screen using algorithms such as DLSS, which stands for Deep Learning Super Sampling. And we've already discussed that deep learning is about neural networks and it basically improves graphics quality via the GPU. And so both Cyberpunk 2077 and Hogwarts Legacy were good examples of real-time rendering and DLSS. um Is that referring to like the draw distance? No, it's referring to if you if you just view cyberpunk. So you and I will be able to view cyberpunk in 4K. That's the actual resolution of the textures on the walls and the players, the NPCs, etc.
00:53:25
Speaker
DLS takes that as an input and basically upscales it in real time based on what's on the screen at any given moment if your GPU supports DLS S3, which is the latest version. So it's quite likely Dan playing on a PC in 4K, his version of the game will look a lot better than our version of the game at the same graphics quality because it's applying DLS S3.0 through you know his GPU. right So it's kind of real time and this is where GPUs have kind of overtaken CPUs and importance when it comes to machine learning. And without diving too deep into the tech stack, CPUs are really good for computing integers. So one plus one equals two and binary storage and very fast at computational basic maths.
00:54:15
Speaker
GPUs are really good at fractions so floating point algorithms and people have heard the term floating panel floating point algorithms that's really what it refers to is anything between a one and a two. Is a floating point number it's a fraction of a number of an integer and GPUs are engineered to deal with maths based on fractions much faster than CPUs. And so when you talk about it, but they can both do it. Yes, absolutely. They they both have and do. I mean, you know, i'll ah yeah before we had GPUs, we had CPUs on graphics cards. So right. Absolutely. um The latest version of real time rendering is neural rendering. um And that's it's a newer example that upscales textures in real time. ah And the reason they do that is they can
00:55:08
Speaker
Let's say they draw 4K or even 8K resolution textures for the actual game files. If they tried to bake that onto a disk to sell on the store shelf, they'd never fit it on a disk. So what they're doing now is they might bake the 8K texture to start with, then they'll downscale it to say, you know, HD, bake it onto the disk as HD to get the file size smaller, and then use neural rendering to upscale it back up to 8K. And it will convincingly do that so that it looks like it might be rendered in 4K or 8K, but in reality, it's a shitty blocky texture. And so the the AI algorithms um that have been trained through machine learning are that smart now that you cannot tell the difference between ah a true 8K texture and one that was a shitty 1K texture, that it was just upscaled in real time. And games that use that, um so God of War, Ragnarok was the latest example that used neural rendering in real time to upscale textures.
00:56:08
Speaker
And it leads straight into ah pipelines and it's it's basically the same concept. So they're using machine learning. Originally, the mod community used machine learning. So it would they would take a game from the 90s with crappy blocky textures, pass it through a new rendering engine, and suddenly you've got a 4K remaster mod for that old game, even though the original game developers didn't participate in that process at all. um And so if you think about things like how do they make Skyrim look so fucking sexy when you apply a million mods, that's how they're doing it.
00:56:44
Speaker
um They're rendering it in real time on your GPU using DLSS or something similar or a neural network algorithm. Ah, those cheeky dogs. And it's gotten so good now that official games manufacturers have started using the exact same thing for the same purposes. So Mass Effect Legendary Edition, yeah they did not recreate all the textures from the original game. They just applied a neural rendering engine over the old shitty textures and now it's a 4K game. I wonder if that's what's going to happen. and there's there's ah There's a big rumor going around at the moment that 343 are going to remaster Halo, the original original trilogy, and then release it on PS5. That would be sick. What if they'll do that? On PS5? Yeah, they're bringing Master Chief over. Wow. That's big news. Yeah. um yeah and And you think about it, if you tie it back to what I said before, well what I said at the start around how machining learning works, machine learning works.
00:57:40
Speaker
All they've got to do is feed it a fuck ton of 8K textures and a fuck ton of 640 by 480 textures and tell it to interpret those six or upscale those 640 by 480 textures up and then compare it to the actual 8K. So it's supervised learning. It can compare it to the actual result that you're looking for and over time it will learn. Oh, I shouldn't have put that gray pixel there. That's a crack. It should have been a black pixel and it learns over time by training it that way. Right. So that's how these algorithms are taught. Is that the same for when you're teaching those generative AI to create art and stuff? Yeah. that You know, so many people do. So why then this is, this is my, obviously my big gripe with a lot of the AI art that you see today. Why can't they do arms? Why is it so spiky? It's, it's still learning and, and. Right.
00:58:33
Speaker
I guess the other thing. That's a different method to what you're talking about, like feeding and all that stuff. No, no, it's the same. It's the same concept around training a machine learning algorithm. It comes down to the quality of the data and how much data you've fed it for it to it learns over time. The more data you give it, the better it gets at doing its job. But the other thing I would say to you is. A lot of those algorithms for AI-generated art are freeware versions, which are two or three generations behind. And you would be amazed at how far like generative AI has come in three years. If you if you plot it, and because I've got a mate in that space, he runs a business in um like business process automation using AI. And so he's constantly on LinkedIn with
00:59:17
Speaker
you know, research articles from MIT and, you know, mapping the latest, the latest chat GBT models. so And the curve is like a fucking, ah it's it's practically vertical at the moment over time, over a three year period. So. Terrifying and exciting at the same time. But you don't, you don't give away your best algorithm for free. You give away, you give away two generations before and then if you want to pay me a million dollars a month and I'll give you access to my best. yeah You know, we're still using chat GBT 3.5. um chatgbt4 is baked into co-pilot for Microsoft's products, so aent OpenAI has licensed it. um Apple's about to license OpenAI, I think it's chatgbt4.5.
00:59:59
Speaker
and right And chat gbt as a conceptual model is nowhere near at the peak compared to some of the other ne ah like natural language models like or large language models so which are things like a algorithm called llama. And there's like 40 versions of llama but some of the best versions of llama run circles around chat gbt in its ability to interact naturally. So right there's a spectrum here and unfortunately there's also a commercial reality that drives our interaction as opposed to where they are in terms of state of the art. um And so just. And that's why I always say thank you to my Google home every time and I ask you something.
01:00:39
Speaker
Yeah, that's actually ah it's an interesting behavior. I found myself even interacting with chat GPT this afternoon to get that script at the intro. I found myself saying, that's excellent. Thank you very much. But can you please do this? ummit nu You just have to train that out of your brain. Well, when I was mucking around with light humans the ah the ah the Microsoft version, the Bing version of um Art Generator that I've been using, ah I started off talking to it like a person. That's a dirty you. I feel like I just walked into a sewer. I think that's the the one that I'm using. ah But it was just the one Kelvin was using, so he he put me onto it. um But I started talking to it like a person and it wasn't giving me the results that I wanted to see.
01:01:34
Speaker
So I stopped talking to it like a person and removed all the extra fluff in between that you. Talk to me, I'm your dirty little slut. Yeah. And basically I wrote things like that. Pay me like one of your French girls. All I heard was daddy wants, Tom. Yeah. And daddy wants hair metal dwarves. Give me now. Yeah. But even that's full of that full of fluff words. You just, yeah you just write, you just write the the most basic things. And I found that the punctuation in a sentence helped a lot as well. So it so have it in two or three words. Which is funny because that means that the future generations of fucks because they don't know how to punctuate. Yeah.
01:02:23
Speaker
For like maybe five or six words in a sentence full stop next sentence yeah full stop and then break it down like that break it down how you would do. the most basic support boss code. you know what But you know what that's telling me, Dan, is that the input engine for those art generators are way behind like a chat GBT. Potentially. Yeah. Well, definitely. I mean, you saw the prompts that I fit it and maybe we'll talk about that at the end and I can read out the actual prompts. But I gave it complex language. I gave it complex context as to what the podcast was about. I said our hosts are, you know, Fred
01:03:02
Speaker
Tom and no, I've Fred Jim and James because they want to give it our real names. I didn't even give it the name of the podcast. And I didn't even specify what type of pop culture we talk about. it's So like I said, Marvel movies and DC movies and comic books, I just said pop culture movies, TV shows, video games and the the pop culture references that spat out were all on fucking point. So it just tells me that those those graphics, AI engines, they've obviously been focusing more of their time on the actual graphics, artificial intelligence, not the interpretation engine to to feed it data. And which version of that we're you using for that, Pete?
01:03:43
Speaker
the latest freeware version of chat gbt i think it's for but i'll tell you in a sec i can i can multitask i promise it's for cool so moving on to the next next video game um machine learning so animation cycles i found this really interesting so um they use it for animation cycles where You know you can walk and shoot or you might walk stop and change direction and they're using machine learning in the animation process on screen to just smooth out the blends between the various animations. So a human animator has to create a walk cycle and that animation is how it walks across the screen and then they have to animate kind of the the pick up the gun out of the holster and point it and shoot it.
01:04:30
Speaker
But all of the stuff in between is all completely machine learning no animator has to actually can craft that and. Ironically we talked about this almost a year ago in the first up late around how smooth the blends where when we change lightsabers in what was the second star wars. um saliva Survivor survivor. And we talked about, we commented on how smooth it was to put the lightsaber together in real time and split it into two pieces from the, that's all AI. No human actually sat there and handcrafted those animation sequences. Interesting.
01:05:05
Speaker
lasting Another thing we've talked about in the podcast before is is using machine learning for accessibility. So the obvious easy easy to use the easy to understand ones are things like real time subtitle generation. People don't have to sit there and listen anymore in seven languages and or type it out in seven languages based on the audio track in a video game anymore. Those humans have now all been made redundant and replaced with machine learning. that And yes, I was about to say like, that's great. But then I was like, oh no, that's bad. There is a human with all frozen yogurt at the store. The yogurt is cursed. That's bad. um Text to speech. So again, a lot of audio. So NPC dialogue is now no longer voice acted. It's actually AI voice acted the same way that our intro was. You just feed it text and it spits out the vocals.
01:05:59
Speaker
um But the other obvious one that makes but it a sense yeah yeah the other one that we have talked about is I think we're talking about in the context of thalassophobia with. um ah Forbidden west but we talked about there was an AI I think in the news I mentioned there was an AI algorithm that is now being made freeware by one of the big studios to analyze frame by frame looking for epilepsy triggers.
AI Impact on Gaming Experience
01:06:23
Speaker
yeah Yeah, and so that's and that's ah that's an open source machine learning based algorithm that that is now out there in the market. So if you're a video game developer, you don't really have to think about epilepsy triggers in your video games anymore, you just feed it to the algorithm and it warns you if you've got any triggers.
01:06:39
Speaker
And that's based on that's good. That's based on machine learning where somebody sat there and fed it thousands of hours of videos that will hit a epilepsy trigger and thousands of hours of videos that will not, but you could misinterpret as hitting an epilepsy trigger and it learns the difference between the two that are quite subtle. ah I know one of the machines going to rise up against us because we just fettered a bunch of videos to give it epilepsy. so one actually you know It's one of the things I really skipped over um that I probably shouldn't have is when we're talking about symbolic AI, one of the important distinctions in video game technology around symbolic AI is there is a big difference between traditional symbolic AI that you might build build into corporate applications and robotics use. Like if you think about goal action,
01:07:29
Speaker
ah what What was it called fucking goal oriented action planning. yeah You could feed that to a robot and say your priority is to get to the top of that beam, but there's another priority. that' and That's like loads and dynamics. Yes, exactly. That's exactly what the picture in my head. The key distinction between video games and real world is video games. The reality is distorted to make it fun. You don't want to teach a real robot. It's fun to shoot people and throw grenades at them. So their game AI is is its own subsidiary of symbolic AI. And they eventually will. we know We know it's coming. yeah yeah You don't want it to learn that on its own either. yeah yeah yeah um but're we're not going We're not intentionally trying to create chappies, but we will. One of the key distinctions there, though, is that what's fun is subjective and you want your AI to drive fun behavior, not necessarily the most efficient path.
01:08:26
Speaker
Which is an objective path. And so game AI is actually, symbolic AI is is kind of oriented towards what's fun, not necessarily what's the most optimal path or else every bot would wipe the floor with you a hundred percent of the fucking time. ah Easily, yeah. It's really interesting because ah the ah the saddest in me, one of the most fun things in RDR is weaning the enemies in the leg first and then executing them. so
01:08:55
Speaker
Let's not teach the machine that. I've actually just gone back. i dan I've, Dan, I've thoroughly committed to RDR1 last week. And it is one of the most relaxing games I've ever played. I just ranched. I ranched for about eight hours yesterday. Just ranched. Everything you enjoy about that game, magnified tenfold is where you get with RDR2. I'll get to it eventually. Yeah. I'm a little bit behind. um um Pete, does that does that is that then that's doing all the stuff with like color blindness? Like when you you mark character, that's all AI? Yep. Cool.
01:09:28
Speaker
Where you well so they may apply filters so um like for example there's gotta go to a Ragnarok you you if you apply the ah hard to so like there's like a filter for like but the color color coefficient it's not even colorblindness it's like it's literally just like I'm I'm actually kind of going blind like I can't I can't differentiate I actually don't know if somebody so crato is always in like ah ah a green and then a trace is always in a red so you know where they are on the screen. ah I would assume... I'm assuming that would be the AI. I think the AI would draw the color filter, but I assume a human would have selected that Kratos should be green and the game, obviously the game engine knows... Oh yeah, I was more saying like in terms of those accessibility features that AI is... Yeah, so that's one example where I'm not yeah i'm not sure if that's ML or if that's a human that says Kratos needs to be red and I've come up with that idea. I genuinely don't know what I have to find out.
01:10:24
Speaker
um The next topic is online matchmaking. so It's really interesting. so xbox Xbox does a lot of this. A lot of Xbox games do this. They observe the actual player behavior and learn its skills in various circumstances over time. so It's not just a plain kill to death ratio kind of scoring system. It actually understands contextually skill sets in various environments and then ranks you and allows it to match make with other players in a similar skill set in various circumstances. So it's kind of the next generation. PlayStation definitely doesn't do that. That is bullshit. because i The amount of times I've played an online game where I've had to do something online and I get matched with absolute pros when I barely struggle to remember where my, you know, my med pack button is.
01:11:14
Speaker
Yeah well i mean there's this your mileage may vary play modeling and analytics where we have talked about this so this is where the using machine learning to to trigger alerts to game developers about possible bugs. Play behavior within their maps where are they focusing where are they concentrating they could do heat maps for. You know, where are the players most interacting? We should put more NPCs in that area. um And we talked about this in the context of Baldur's Gate 3, where they were actually testing encounter difficulty by having a heat map of how many player teams wiped in any given fight.
01:11:54
Speaker
And so they' we're able to see a hot spot where that underground boring motherfucker in the under dark in bolders k3 was just annihilating people and they tweak the difficulty slightly as a result of a machine learning algorithm going, there's a problem here guys. Yeah, this is tuned too difficult. um And like kind of following on, and we're're we're almost at the end, anti-cheat algorithms are very similar in that sense, in the in that it's constantly watching player behavior. So you you can imagine machine learning as a robot staring at a screen watching footage, because it is kind of very similar to that. It feeds constant footage in, and it can now do it in real time. And that's how things like Valve's anti-cheat system, um Blizzard's Warden,
01:12:42
Speaker
epic Epic's Fortnite AI, Riot Games, Vanguard systems, they all work the same way. They watch players in real time and figure out if they're aiming through walls. by me Where was that 15 years ago in Counter Strike? Grand Theft Auto Online. I could pretty much ruin that game. um What's interesting now though is with the advent of large language learning models, or large language models, LLMs, and real time voice to text, they're now able to do toxicity detection as an extension of anti-cheat algorithms. So if you're calling someone an FNN, it's in real time capable of turning that into text and then running it through the same algorithms for this is, you know, unacceptable behavior. ah In a real world example of that. So we had we had a group
01:13:34
Speaker
When I was playing Battlefront 1 recently, I went back and did that before I went to America. and like We had a but a big group. it was It was a trophy hunting group. so you know they were we would We were essentially picking certain game modes that we could boost. just You needed to get to a certain level. It was just the easiest way to get XP. um we had
01:13:55
Speaker
Eight of the 12 people in the in the PlayStation chat we had got reported for anti-cheating. like things, and it turned out this particular session they did, they had some guy jump in, some random jumped into it, um said, hey, can you guys do this? Someone goes, no, man, we've we're doing but doing whatever. Can you fuck off? like We've planned this. It's been planned for like a week, so you can actually fuck off.
01:14:27
Speaker
The guy left, then instantly they all got reported. They all got a like a year on week reporting thing. i so I wonder if it was that. No, I suspect that guy reported him because he was just a little bitch. This is more like using the n-word in Modern Warfare 2 and Modern Warfare 3. This is like horribly toxic behavior. No, not Call of Duty that I apparently play every day. so The beta for this um this algorithm was actually driven into metal Modern Warfare 2 and Warzone, and the final version- That is the least surprising bit of thing you've said all day. The the final version of of the Activision Blizzard's version of that utility is in Modern Warfare 3. If you call someone an F and N word, expect to get instant banned. It doesn't report it to a human to check the footage, it just bans you straight up.
01:15:21
Speaker
yeah And I think that's a really good thing because and ah what will, interestingly, improve over time. Well, there's no place for the n-word, man. It's pretty much the only place you see it. Unless you are a person of color who can use that word. ah yeah that That, I think, is where it does that factoring. yeah um Certainly, large language models are capable of understanding context. And so, I don't know if it's capable yet of understanding the context between the ah ER and the A and the intent behind the words and the, you know, you got to be careful. And one of the things that that is out of scope of this topic today is,
01:16:00
Speaker
is ah racial context in languages that you want to make sure that you don't train a machine language or a machine learning model to be racist, yeah accidentally or on purpose. And there have been examples of Google had to shut down a chat GBT style bot a couple of years ago because it was spouting Nazi fucking propaganda at people because dumb fuck humans had fed it so much language around it, it had learned that behavior. So, you you kind of treat some of those early systems like a six year old, it's going to parrot back and it's going to form its world view around what you feed it and sometimes. But again, there are certain times when you can use that language if you are a person of that culture where it's like you shouldn't get banned for using it. But like maybe like maybe it's it is contextual, it's like if you're if you're playing with a group of people that are all people of colour, you can use that word with
01:16:55
Speaker
So it's going to sound like a joke, but it's not actually intended to be so like, it's not like it's going to say, well, his name's Leroy, so it's probably okay. And that's a good example of you have to be really careful when you train this stuff that you don't accidentally train it to be racist. Yeah. Yeah. And it's, if it's got to be, uh, language-based too, it will, or a culture-based like an Aussie could go, good shot. Can't. And yeah, pencil in yeah that's all right. I thought I had to drop it in there. Yeah, it's been a while. But I mean, that's so. Socially acceptable. among Yeah. Socially acceptable. In terms of endearment in our culture. Exactly. So that that would be something that would have to be based in. And and this is and it's something I've also definitely said on mic to people. Yeah.
01:17:46
Speaker
With love. It's got to be one of those things, especially in that space where it can't be a hard band. it has to be There has to be some sort of sound bite that's grabbed by the AI where somebody can go, hey, Then you have the human judges come in. I now need to appeal to the company and the person saying, you need to review this for me. I've just lost access to like all of my thousands of dollars or all microtransactions that I've paid you guys. I can no longer access that account because
01:18:21
Speaker
your AI has determined that I'm toxic when it's just a socially acceptable this is how i talk to make terminology in in my country in content then making ah yeah and they can say yay or nay. But I know they're not doing that at the moment and there's a lot of hard bands around that in those sort of spaces. But wait, there's a lot of that machine learning stuff. but That would definitely use like the trophy data. But when you're talking about like Baldur's Gate 3, you'd assume that's you using a lot of that. No, so it was literally just watching. it's I mean, the ah that particular example for BG3 was it would watch every combat encounter outcome across the entire game.
01:19:00
Speaker
And every time there was a full party wife, it would, it would basically create a hotspot. And if you had a a million party wipes in one encounter in one area, it, you know, it was obvious for the developers just to scan through the, ah like an overview of the map and go, shit, there's a whole lot of white there. That's white hot. Something's going on there. We'll delve into it. Is that one of the reasons for why developers now are really pushing for constant online connections? No, I think that's pretty much just rubbing their hands together in greed, to be honest with you. Okay, fair enough. But ML will become more and more of a factor in those types of games because a lot of the processing you can offload to the cloud that way with live services, whereas if it's entirely running on your console. Oh, not live services. I was talking about like,
01:19:49
Speaker
you think about your Last of Us. There's there's a few games coming up that ah that the single player- Still needs to be internet connected. it's it's It's now saying, no, there needs a constant internet connection to play this single player thing. When you don't, you wouldn't think you'd need to be like connected for that. Yeah. And the play look the the honest answer is I don't know, but what I what i do know is you will you will increasingly see a reliance on an internet connection for some of the other stuff we're going to talk about in terms of current research in the future. ah come The last piece I wanted to talk about was just really basic is um code generation itself. So programmers sitting down and actually writing code for the game. You almost can't code these days without real time AI tools popping up and saying,
01:20:34
Speaker
You're using this method wrong. You're not writing this code as efficiently as you can. You have a myriad of, of utilities. So the more annoying version of Clippy. A hundred percent while writing code. But Clippy is a really good example of where large language models are going. But, but like, you know, I'm, I'm working with some developers at the moment at work and they basically pass all of their code. So you've got Microsoft's development environment, which will tell you in real time, Hey, you just typed that algorithm shit. Try again like here's some pointers and they've not got copilot which is chat chat 4.5 real time interpreting your code and telling you if you're you're using their coding language properly but they feed the whole code chunk.
01:21:20
Speaker
Into a third party utility the users I to look at contextually how does it work is it optimized can you rewrite that code better. And pass you back comments and so a lot of video game writers and it's probably more important in video games where the hardware shit. Like the PlayStation 5 is what, three, four years old now? Five years, I think technically. The hardware spec was settled on seven years ago then. Yeah, yeah. Right. So, so writing optimized code for consoles is probably more important than professional, you know,
01:21:53
Speaker
utilities in the in the commercial world because most utilities are running in the cloud these days so you just want this company that's where you don't always always succeeded is that they. They always released a last this game towards the end of a playstation console life. because I don't want the crap out of it yeah yeah. yeah So that's it for kind of taking you through top to towel artificial intelligence in video games from kind of symbolic AI and and machine learning and all their branches and offshoots. um I guess the only other thing I wanted to talk about was and before we talk about the future, because I think that's where this is exciting and where we kind of started a couple of episodes ago. um There has been some some videos and I've only really looked at one.
01:22:40
Speaker
around researching how to better integrate more modern AI systems into video games. so i There's a bunch of Ubisoft videos. So Ubisoft is spending a shitload of money at the moment in R and&D um in baking chat LGBT-style models into their video games. so but and And there's a bunch of videos now where they invited YouTubers to come into their studio in Montreal and interact with their demos, their tech demos. And what it does is your NPC and it's a traditional RPG style engagement like ah a conversation with somebody in cyberpunk.
01:23:17
Speaker
Yeah, um but instead of picking answers from a dropdown box, you talk and it converts your text to actual language and then interprets that actual actual language in context. and then interacts with the NPC. I could just already picture how much we would ruin that. oh it's It's already been done. Just just us three. Just like like just dropping sea bombs left, right and center and some poor NPC. The NPC and the NPC learns. so it yeah and And so the game builds up a picture of your, I'm going to say persona.
01:23:53
Speaker
Because the the fear will be over time it'll be your actual personality in these video games will actually know quite a lot about you as an actual person. well it's like the you the game a bard's tale you apply that it was like p s two time p s two zero yeah yeah it was a pc game PC and PS2, but it was that you got to play as snarky or nice yeah as this like RPG thing. And you could just all the time pick the really, really jerk options to talk to back to the NPCs. It'd be 100% be that for us. so So in this tech demo, this guy says, you know, there's some prompts and it says, you know, you get extra optional, um and optional mission credit if you ask him about the resistance and you ask him about the other people in the resistance. You got all these prompts that kind of tell you what to ask about.
01:24:39
Speaker
The guy goes right there in the guy goes so so you know what are you into what do you what you know what are your hobbies and stuff in the npc comes up with an answering goes are cool i would just really like eating hamburgers. He's like, OK, cool. So the resistance, are you're going to join. it And he goes, no, I mainly just want to go out for some burgers. Do you know any good hamburger joints around here? And he goes, well, not really, but can we focus on the resistance? We're in the middle of a fucking war at the moment. And maybe if we can beat these guys back, then we'll go out with so for some celebratory burgers. And so it it wove the dude like the human's language back into its answers and it ended up being
01:25:17
Speaker
Probably more sarcastic and quippy as a result of the yeah bullshit answers the human was giving. And so this is just a tech demo, but this is kind of what we were talking about in terms of the future of games several episodes ago, is having NPCs being able to interact with you in real time and and develop a personality profile on you or your persona in the game. And this is kind of where I wanted to throw it to you guys. We can't do that. I already have enough of a hero complex and I pay enough of my therapist a month to get through that. I can't be having me talking to NPCs trying to be even more of a hero. Like, come on.
01:25:57
Speaker
No, but what you might have is is LLM baked interview. i Look, I expect that the NPC interaction in RPGs will 100% be chat LGBT interactions moving forward. Yeah. As soon as they're able to bake bake that in. what what Weren't we already talking off chat about that a few months ago, about how they're using chat GBT now to feed the conversations that NPCs have with you? I think we were postulating. Rather than using a writer for that. No, but that was on episode we talked about that. that's actually That was the seed for this episode, really.
01:26:32
Speaker
Well done, Tom. You paid attention. ah um that do what what what I know Tom was just joking about what he was saying about the the talking to the MVCs and hero complex. but that I wasn't. but that's That's the crux of my therapy sessions every month. But yes, go ahead. It could quite it could quite be a dangerous thing for some people if they get into this ah Get into this. i'm I'm here to save the world and they're communicating with an AI in a game in real time. And then they go out into the real world and people don't react to them.
01:27:08
Speaker
Like, because obviously there's there's still going to be borders, boundaries and parameters set in place to drive stories forwards and things like that. So if you're the Dragonborn in Skyrim, you're always the chosen one. You're always the Dragonborn you're always going to be spoken to in a certain way to drive things. There are enough people in the world that Mummy didn't tell them they were special enough that already go off and do horrible things. Exactly. I mean, you start getting things like that out there and they go into a cafe and talk to somebody behind the counter how they can freely talk to an AI in the game and let fly whatever they want to let it fly. It could be a really bad thing for...
01:27:44
Speaker
um I wouldn't say a lot of people, but there would be definitely be a number of, ah ah ah probably a higher percentage of people out there we would like to think than we would like to think that would react in that way. That's an extreme example, but what about the basic example of lonely video gamers forming relationships that are meaningful to them with NPCs in RPGs? I was actually about to say that. Because they do respond in such a human way in context and they remember who you are. But also then going to Dan's example of those people then trying to go out to a cafe and they talk to someone behind the counter and in a game, they can woo that person behind the counter. Well, I'd be less worried about that. Well, no, I was just saying that it might be an extreme example, but they go to a cafe, they can't talk to the person behind the counter the same way they can talk to the AI.
01:28:38
Speaker
They get frustrated something happens, you know, and it that that's pretty common. What about the NPC that this this human had fallen in love with getting killed as part of the script? That's a totally different level of emotional impact. Absolutely. Yeah, I didn't. Yeah. So, you know, that's I mean, taking taking video game context as it nears human psychology back into the real world is one problem and extreme to deal with. But what about just the dude sitting in his basement whose only meaningful relationship is with an NPC in a video game and the game ends even if they don't end?
01:29:14
Speaker
So, yeah, I mean, example of me playing Red Dead Redemption last night, I had to defend the dog Charlie as we went around McFarland Ranch. I would have died for Charlie. I would have killed everyone in this entire neighborhood for Charlie. Yeah. But yeah, I mean, look, we've we've got a lot of hurdles to to jump in the next couple of years as a society, less specifically around video games, but that is part of it. But certainly the ethics of machine learning and and large language models as we kind of approach. I mean, we we are already at a human level of interaction, frankly.
Ethical and Social Implications of AI
01:29:51
Speaker
I mean, the conversation. Would you would you say, Pete, that video games are the perfect like testing ground for these kind of things and like compared to say like using Siri on your iPhone or
01:30:00
Speaker
Whatever I use for your phone, I would say that they're the perfect microcosm for the issues that we're likely to face in the broader community. yeah um Testing, I think, yes, they're a good good petri dish, but they're also a good representation of the broader population issues we're likely to face. um The bigger challenge and then I was talking about this yesterday, the bigger challenge will be Western cultures likely to try and actively legislate some rule sets around artificial intelligence in the next year or so. And frankly, it's moving so quickly. They fucking need to do it in the next year or so. And yet there will be other cultures where or other areas of the world where they won't. And those are probably the scarier places.
01:30:46
Speaker
Where you have, yeah you have a context aware artificial intelligence that charges you more for your coffee because you're a bad person than that person last week as captured on video surveillance. Thank you, CCP. Right. We have plenty of countries in the world that are actively surveilling their their people and and these types of algorithms can be used for pretty dark purposes very, very easily. I think it's also a very a very good point that the people who are going to inevitably enact these rules against what you're talking about and like the lords around it.
01:31:23
Speaker
Can I just have some younger people on those panels? Because I don't know if you ever watch the US Senate versus TikTok. None of those fuckers have ever existed in the real world. They've grown with a silver spin up their ass and a gold one up them in their mouth. No, but they're also the people who make those rules, and especially in the US. But if you watch any of the trial of of TikTok versus the US Senate, and it's just the old man being like, how is the TikTok in my phone? like ah fuck Well, they asked the they they asked the the guy that was born in Hong Kong, who is a US Senator, if he's a part of the Chinese National Party, like 11 fucking times. He goes, yeah. Like, and it was so obvious. He was just ready to say, listen, just because I look fucking Chinese, like, you're the racist for asking me this question for the eighth fucking time. Yeah.
01:32:10
Speaker
Anyway, I'm going to open the floor now. Like i've um I'm done with my, my, uh, my lecture around artificial intelligence and video games. So I guess it was very good. Thank you for Professor Pete. Kind of throw it open to, to what are our thoughts about the future? Cause I think it's, before we get into that, I'm going to review my second beer for this episode because I've really been enjoying it as Dan has, I'm sure seen with my factual reaction. Will you do that while I take a leak? So I'll be right back. Um, I decided to crack open the Dick Tracy, uh, 10 years of Ben spoke.
Craft Beer Review
01:32:41
Speaker
So this is their 10th anniversary. Um, American Brown owl. And it is a Dick Tracy. Okay. Yeah. It is a fan. It is like a.
01:32:54
Speaker
pinpoint accurate American brown owl. It has that dankness coming through. It has that really rich molt character coming through. Um, the, the like robust bitterness that you would normally attribute to like a black IPA and that roastiness is, is just sitting on the top of it. It is a true testament to 10 years of a brewery and it's, it's amazing. It's a 6%. yeah Those who are unaware, ah Dick Tracy, not only is it a pop culture reference, it's a reference to Richard and Tracy who set the brewery up. Yes. ah This was Ben Spoke's first brew in 2014 as well. It says on the, so this is again why it's their 10th anniversary brew. Happy 10 years. Fantastic way to celebrate it. This is this is the beer I would
01:33:44
Speaker
demand if I was in Canberra, especially right now, like on a chilly day, like today, like, yeah, it's, it's sensational. This is, this is one of the best Ben spoke beers I've had in a long time. That isn't part of their core range. Like crankshaft is, uh, my go-to IPA. If I see crankshaft on tap, I grab it. Um, but the fact that they can still bring out bangers like this, like that's a, that's a true Testament to a brewery. So yes. Well, well done, Ben spoke. I was talking to Richard at Gabs and I wasn't drinking at Gabs because I was working and he made me try his non-alcoholic. I'm like, I don't like non-alcoholics, man. And he's like, no, just have it. it was It was good. It was actually really good. I shouldn't have been surprised, ah but it was by far the best non-alcoholic beer that I had. It had good...
01:34:36
Speaker
Not strong hop character to it, but hop character to it. And, but it even had a multi, a multi backbone to it as it finished. Um, which is rare, which is rare. and an nonahoic yeah That's exactly it. So I was talking to him about how he was, how he made it in that. So he wants to, uh, get more hop character in it as well to boost it up even more, but, um, absolute, um, cracker of a brewer, which it is always has been, there you go. Pete, I will put my beer notes in the things for you. Thank you, sir. You're welcome. And I will add it to my beer cartel cart. I think I'm just going back to like the future of video games.
Creative Applications of AI
01:35:17
Speaker
it's What's interesting for me is researching for this podcast, you know I had a clear picture in my head because I interact with machine learning at work. And i have to I've had to do some research for what the future might hold for the
01:35:32
Speaker
you know the specific application we use AI for at work or we may use in the future. Things like looking at data about customer behavior without being identifiable, looking at um machine learning for visual recognition systems and you can you can teach a camera to do a fuckload these days with AI and and ML specifically. and And it really does allow you to be quite creative about the applications for machine learning in in those spaces professionally. But I think so, you know, my focus was more on the future of video games and could we have NPCs that don't repeat themselves over and over again with exactly the same predetermined line because you can now have it one, contextually aware to come up with 500 ways to say the same thing.
01:36:21
Speaker
And three, pass that into a natural language model and have it spit it out the way we did the the start of our episode. But the more research I did, the more I realized that video games already use a fuck ton of AI and we have been since like the 80s. Like almost everything in a video game is AI. I just think with what you would, especially what you were just saying now about like, um when they when they spit out the same line, like you walk past an NPC that you're not necessarily going to... Which was the example we talked about in the podcast, right? I was even more thinking about like ah an NPC... per example last night playing rdr1 and i was doing one of the stranger quests and it was just like this woman was standing at the sheriff's station just yelling mr. mr. come over here
01:37:12
Speaker
Initially, I was like, fuck off. I don't have time for you right now. from I'm doing something else. But if she'd been able to change the way she was like requesting my presence, I might have engaged with her earlier than I did. um I think that's probably the the way that I would, if I was a developer, I would be looking at using this AI to to to fix that. like Well, how about scripting memory? thats So so they might write the first encounter is a Mr. Mr. Please come and help me. And then you might encounter her later on in the game. And she's like, Hey, play a name and call you out by name. Yeah, help me with this last thing. Or you didn't help me with this last thing. This one's more important. Your fame's gone up now. Your fame ranks higher. But being contextually aware of Mr. Marston, please come over and help me. Like I know you help people kind of thing. Yeah. But being contextually aware not just that you, because we've we've already had games where they're that the NPCs are aware that you finished a quest, or that you you have a certain honorary rank. But being able to limit the the their awareness to context, how you did this in my town, but I'm not aware of what you did in the neighbouring kingdom.
01:38:25
Speaker
adds a lot of realism to a lot of these types of games. yeah I think there's a lot of scope there to kind of explore a more realistic RPG. And and a lot of it to for me, I mean, we've talked a lot about first-person shooters and those types of, you know, the the symbolic AI that goes into those. But RPGs, I think, is really where large language models are going to really unlock a capability we don't have. Absolutely. Yeah. I even think back to what we were just saying before about Shadow of Mordor, and it's like, there's definitely like, you know, sure, the orcs talk, but you kill someone in one side of the map of Mordor in a certain way. How the how the fuck does orc JoBlo over here know that you, you burnt this guy to a crisp and he's like, hey, grave walker, you won't set fire to me. It's like, it's when it.
01:39:15
Speaker
that AI goes a little bit, it's too predictive in a way. It's too fast. Yeah. yeah yeah I like what you were saying in terms of, especially with RPGs of like, yeah, the the kingdoms don't necessarily talk to each other. What does Dan keep dropping? I keep playing with the ballcat. You fucking fidget spinner. I do. and I always need to fidget with something and I cleaned my desk before and the stuff I usually fidget with isn't around. it's Get the boy a Rubik's Cube. it's It's funny, I've been so I've been reading the Horace Harrison novels from Warhammer, well, 40k but sent in the 31st century for the Horace Harrison. And kind of as I've been researching for this podcast episode and and thinking about AI models,
01:39:59
Speaker
Like, you can easily imagine a world like the Horus Heresy novels where, you know, various legions come together in certain engagements and they form these strong bonds between battles in brother, like, brothers in battle, rather. I've had a couple of Bali ones now. um And they have this this really strong connection cuz i survived something that they probably shouldn't survive that was against their odds. And then they they fall like they split apart and they don't do their own thing for a hundred years and come back together and they still have. Those bonds and those interactions and i think those are the sorts of things that you could easily bake in even to a first person shooter to have an NPC say hey i was actually with you in the trenches.
01:40:42
Speaker
in Normandy when we landed at Omaha Beach or whatever, yeah just to have a random NPC interact with you and know contextually, and I was there when you mowed down, you threw the grenade into the pillbox and killed all those Germans, as opposed to the next player who would never have that interaction because that's not what they did. Yeah, that's where immersiveness really you know takes a ah new turn or or and it takes advantage of this technology. I think that's some really exciting shit that's going to come out of this. If we can successfully skirt around some of the
01:41:16
Speaker
issues that we've been talking about very lightly. Well I think it's also like yeah just in going further on what you're saying there of like you know you know a certain player character does this this act in this mission compared to someone else but like even thinking certain games that I played recently where it's like I just favored one type of weapon and they're like cool we're gonna throw a bunch of enemies at you now that are resistant to that type of weapon. Like, you know, if I'm, ah if I'm using a long rifle, um for whatever reason, it's just like, you know what, this is actually better, you're better suited now to, you know, use a short range, like a pistol with a higher caliber, or something like that. So just like pierces that armor kind of thing. But just just just to change it up, so make sure that I'm using the full
01:42:05
Speaker
the full scope of things that are created for that game. I think that's a super interesting way you could take it in terms of using that AI. Yeah. Yeah, completely agree. Yeah. I've got lots of opinions, but I can't formulate any of them at the moment there. But I think mostly what we're talking about here is, I mean, obviously Pete, you just said the the big thing is around RPGs. I completely agree with that. be Everything you guys have all mentioned other than sort of the weapons just there is mostly focused around NPCs. rather than, um, behavior creation PCs and yeah. No, no, we sort of world creation. I mean, I know we've got the, we've currently got a procedural generation. We've currently got the the things that happen within no man's sky. I look at like, um, return to Moria where there's, it's a, it's a seeded world. So once you load into that world that the seed is created, uh, it's procedural when it creates that seed, but.
01:43:04
Speaker
Further on, it's a lot of just reused assets. It's not, it's not you get into, I know it's ah it's a mine, so you could limit it on the the biomes that they're they're doing in this um survival game. but It's literally reused rooms in the third zone. It might be a little more vertical, but you're you're walking through two or three rooms that are exactly the same and structured the same as what they are in the first, in the first biome. Yeah. I think that's.
01:43:38
Speaker
a good area where people, if they, smaller teams could be creatively using AI to make it feel like a bigger teamed game than what it actually is. If yeah if if you fed a machine learning algorithm a thousand drawings of a lantern, it could come up with a million drawings of a lantern. and Exactly. so right yeah But actually, counterpoint at the same time, if you're going into the third room of Amoria Mine, it would look the same. No, but- In the universe. Yeah, but it would be variations of a theme. It may not necessarily be identical, although unless there's a reason for it to be identical.
01:44:21
Speaker
I'm not looked talking about the um the aesthetic of the things that are in there. Yeah, cool. They're therere they're a culture. Everything that they have is going to look similar. I'm talking about they are exact mirror images like of rooms. And you can walk in a loop of three rooms, mine through a wall, and it is i get the exact room as you go along. So and what you're talking about is is art asset generation through machine learning where I can iterate on themes and I can understand the context of those themes. So I think that's a really cool place to explore as well. here here is your Here is your concept artist designed concepts. It needs to stay within these parameters.
01:45:08
Speaker
make every biome that the or every level of the mind that the player goes to within these aesthetic parameters, make every single thing that the player is going to interact with within a room unique. So every room they go through unique. That's a good example of reinforcement learning where you can set a, you can give it data as an input, let it iterate and come up with its own. And then you can give it feedback to say, well, no, that's deviating too far from the aesthetic. That's deviating too from the far from the aesthetic, but all of this stuff is good. And it will then iterate on that as as a concept as well. And it will learn the borders of what you're trying to drive towards pretty quickly and still be able to come up with stuff that's slightly unique each time.
01:45:55
Speaker
And so those those types of algorithms already exist in the cloud. What I think is going to be interesting and what i what I predict you're likely to see over or where likely to see over the next 12 to 18 months because this is a field that is moving very fucking quickly, very quickly. But I think over the next 12 to 18 months, you're going to see indie developers and smaller studios from large game devs experiment with this stuff and there's going to be some flops and there's going to be some standout moments we go it's a shit game but it's got this really cool idea that they've baked in and they will play with the technology because this tech is really only been out for 18 months.
01:46:36
Speaker
The large language model i was gonna say is that is that a long enough period to start see that in game development like i think you well i think you will from the india developers who tend to move quickly and there's lots of really small teams who get to fuck with this stuff and to be honest about that some of the big developers that release annual annual releases well this is what i think is going to be interesting. Because machine learning will get to a point where it will probably be the great equalizer between those smaller studios who don't have billions of dollars to pump in, but they've got an algorithm that can do it.
01:47:09
Speaker
versus the big studios who pay thousands of developers to manually sit there and see some of those really good stories come through, like, yeah. And then those big dev houses will steal those ideas from the developers. It could also drive the, I mean, looking at a negative that could lead to a
AI and Industry Evolution
01:47:30
Speaker
positive. So if those bigger houses implement more of these AIs and downsize their teams, you'd you'd want to hope those teams now go, well, I'm no longer tied to this, making a call of duty every year, making an Assassin's Creed every year. I'm taking what I've learned there. I'm going to also take the, what I've learned with the AI to my own studio. And we're going to make a fantastic game that is going to rival
01:47:56
Speaker
That because of what we've learned in the in the i help throw out the undoubtedly the big guys are downsizing the downsizing and when this when this one small there obviously gonna downsize even more but. like I guess look at the sort of aligning there. I might get put these other people out into the world if they can get- It goes back to what Pete's saying is the AI is creating stuff, but it's not creative. it's Correct. not yeah Yeah. You're always going to need a human to feed the initial ideas and set the parameters and and define what's fun and what the background lore and story is. And I think
01:48:35
Speaker
You know, when we develop artificial general intelligence where it can actually truly be creative, we're not going to be worried about video games. We're going to be dodging nukes anyway. So and between then and now, ah there's still a place for, you know, the creatives and the art directors and and the artists themselves to define the parameters of a world and feed that as an input into these systems. I just think these systems facilitate like a magnifi like ah a magnifier on it on output. They're a multiplier on output. So rather than drawing 10 assets a day, I can draw 10 assets, feed it into a system and get 100 assets on day two. That that allows an indie developer a team of one or three or 10,
01:49:20
Speaker
To produce the same output as a massive studio would have the traditional way and it's actually the big studios I suspect will have to adjust more than the small studios here to these tech. like yeah technology yeah Also, like the I play plenty of big studio games where you literally fight the same NPC character. We're 100% going to be the beneficiaries of this technology. There's no question about it. Yeah, exactly. We've talked about that. We've joked about it before. like I've killed Joe 18 times in this one mission.
01:49:53
Speaker
like oh Just every time you fucking exit from a flu in Hogwarts legacy, you had the same three lines of dialogue. Like yeah it drove us nuts. and it And it stood out as poor game design because it was so repetitive and feeding that into an LLM and just having it spit out a different variation of words in that voice or just knowing you just sometimes just need to shut the fuck up. So the balance is going to be we are depriving people of their jobs because you're not going to need somebody to voice that line of dialogue anymore. Yeah. But at the same time, we as gamers benefit from not having to listen to that one line of dialogue prerecorded a million times over. And so I think the game industry is going to have to adjust a little bit there. and yeah I think it goes back to what you were saying before about that will then
01:50:43
Speaker
quite quite possibly inspire people to start their own studios and and and smaller studios and stuff. one hundred percent and look Because you have the tech that to build you up that you didn't have before, um which you don't have to pay like people for. Yeah. So I've been in the the technology industry my whole career, so 24 odd years, there's been maybe a dozen, like maybe 12 waves of technology where staff in various industries got shit scared they were going to lose their jobs as a result of the next big wave of technology. We had virtual machines, replaced hardware, all those old hardware engineers were going to be made redundant. You've got CIOs who are scared of losing their job because everything's moved to an outsourcer. And even go back to nobel network, right? That that technology did die. In almost no
01:51:33
Speaker
In almost none of those circumstances have ah have people actually just not been able to get work. it's It's very rare that a technology displaces people or it all entirely replaces people. It requires people to kind of take a sidestep and and yeah adjust and adapt to a changing technology. But I'm no longer in that kind of panic button mode of, holy shit, this is going to replace 50 people doing voice acting. It's just going to change what they do. exactly
01:52:05
Speaker
Well, instead of, instead of Joe Blow coming in and recording 50 lines in a day of the same monotonous shit, he might need to do a half day. Yeah, exactly. yeah He just, he has to record it, he has to record it like eight ways and then they can manipulate it. But the challenge that we will face is what the SAG after was, um, ah um, what do you call it? Um, protesting against and that is to take 10 lines of dialogue and be able to replace the actor for the rest of their lives is problematic. Absolutely. yeah Being able to visually record them from in a 3D recording studio to be able to take their likeness and even need to employ that actor again is problematic.
01:52:51
Speaker
I think the perfect example is our intro you did, like, that's not Arnie 20 years ago. No. Like, that's that's that's an older Arnie we were using, you know. Yeah. You know, and then the same thing with Mark Hamill's Joker, like, that voice has slightly changed over the years as he's done it. But it is. Because he's aged and his voice boxes has has aged with him, like. And that's and that's crappy. Well, I shouldn't say crappy, but it is dodgy. freeware that, yeah you know, trawled my computer for fucking data. um a A professional engine that's actually licensed for Hollywood would be able to do that stuff perfectly. And so.
01:53:29
Speaker
You know, the application. I think what I was more saying was like those voices do change over time. 100%. Yeah. But it's, it's not hard to- James L Jones is the perfect example. Like Vader's voice has changed as he's done it over time. But how many times do you have to feed a machine learning algorithm? This is- Well, the problem, the problem actually with James L Jones is he's now licensed his voice to Disney after he's died. But my point is how many times do you have to feed an ML engine? This is, This is what he sounded like when he was 20, 21, 25, 30. And in the the ML Engine will be able to emulate anyone's voice at any age pretty quickly. That's not it not a challenge it's it that would be difficult for it to overcome. Anyway. It was good. Entirely fascinating stuff, Peter. oh Thank you very much. um Thank you. I just wanted to finish with a bit of a public service announcement for for using
01:54:28
Speaker
Algorithms like chat GBT at work, just don't.
AI in Work and Humor
01:54:32
Speaker
like Honestly, honestly we're we're already at a point where AI is being used by bad people and is being battled by AI in antivirus software. And so the the battlefield is already devoid of humans, right? so So it's important to understand that when you play with chat GPT and it's a fun utility to fuck around with and give it data and type in stuff and it can write your essay for uni if you don't get caught. Just be aware that in a corporate world, every time you upload a document,
01:55:05
Speaker
The licensing model between the free version of chat GPT is that they own that data forever. And so it's very easy to very innocently leak corporate sensitive information into the internet through these utilities because it gives you a cool output and you don't really think about that stuff. So even when I was typing in the prompts for our intro, I didn't tell it our names, I didn't tell it the name of the podcast. I gave it a bunch of information and surprising how little information I needed to give it for it to give us a quite a cool intro. And then I tweaked it a little bit and said, actually, I want a bit of, there's no short break at the end. We're not doing kind of talking, then intro sequence, then beers just lead straight into beers. And I want some quippy pop culture references. And it spat out a new result, which was really cool. But at no point did I give it our names. I gave it Fred, Jack and Bill.
01:55:58
Speaker
So just just be wary with ah the fact that every time you talk to one of these large language models, you're actually training it and it remembers that information forever. Even co-pilot can be a bit sketchy and that's a licensed Microsoft product. um Law firms in Australia will not use co-pilot most of the time. They will not feed it contractual details or not feed it client details because they can't control where that data is actually going. So just be a little bit wary about this new technology, particularly since it's emerging and that's where the the risks really sit. People don't know how it can be misused or don't think about it. So just be careful.
01:56:39
Speaker
And again, just remember that AI can't lift a keg for you. So if you need that, call me. No, but Boston Dynamics can probably do that. I'll fight that robot. I will fight it hard. I think you'd lose, man. That shit's scary. Nah, I got him. I got him. I got him on the back burner. I got him. Let's go. Thanks for listening. Thank you. Thank you. bye
01:57:22
Speaker
right deep into artificial intelligence in video games. From NPCs behaving like they've got Batman's detective skills to machine learning crafting worlds as immersive at Tolkien's Middle East, we covered it all. Middle East? What? I can't let that go. Oh, we're definitely doing outtakes in this episode. I thought you actually did that on purpose. No, I'm just drunk. I've driven. driven and driven controlron I've driven seven barley wines into my face. I've now had an imperial stout of 11.3%. I'm quite trashed. G'day gamers.