WEBVTT 00:00.000 --> 00:10.440 Hello everybody, I'm here to present a software project I've been working on called 00:10.440 --> 00:14.600 Nobody Who. 00:14.600 --> 00:16.160 This is a structure of the talk. 00:16.160 --> 00:20.160 First I'll give you a brief overview of what nobody who is, what it will do for you. 00:20.160 --> 00:24.440 Then I'll get into the motivation of why nobody who is a good idea, why we actually 00:24.440 --> 00:27.440 need this, and why we built it, how we built it. 00:27.440 --> 00:31.120 Then I'll get into the details of how it works, and how you can use it, some of the 00:31.120 --> 00:35.000 technology choices we made, and then at the end I'll get into some demos, I'll show you 00:35.000 --> 00:42.080 some games that people built with nobody who, because nobody who, what is it? 00:42.080 --> 00:48.320 It is a plugin for goodo, natively compiled written and rust for running local large language 00:48.320 --> 00:49.720 models. 00:49.720 --> 00:53.800 A common misconception that I hear all of the time when I present to nobody who like this is 00:53.800 --> 00:56.920 people are going like, oh, so it'll write the GD script code for you and make the game. 00:57.000 --> 00:58.440 This is not the point. 00:58.440 --> 01:04.040 The point of nobody who is to run the large language models at runtime to provide an enriched 01:04.040 --> 01:07.240 role-playing interactive fiction experience. 01:07.240 --> 01:10.800 It's sort of the core use case, there's lots of different applications of it, but the core 01:10.800 --> 01:16.600 use case is to have richer NPCs, it's 2025, that's not to multiple choice dialogue 01:16.600 --> 01:20.160 in RPGs anymore. 01:20.160 --> 01:24.200 This is sort of what the stack looks like in the middle there, just under my cursor, 01:24.200 --> 01:30.200 so nobody who logo, we depend on LMS CPT, if you work with local LLMs, you're probably 01:30.200 --> 01:35.440 familiar with this library, it's really cool, it goes really fast, and it has uses an open 01:35.440 --> 01:40.760 specification for the language models, you can use all sorts of open weights models, not 01:40.760 --> 01:42.800 just meta-slama. 01:42.800 --> 01:48.120 We do hardware acceleration with Vulcan, so it'll work on your GPU, no matter if it's an 01:48.120 --> 01:52.280 Nvidia or indeed or an IGPU, and if you don't have a GPU it'll go as fast as a camera 01:52.280 --> 01:59.160 on the CPU, which is, in my experience, fast enough, it all works in good though, and 01:59.160 --> 02:07.480 we sort of ship these pre-compiled binaries that are linked into the good though game, and 02:07.480 --> 02:10.040 it's just one big fat binary. 02:10.040 --> 02:17.560 We ship for Linux and Mac and Windows, both on Armin, X86, and we have sort of an experimental 02:17.560 --> 02:21.800 Android build, it's kind of fun. 02:21.800 --> 02:22.800 So why? 02:22.800 --> 02:25.720 Now I'll try and convince you that this is actually a good idea, this is something 02:25.720 --> 02:32.320 that games developers could make their games better with. 02:32.320 --> 02:38.240 First of all, why would you want to use language models in games at all? 02:38.240 --> 02:42.880 And I think my core argument here is that in real-playing games the state of the art is sort 02:42.880 --> 02:50.920 of having these multiple choice dialog trees, and I think they kind of suck. 02:50.920 --> 02:54.540 If I'm being generous, I would say it'd be very useful because you can force a character 02:54.540 --> 02:59.680 into taking specific decisions, but that also takes out a lot of the fun of role play. 02:59.680 --> 03:03.680 A comparison is you can sort of think of if you're playing Baldur's Gate 3, which has 03:03.680 --> 03:09.160 D&D rules versus playing actual D&D, you can be much more creative in actual D&D, because 03:09.160 --> 03:11.800 you could sort of bring more of your self into it, you could bring more role playing 03:11.800 --> 03:12.800 into it. 03:12.800 --> 03:14.040 It's a very, very common experience. 03:14.040 --> 03:18.280 If you've ever played one of these RPGs with multiple choice dialogs, that you reach 03:18.280 --> 03:22.280 a point where you're like, oh, but none of these three options are actually what I want 03:22.280 --> 03:24.280 my character to do. 03:24.280 --> 03:30.560 But if you have free text input or voice, if you want that, then you can sort of say whatever 03:30.560 --> 03:35.000 you want, and the characters in the game will improvise along with you and sort of make things 03:35.000 --> 03:37.200 work. 03:37.200 --> 03:41.760 This is another case if you've ever played Skyrim, you are familiar with this dialogue option, 03:41.760 --> 03:44.520 like it sort of gets very, very repetitive. 03:44.520 --> 03:49.680 What we want is emergent gameplay, so you can role play in a universe for a long time, 03:49.680 --> 03:54.440 and with pre-written dialogue, it takes a lot of time from games writers, and even when 03:54.440 --> 03:59.520 it takes long time from games writers, with AAA projects like this, it still ends up 03:59.520 --> 04:03.440 being sort of repetitive, and it doesn't feel that that immerses, and we want more 04:03.440 --> 04:06.600 immersion in our games. 04:06.600 --> 04:10.440 There's sort of the other thing around, I don't think we should be ever be applying 04:10.440 --> 04:17.880 a lens just for the case of playing a lens, but games are also one of the best places 04:17.880 --> 04:22.200 to apply a lens. 04:22.200 --> 04:27.080 If you're thinking about harmful uses of a lens, it's sort of these things where the 04:27.080 --> 04:30.120 risk of the LLM hallucinating can be very, very bad. 04:30.120 --> 04:34.920 If you're looking at like parola provost for correctional facilities or screening job interviews 04:34.960 --> 04:43.160 have you never want hallucinations, but interactive fiction hallucinations are not that bad. 04:43.160 --> 04:50.600 I think there's actually a case to be made that you a lot of the time do want hallucinations. 04:50.600 --> 04:54.400 I'm thinking of like, if I walk up to a character, the standing on the court of the street, 04:54.400 --> 04:57.400 and then I say, what did you have for breakfast, right? 04:57.400 --> 05:02.240 The character didn't have breakfast, and probably was standing on the court of the street 05:02.320 --> 05:09.280 the whole time, but a sort of willing improvisational partner will say, oh, I should 05:09.280 --> 05:13.040 yes, and then I should go over breakfast, I had big kinetics, right? 05:13.040 --> 05:17.120 So we actually, actually, a lot of the time do want it to just go along with whatever 05:17.120 --> 05:20.320 you're presenting. 05:20.320 --> 05:24.400 Another case for why games are a good application of the lens is that machines that 05:24.400 --> 05:29.760 are built for running video games have graphics cards. 05:29.760 --> 05:33.360 This is exactly the kind of hardware you want if you want to run local elements in the 05:33.360 --> 05:34.880 fast. 05:34.880 --> 05:39.040 So when looking at like, oh, what's the compromise of local elements or the machines 05:39.040 --> 05:39.840 good enough? 05:39.840 --> 05:44.320 The machines that are good enough for the machines that are built for running video games. 05:44.320 --> 05:48.480 So why do we want to do it local? 05:48.480 --> 05:52.400 A big one is it fits really well with the current market of game development. 05:52.400 --> 05:56.400 If you're sort of looking at how do games developers make money, what they do is they put 05:56.400 --> 05:59.680 their game on Steam, you pay 10 euros, you get the game, and you as a consumer sort of 05:59.760 --> 06:03.200 expect for this one-time payment to keep playing the game. 06:03.200 --> 06:07.440 If you want to be able to play the game in a year or in 10 years, and if you're relying on 06:07.440 --> 06:14.160 sort of the easy thing to do, which is to use like GBT4 and the OpenAI, it's to PAPI, 06:14.160 --> 06:17.920 it's very quickly get something going, but you'll end up being paying for usage, 06:17.920 --> 06:19.920 and this doesn't work that well. 06:19.920 --> 06:24.240 The games that do this, they typically do it by keeping the product sales business model, 06:24.240 --> 06:26.720 but then imposing a usage limit. 06:26.720 --> 06:32.000 And then you sort of, the game just shuts down after you're playing it for five hours or something like this, 06:32.000 --> 06:36.960 because the game developers can only afford processing this many tokens for one sale. 06:36.960 --> 06:40.480 And that sucks. I want to be able to buy a game and then be able to keep running it. 06:41.520 --> 06:48.080 There's a privacy aspect. If I'm a role-playing, I don't always want somebody to pick in, right? 06:48.080 --> 06:52.320 This is a thing that's familiar to a pen and paper, a bunch of dragons players, right? 06:52.320 --> 06:58.080 You're very, there's a lot of trust building in order to be able to improvise with other people. 06:58.080 --> 07:01.920 So it can be uncomfortable this idea of other people peaking on your conversations. 07:02.720 --> 07:06.960 Offline capabilities, just a great feature. When I go on the airplane home from Boston, 07:06.960 --> 07:09.600 I want to be able to take out my steam deck and play a few games. 07:09.600 --> 07:13.200 I can't play the games that require an online API. 07:13.200 --> 07:18.400 So this is a cool new feature that you only can get with the alarms if you're doing local alarms. 07:18.880 --> 07:24.720 There's reliability. If you're depending on an external API to run your game, 07:24.720 --> 07:28.560 then that API can change. It's sort of a bunch of software that's out of your control. 07:29.360 --> 07:33.120 If you look at the history of OpenAI's or OpenAI's APIs, 07:34.480 --> 07:37.200 they've sort of changed how the models respond over time. 07:37.200 --> 07:40.400 So there's something that the model might want to play along with one week. 07:40.400 --> 07:44.880 Could in the next week it could just go like, oh, as an AI assistant, I cannot. 07:45.840 --> 07:51.120 And you don't want this. You want things to keep working in the same way as they did when you shipped your game. 07:52.800 --> 07:56.800 This one, I think, is really the big and important one. 07:57.520 --> 08:02.320 Games are art, games are culture. We want to be able to preserve this for future generations. 08:03.200 --> 08:06.320 Games are a cable. There's people working with games are a cable professionally, 08:06.320 --> 08:10.320 and they have a really, really hard time making things running or keeping things running. 08:11.200 --> 08:14.240 Players Rocket League will notice that they've dedicated Linux support a few years ago, 08:14.240 --> 08:17.920 and then like, oh, this game that I bought, it doesn't even run on my computer anymore. 08:17.920 --> 08:22.160 Well, it will, but I can't play with my friends because it requires a server for online play. 08:22.720 --> 08:27.440 So that kind of sucks. But with local elements, we can sort of make the type of game that I can 08:27.440 --> 08:31.040 put on my laptop with my laptop and a vacuum chamber and pull it out a hundred years later, 08:31.040 --> 08:34.080 and it'll still work. So I think that's cool. 08:35.600 --> 08:41.280 Okay, so how are we building this? I'll get back to this slide that we saw before. 08:42.000 --> 08:48.560 This is Lama CBP with Vulcan Acceleration, so we get good hardware support and runs fast. 08:50.480 --> 08:54.400 You'll sort of notice, if you know Codell, well, you'll notice that's a C++ project, 08:54.400 --> 08:59.200 and Lama CBP is a C++ project from home, but then we wedged some rust in there. 08:59.920 --> 09:06.640 That sort of gives us a lot more confidence in writing correct code and it runs fast like C++ does. 09:07.680 --> 09:10.720 Another thing that sort of sticks us out a little bit in the ecosystem is that 09:11.280 --> 09:14.400 looking upwards in the stack and downwards in the stack, and everything is. 09:14.400 --> 09:18.400 It might see lessons. We cover left license, nobody who. 09:20.320 --> 09:22.880 For reasons I don't have to explain to this obvious. 09:24.560 --> 09:28.560 I think that we notice a lot of the time with new people picking up this project is that they 09:29.760 --> 09:32.960 are going like, oh, copy left license, that's scary because I want to make a proprietary game. 09:34.000 --> 09:38.480 So we have to be sort of very explicit in the read me saying like, you can link into this 09:38.480 --> 09:42.880 copy left dependency, you just can't make proprietary forps. It's not something that the 09:42.880 --> 09:44.960 game dev communities totally privy to. 09:47.280 --> 09:52.880 Yeah, so because of we're using Lama CBP, Lama CBP exposes or increments an open 09:52.880 --> 09:57.280 specification for the model files, which means that we can use lots of different models. 09:57.280 --> 10:01.760 Whenever a new fancy model comes out very quickly, there's a GDUF, which is a file format, 10:03.520 --> 10:07.600 a quantization available for it, and you can sort of just pull these down from the internet. 10:07.680 --> 10:12.000 So we don't ship a specific model. We're expected developers to sort of figure out what model 10:12.000 --> 10:17.280 suits their needs. So if you want to just run the state-of-the-art game out or deep-seek models, 10:17.280 --> 10:21.680 that's fine. If you want to run something that's trained on a specific data set or doesn't come 10:21.680 --> 10:24.960 out of big tech and Silicon Valley for some reason, then you can go and use the 10:24.960 --> 10:29.120 sell-and-mental models from Barcelona University or the Luther AI stuff for whatever you want. 10:29.120 --> 10:33.360 So we don't have a strong opinion on what the open weights model should look like and sort of just 10:33.360 --> 10:38.640 use any model that's in this format or fine-tune your own thing if you want to. Let's get into 10:38.640 --> 10:45.280 the code. We expose them very, very simple API and look sort of like this. I'll just go through 10:45.280 --> 10:51.280 all of it. We first configure, like set the model, this is the file of GDUF, and you select 10:51.280 --> 10:55.200 them from the last slide. There's a prompt, says you're an evil wizard, you always curse whoever, 10:56.160 --> 11:01.120 then you call say, which is the user saying something to the character, and then it'll sort of 11:01.120 --> 11:04.640 in that thread in the background generator response, and if you're familiar with GDUF idioms, 11:04.640 --> 11:09.360 you'll know about signals, and which is sort of like a pub sub-tech thing you can 11:10.080 --> 11:15.200 await response. You can also wait for the tokens as they come out when at a time, which makes 11:15.200 --> 11:21.920 a much better experience. We'll see this in the next couple of demos. We also support embeddings 11:21.920 --> 11:28.000 by the bar. I don't only have three minutes, so I'll talk too much about this. We support all 11:28.080 --> 11:32.080 of the sample methods that allow us to be personalized, so you can do what you need to in terms 11:32.080 --> 11:38.480 of controlling the model. If you ever want to penalize some tokens, compared to others, if you 11:38.480 --> 11:45.280 want to greedy or a mere step-sign for them in P, whatever we can do all of this. This slide is too 11:45.280 --> 11:56.080 early. Let's see some games. This game was made as a demo by a friend of mine. I like to think 11:56.160 --> 12:01.200 of it. It's sort of like papers, please, in reverse. You're suspect of a crime. You're talking 12:01.200 --> 12:06.720 to an interrogator who is sensing you with evidence, and you sort of have to provide convincing 12:06.720 --> 12:11.280 explanations and elevations for each piece of evidence that they're presenting with. This one is 12:11.280 --> 12:15.920 using two different model contexts, one that's sort of prompted to in a structured way using 12:15.920 --> 12:22.160 JSON, assess the viability of your, of your, of your, of your, elevating your explanation. And then 12:22.240 --> 12:26.560 there's another model context using the same model loaded into VRAM, which is doing the role 12:26.560 --> 12:33.040 play of the interrogator. And you can see there's sort of a credibility meter down there. I think 12:33.040 --> 12:44.000 that's really fun. Next one. This one is not a full game. It's a small demo I made to show that 12:44.000 --> 12:49.600 you can have the element direct back to your game code. This is like a potion shopkeeper. I named 12:49.680 --> 12:56.320 her a fubar. You can sort of ask her what potions tell for sale. What about that orange one? What 12:56.320 --> 13:00.480 about the blue one? What does it do? I need a strength potion. What would you recommend? And then 13:00.480 --> 13:08.720 you can also ask her like, okay, sure, let me get the strength potion. And it'll give you 13:08.720 --> 13:12.320 like a pop-up. And this is sort of showing that you can use this to trigger your actual game code. 13:12.320 --> 13:17.360 And then the strength potion disappears from the inventory. You can also do stuff like ask her like, 13:17.440 --> 13:20.640 what do they taste like? Or what do you have for breakfast? And she'll sort of play along. 13:24.640 --> 13:30.640 But this next one was something that somebody made for global game jam last week. 13:32.320 --> 13:36.960 And it's another detective game. This one you play sort of as the detective, you arrive at a 13:36.960 --> 13:40.800 space station that's binding helium 3. And there was an incident recently you have to figure out 13:40.800 --> 13:45.600 what happened. And there's some undertones that like probably the corporate overloads for running the 13:45.680 --> 13:50.320 station are trying to cover stuff up. And there's a bunch of different characters and they all 13:50.320 --> 13:55.680 have different tones of voice and like they have secrets that you can sort of transcess out. 13:56.880 --> 14:02.960 And that's kind of fun. And you can tell all of these demos by the way are running on a laptop, 14:02.960 --> 14:09.440 right? So, and so it doesn't require like insane the gamer rigs or something. This is just a 14:09.440 --> 14:18.320 framework 16. And the tokens come up reasonably fast. Who, one minute? It's good because I think 14:18.320 --> 14:26.000 of one more demo. This one is sort of a dungeon crawler. I really like the, well I really like the 14:26.000 --> 14:31.760 aesthetics of this game, but I really like the ambitions of this guy. Because there are some projects 14:31.760 --> 14:37.200 that are like LLM power dungeon crawlers. But a lot of them work like collaborative storytelling 14:37.280 --> 14:40.400 exercises. You can sort of go with them. Then I pull out my two double flame swords. And to 14:40.400 --> 14:44.960 capitate tenogers in the game, we're just going to make sure you do another old debt. This one's 14:44.960 --> 14:50.160 a bit more intelligent and works like an actual game. Like the power of your sword will assess the 14:50.160 --> 14:54.400 probability if you actually feeling a certain number of damage. And you could still be creative, 14:54.400 --> 14:59.040 like a pan in the, in the indecession and saying, okay, I use my shield to bash the barrel and make 14:59.040 --> 15:04.160 it find the direction of the skeleton. And then it like SSO is absurd, false snow. This is actually 15:04.240 --> 15:08.480 a barrel in the room. So it actually works like a real game, but you get to bring a lot of 15:08.480 --> 15:17.440 your own creativity into it. That's it, I guess. Yes, you're out of time. Thank you, links. 15:22.560 --> 15:24.400 Is Daniel around on next speaker?