1 00:00:00,000 --> 00:00:06,640 So let's talk about James, dancing Greek statues, had one and a half hour talk with him. 2 00:00:06,640 --> 00:00:09,600 Now he's a master of control nuts inside of stable diffusion. 3 00:00:09,600 --> 00:00:30,800 And Elon Musk replied, James is like we had to reply immediately back. 4 00:00:30,800 --> 00:00:34,160 Ladies and gentlemen, boys and girls, children of all ages, dogs, cats, 5 00:00:34,160 --> 00:00:40,320 robots and everybody in between, especially you first time viewers on YouTube. 6 00:00:40,320 --> 00:00:42,000 That's right. We got a live podcast. 7 00:00:42,000 --> 00:00:46,400 Everybody welcome to HTTTA how to talk to AI. 8 00:00:46,400 --> 00:00:53,920 I am your host in the flesh, the synth buying West and it's all I am joined by the gush, 9 00:00:53,920 --> 00:01:00,800 the gracious, the globe trotting, the galvanizing force of nature herself. 10 00:01:00,800 --> 00:01:04,400 Never ghastly, always graceful, the miss go to go herself. 11 00:01:04,400 --> 00:01:05,440 Gee, how are you? 12 00:01:05,440 --> 00:01:07,360 I'm great. Thank you so much. 13 00:01:07,360 --> 00:01:10,800 You know, I'm always giggling with a huge smile. 14 00:01:10,800 --> 00:01:14,320 You're amazing. Your introductions are fantastic. 15 00:01:14,320 --> 00:01:16,480 I'm excited about this video. 16 00:01:16,480 --> 00:01:18,080 Let's see how that goes. 17 00:01:18,080 --> 00:01:19,600 That's definitely something new. 18 00:01:20,080 --> 00:01:20,400 Yes. 19 00:01:20,400 --> 00:01:22,400 And not entirely new. 20 00:01:22,400 --> 00:01:24,960 You have a little more experience in this realm than me. 21 00:01:24,960 --> 00:01:26,960 Yeah, but it doesn't get easier. 22 00:01:26,960 --> 00:01:28,320 So right to that. 23 00:01:28,320 --> 00:01:29,920 Hey, let's jump right in. 24 00:01:29,920 --> 00:01:38,640 I recall a few months back, you did a video on RunwayML's gen one text to image generation. 25 00:01:38,640 --> 00:01:42,320 Do you want to just fill people in on kind of like your experience using that? 26 00:01:42,320 --> 00:01:45,120 Because we have some news that gen two is out now. 27 00:01:45,120 --> 00:01:47,760 Right. So I tested gen one. 28 00:01:47,760 --> 00:01:52,800 I don't know if we talked about it, but my experience was that it was definitely not 29 00:01:52,800 --> 00:01:57,120 easy with gen one, but it was also beta version in this core. 30 00:01:57,120 --> 00:02:03,440 So to get something consistent, like suddenly, for example, you're prompting something like 31 00:02:03,440 --> 00:02:07,200 a character and suddenly I would get a man with a beard. 32 00:02:07,200 --> 00:02:10,560 And I was like, is it trying to tell me something? 33 00:02:10,560 --> 00:02:18,080 But that being said, gen two came out and that's definitely a little bit different right now. 34 00:02:18,080 --> 00:02:20,080 And everyone is now snapping fingers. 35 00:02:20,080 --> 00:02:21,440 It's so funny. 36 00:02:21,440 --> 00:02:22,240 It is pretty funny. 37 00:02:22,240 --> 00:02:27,920 Runway is offering on their website, a little tutorial video on how to use gen one. 38 00:02:27,920 --> 00:02:31,600 And I think some of the flow, some of the content is even more compelling. 39 00:02:31,600 --> 00:02:35,440 And you see what's the difference because now I think on the paid version, because I 40 00:02:35,440 --> 00:02:41,600 tried free, I'm going to check out paid one too, but there's like sliders, the important 41 00:02:41,600 --> 00:02:45,680 part, this like connectivity with the footage that you prompt. 42 00:02:45,680 --> 00:02:50,160 If you see all these videos, they are all based on original footage. 43 00:02:50,160 --> 00:02:56,000 So you upload a video and then you have a slider on how much of that video it takes 44 00:02:56,000 --> 00:02:57,360 as a reference point. 45 00:02:57,360 --> 00:03:02,480 This is like in a stable diffusion, you know, when you control opacity of how much image 46 00:03:02,480 --> 00:03:06,000 is influencing your final image versus text prompt. 47 00:03:06,000 --> 00:03:06,400 Okay. 48 00:03:06,400 --> 00:03:13,120 I think what is driving this type of results, but you can actually see faces and body 49 00:03:13,120 --> 00:03:19,280 structures or even fingers more accurate as if you just prompted, you have to have like 50 00:03:19,280 --> 00:03:20,720 a video as a reference. 51 00:03:20,720 --> 00:03:26,880 So image to image, video to video, something to kind of give the model. 52 00:03:26,880 --> 00:03:27,360 Yeah. 53 00:03:27,360 --> 00:03:28,320 Some, some lifting. 54 00:03:28,320 --> 00:03:28,960 And you know what? 55 00:03:28,960 --> 00:03:34,880 I will say in prior image generations where I'm looking for very specific details, 56 00:03:34,880 --> 00:03:39,920 someone's holding something a specific way, the lighting conditions need to be such a 57 00:03:39,920 --> 00:03:44,320 way I know in mid journey stable diffusion, you could do an image sample in addition to 58 00:03:44,320 --> 00:03:45,120 your prompt. 59 00:03:45,120 --> 00:03:49,440 And oftentimes the model just goes, okay, I got it. 60 00:03:49,440 --> 00:03:53,040 That, that helps me, you know, not try to have to figure everything out. 61 00:03:53,040 --> 00:03:59,360 You know, it's interesting because with AI video, it's kind of this two paths of the 62 00:03:59,360 --> 00:04:04,800 ones which are using video as a reference of you and kind of overlaying whatever you 63 00:04:04,800 --> 00:04:05,440 want. 64 00:04:05,440 --> 00:04:12,800 And another one is this AI generated avatars where you take the image or the model trained 65 00:04:12,800 --> 00:04:16,960 on your realistic images, and then you feed the text. 66 00:04:16,960 --> 00:04:20,640 It's two different directions of the control of the output. 67 00:04:20,640 --> 00:04:25,120 I mean, it's, it seems like at some point you're going to have to like go, all right, 68 00:04:25,120 --> 00:04:32,000 I have to lean on the model to generate something compelling, but I also as a director, as 69 00:04:32,000 --> 00:04:38,160 producer, as the person creating this, I have a specific vision and I do need some deterministic 70 00:04:38,160 --> 00:04:40,720 lever in the results to achieve that. 71 00:04:40,720 --> 00:04:45,840 I wanted to show that you was kind of back to reality experience. 72 00:04:45,840 --> 00:04:48,960 So let's talk about James, Dancing Greek Statues. 73 00:04:49,840 --> 00:04:52,480 Had one and a half hour talk with him. 74 00:04:52,480 --> 00:04:54,880 He's amazing artist and we were just geeking out. 75 00:04:54,880 --> 00:04:57,120 He was like, I didn't know that you're painting. 76 00:04:57,120 --> 00:04:58,400 He went on my Instagram. 77 00:04:58,400 --> 00:05:02,800 Check out last two weeks newsletters for a little bit more James's work from episode 78 00:05:02,800 --> 00:05:06,800 nine and, and some of Goda's art that we just talked about last week. 79 00:05:06,800 --> 00:05:13,440 Right, so, so we kind of like bonded in this artistic style and he's just, yeah, brilliant 80 00:05:13,440 --> 00:05:18,000 and it's so amazing to hear, you know, that his work took off like that. 81 00:05:18,560 --> 00:05:21,840 But then I was showing his work to everybody. 82 00:05:21,840 --> 00:05:24,320 This is just how I was just like, you have to see this. 83 00:05:24,320 --> 00:05:29,680 And we showed it to our family, the parents, parents-in-law, we were like, 84 00:05:29,680 --> 00:05:32,640 we watched it and they're like, okay, it's nice animation. 85 00:05:32,640 --> 00:05:36,320 And I was like, you don't understand. 86 00:05:36,320 --> 00:05:38,320 This is not animation. 87 00:05:38,320 --> 00:05:40,320 Just Elon Musk replied. 88 00:05:41,120 --> 00:05:47,200 He was like, we had to reply immediately back, but we don't know if Elon Musk saw that, 89 00:05:47,920 --> 00:05:51,840 you know, his version because we did with Elon Musk face on it. 90 00:05:51,840 --> 00:05:52,240 Yep. 91 00:05:52,240 --> 00:05:54,000 You know, that's some good coverage. 92 00:05:54,000 --> 00:05:56,800 Unlike our little video, this is our, our pal James. 93 00:05:56,800 --> 00:05:58,240 We talked about it for a couple of weeks now. 94 00:05:58,240 --> 00:06:01,280 He's a master of control nuts inside of stable diffusion and 95 00:06:01,280 --> 00:06:04,960 just so many possibilities, such, such fun stuff there. 96 00:06:04,960 --> 00:06:09,840 And as we say often on here, this, these kinds of things, they're only going to get better. 97 00:06:09,840 --> 00:06:11,680 This is the worst they'll ever be. 98 00:06:11,680 --> 00:06:13,120 And they're already compelling. 99 00:06:13,120 --> 00:06:16,880 His stuff has been viewed in the last two, three weeks, over 200 million times. 100 00:06:16,880 --> 00:06:21,680 But going back to this kind of, you know, stepping outside of the AI bubble, 101 00:06:21,680 --> 00:06:25,440 it is interesting to me how the outside world perceives these things, 102 00:06:25,440 --> 00:06:29,120 because when at the end of the day, there is going to be value attached to that. 103 00:06:29,120 --> 00:06:29,360 Right. 104 00:06:29,360 --> 00:06:29,760 Yeah. 105 00:06:29,760 --> 00:06:34,480 So while we are geeking out and we're like, oh my God, like I will just pay you as much as possible. 106 00:06:34,480 --> 00:06:36,240 I want something like that. 107 00:06:36,240 --> 00:06:43,360 The public sees this as we said, oh, nice animation as it's like some sort of just, 108 00:06:43,360 --> 00:06:45,360 I don't know, premiere pro package. 109 00:06:45,360 --> 00:06:46,320 Oh, yeah. 110 00:06:46,320 --> 00:06:47,040 That's some frame. 111 00:06:47,040 --> 00:06:48,720 That's the only frame of reference they have. 112 00:06:48,720 --> 00:06:49,200 I don't know. 113 00:06:49,200 --> 00:06:53,600 For me, that was kind of interesting insight because yeah, at the end of the day, 114 00:06:54,400 --> 00:06:58,640 these things will be generalized and actually made available. 115 00:06:58,640 --> 00:07:02,720 So, you know, not to minimize the quality and value of James work. 116 00:07:02,720 --> 00:07:07,600 But I also, because of that, I saw, I think two apps popping up saying that, 117 00:07:07,600 --> 00:07:10,960 oh, I just uploaded a video and we generate AI video. 118 00:07:10,960 --> 00:07:15,520 And it is just kind of, it is this visual effect applied to it. 119 00:07:15,520 --> 00:07:18,640 I really don't believe that they run the whole thing. 120 00:07:18,640 --> 00:07:23,120 Because when we talk to James, like if people think that this is just like 121 00:07:23,120 --> 00:07:26,560 one click prompt easy, so wrong. 122 00:07:26,560 --> 00:07:27,200 Not right now. 123 00:07:27,200 --> 00:07:35,760 No, it is so much work, so much iteration and it costs a lot of money just to run with things. 124 00:07:35,760 --> 00:07:40,400 No one has a concept of for every frame of those little clips of James, 125 00:07:40,400 --> 00:07:44,800 if it's running at 60 frames a second, which some of those are, because they're pretty smooth, 126 00:07:44,800 --> 00:07:49,760 you know, seven second video, that's hundreds and hundreds of frames that individually need 127 00:07:49,760 --> 00:07:51,600 to be rendered and stitched together. 128 00:07:51,600 --> 00:07:57,520 Unless you have a pretty beefy GPU or paying for some cloud server. 129 00:07:57,520 --> 00:07:59,200 Or both. 130 00:07:59,200 --> 00:08:00,080 Yeah, or both. 131 00:08:00,640 --> 00:08:03,520 That takes an inordinately long amount of time. 132 00:08:03,520 --> 00:08:06,000 So talking about GPUs. 133 00:08:06,000 --> 00:08:07,920 We could definitely talk about GPUs. 134 00:08:07,920 --> 00:08:11,840 The biggest transition to that, my video I did about mid journey, 135 00:08:11,840 --> 00:08:16,000 where founder was talking about this two futures of AI. 136 00:08:16,000 --> 00:08:17,680 Did you, I hope you watched it. 137 00:08:17,680 --> 00:08:18,240 No, don't. 138 00:08:18,240 --> 00:08:20,240 Of course, of course I watched your video. 139 00:08:20,240 --> 00:08:22,400 Yes, of course, I am a natural professional. 140 00:08:23,200 --> 00:08:30,480 That's one of the older ones, but interesting thing he had this mention of GPUs and he was 141 00:08:30,480 --> 00:08:32,320 saying there's two realities. 142 00:08:32,320 --> 00:08:39,200 One is that to scale anything, we are just bound by physics as a first principle, just as a data 143 00:08:39,200 --> 00:08:43,200 centers and chips that we can produce physically. 144 00:08:43,200 --> 00:08:45,520 And it will take us seven years to scale. 145 00:08:45,520 --> 00:08:49,680 And that was a, that is a reason why mid journey is not marketed. 146 00:08:49,680 --> 00:08:54,400 You are not going to see ads for mid journey or like a, hey, there is an app for mid journey 147 00:08:54,400 --> 00:08:58,880 because there's simply no ability to scale to sustain such user base. 148 00:08:58,880 --> 00:09:05,120 But when he talked about different path, if we invent new chips or new abilities or integrate 149 00:09:05,120 --> 00:09:08,480 neural nets and chips, it's out of my domain knowledge. 150 00:09:08,480 --> 00:09:12,720 But, and then he said that if this happens, then we scale in a year. 151 00:09:12,720 --> 00:09:13,120 Yeah. 152 00:09:13,120 --> 00:09:16,720 So think what the craziness what's happening right now. 153 00:09:16,720 --> 00:09:22,240 And then if there is next technological breakthrough in the physical department, the 154 00:09:22,240 --> 00:09:22,800 hardware. 155 00:09:23,680 --> 00:09:26,320 Well, speaking of hardware, let's talk. 156 00:09:26,320 --> 00:09:26,880 Yeah. 157 00:09:26,880 --> 00:09:31,040 Since we're talking chips and it is kind of wild to think about like, you know, that's why GPT-4 158 00:09:31,040 --> 00:09:34,080 is constrained in terms of like how fast it generates. 159 00:09:34,080 --> 00:09:35,920 It's literally the compute. 160 00:09:35,920 --> 00:09:40,960 So what's starting to take place is a lot of these big companies are completely rethinking 161 00:09:40,960 --> 00:09:42,400 how compute happens. 162 00:09:42,960 --> 00:09:46,560 You know, I'm sure people, while they may not know the ins and outs of it, 163 00:09:46,560 --> 00:09:52,240 you can conceptualize that all of our stuff, our websites, the apps be like clouds. 164 00:09:52,240 --> 00:09:54,240 They're running in these big data centers. 165 00:09:54,240 --> 00:09:58,640 These data centers are power hungry and inside of each one of these servers is essentially 166 00:09:58,640 --> 00:10:02,880 a different type of computer that is CPU centric, right? 167 00:10:02,880 --> 00:10:05,360 It has a central processing unit. 168 00:10:05,360 --> 00:10:11,040 So one of the things that gets inefficient about these is they spend as much time kind 169 00:10:11,040 --> 00:10:16,560 of moving around data as they do actually like computing, like doing computations. 170 00:10:16,560 --> 00:10:21,680 So that tends to tax and take longer for processes to take place. 171 00:10:21,680 --> 00:10:26,640 But with the onset of some of these generative AI technologies, 172 00:10:26,640 --> 00:10:32,160 the video who's the world's leading chip maker has just put out, it's totally worth the YouTube 173 00:10:32,160 --> 00:10:32,560 watch. 174 00:10:32,560 --> 00:10:36,880 Even if you don't geek out a bunch of this hardware stuff, it is pretty incredible in 175 00:10:36,880 --> 00:10:38,480 terms of some of the things they demo. 176 00:10:38,480 --> 00:10:44,000 So this is the Brace Hopper AI supercomputer. 177 00:10:44,000 --> 00:10:52,400 It's basically a singular GPU, the size of a school bus, 150 miles of fiber optic cable 178 00:10:52,400 --> 00:10:53,200 inside of it. 179 00:10:53,200 --> 00:10:59,200 But this singular row of servers, granted it's a lot of servers, are GPU based. 180 00:10:59,200 --> 00:11:06,160 So it uses something to the tune of like a 44th of the amount of power that to achieve 181 00:11:06,160 --> 00:11:12,240 the same level of compute as this GPU server would, it would need 45 times more power in 182 00:11:12,240 --> 00:11:14,000 our conventional servers right now. 183 00:11:14,000 --> 00:11:19,200 So you can get way more cost effective bang for your buck, so to speak, train a generative 184 00:11:19,200 --> 00:11:25,120 image model, train a language model in fractions of the time, just because of how efficiently 185 00:11:25,120 --> 00:11:26,080 this can handle. 186 00:11:26,080 --> 00:11:32,320 Like this is a 140 terabytes of GPU compute power that is all running in series. 187 00:11:32,320 --> 00:11:34,080 So it works as one. 188 00:11:34,080 --> 00:11:38,320 These are the kind of things that when all these data center transformations happen, 189 00:11:38,320 --> 00:11:43,200 not only will we be able to use AI like we want to, these are the kind of things that 190 00:11:43,200 --> 00:11:48,800 you need to say, okay, computer, supercomputer, we've got 20 of these chained together, please 191 00:11:48,800 --> 00:11:50,320 cure all diseases for us. 192 00:11:50,320 --> 00:11:56,000 And they just can start processing these huge, inconceivably large problems. 193 00:11:56,000 --> 00:11:57,680 I was listening to you. 194 00:11:57,680 --> 00:12:02,320 This is fascinating, but I was checking stock price of any media. 195 00:12:02,320 --> 00:12:04,240 It definitely shot up. 196 00:12:04,240 --> 00:12:09,360 And what's that part is yesterday, I was like, I need to buy some shares of new media. 197 00:12:09,360 --> 00:12:13,200 And then something happened and this is something because it's on mobile. 198 00:12:13,200 --> 00:12:15,760 So it's almost like, oh yeah, let me know. 199 00:12:15,760 --> 00:12:17,040 Now it sounds bad. 200 00:12:17,040 --> 00:12:21,360 It's not like you are just casually buying shares, but I think the media is the place 201 00:12:21,360 --> 00:12:22,640 where it makes sense. 202 00:12:22,640 --> 00:12:30,880 They already control something like 85% of the GPU chip market, which is dangerous. 203 00:12:30,880 --> 00:12:32,480 It is, it totally is. 204 00:12:32,480 --> 00:12:36,160 And I think there's going to be antitrust stuff coming down the pike eventually that, 205 00:12:36,160 --> 00:12:40,480 you know, Microsoft had to face the same thing when they were, you know, this huge software 206 00:12:40,480 --> 00:12:43,040 company in the late nineties, early two thousands. 207 00:12:43,040 --> 00:12:47,280 Yeah, you're going to have this AI arms race that no doubt is going to emerge between these 208 00:12:47,280 --> 00:12:53,280 big players like Microsoft, like Google, like Meta, but they all still need this compute. 209 00:12:53,280 --> 00:12:59,520 So this huge new GPU central server that the videos invented and put together, it has 210 00:12:59,520 --> 00:13:02,720 something like six of these chips inside of each server blade. 211 00:13:02,720 --> 00:13:07,840 The chips cost $200,000 each and there's over a thousand server blades inside of one of 212 00:13:07,840 --> 00:13:09,600 these exa-flock machines. 213 00:13:09,600 --> 00:13:14,560 So that's not something that anybody other than like fortune 50, fortune 500 companies 214 00:13:14,560 --> 00:13:17,600 can even afford to have access to at this point. 215 00:13:17,600 --> 00:13:22,400 But considering we've already blown Moore's law out of the water with some of the, some 216 00:13:22,400 --> 00:13:27,200 of the level of innovations that are happening, you know, it's only going to get cheaper. 217 00:13:27,200 --> 00:13:33,600 I remember I saw this graph learning from history is always a good way to go and was 218 00:13:33,600 --> 00:13:38,560 looking at what happened with semiconductors and SaaS companies, like, you know, big tech 219 00:13:38,560 --> 00:13:43,760 that historically SaaS companies are the ones who win and not the hardware. 220 00:13:43,760 --> 00:13:44,560 Yeah. 221 00:13:44,560 --> 00:13:50,240 But I think when Nvidia case is interesting because as much as they are hardware controlling 222 00:13:50,240 --> 00:13:55,520 market, they are coming with their larger language models like Neo, right? 223 00:13:55,520 --> 00:14:02,960 So it's because we can really attack from all sides and do all this collaborations with 224 00:14:02,960 --> 00:14:04,000 the companies. 225 00:14:04,640 --> 00:14:11,520 It is even like scary to think when we live in a world where all compute is in one company's 226 00:14:11,520 --> 00:14:12,080 hand. 227 00:14:12,080 --> 00:14:13,120 Totally crazy thing. 228 00:14:13,120 --> 00:14:15,440 But I got two computers right here in front of me. 229 00:14:15,440 --> 00:14:17,600 I've got Nvidia GPUs in it. 230 00:14:17,600 --> 00:14:18,400 Right. 231 00:14:18,400 --> 00:14:20,480 Everything else is different and diversified. 232 00:14:20,480 --> 00:14:22,480 The GPUs are both Nvidia. 233 00:14:22,480 --> 00:14:23,280 That's interesting. 234 00:14:23,280 --> 00:14:24,400 I just got Mac. 235 00:14:24,400 --> 00:14:30,080 So remember we were talking about if to get Nvidia and this one has M2 and I'm very happy 236 00:14:30,080 --> 00:14:30,720 with it. 237 00:14:30,720 --> 00:14:35,600 But yeah, that's another thing like maybe Apple, especially with their, you know, Vision 238 00:14:35,600 --> 00:14:39,680 Pro and how much compute that but that's completely two ball games. 239 00:14:39,680 --> 00:14:43,280 Nvidia is on the whole other racetrack. 240 00:14:43,280 --> 00:14:50,640 And what's always made Mac run so efficiently is since they are developing the silicon and 241 00:14:50,640 --> 00:14:56,000 the software in that case, they are able to develop processes, system architectures that 242 00:14:56,000 --> 00:14:58,880 run super efficiently on their chips. 243 00:14:58,880 --> 00:15:02,080 So they're able to kind of optimize both things. 244 00:15:02,640 --> 00:15:09,120 You know, so in many instances, an M2 chip does some of the same things that a big beefy 245 00:15:09,120 --> 00:15:12,400 GPU could do for machine learning purposes. 246 00:15:12,400 --> 00:15:18,000 But it's only going to enable more creativity, more problems being solved. 247 00:15:18,000 --> 00:15:19,600 It's an exciting thing to think about. 248 00:15:19,600 --> 00:15:22,960 You know, on this kind of mass global power. 249 00:15:23,760 --> 00:15:29,680 Now with Sam Altman and you know, the leaders of AI labs on Trophy Cloud, they're on this 250 00:15:29,680 --> 00:15:32,000 world tour for some of them. 251 00:15:32,000 --> 00:15:38,320 And you see these pictures and every time we see it, Sam Altman meeting some leaders 252 00:15:38,320 --> 00:15:42,640 of our presidents or prime minister, those looks like he's a president now. 253 00:15:42,640 --> 00:15:44,720 Didn't it cost you? 254 00:15:44,720 --> 00:15:46,560 He wields a lot of power. 255 00:15:46,560 --> 00:15:51,200 Yeah, and when you look at this picture, you're like, who in this picture is actually the 256 00:15:51,200 --> 00:15:52,640 man of power now? 257 00:15:52,640 --> 00:15:53,280 Yeah. 258 00:15:53,280 --> 00:15:58,400 Speaking of power, one of the things that is an interesting kind of discussion point 259 00:15:58,400 --> 00:16:03,040 as we're talking about more efficient GPUs and chips and all the things that run behind 260 00:16:03,040 --> 00:16:08,720 the scenes to enable these really wonderful AI technological innovations. 261 00:16:08,720 --> 00:16:12,240 I didn't realize this week at the Good Castle came out one of the talks that 262 00:16:12,240 --> 00:16:18,160 Sam Altman's a big investor into fusion power research and maintains even talked about 263 00:16:18,160 --> 00:16:18,640 a little bit. 264 00:16:18,640 --> 00:16:22,960 Some of the recent talks this week, we'll link a few in the description on the newsletter, 265 00:16:22,960 --> 00:16:28,640 talked about how to achieve things like GPT-567 where you're going to need just so much compute, 266 00:16:28,640 --> 00:16:31,840 not only to train, but then to run it, let everybody run it. 267 00:16:31,840 --> 00:16:35,200 It's going to need fusion power to achieve. 268 00:16:35,200 --> 00:16:39,600 So, you know, getting to a point where it's a tenth of the cost of any energy and then 269 00:16:39,600 --> 00:16:44,800 we can make enough of them to go around the world is pretty incredible prospect. 270 00:16:44,800 --> 00:16:50,080 And, you know, it's been a promise of, I've heard it my whole life, fusions 30 years away 271 00:16:50,080 --> 00:16:52,640 or 40 years away, you know, it's going to happen in the future. 272 00:16:52,640 --> 00:16:56,560 But a lot of things that I hear when it comes up, it's like five years away. 273 00:16:56,560 --> 00:16:57,200 Right. 274 00:16:57,200 --> 00:16:59,440 It kind of gives me an idea of it. 275 00:16:59,440 --> 00:17:03,120 What Sam Altman is not part of or invested. 276 00:17:03,120 --> 00:17:07,520 World coin, like scanning your eyeballs, universal basic income. 277 00:17:07,520 --> 00:17:10,320 And probably list goes on. 278 00:17:10,320 --> 00:17:11,680 It's incredible. 279 00:17:11,680 --> 00:17:14,080 The power dynamics are very interesting. 280 00:17:14,080 --> 00:17:17,920 He'll be our main ambassador to the AI overlords later on. 281 00:17:17,920 --> 00:17:23,520 But on the flip side, there is, you know, again, if we look at historically first movers 282 00:17:23,520 --> 00:17:29,120 in the market and not necessarily the ones leading the market at the end of the day. 283 00:17:29,120 --> 00:17:34,320 I'm kind of always running with that thought, but if it's not open AI, who is coming? 284 00:17:34,320 --> 00:17:34,960 Yeah. 285 00:17:34,960 --> 00:17:40,720 I will say though, just having, you know, worked and been exploring the space for six months. 286 00:17:40,720 --> 00:17:43,520 There is something about the first mover advantage. 287 00:17:43,520 --> 00:17:48,480 That's a little just kind of flipped on its head with some, some of these things, 288 00:17:48,480 --> 00:17:52,480 because people are inventing and creating so much and stitching these AI technologies 289 00:17:52,480 --> 00:17:56,320 together to form innovative new products and solutions. 290 00:17:56,320 --> 00:18:02,320 But at the end of the day, it's to solve very similar problems that people can kind of conceive 291 00:18:02,320 --> 00:18:05,200 of themselves that can go, Oh, I see that. 292 00:18:05,200 --> 00:18:09,840 I see that they're not using any, all using the same technology, all using the same language 293 00:18:09,840 --> 00:18:10,640 models. 294 00:18:10,640 --> 00:18:11,040 Okay. 295 00:18:11,040 --> 00:18:13,920 So it's really not a game about whose tech is better. 296 00:18:13,920 --> 00:18:15,440 It's about who can market it better. 297 00:18:15,440 --> 00:18:21,280 So in a lot of instances I've seen, you know, just as long as you're kind of there to the 298 00:18:21,280 --> 00:18:26,320 table first with an idea that might be enough in some instances. 299 00:18:26,320 --> 00:18:29,120 I'll give you an example as it relates to prompt engineering. 300 00:18:29,120 --> 00:18:31,280 You know, I happened to start selling products. 301 00:18:31,280 --> 00:18:35,680 I happened to start selling some prompts on prompt base earlier this year. 302 00:18:35,680 --> 00:18:36,400 Right. 303 00:18:36,400 --> 00:18:41,760 It really didn't start taking off till like March or April, but what that's done because 304 00:18:41,760 --> 00:18:44,960 of things like Amazon and how we online shop before. 305 00:18:44,960 --> 00:18:45,760 Well, what do you do? 306 00:18:45,760 --> 00:18:48,720 The first thing you do when you get on Amazon, Hey, you're looking for a toothbrush. 307 00:18:48,720 --> 00:18:52,080 Hey, let me sort by the five stars with the most sales. 308 00:18:52,080 --> 00:18:52,880 Right. 309 00:18:52,880 --> 00:18:54,000 That's just like our condition. 310 00:18:54,000 --> 00:18:55,920 I want the best thing that's out there. 311 00:18:55,920 --> 00:18:57,200 I do that all the time. 312 00:18:57,200 --> 00:18:58,320 Read reviews. 313 00:18:58,320 --> 00:18:59,200 Right. 314 00:18:59,200 --> 00:19:04,240 Something like this, like I feel like in some instances I've unfairly benefited from that 315 00:19:04,240 --> 00:19:09,360 just because I had plenty of sales before this started taking off a little bit more. 316 00:19:09,920 --> 00:19:11,200 And what do people do? 317 00:19:11,200 --> 00:19:13,360 Hey, what's the top selling in this category? 318 00:19:13,360 --> 00:19:14,160 All right. 319 00:19:14,160 --> 00:19:19,680 You know, by virtue of being to the table first, that probably benefits me, you know, 320 00:19:19,680 --> 00:19:22,720 in that instance, it doesn't, it doesn't for other things. 321 00:19:22,720 --> 00:19:26,560 And I would hope that the creators of prop base or some of these marketplace do things 322 00:19:26,560 --> 00:19:30,160 to encourage new creators to put themselves out there. 323 00:19:30,160 --> 00:19:35,680 That makes me think I had this dilemma when I made the first videos going about 324 00:19:35,680 --> 00:19:38,880 chadgpt and learn prompting and prompt engineering. 325 00:19:38,880 --> 00:19:44,640 And I remember there was this moment I just made a video about how chadgpt works 326 00:19:44,640 --> 00:19:46,640 and explaining, you know, the model behind it. 327 00:19:46,640 --> 00:19:48,160 And it was 3.5. 328 00:19:48,160 --> 00:19:51,680 I released this video in a week and a half where is GPT-4. 329 00:19:51,680 --> 00:19:53,680 And I was like, oh, okay. 330 00:19:53,680 --> 00:19:58,080 So if I make a video about something which is happening, and this is the thing why 331 00:19:58,080 --> 00:20:03,680 I don't like to go news direction because our video is not relevant next week. 332 00:20:03,680 --> 00:20:04,000 Yeah. 333 00:20:04,000 --> 00:20:08,000 But, and I was thinking at the time, like, oh, I need to really, you know, 334 00:20:08,000 --> 00:20:11,520 think about videos which live long, very meaningful. 335 00:20:12,320 --> 00:20:18,400 The thing is that that video plus prompt engineering are getting views every single day. 336 00:20:19,040 --> 00:20:20,000 Like the most. 337 00:20:20,000 --> 00:20:24,560 And I'm like, in my world, I'm like, everyone knows already this. 338 00:20:24,560 --> 00:20:28,800 And I still read comments from people like, oh, I just discovered this. 339 00:20:28,800 --> 00:20:29,280 Wow. 340 00:20:29,280 --> 00:20:30,320 I didn't know this. 341 00:20:30,320 --> 00:20:34,480 I'm like, it feels like not three months, but three years ago. 342 00:20:34,480 --> 00:20:38,640 And that just makes me think this is why, you know, I was sharing this story, 343 00:20:38,640 --> 00:20:45,520 bringing these AI tools outside of the bubble and actually getting real world perspective. 344 00:20:45,520 --> 00:20:51,040 You know, you have to see how the real world actually adapts this fix. 345 00:20:51,040 --> 00:20:53,360 It's still, still trickling out there. 346 00:20:53,360 --> 00:20:53,840 Yeah. 347 00:20:53,840 --> 00:20:59,920 And I think with Chai GPT, when I was doing, you know, the, the keynote in embassy, 348 00:20:59,920 --> 00:21:03,600 three people came to me saying, oh yeah, yeah, I know it. 349 00:21:03,600 --> 00:21:07,280 I tried it, but I'm so disappointed with the results. 350 00:21:07,280 --> 00:21:08,480 Just so basic. 351 00:21:08,480 --> 00:21:09,040 Yep. 352 00:21:09,040 --> 00:21:13,840 And after my presentation, they're like, oh, so it can not like, you can actually 353 00:21:13,840 --> 00:21:18,000 control it and make it interesting and it can write in different things. 354 00:21:18,000 --> 00:21:25,760 And I was like, to me, it was like, yeah, of course, but people had tried and left it off. 355 00:21:25,760 --> 00:21:28,240 And I think it also happens with image tools. 356 00:21:28,240 --> 00:21:30,800 You can get very quickly discouraged. 357 00:21:30,800 --> 00:21:35,280 And I think this learning prompt engineering, that's why I think it's relevant and it's 358 00:21:35,280 --> 00:21:36,320 going to be relevant. 359 00:21:36,320 --> 00:21:36,960 Definitely. 360 00:21:36,960 --> 00:21:37,280 Yeah. 361 00:21:37,280 --> 00:21:41,760 I mean, it kind of boggles my mind that some people don't necessarily see the 362 00:21:41,760 --> 00:21:42,800 possibilities of it. 363 00:21:42,800 --> 00:21:47,360 But I also, people come at this with different needs and different expectations, you know, 364 00:21:47,360 --> 00:21:52,000 like obviously I'm coming at it from someone who's an avid consumer of anything 365 00:21:52,000 --> 00:21:53,520 electronic or technological. 366 00:21:53,520 --> 00:21:55,520 Does one need eight keyboards? 367 00:21:55,520 --> 00:21:58,000 My wife would say no, but I say otherwise. 368 00:21:58,000 --> 00:21:58,400 Yeah. 369 00:21:58,400 --> 00:22:01,600 I mean, I remember the first time my dad tried chat GPT. 370 00:22:01,600 --> 00:22:03,920 He's like, yeah, I put myself into it. 371 00:22:03,920 --> 00:22:05,920 Came up, didn't know anything about me. 372 00:22:05,920 --> 00:22:07,120 It's kind of milk toast. 373 00:22:07,120 --> 00:22:11,520 And I'm like, this is not how large language models work. 374 00:22:11,520 --> 00:22:16,800 Let me tell you about how it's only trained up till September, 2021 and about hallucinations, 375 00:22:16,800 --> 00:22:19,440 you know, so basic things you need to know. 376 00:22:19,440 --> 00:22:21,920 But like, but that's not on the front page at all. 377 00:22:21,920 --> 00:22:25,440 No, it's at the bottom, tiny letters. 378 00:22:25,440 --> 00:22:25,920 Yeah. 379 00:22:25,920 --> 00:22:27,760 This might be totally made up. 380 00:22:27,760 --> 00:22:28,240 Yeah. 381 00:22:28,240 --> 00:22:31,360 Not factual, wrong facts information. 382 00:22:31,360 --> 00:22:32,160 Oh yeah. 383 00:22:32,160 --> 00:22:33,440 That's the whole other thing. 384 00:22:33,440 --> 00:22:38,880 You know, we already talked about this, that there is certain level of responsibility when 385 00:22:38,880 --> 00:22:41,680 you drop a product like that to a mass market. 386 00:22:41,680 --> 00:22:42,080 Yeah. 387 00:22:42,080 --> 00:22:46,480 No information, what it is, how it is, nothing. 388 00:22:46,480 --> 00:22:48,720 It's just like, Hey, chatbot, that's it. 389 00:22:48,720 --> 00:22:54,800 But, but I think some of it is like, Herod, let's let people play around with it and 390 00:22:54,800 --> 00:22:59,840 explore while the stakes are low, as opposed to have this AGI that's ready. 391 00:22:59,840 --> 00:23:03,840 And then it just, they drop it on us and there's this huge shift that needs to happen. 392 00:23:03,840 --> 00:23:08,640 There's probably some value in it slowly, gradually, but it's not like, 393 00:23:08,640 --> 00:23:14,560 granted how slowly can you say it's still the fastest growing app or service in history. 394 00:23:15,120 --> 00:23:20,480 But, you know, it's still probably some value in letting it kind of trickle down. 395 00:23:20,480 --> 00:23:20,800 Yeah. 396 00:23:20,800 --> 00:23:22,320 Who values from that? 397 00:23:22,320 --> 00:23:27,760 I mean, it's probably the altruism of our great ambassador overlord, Sam Altman, 398 00:23:28,560 --> 00:23:30,800 you know, and those kind of investors. 399 00:23:31,360 --> 00:23:33,680 I hope people learned one thing. 400 00:23:33,680 --> 00:23:35,520 Nothing on internet is free. 401 00:23:36,080 --> 00:23:36,400 Yeah. 402 00:23:36,400 --> 00:23:39,360 So if there is something free, it's not free. 403 00:23:39,360 --> 00:23:42,080 Like you are trading off different value. 404 00:23:42,080 --> 00:23:48,640 And I worked in big data analytics company, the amount people pay for data and especially 405 00:23:48,640 --> 00:23:51,840 investment banks, it's insane. 406 00:23:52,320 --> 00:23:54,560 It's millions upon millions. 407 00:23:55,120 --> 00:24:02,320 So I think that's what people don't realize that that just the volume of data from this type of 408 00:24:02,320 --> 00:24:08,960 usage, but what you said, it's a number one, like most trending website or something like that. 409 00:24:09,680 --> 00:24:11,120 Literally vertical growth. 410 00:24:11,120 --> 00:24:12,960 That's if you map it against anything else. 411 00:24:13,840 --> 00:24:17,200 No, but let's talk something prompting. 412 00:24:17,200 --> 00:24:17,360 Yes. 413 00:24:17,360 --> 00:24:22,000 Now that we're leading to that, I wanted to share something with you coming back to the, 414 00:24:22,640 --> 00:24:26,880 my past afterlife in architecture and design. 415 00:24:26,880 --> 00:24:27,360 Yeah. 416 00:24:27,360 --> 00:24:38,000 I saw this design piece, it's foldable sofa, which can fit in the envelope designed by AI. 417 00:24:38,000 --> 00:24:40,800 It's studio 10 in Copenhagen. 418 00:24:40,800 --> 00:24:47,520 I actually know the studio and I think I'm almost sure I've been there, but anyway, 419 00:24:47,520 --> 00:24:50,720 so it's developed closely with IKEA. 420 00:24:50,720 --> 00:24:56,160 So you see this modular system and it's around 10 kilos. 421 00:24:56,160 --> 00:25:02,160 But the interesting thing that they went with first talking with Chad GPT, for example, 422 00:25:02,160 --> 00:25:07,200 one of them is saying, I actually wrote into Chad GPT, could a couch fit into an envelope? 423 00:25:07,200 --> 00:25:08,800 And of course it said no. 424 00:25:08,800 --> 00:25:14,080 And then we went to Midjourney and Delhi and we're spending like hundreds of prompts 425 00:25:14,080 --> 00:25:20,400 testing different words, terminology, and nothing worked really to achieve what they 426 00:25:20,400 --> 00:25:26,160 were looking for. And it was that you on a sofa, you're not front facing like to the TV, 427 00:25:26,160 --> 00:25:28,960 but you're actually sitting in front facing each other. 428 00:25:28,960 --> 00:25:35,600 And the key word which we discovered that actually led to that was conversation pit. 429 00:25:36,560 --> 00:25:37,520 Conversation pit. 430 00:25:38,080 --> 00:25:40,400 Right. So now we build it. 431 00:25:40,400 --> 00:25:47,520 So I would think that something like that is product of a machine learning or a labeling AI 432 00:25:47,520 --> 00:25:54,400 where it takes in tons of different designs and analyzes it and goes, this is the most ergonomic 433 00:25:54,400 --> 00:26:00,880 way to do this, but it's from a language set, generative language model, just kind of going, 434 00:26:00,880 --> 00:26:02,400 this is what I think it would look like. 435 00:26:02,400 --> 00:26:09,440 Yeah, but generative models is actually nothing new in AI space, especially when you look in a 436 00:26:09,440 --> 00:26:15,760 parametric design. So this, like even when I was doing my bachelor thesis, I was actually using 437 00:26:15,760 --> 00:26:22,080 this grasshopper plugins in the rhinos model to do parametric design. And then you can play with some 438 00:26:22,080 --> 00:26:28,400 generative design pieces, but the diffusion models is different. That's something new. 439 00:26:28,400 --> 00:26:34,720 And the generative AI, you have, in a way you have more control because it's like more like 440 00:26:34,720 --> 00:26:38,720 real database. And I could be completely wrong. Like this is just, you know, 441 00:26:38,720 --> 00:26:45,840 I'm kind of surface sliding through this topic right now, but diffusion models in architecture 442 00:26:45,840 --> 00:26:52,720 for render amazing. But then you want to real design vector based because, you know, you need 443 00:26:53,360 --> 00:26:59,680 centimeter millimeter control of the dimensions. That's kind of a different realm. I'm keeping a 444 00:26:59,680 --> 00:27:05,600 close eye on what's happening in architecture space design, because, you know, imagine like 445 00:27:05,600 --> 00:27:09,200 if it goes that route explosion, another explosion. 446 00:27:09,200 --> 00:27:15,840 There's, there's a huge market. I've seen some things for people exploring it from prompting to 447 00:27:15,840 --> 00:27:22,320 some 3d design prompting to vectors. No one has cracked it yet, but you know, there's a few things 448 00:27:22,320 --> 00:27:26,240 I think we brought up on the podcast before we have people that are doing generative AI be a 449 00:27:26,240 --> 00:27:32,000 prompting to create assets for computer games that can be used inside of Unreal five engine. 450 00:27:32,000 --> 00:27:37,760 I'll try to see if I can find them right now, but I know there's a company that does text prompt to 451 00:27:37,760 --> 00:27:43,760 3d STL model. So you can basically just type in a prompt and then have what you would need to 452 00:27:43,760 --> 00:27:48,720 3d print it, you know, throw it into the slicer and then, you know, put it through. 453 00:27:49,440 --> 00:27:56,320 Leonardo AI for anybody looking for a game development, like what the guys behind it, 454 00:27:56,320 --> 00:28:01,280 like I think I told you about Ethan, one of the co-founders, we were, you know, back in the days 455 00:28:01,280 --> 00:28:07,200 before Leonardo, we were just geeking out in the discord channels. It's just incredible. It's, 456 00:28:07,200 --> 00:28:14,640 I think they are reaching millions of users now in such short time. And that also makes me think 457 00:28:15,280 --> 00:28:22,880 a year ago, I made this video, which is who's going to build in metaverse AI or architects. 458 00:28:22,880 --> 00:28:28,480 And in that video, I made completely like just, you know, going down the rabbit hole of this idea. 459 00:28:28,480 --> 00:28:36,560 I was like, in the future, you just prompt and say, I want them desert with flying buildings made 460 00:28:36,560 --> 00:28:44,400 of diamonds, wherever space and boom, AI generates that 3d environment space for you. And now a year 461 00:28:44,400 --> 00:28:51,920 later, we have VR spaces generated by AI. That's almond down the pike. Back then, I remember when 462 00:28:51,920 --> 00:28:57,520 I wrote it, I was like, Oh, probably people will be like, what she's talking about. It's crazy. 463 00:28:57,520 --> 00:29:04,800 Yeah. Well, I mean, we're every day, like I said, it must be, but it felt like to see an airplane 464 00:29:04,800 --> 00:29:10,400 flying the sky for the first time, something so divergent from your experiences up into this date. 465 00:29:10,400 --> 00:29:15,440 But like speaking of things that we haven't seen before, you know, a natural progression, 466 00:29:15,440 --> 00:29:19,840 we've talked about some text to video, we've talked about some of our friends that are, 467 00:29:19,840 --> 00:29:24,960 you know, some luminaries and thought leaders in the space. How long do you think it's going to be 468 00:29:24,960 --> 00:29:34,560 until we see a full on prompted movie? I think it's in the workings right now. Definitely. I mean, 469 00:29:34,560 --> 00:29:40,400 so this was set up question. You're like, come on, I got an agenda here. We gotta, we just gotta 470 00:29:40,400 --> 00:29:45,840 try to be like, you know, professional podcasters here. Yeah. Okay. So I was more thinking about 471 00:29:45,840 --> 00:29:54,320 like full length cinema, but of course this piece from the frost is already like, takes you a step 472 00:29:54,320 --> 00:30:00,640 back to think what is the future of animation and not even animation. You know, I think, look, 473 00:30:00,640 --> 00:30:05,280 make no mistake. There's still some uncanny Valley. Our brains are so used to looking at 474 00:30:05,280 --> 00:30:11,600 images of, you know, not images, but just looking at other people. So anytime their face or hands 475 00:30:11,600 --> 00:30:16,480 are doing something that's not natural to what we're just used to seeing all the time, we notice it. 476 00:30:16,480 --> 00:30:22,800 So that's definitely apparent in these types of things too, but this is a 13 minute completely 477 00:30:22,800 --> 00:30:28,560 generated, you know, AI movie and we're going to include the link to it, but you know, it, 478 00:30:28,560 --> 00:30:34,800 you could follow the narrative structure easily. That's not lost at all. And in some ways, you know, 479 00:30:34,800 --> 00:30:39,280 some of the choices they've made because of the current technological limitations could just be 480 00:30:39,280 --> 00:30:45,120 stylistic ones that, you know, might be said, Hey, we made this deliberately, this kind of blocky 481 00:30:45,120 --> 00:30:50,640 chunky kind of movements and stuff like that. I'd still say like half of these shots, I think the, 482 00:30:50,640 --> 00:30:56,880 the human ones admittedly are hard, but like half of them are still like ones that, you know, I, 483 00:30:56,880 --> 00:31:00,960 I'd be like, all right, that'd be something I'd see in a movie. 484 00:31:00,960 --> 00:31:05,120 I want to kind of take the different camp on this one. 485 00:31:05,120 --> 00:31:05,600 Please do. 486 00:31:06,160 --> 00:31:13,360 Especially as a content creator and editor, my own videos, when you see the composition, 487 00:31:13,360 --> 00:31:20,400 the color grading, the transitions, because I, it was edited by a human, the creative decisions 488 00:31:20,400 --> 00:31:26,720 of the storytelling of the scenes, that central composition leading lines, all this matters 489 00:31:27,280 --> 00:31:34,800 were still decided by a human. Or if you tell me that this was completely just a prompt in the 490 00:31:34,800 --> 00:31:43,040 composition, the mood, like all these things are deliberately there, the color theory, and it's 491 00:31:43,040 --> 00:31:49,520 human who makes those decisions. But of course, for example, Major and he makes incredible images 492 00:31:49,520 --> 00:31:55,680 and you can see this composition decisions just adapted. But do you think in this case was it too? 493 00:31:55,680 --> 00:32:01,120 I mean, I think it probably was in a lot of instances, but you know, you're absolutely 494 00:32:01,120 --> 00:32:05,760 right. The true power here is when you have someone with expertise and creativity, 495 00:32:05,760 --> 00:32:11,120 you know, working alongside AI. Right now, some of these things like this is not, 496 00:32:11,120 --> 00:32:16,240 they wouldn't pass for a Hollywood movie at this point, but it's not too far off. 497 00:32:16,240 --> 00:32:20,720 You know, it looks kind of like 80s or 90s special effects in some degree. 498 00:32:20,720 --> 00:32:27,200 So I'm not saying that it's not good, but I'm still want to think that final creative decisions 499 00:32:27,200 --> 00:32:35,440 is going to be made by human. Or maybe if we go like what I did a year ago, and we say that, 500 00:32:35,440 --> 00:32:40,880 okay, in a year or two, the whole movies are going to be generated by AI. I think in Black 501 00:32:40,880 --> 00:32:48,320 Mirror episode, they teased this moment where you're watching kind of Netflix and seeing 502 00:32:48,320 --> 00:32:53,680 representation of yourself. And the character is like, oh, it's like, she's just like me talks like 503 00:32:53,680 --> 00:33:00,560 me looks like me. So where did they get this data from? Well, kind of on the same vein as you know, 504 00:33:00,560 --> 00:33:07,440 what we're discussing here about creativity, and still needing to have that final creative decision. 505 00:33:07,440 --> 00:33:12,560 One thing I don't think we've brought up on the pod yet is we've touched on it a little bit in terms 506 00:33:12,560 --> 00:33:18,720 of some of the legal implications for these generative models, but they've been trained on. 507 00:33:18,720 --> 00:33:23,520 And if folks aren't aware, there's, you know, it's going to be a fight in a lot of different ways. 508 00:33:23,520 --> 00:33:29,280 And the first one that's going to be the waging the war during the battle is stable diffusion. 509 00:33:29,280 --> 00:33:33,040 This article right here is back from February, but just in the past week, 510 00:33:33,040 --> 00:33:38,320 they released that this is going to trial. But I mean, Getty Images suing stable diffusion for 511 00:33:39,200 --> 00:33:47,280 the landmark amount of 1.8 trillion, the largest lawsuit ever for the 12 million images that they 512 00:33:47,280 --> 00:33:53,040 claim were in the sample database. And they're claiming damages of 150 grand per image. 513 00:33:53,680 --> 00:34:00,240 Yeah. And oh my God, we can dive into this because I was looking into this training data model, 514 00:34:00,240 --> 00:34:06,800 which was stable diffusion, but it just Lyon it's called. And we will include in the newsletter, 515 00:34:06,800 --> 00:34:13,120 there is a link where you can actually upload and check if what is the chance that your image 516 00:34:13,120 --> 00:34:20,160 was used in the training data, but the amounts are staggering. And it's known that they did use 517 00:34:20,720 --> 00:34:26,960 this data. And it's just interesting that I read somewhere that there was shutter stock in Getty 518 00:34:26,960 --> 00:34:34,480 Images was striking a deal with these to get paid. And it just looks to me like that deal maybe 519 00:34:34,480 --> 00:34:40,880 didn't go through. And with the sentiment from art community, they just went full on. If you go in a 520 00:34:40,880 --> 00:34:47,440 law case and you say like, yeah, none of the images, new produced images actually have exactly, 521 00:34:47,440 --> 00:34:53,920 exactly copy of your image. And it's a tricky argument they are using, you know, but there's 522 00:34:53,920 --> 00:35:00,880 an easy solution to that. What's that just develop like a coordinate system or metadata system or 523 00:35:00,880 --> 00:35:07,600 like, you know, licensing or copyright thing in the image and just define the radius that okay, 524 00:35:07,600 --> 00:35:14,000 there is a certain spectrum, everything is packed from nothing is like just true or not. So if my 525 00:35:14,000 --> 00:35:22,240 image and the spectrum, the images get close to my style, to, to my original image, I get a royalty. 526 00:35:22,240 --> 00:35:29,440 I mean, free open source model. It's not quite that simple because no, when you look at that, 527 00:35:29,440 --> 00:35:35,440 when you look at that globe there, right? That is an undirected network. What that means in network 528 00:35:35,440 --> 00:35:41,520 science context. It's not like you start with one thing and it makes these decisions linearly. 529 00:35:41,520 --> 00:35:48,000 The way these diffusion models work is they not only can identify what images are, they understand 530 00:35:48,000 --> 00:35:54,240 how they interact with each other. So it won't necessarily be this linear decision of, okay, 531 00:35:54,240 --> 00:35:58,240 we use a little bit from this sample image, then went over here and use some of this and came back 532 00:35:58,240 --> 00:36:02,960 to this, this neighborhood of cluster, you know, and on the way we have to fill in some stuff that 533 00:36:02,960 --> 00:36:09,600 we, we didn't have any samples for. It's more so, it just kind of does it all at once. So I think 534 00:36:09,600 --> 00:36:17,360 that still being able to track it in a way that's methodical where you can even say that like, hey, 535 00:36:17,360 --> 00:36:24,400 75% of this is, is the same. Well, all right, let's use that argument. So say that's the case. 536 00:36:24,400 --> 00:36:30,000 If I'm not mistaken, I am no legal expert, but I found this kind of surprising in my beloved country 537 00:36:30,000 --> 00:36:36,240 to the North Canada, with my understanding that for using a trademark logo, for example, if you 538 00:36:36,240 --> 00:36:42,480 make one change to that logo, change one color, add something that is now a derivative and it's 539 00:36:42,480 --> 00:36:50,880 a separate thing. Something is 75% the same, but 25% different. That's a derivative. Yeah. 540 00:36:50,880 --> 00:36:58,480 And this is absolutely true. And I'm looking at another law case, which actually this artist, 541 00:36:58,480 --> 00:37:05,920 Carol Heintz, met, and this was in 2016. She sued back then Getty images for 1 billion 542 00:37:06,480 --> 00:37:11,760 because they were using her images and actually send her, this is a funny story. She didn't win, 543 00:37:11,760 --> 00:37:18,800 but the funny story is they send her a bill for license, but she's using license of her own work 544 00:37:18,800 --> 00:37:24,880 illegally. And this is how she found out that they've been stealing her work for years and 545 00:37:24,880 --> 00:37:31,440 licensing and charging people. And now we try to charge author herself. Oops. Yeah. So what, 546 00:37:31,440 --> 00:37:39,360 so she went like 1 billion damages. Yeah. We need to have a lawyer on here. It's such 547 00:37:39,360 --> 00:37:44,080 interesting decisions. I'm hopeful, but also the people that are going to be making these decisions 548 00:37:44,080 --> 00:37:49,760 around the world are not of our generation of conqueror. I mean, these are people that 549 00:37:49,760 --> 00:37:55,440 may not even be able to concede you and understand and have an appreciation for some of these things. 550 00:37:55,440 --> 00:38:00,320 So, I mean, I don't envy them, but it's going to be very challenging kind of decisions and 551 00:38:00,320 --> 00:38:06,720 complexities to sort through. I'm just running in my mind about like all the network of legal. 552 00:38:06,720 --> 00:38:12,320 Even if I know a lawyer, I would be like, does this lawyer know the intricacies of this? 553 00:38:12,320 --> 00:38:13,520 We need an AI lawyer. 554 00:38:13,520 --> 00:38:19,840 This is so new. Yeah. It's funny. Like we have a contact in the patent department. So now I'm 555 00:38:19,840 --> 00:38:24,000 thinking that would be interesting conversation, but yeah, we'll be bringing up guests. 556 00:38:24,000 --> 00:38:30,480 We have a lot of, a lot more of that on the horizon in future weeks of filmed video, 557 00:38:30,480 --> 00:38:37,840 HTTTA. So I think with that, this is a good enough place to call it quits for the week. 558 00:38:37,840 --> 00:38:41,920 So for GoToGo, I am Wes the SynthMind saying happy prompting everybody. 559 00:38:42,560 --> 00:38:44,480 Happy prompting everybody. 560 00:38:45,680 --> 00:38:52,320 Thanks for listening to how to talk to AI with your hosts, GoToGo and Wes the SynthMind. 561 00:38:53,200 --> 00:38:59,040 As always, you can check out the show notes and links at howtotalkto.ai. 562 00:38:59,040 --> 00:39:02,720 That's all for this week's episode. Happy prompting everyone.