1 00:00:04,120 --> 00:00:07,200 Speaker 1: Get in touch with technology with tech Stuff from half 2 00:00:07,200 --> 00:00:13,680 Speaker 1: stuff works dot com. Hey there, and welcome to tech Stuff. 3 00:00:13,720 --> 00:00:16,480 Speaker 1: I'm your host, Jonathan Strickland. I'm an executive producer with 4 00:00:16,520 --> 00:00:19,639 Speaker 1: how Stuff Works in a love all things tech, and 5 00:00:19,680 --> 00:00:23,920 Speaker 1: today we're going to tackle a story that recently unfolded recently. 6 00:00:24,000 --> 00:00:26,480 Speaker 1: As of the recording of this show, I'm sitting in 7 00:00:26,560 --> 00:00:30,639 Speaker 1: the recording studio on October two, thousand eighteen. It's not 8 00:00:30,680 --> 00:00:33,159 Speaker 1: my normal studio either, So if you hear other noises, 9 00:00:33,360 --> 00:00:36,240 Speaker 1: that's because we've got noisy people walking around the office 10 00:00:36,240 --> 00:00:39,400 Speaker 1: and I'm in a different studio. That's commentary. But this 11 00:00:39,479 --> 00:00:44,120 Speaker 1: story unfolded just at the very end of October. That 12 00:00:44,280 --> 00:00:48,440 Speaker 1: was when the auction house Christie's, put a special item 13 00:00:48,560 --> 00:00:52,200 Speaker 1: up on the auctioning block. It was a somewhat blurry 14 00:00:52,280 --> 00:00:56,480 Speaker 1: portrait of a man dressed in antiquated clothing. It looked 15 00:00:56,520 --> 00:00:58,560 Speaker 1: like a painting that could have come from the eighteenth 16 00:00:58,680 --> 00:01:02,240 Speaker 1: century from one of any number of artists, but it 17 00:01:02,360 --> 00:01:06,279 Speaker 1: was in fact a much more recent painting. The artist 18 00:01:06,360 --> 00:01:09,880 Speaker 1: was not a famous painter. In fact, the artist wasn't 19 00:01:10,000 --> 00:01:14,600 Speaker 1: a person. It was an artificially intelligent algorithm that created 20 00:01:14,640 --> 00:01:18,800 Speaker 1: the portrait through the process of machine learning. And what's more, 21 00:01:19,280 --> 00:01:23,360 Speaker 1: the group of human artists who supplied the AI generated 22 00:01:23,400 --> 00:01:27,440 Speaker 1: portrait had taken a great deal of direction, let's say, 23 00:01:27,560 --> 00:01:31,280 Speaker 1: from a different computer programmer, but perhaps did not do 24 00:01:31,720 --> 00:01:36,040 Speaker 1: as much to attribute that coder's work to the creation 25 00:01:36,080 --> 00:01:39,039 Speaker 1: of this portrait that they should have done. So what 26 00:01:39,160 --> 00:01:41,919 Speaker 1: we have here sounds a bit like a twenty first 27 00:01:42,040 --> 00:01:45,800 Speaker 1: century futuristic art heist, only this isn't about stealing a 28 00:01:45,840 --> 00:01:50,440 Speaker 1: work of art, but rather a means of generating art itself, 29 00:01:50,720 --> 00:01:54,240 Speaker 1: and it's creating a lot of interesting conversations about concepts, 30 00:01:54,360 --> 00:01:57,600 Speaker 1: ranging from what is art in the first place, to 31 00:01:57,680 --> 00:02:00,960 Speaker 1: the practical applications of machine learning to the nature of 32 00:02:01,040 --> 00:02:04,840 Speaker 1: open source code. So let's dive down into this, because 33 00:02:04,880 --> 00:02:08,520 Speaker 1: when it comes to discussing our how technology interacts with 34 00:02:08,520 --> 00:02:11,720 Speaker 1: our lives, this is a doozy of a story. It 35 00:02:11,800 --> 00:02:15,000 Speaker 1: highlights not just technological issues but human ones that just 36 00:02:15,120 --> 00:02:18,880 Speaker 1: happened to intersect with technology. So to begin with, let's 37 00:02:18,919 --> 00:02:22,440 Speaker 1: talk about the tech behind generating this portrait in the 38 00:02:22,480 --> 00:02:27,119 Speaker 1: first place. It is an application of machine learning. That's 39 00:02:27,120 --> 00:02:29,880 Speaker 1: one of those topics we've talked about a lot on 40 00:02:30,040 --> 00:02:34,960 Speaker 1: tech stuff, especially recently. But basically, machine learning is all 41 00:02:35,000 --> 00:02:38,920 Speaker 1: about designing processes that allow machines to parse data in 42 00:02:38,960 --> 00:02:42,480 Speaker 1: some useful way and then apply the results of those 43 00:02:42,520 --> 00:02:46,280 Speaker 1: operations to future problems. But that's pretty darn vague, right, 44 00:02:46,360 --> 00:02:48,640 Speaker 1: that's not that doesn't really tell you anything useful if 45 00:02:48,639 --> 00:02:51,280 Speaker 1: you dive down a bit further, it's about creating a 46 00:02:51,320 --> 00:02:55,160 Speaker 1: framework within which machines can learn to perform a task 47 00:02:55,600 --> 00:02:58,800 Speaker 1: without having to be programmed to do it. So let's 48 00:02:58,880 --> 00:03:01,280 Speaker 1: use an example, and it's one I've talked about a 49 00:03:01,280 --> 00:03:03,600 Speaker 1: lot because it was one of the early examples of 50 00:03:03,600 --> 00:03:06,240 Speaker 1: what machine learning could do once it reached a certain 51 00:03:06,320 --> 00:03:10,280 Speaker 1: level of sophistication. Back in two thousand twelve, Google showed 52 00:03:10,320 --> 00:03:14,799 Speaker 1: how their computer scientists teams had taught an AI algorithm 53 00:03:14,919 --> 00:03:19,840 Speaker 1: or neural network to recognize images of cats. Now, this 54 00:03:19,919 --> 00:03:21,920 Speaker 1: was perhaps a funny way of showing an approach to 55 00:03:21,960 --> 00:03:25,200 Speaker 1: a difficult problem. So if you want a computer to 56 00:03:25,280 --> 00:03:28,720 Speaker 1: recognize an image of a cat, if it's a specific 57 00:03:28,760 --> 00:03:31,079 Speaker 1: image of a cat, you have a couple of different options. 58 00:03:31,360 --> 00:03:34,280 Speaker 1: One is, you can program the computer so that when 59 00:03:34,320 --> 00:03:38,600 Speaker 1: it encounters a specific arrangement of pixels for this particular image, 60 00:03:38,880 --> 00:03:41,880 Speaker 1: it recognizes that as the image of a cat, and 61 00:03:41,920 --> 00:03:45,440 Speaker 1: that you have programmed the computer to say, when you 62 00:03:45,840 --> 00:03:50,080 Speaker 1: see this arrangement of pixels, then that means this is 63 00:03:50,200 --> 00:03:52,720 Speaker 1: a cat. The computer doesn't understand what a cat is, 64 00:03:52,800 --> 00:03:56,400 Speaker 1: it doesn't have any context. It doesn't understand what any 65 00:03:56,440 --> 00:03:59,280 Speaker 1: other picture of a cat might be because that would 66 00:03:59,320 --> 00:04:03,600 Speaker 1: be a different arrangement of pixels. So you could program 67 00:04:03,600 --> 00:04:05,480 Speaker 1: a computer to do this and it would be able 68 00:04:05,520 --> 00:04:07,880 Speaker 1: to do it with that one image. But if you 69 00:04:07,880 --> 00:04:09,680 Speaker 1: gave it a different image of a cat, or even 70 00:04:09,680 --> 00:04:12,320 Speaker 1: an image of the same cat, but it's a different picture, 71 00:04:12,880 --> 00:04:14,960 Speaker 1: the computer would not be able to identify it. You 72 00:04:14,960 --> 00:04:18,560 Speaker 1: would have to repeat the entire process from beginning to 73 00:04:18,720 --> 00:04:21,400 Speaker 1: end to get the same result. And once you start 74 00:04:21,400 --> 00:04:25,440 Speaker 1: adding up images, you realize this is not really an 75 00:04:25,440 --> 00:04:30,680 Speaker 1: efficient means of teaching a computer anything. Or you could 76 00:04:30,680 --> 00:04:34,560 Speaker 1: create an artificial neural network that examines the pixels in 77 00:04:34,600 --> 00:04:37,719 Speaker 1: an image, and each neuron might be looking at a 78 00:04:37,760 --> 00:04:40,440 Speaker 1: different element of the data to determine if that data 79 00:04:40,600 --> 00:04:44,200 Speaker 1: was consistent with images of cat pictures. So we've talked 80 00:04:44,200 --> 00:04:48,160 Speaker 1: about this recently too, and artificial neuron can take in 81 00:04:48,320 --> 00:04:53,080 Speaker 1: multiple bin binary points of data's euros and ones and 82 00:04:53,080 --> 00:04:56,920 Speaker 1: then create a single binary output. So it might be 83 00:04:56,920 --> 00:05:00,840 Speaker 1: looking at specific features that might have to do with ears, 84 00:05:00,880 --> 00:05:03,600 Speaker 1: for example, and if it detects that the ears are 85 00:05:03,680 --> 00:05:06,840 Speaker 1: consistent with those of a cat, it might pass a 86 00:05:06,880 --> 00:05:10,080 Speaker 1: positive response further down the neural network, and a full 87 00:05:10,120 --> 00:05:12,839 Speaker 1: collection of all these looking at multiple points of data 88 00:05:13,160 --> 00:05:17,320 Speaker 1: would allow the computer to come to a decision does 89 00:05:17,400 --> 00:05:21,640 Speaker 1: this image represent a cat or does it represent something else. So, 90 00:05:21,720 --> 00:05:26,120 Speaker 1: in this way, by feeding thousands or tens of thousands 91 00:05:26,200 --> 00:05:29,320 Speaker 1: or hundreds of thousands of images to a computer, you 92 00:05:29,360 --> 00:05:32,160 Speaker 1: can train it to recognize cats. And the more you 93 00:05:32,240 --> 00:05:35,560 Speaker 1: train it and the more closely you're able to tweak 94 00:05:35,720 --> 00:05:39,600 Speaker 1: the network so that it waits certain elements more than others, 95 00:05:40,240 --> 00:05:43,640 Speaker 1: the better it gets. So the tweaking makes the network 96 00:05:43,680 --> 00:05:47,400 Speaker 1: more capable and eventually get to a point where it 97 00:05:47,480 --> 00:05:50,840 Speaker 1: can identify a picture as either being a cat or 98 00:05:50,920 --> 00:05:56,080 Speaker 1: not a cat with pretty good results. Um Back in 99 00:05:56,120 --> 00:05:59,719 Speaker 1: two thousand twelve when Google was talking about this, it 100 00:05:59,839 --> 00:06:03,720 Speaker 1: was still a little jankie. It could sometimes recognize a cat, 101 00:06:04,000 --> 00:06:06,479 Speaker 1: and sometimes it would think that a person was a 102 00:06:06,480 --> 00:06:09,240 Speaker 1: cat or that a cat was a person, So it 103 00:06:09,400 --> 00:06:13,799 Speaker 1: was not infallible, but it was pretty good. Now, because 104 00:06:13,839 --> 00:06:16,920 Speaker 1: I've covered artificial neural networks in recent episodes of tech Stuff, 105 00:06:17,240 --> 00:06:20,119 Speaker 1: I'm not gonna go through the whole thing all over again. 106 00:06:20,160 --> 00:06:22,279 Speaker 1: That high level I just gave you that's a pretty 107 00:06:22,279 --> 00:06:24,919 Speaker 1: good starting point. It's just important to remember that the 108 00:06:25,000 --> 00:06:28,560 Speaker 1: general output here is through training and network using that 109 00:06:29,120 --> 00:06:32,000 Speaker 1: input data set in this case or in the case 110 00:06:32,040 --> 00:06:35,080 Speaker 1: of that example, hundreds of thousands of images of cats. 111 00:06:36,000 --> 00:06:40,400 Speaker 1: Machine learning can actually take a few different approaches. The 112 00:06:40,440 --> 00:06:44,120 Speaker 1: one that I sort of outlined earlier would kind of 113 00:06:44,160 --> 00:06:48,040 Speaker 1: fall into the category of supervised machine learning. See in 114 00:06:48,120 --> 00:06:50,880 Speaker 1: that approach, we human beings are trying to teach a 115 00:06:50,920 --> 00:06:56,640 Speaker 1: machine through algorithms and data sets two recognize something that 116 00:06:56,680 --> 00:07:00,000 Speaker 1: we already know the answer for. Right, you can look 117 00:07:00,000 --> 00:07:02,640 Speaker 1: get a picture, and you can recognize whether that picture 118 00:07:02,680 --> 00:07:05,120 Speaker 1: is of a cat or not, so you already know 119 00:07:05,200 --> 00:07:07,320 Speaker 1: the answer. You're not asking the computer to give you 120 00:07:07,400 --> 00:07:10,200 Speaker 1: new information. You're trying to teach the computer to do 121 00:07:10,280 --> 00:07:16,360 Speaker 1: something that you already can do. So we human beings 122 00:07:16,440 --> 00:07:20,160 Speaker 1: are able to supervise the machine as it is learning 123 00:07:20,200 --> 00:07:23,560 Speaker 1: this process and make those minor adjust adjustments that are 124 00:07:23,600 --> 00:07:26,160 Speaker 1: needed throughout the system in order for it to get 125 00:07:26,200 --> 00:07:29,920 Speaker 1: better at its job. That is supervised machine learning. We 126 00:07:29,960 --> 00:07:32,280 Speaker 1: can keep working with it until it reaches what we 127 00:07:32,320 --> 00:07:36,080 Speaker 1: consider to be an acceptable level of success, which doesn't 128 00:07:36,080 --> 00:07:37,480 Speaker 1: mean it has to be perfect. It just has to 129 00:07:37,480 --> 00:07:39,840 Speaker 1: be good enough for whatever it is we're building it for. 130 00:07:40,480 --> 00:07:46,160 Speaker 1: But there's another approach called unsupervised machine learning, and as 131 00:07:46,400 --> 00:07:50,040 Speaker 1: you might imagine, this is different from the previous one. 132 00:07:50,160 --> 00:07:53,520 Speaker 1: On this approach, you only have input data and your 133 00:07:53,520 --> 00:07:56,120 Speaker 1: goal as a human is to learn more about that 134 00:07:56,240 --> 00:07:59,640 Speaker 1: data itself. So you don't have a correct answer in mind. 135 00:08:00,040 --> 00:08:03,400 Speaker 1: You don't already know that the data represents, say a 136 00:08:03,520 --> 00:08:06,360 Speaker 1: cat in a photo. It's a different type of problem 137 00:08:06,400 --> 00:08:09,680 Speaker 1: you're looking at. Uh. The machine is learning about the 138 00:08:09,760 --> 00:08:13,600 Speaker 1: nature of the information itself, including how different points of 139 00:08:13,680 --> 00:08:17,360 Speaker 1: data relate to one another or correspond with other data, 140 00:08:17,680 --> 00:08:21,080 Speaker 1: and you in turn can learn more about the information 141 00:08:21,120 --> 00:08:24,000 Speaker 1: as well. So within this category you have a couple 142 00:08:24,240 --> 00:08:29,040 Speaker 1: of subcategories. There are clustering problems. With a clustering problem, 143 00:08:29,120 --> 00:08:32,800 Speaker 1: you're learning about the groupings within data. So one example 144 00:08:32,880 --> 00:08:35,079 Speaker 1: might be that you have a population of customers. Let's 145 00:08:35,080 --> 00:08:37,520 Speaker 1: say you own a business. You've got customers. You have 146 00:08:37,600 --> 00:08:40,920 Speaker 1: data that represents all these different customers, and you're using 147 00:08:40,920 --> 00:08:45,080 Speaker 1: the collective behaviors of those customers to sort them into 148 00:08:45,160 --> 00:08:48,360 Speaker 1: meaningful groups so that you can better serve each of 149 00:08:48,400 --> 00:08:52,600 Speaker 1: those groups. Maybe you learn that there are four basic 150 00:08:52,679 --> 00:08:55,439 Speaker 1: types of customers, and that helps you plan out your 151 00:08:55,440 --> 00:08:59,160 Speaker 1: business so that you can cater it to those four types. 152 00:09:00,000 --> 00:09:03,280 Speaker 1: But another type of problem in unsupervised machine learning is 153 00:09:03,280 --> 00:09:06,960 Speaker 1: called an association problem. Now, in those problems, you want 154 00:09:07,000 --> 00:09:09,880 Speaker 1: to learn rules that describe large parts of the data 155 00:09:09,960 --> 00:09:12,440 Speaker 1: that you're feeding into the system. So, for example, let's 156 00:09:12,440 --> 00:09:15,000 Speaker 1: go back to you run a business. You've got this 157 00:09:15,040 --> 00:09:17,880 Speaker 1: big pool of customers, and you're feeding all the customer 158 00:09:18,120 --> 00:09:22,280 Speaker 1: behavior data into your system. It might tell you that, hey, 159 00:09:22,520 --> 00:09:26,280 Speaker 1: it turns out that of the customers who are buying 160 00:09:26,880 --> 00:09:30,840 Speaker 1: widgets go on to buy sprockets. So that would tell you, hey, 161 00:09:31,000 --> 00:09:33,320 Speaker 1: now I know more information. I know that if I 162 00:09:33,360 --> 00:09:35,480 Speaker 1: sell a widget to someone, there's a good chance I 163 00:09:35,520 --> 00:09:38,600 Speaker 1: can upsell that and include a Sprocket as well. So 164 00:09:38,640 --> 00:09:41,760 Speaker 1: I'm going to tailor my business approach to try and 165 00:09:41,800 --> 00:09:44,440 Speaker 1: take advantage of that. Now, the reason I went through 166 00:09:44,480 --> 00:09:47,240 Speaker 1: all of this is to explain that the type of 167 00:09:47,320 --> 00:09:51,439 Speaker 1: artificial intelligence algorithm that was used to produce the painting 168 00:09:51,559 --> 00:09:53,360 Speaker 1: I was talking about at the top of the show, 169 00:09:53,920 --> 00:09:58,400 Speaker 1: falls into a group called generative adversarial networks or g 170 00:09:58,559 --> 00:10:03,120 Speaker 1: a N or a GAN. These are used in unsupervised 171 00:10:03,240 --> 00:10:07,400 Speaker 1: machine learning applications. So it's in that second category I 172 00:10:07,480 --> 00:10:11,360 Speaker 1: was just talking about. So what is with this name? 173 00:10:11,440 --> 00:10:17,720 Speaker 1: What is a generative adversarial network? Well, for one thing, 174 00:10:18,360 --> 00:10:23,120 Speaker 1: it actually uses a pair of deep neural net architecture networks. 175 00:10:23,600 --> 00:10:27,200 Speaker 1: These two nets are in competition with one another. That's 176 00:10:27,200 --> 00:10:31,480 Speaker 1: why it's called an adversarial network. You have these two 177 00:10:31,520 --> 00:10:37,800 Speaker 1: different constructs that are working against each other. The approach 178 00:10:37,880 --> 00:10:40,840 Speaker 1: was first proposed by researchers at the University of Montreal, 179 00:10:41,240 --> 00:10:44,559 Speaker 1: and we chiefly associate the concept with a guy named 180 00:10:44,600 --> 00:10:49,560 Speaker 1: Ian Goodfellow. Ian Goodfellow wrote the definitive paper on the 181 00:10:49,600 --> 00:10:53,559 Speaker 1: subject back in two thousand and fourteen, and it is fascinating. 182 00:10:53,679 --> 00:10:56,480 Speaker 1: So from a very high level, what's happening is that 183 00:10:57,160 --> 00:11:00,320 Speaker 1: you have a neural network called the generator and you 184 00:11:00,360 --> 00:11:04,120 Speaker 1: have a second year old network called the discriminator. So 185 00:11:04,280 --> 00:11:08,840 Speaker 1: you're feeding the discriminator your input data. Let's again go 186 00:11:08,960 --> 00:11:12,880 Speaker 1: with pictures of cats, So actual pictures of cats photographs 187 00:11:12,960 --> 00:11:16,040 Speaker 1: of cats. If you will, you're you're feeding photographs of 188 00:11:16,080 --> 00:11:20,360 Speaker 1: cats to the discriminator. The generator's job is to create 189 00:11:21,280 --> 00:11:26,120 Speaker 1: a an image that fools the discriminator into thinking that 190 00:11:26,120 --> 00:11:29,520 Speaker 1: that's a legitimate photograph of a cat, but in fact 191 00:11:29,600 --> 00:11:34,040 Speaker 1: it was created or generated by the generator. So you've 192 00:11:34,040 --> 00:11:37,760 Speaker 1: got two processes going on at the same time. The 193 00:11:37,840 --> 00:11:41,280 Speaker 1: generator is trying to create essentially a forgery or a counterfeit. 194 00:11:41,880 --> 00:11:46,720 Speaker 1: It's it's creating something from scratch to fool the discriminator 195 00:11:46,760 --> 00:11:50,880 Speaker 1: into thinking this is a legitimate piece of data from 196 00:11:50,920 --> 00:11:55,320 Speaker 1: the training data set. The discriminator is looking at each 197 00:11:55,360 --> 00:11:58,360 Speaker 1: image and thinking, all right, now does this represent a 198 00:11:58,440 --> 00:12:01,800 Speaker 1: real picture or is this something that is coming from 199 00:12:01,840 --> 00:12:04,800 Speaker 1: the generator that's designed to fool me, And the two 200 00:12:04,880 --> 00:12:08,240 Speaker 1: are working against each other. Both networks learn as this 201 00:12:08,280 --> 00:12:11,199 Speaker 1: goes on. If the discriminator gets an image and rejects it, 202 00:12:11,720 --> 00:12:15,160 Speaker 1: that becomes a feedback to the generator and the messages. Essentially, 203 00:12:15,800 --> 00:12:18,360 Speaker 1: this was not good enough, and the generator starts to 204 00:12:18,800 --> 00:12:22,960 Speaker 1: try again, taking a slightly different approach. If the discriminator 205 00:12:23,040 --> 00:12:25,840 Speaker 1: accepts it, the generator says, ah ha, you're onto something. 206 00:12:26,160 --> 00:12:31,240 Speaker 1: But then you can tweak the discriminator and say this 207 00:12:31,320 --> 00:12:33,440 Speaker 1: was wrong. You you got this part wrong, and it 208 00:12:33,480 --> 00:12:36,000 Speaker 1: can start to try and look for signs that might 209 00:12:36,080 --> 00:12:40,440 Speaker 1: otherwise fool it. The goal here is that you are 210 00:12:40,480 --> 00:12:44,320 Speaker 1: going to have a generator producing better and better versions 211 00:12:44,480 --> 00:12:48,440 Speaker 1: of whatever it is you're trying to create. And that 212 00:12:48,520 --> 00:12:52,520 Speaker 1: could be a picture, it could be text, it could 213 00:12:52,559 --> 00:12:56,400 Speaker 1: be music. You could feed any sort of data to 214 00:12:56,559 --> 00:13:00,520 Speaker 1: both of these systems in an effort to deuce a 215 00:13:00,600 --> 00:13:05,080 Speaker 1: computer generated version of that thing, and as long as 216 00:13:05,080 --> 00:13:08,680 Speaker 1: it reached a certain level of quality, the discriminator won't 217 00:13:08,679 --> 00:13:10,600 Speaker 1: be able to tell the difference, and then you've got 218 00:13:10,640 --> 00:13:14,480 Speaker 1: yourself a computer generated whatever it might be, in this case, 219 00:13:15,200 --> 00:13:18,480 Speaker 1: a painting. I'll explain more about the specifics of this 220 00:13:18,559 --> 00:13:20,560 Speaker 1: case in just a moment, but first let's take a 221 00:13:20,640 --> 00:13:30,640 Speaker 1: quick break to thank our sponsor. So a couple of 222 00:13:30,720 --> 00:13:34,320 Speaker 1: years ago, there were computer scientists at Microsoft as well 223 00:13:34,360 --> 00:13:38,200 Speaker 1: as tu Deft University, and they were working together with 224 00:13:38,280 --> 00:13:41,080 Speaker 1: a banking company I n G to create a brand 225 00:13:41,080 --> 00:13:45,440 Speaker 1: new painting in the style of the painter Rembrandt. This 226 00:13:45,559 --> 00:13:49,640 Speaker 1: project involved processing high resolution digital scans of three hundred 227 00:13:49,840 --> 00:13:56,320 Speaker 1: forty six different images of Rembrandt's works, specifically portraits of men. 228 00:13:56,840 --> 00:14:00,480 Speaker 1: That information was fed to a deep learning algorithm that 229 00:14:00,600 --> 00:14:05,560 Speaker 1: analyzed Rembrandt's style and also the techniques that were common 230 00:14:05,640 --> 00:14:08,160 Speaker 1: across all the images. What were the common elements that 231 00:14:08,200 --> 00:14:12,600 Speaker 1: were found in those numerous paintings, And eventually this machine 232 00:14:12,679 --> 00:14:15,480 Speaker 1: was told, or this system was told to produce a 233 00:14:15,600 --> 00:14:20,680 Speaker 1: new painting based on those uh those common factors. And 234 00:14:20,720 --> 00:14:23,160 Speaker 1: so it narrowed down the approach to be a portrait 235 00:14:23,480 --> 00:14:26,320 Speaker 1: of a Caucasian white male because that's what most of 236 00:14:26,360 --> 00:14:30,760 Speaker 1: Rembrandt's portraits were of, somewhere between the ages of thirty 237 00:14:30,760 --> 00:14:33,720 Speaker 1: and forty, wearing white and black clothing, because again that 238 00:14:33,800 --> 00:14:37,600 Speaker 1: was the vast majority of the portraits that Rembrandt created, 239 00:14:37,880 --> 00:14:42,640 Speaker 1: and the focus of the subject was off to the right, 240 00:14:42,720 --> 00:14:46,280 Speaker 1: like looking slightly off to the right, because a lot 241 00:14:46,280 --> 00:14:48,480 Speaker 1: of the subjects in the other paintings were doing the same. 242 00:14:49,040 --> 00:14:52,520 Speaker 1: The algorithm also analyzed the faces of all those portraits 243 00:14:52,520 --> 00:14:54,560 Speaker 1: and came up was sort of a kind of a 244 00:14:54,640 --> 00:14:57,600 Speaker 1: mishmash average of them to produce the face of the 245 00:14:57,640 --> 00:15:00,840 Speaker 1: fictional Dutch gentleman in the new painting. To go a 246 00:15:00,880 --> 00:15:04,400 Speaker 1: step further, the team then added depth to this painting. 247 00:15:04,440 --> 00:15:06,800 Speaker 1: It was a two dimensional image, and then they decided 248 00:15:06,840 --> 00:15:09,000 Speaker 1: to add some depth. They included some ridges and some 249 00:15:09,080 --> 00:15:13,080 Speaker 1: bumps that would have been created from brush strokes onto 250 00:15:13,240 --> 00:15:17,320 Speaker 1: a two dimensional surface. So if you're using paint, then 251 00:15:17,560 --> 00:15:19,800 Speaker 1: it's actually a three dimensional image. You know, if you 252 00:15:19,840 --> 00:15:23,120 Speaker 1: get super close enough, you can see raised areas and 253 00:15:23,560 --> 00:15:26,680 Speaker 1: dips and trenches and stuff like that that the brush leaves. 254 00:15:26,800 --> 00:15:31,560 Speaker 1: And it all depends upon your painting technique how these 255 00:15:31,600 --> 00:15:35,280 Speaker 1: get laid out on canvas. So the team added those 256 00:15:35,320 --> 00:15:39,640 Speaker 1: details in to make it look even more authentic. Ultimately, 257 00:15:39,720 --> 00:15:43,640 Speaker 1: the design was printed using thirteen layers of ultra violet 258 00:15:43,720 --> 00:15:46,840 Speaker 1: based inc and the result is a work that looks 259 00:15:46,880 --> 00:15:49,600 Speaker 1: like it could have come from Rembrandt, complete with techniques 260 00:15:49,600 --> 00:15:53,760 Speaker 1: Rembrandt used in actually making his brushstrokes. And that's just 261 00:15:53,880 --> 00:15:57,480 Speaker 1: one high profile example of computers generating paintings after being 262 00:15:57,520 --> 00:16:01,040 Speaker 1: fed information about works that human artists have created. Now, 263 00:16:01,040 --> 00:16:05,040 Speaker 1: as get back to the story of the recently auctioned painting. Now, 264 00:16:05,600 --> 00:16:07,440 Speaker 1: to do that, we have to talk about a young 265 00:16:07,480 --> 00:16:12,560 Speaker 1: man named Robbie Barrett. Barrett is nineteen years old and 266 00:16:12,680 --> 00:16:16,200 Speaker 1: is attending Stanford and has been doing some really interesting 267 00:16:16,240 --> 00:16:19,840 Speaker 1: work in machine learning. It was his code that would 268 00:16:19,840 --> 00:16:22,640 Speaker 1: be the basis for the computer generated portrait that was 269 00:16:22,760 --> 00:16:26,040 Speaker 1: recently auctioned off. Barrett's work was going a step further 270 00:16:26,560 --> 00:16:30,800 Speaker 1: than copying the style of an established artist. Barrett's algorithms 271 00:16:30,920 --> 00:16:34,640 Speaker 1: would work to create new images after having analyzed numerous 272 00:16:34,720 --> 00:16:38,120 Speaker 1: real world examples. So just a couple of years ago, 273 00:16:38,480 --> 00:16:42,000 Speaker 1: the state of the art in GAN networks or GN 274 00:16:42,040 --> 00:16:46,680 Speaker 1: networks might produce some really disturbing images, like there are 275 00:16:46,720 --> 00:16:50,200 Speaker 1: early pictures of GAN attempts at making realistic human faces 276 00:16:50,680 --> 00:16:53,960 Speaker 1: that were not terribly successful, and that's because those networks 277 00:16:53,960 --> 00:16:57,560 Speaker 1: were able to recognize certain basic visual elements and images, 278 00:16:58,160 --> 00:17:02,880 Speaker 1: but not understand the reation ships between multiple elements within 279 00:17:02,960 --> 00:17:05,200 Speaker 1: an image, so you could end up with a face 280 00:17:05,480 --> 00:17:11,040 Speaker 1: with really extreme features like pronounced asymmetry. But over just 281 00:17:11,080 --> 00:17:13,040 Speaker 1: a short amount of time, people have developed much more 282 00:17:13,040 --> 00:17:17,160 Speaker 1: sophisticated GAN algorithms and performance has improved, and there of 283 00:17:17,160 --> 00:17:20,440 Speaker 1: course artists who have gone in a different approach, specifically 284 00:17:21,240 --> 00:17:25,600 Speaker 1: emphasizing some of these more absurd elements in order to 285 00:17:25,640 --> 00:17:29,920 Speaker 1: get that kind of a result when you're actually producing art. 286 00:17:30,560 --> 00:17:33,439 Speaker 1: Verrett created GAN algorithms that could generate all sorts of 287 00:17:33,440 --> 00:17:37,800 Speaker 1: interesting images. He was enabling computers to make art themselves. 288 00:17:38,240 --> 00:17:41,760 Speaker 1: And sure, these computers were learning to create art after 289 00:17:41,800 --> 00:17:45,679 Speaker 1: being fed numerous paintings and images from human artists. But 290 00:17:45,800 --> 00:17:47,920 Speaker 1: you could argue that if you want to become a 291 00:17:48,000 --> 00:17:50,639 Speaker 1: human artist, you have to do the same thing. You 292 00:17:50,680 --> 00:17:53,240 Speaker 1: have to study art that was created by other people. 293 00:17:53,440 --> 00:17:57,960 Speaker 1: So computers are no different. The computers weren't replicating specific works, 294 00:17:57,960 --> 00:18:00,840 Speaker 1: they weren't trying to make a copy. They were learning 295 00:18:01,160 --> 00:18:07,280 Speaker 1: various styles. Barrett would frequently put these images and also 296 00:18:07,320 --> 00:18:10,439 Speaker 1: the algorithms he used to create those images up on 297 00:18:10,560 --> 00:18:13,920 Speaker 1: get hub for free and open source. He also had 298 00:18:14,560 --> 00:18:18,760 Speaker 1: uh people download these and upload their own art, and 299 00:18:18,800 --> 00:18:21,720 Speaker 1: it was all in the spirit of this open source community. 300 00:18:23,200 --> 00:18:25,439 Speaker 1: This way, not only could people use the tools that 301 00:18:25,480 --> 00:18:28,399 Speaker 1: Barrett had created, they could understand how those tools worked, 302 00:18:28,840 --> 00:18:31,440 Speaker 1: and perhaps in the future they can make their own tools, 303 00:18:32,000 --> 00:18:36,640 Speaker 1: tweaking the approach the Barrett had used, maybe making art 304 00:18:36,720 --> 00:18:41,639 Speaker 1: that was even more indistinguishable from human art, or perhaps 305 00:18:41,640 --> 00:18:44,760 Speaker 1: going in a totally different direction, making something truly new 306 00:18:44,760 --> 00:18:47,560 Speaker 1: and alien. By the way, some of the images created 307 00:18:47,560 --> 00:18:51,320 Speaker 1: by Barrett's algorithms are a little unsettling. They can be 308 00:18:51,440 --> 00:18:54,359 Speaker 1: surreal and absurd, and some of them even come across 309 00:18:54,359 --> 00:18:58,000 Speaker 1: a little sinister to me. But that's my own interpretation. 310 00:18:58,040 --> 00:18:59,919 Speaker 1: I mean, that is what art is all about, is 311 00:19:00,040 --> 00:19:02,520 Speaker 1: the interpretation of the person looking at art. But they 312 00:19:02,560 --> 00:19:05,679 Speaker 1: remind me of some of the horror movie effects you 313 00:19:05,760 --> 00:19:08,639 Speaker 1: might see where the visual effects artists will distort a 314 00:19:08,680 --> 00:19:10,840 Speaker 1: person's face for the effect of horror, like in the 315 00:19:10,880 --> 00:19:16,000 Speaker 1: movie The Ring. Anyway, Barrett created several GAN algorithms and 316 00:19:16,040 --> 00:19:18,600 Speaker 1: put them up online for others to use, and this 317 00:19:18,720 --> 00:19:21,520 Speaker 1: in itself was not unusual. There are many in the 318 00:19:21,560 --> 00:19:24,800 Speaker 1: digital art field who work on AI who have done 319 00:19:24,840 --> 00:19:29,760 Speaker 1: similar things. Now he creates this code, Let's take a 320 00:19:29,800 --> 00:19:33,160 Speaker 1: trip across the world from Stanford over to France. That's 321 00:19:33,160 --> 00:19:37,000 Speaker 1: where three artists in their mid twenties were working in 322 00:19:37,040 --> 00:19:40,920 Speaker 1: a group they had called Obvious and their stated goal 323 00:19:41,119 --> 00:19:45,280 Speaker 1: is to promote ganism, that is, the art that has 324 00:19:45,320 --> 00:19:50,040 Speaker 1: been generated through AI algorithms running on this GAN approach. Now, 325 00:19:50,080 --> 00:19:53,159 Speaker 1: according to an article on Medium written by one of 326 00:19:53,200 --> 00:19:57,520 Speaker 1: these artists, they quote want to send out an update 327 00:19:57,640 --> 00:20:00,600 Speaker 1: of the state of the research and AI end quote. 328 00:20:01,200 --> 00:20:03,879 Speaker 1: They want to do this they want to tell the 329 00:20:03,880 --> 00:20:06,560 Speaker 1: world what is going on in the world of AI 330 00:20:06,680 --> 00:20:10,040 Speaker 1: research through showing off artwork made by AI, so kind 331 00:20:10,040 --> 00:20:14,159 Speaker 1: of a creative artistic way of talking about artificial intelligence. 332 00:20:14,960 --> 00:20:18,000 Speaker 1: The group says that the value of the art may 333 00:20:18,040 --> 00:20:21,280 Speaker 1: not be in the art itself, but rather the discussions 334 00:20:21,359 --> 00:20:25,040 Speaker 1: that the art inspires, like what is it that makes 335 00:20:25,240 --> 00:20:30,720 Speaker 1: art art? Can machines be creative? Who ultimately would you 336 00:20:30,800 --> 00:20:33,199 Speaker 1: say is the artist in a work that was created 337 00:20:33,240 --> 00:20:37,160 Speaker 1: by a machine? What does that art mean? Who does 338 00:20:37,160 --> 00:20:40,640 Speaker 1: it belong to? That's a big one. So the artists 339 00:20:40,720 --> 00:20:44,200 Speaker 1: reached out to Barrett when they were tackling this project. 340 00:20:44,560 --> 00:20:47,800 Speaker 1: They wanted to use a gain algorithm to generate a 341 00:20:47,840 --> 00:20:50,480 Speaker 1: portrait in a style similar to what you see in 342 00:20:50,600 --> 00:20:54,120 Speaker 1: eighteenth century paintings out of Europe. The students have made 343 00:20:54,119 --> 00:20:56,720 Speaker 1: it clear that Barrett had been a big part of 344 00:20:56,760 --> 00:20:59,880 Speaker 1: their inspiration. More on that in just a second now. 345 00:21:00,080 --> 00:21:03,440 Speaker 1: Members of Obvious began using gan code to generate portraits, 346 00:21:03,840 --> 00:21:06,760 Speaker 1: and they created several of them, eleven in fact of 347 00:21:06,800 --> 00:21:14,119 Speaker 1: a fictional noble family they named the Bellamy family B. E. L. A. M. Y. 348 00:21:14,280 --> 00:21:16,600 Speaker 1: The name Bellamy itself was a bit of a pun 349 00:21:16,720 --> 00:21:19,919 Speaker 1: and a reference to Ian Goodfellow, the guy who wrote 350 00:21:19,960 --> 00:21:23,520 Speaker 1: that main paper on gangs. In the first place, Bellamy 351 00:21:23,680 --> 00:21:27,920 Speaker 1: can be broken down into bell and Amy. That would 352 00:21:27,920 --> 00:21:30,679 Speaker 1: mean all the different spellings. It would mean good friend 353 00:21:30,880 --> 00:21:34,120 Speaker 1: or good fellow, which is kind of cute. Right. Well, 354 00:21:34,119 --> 00:21:38,320 Speaker 1: the artists produced these portraits, and they are all of 355 00:21:38,440 --> 00:21:42,680 Speaker 1: hollow eyed nobles that will stare right into the void 356 00:21:42,720 --> 00:21:45,920 Speaker 1: in a way that actually that's getting off track. Never 357 00:21:45,960 --> 00:21:48,440 Speaker 1: mind it. It creates me out a little bit. But 358 00:21:48,520 --> 00:21:51,600 Speaker 1: the last in the line of portraits would be Edmund 359 00:21:51,840 --> 00:21:55,480 Speaker 1: do Bellamy, the fictional noble whose portrait would go up 360 00:21:55,520 --> 00:22:00,640 Speaker 1: on auction in October and fetched way more money than 361 00:22:00,920 --> 00:22:05,160 Speaker 1: was anticipated and so obvious had fed to the algorithms 362 00:22:05,600 --> 00:22:09,520 Speaker 1: numerous paintings from the eighteenth century to guide its efforts, 363 00:22:10,160 --> 00:22:13,160 Speaker 1: and once they started producing these, they had each one 364 00:22:13,240 --> 00:22:16,680 Speaker 1: signed with a line of code referencing the algorithm. They 365 00:22:16,760 --> 00:22:20,600 Speaker 1: framed the machine generated portraits in golden frames, and when 366 00:22:20,760 --> 00:22:23,720 Speaker 1: Edmund de Bellamy went up for auction, the best guess 367 00:22:23,720 --> 00:22:26,320 Speaker 1: was that it would probably fetch between seven thousand and 368 00:22:26,359 --> 00:22:31,040 Speaker 1: eleven thousand dollars. Instead, the winning bid was for more 369 00:22:31,119 --> 00:22:37,199 Speaker 1: than four hundred thirty thousand dollars. So that raises a 370 00:22:37,280 --> 00:22:41,400 Speaker 1: good question who the heck should get that money. Who 371 00:22:41,600 --> 00:22:46,840 Speaker 1: was responsible for this painting and that would become something 372 00:22:46,960 --> 00:22:49,480 Speaker 1: of a controversy. I'll explain more in just a second, 373 00:22:49,520 --> 00:22:52,600 Speaker 1: but first let's take another quick break to thank our sponsor. 374 00:23:00,560 --> 00:23:04,639 Speaker 1: So as the group Obvious was getting press coverage for 375 00:23:04,720 --> 00:23:08,040 Speaker 1: the AI produced Bellamy portraits, this is before they had 376 00:23:08,119 --> 00:23:11,720 Speaker 1: even put one up for auction, some people, including Barratt, 377 00:23:12,920 --> 00:23:17,239 Speaker 1: express some disappointment with the group. They said that it 378 00:23:17,280 --> 00:23:21,480 Speaker 1: looked like they had used Barrett's code to produce these portraits, 379 00:23:21,520 --> 00:23:24,440 Speaker 1: and yet they weren't quick to attribute him. They didn't 380 00:23:24,440 --> 00:23:29,560 Speaker 1: give him credit, at least not readily and not visibly 381 00:23:29,760 --> 00:23:33,440 Speaker 1: in a lot of locations. And so his code, while 382 00:23:33,440 --> 00:23:36,920 Speaker 1: it was open source and he didn't begrudge anyone from 383 00:23:37,119 --> 00:23:40,240 Speaker 1: being able to use it, would have usually meant that 384 00:23:40,240 --> 00:23:44,360 Speaker 1: people would give him credit. Typically in the open source community, 385 00:23:44,359 --> 00:23:47,679 Speaker 1: it's considered bad form or even ghosh if you prefer 386 00:23:48,040 --> 00:23:52,080 Speaker 1: to not give credit where credit is due. As to 387 00:23:52,119 --> 00:23:55,880 Speaker 1: how much of the code was actually used unaltered, that 388 00:23:56,200 --> 00:23:58,840 Speaker 1: is a bit of an open question. The artists that 389 00:23:58,920 --> 00:24:01,600 Speaker 1: Obvious have admitted that they did use his code and 390 00:24:01,640 --> 00:24:05,520 Speaker 1: they changed it a little bit. Some other artists say 391 00:24:05,520 --> 00:24:09,560 Speaker 1: they believe that or more of the code was unaltered. 392 00:24:10,200 --> 00:24:13,200 Speaker 1: One such artist, a New Zealander named Tom White, said 393 00:24:13,240 --> 00:24:17,280 Speaker 1: he downloaded Barrett's code and ran it unaltered to see 394 00:24:17,280 --> 00:24:20,640 Speaker 1: if he could produce images similar to those that Obvious 395 00:24:20,680 --> 00:24:24,439 Speaker 1: had generated, and he said they look pretty close. So 396 00:24:24,480 --> 00:24:26,320 Speaker 1: I took a look at as well. I would say 397 00:24:26,400 --> 00:24:29,760 Speaker 1: that the ones that that White had created with that 398 00:24:29,880 --> 00:24:33,160 Speaker 1: AI have a little bit more of the weird facial 399 00:24:33,240 --> 00:24:35,520 Speaker 1: distortion thing going on than the ones that were made 400 00:24:35,560 --> 00:24:41,080 Speaker 1: by Obvious, but they are fairly similar. Throughout the project, 401 00:24:41,440 --> 00:24:44,280 Speaker 1: members of Obvious reached out to brot to for for 402 00:24:44,400 --> 00:24:48,119 Speaker 1: help and getting the GAN algorithms to run properly on computers. 403 00:24:48,480 --> 00:24:50,840 Speaker 1: Those communications are up on geth hubs, so I mean 404 00:24:51,440 --> 00:24:54,679 Speaker 1: they definitely happened. Anyone can see them. So that's definitely 405 00:24:54,720 --> 00:24:57,720 Speaker 1: a sign that a significant portion of the code used 406 00:24:57,720 --> 00:25:01,879 Speaker 1: to create the expensive painting came from ROT. So we 407 00:25:01,920 --> 00:25:06,200 Speaker 1: get into that tricky question who owns the art before 408 00:25:06,400 --> 00:25:10,440 Speaker 1: it gets purchased at auction? Obviously, so does the computer 409 00:25:10,520 --> 00:25:14,440 Speaker 1: scientist who created the code own anything that the code produces. 410 00:25:15,000 --> 00:25:17,320 Speaker 1: I mean, the code has to have a programmer. Without 411 00:25:17,359 --> 00:25:20,960 Speaker 1: a programmer, there's no code. So without the code, you 412 00:25:21,000 --> 00:25:25,080 Speaker 1: get no artistic output. But then again, you could say 413 00:25:25,119 --> 00:25:28,840 Speaker 1: that human artists learn from their teachers. There's a long 414 00:25:29,000 --> 00:25:33,200 Speaker 1: history of artists taking on apprentices, and those apprentices later 415 00:25:33,240 --> 00:25:35,920 Speaker 1: on go on to become great artists of their own. 416 00:25:36,480 --> 00:25:38,840 Speaker 1: So maybe you could argue that Brought was a teacher 417 00:25:39,200 --> 00:25:43,440 Speaker 1: and the AI was the student, and therefore Brought wouldn't 418 00:25:43,440 --> 00:25:46,080 Speaker 1: own the art. He didn't make it. He just taught 419 00:25:46,119 --> 00:25:50,119 Speaker 1: the student how to make art, not in a traditional sense, 420 00:25:50,359 --> 00:25:56,359 Speaker 1: but that's how it happened. But here's another problem. AI 421 00:25:56,520 --> 00:26:01,199 Speaker 1: cannot own stuff. Artificial intelligence can't have property. We have 422 00:26:01,280 --> 00:26:05,560 Speaker 1: no legal means to assign ownership, so that a program, 423 00:26:05,640 --> 00:26:09,920 Speaker 1: or an algorithm or an artificial neural network could own property. 424 00:26:10,000 --> 00:26:12,239 Speaker 1: And even if we did, what good would it do. 425 00:26:12,400 --> 00:26:16,000 Speaker 1: The AI doesn't want or need anything. It doesn't even 426 00:26:16,040 --> 00:26:21,000 Speaker 1: have will or self awareness. So maybe Obvious could claim 427 00:26:21,040 --> 00:26:25,199 Speaker 1: ownership because they were the ones who fed the information 428 00:26:25,240 --> 00:26:28,520 Speaker 1: to the algorithm. They're the ones who gave the algorithm 429 00:26:28,640 --> 00:26:31,880 Speaker 1: the access to all the different portraits. They made some 430 00:26:32,040 --> 00:26:35,520 Speaker 1: changes to the code, and the algorithms ran on computers 431 00:26:35,560 --> 00:26:40,080 Speaker 1: that they controlled, so if the code was using their assets, 432 00:26:40,600 --> 00:26:43,760 Speaker 1: maybe they own the output. But this is also complicated. 433 00:26:43,800 --> 00:26:46,800 Speaker 1: They didn't build the algorithm. They made use of it, 434 00:26:47,240 --> 00:26:50,639 Speaker 1: but they didn't design it from the ground up. But 435 00:26:50,680 --> 00:26:53,199 Speaker 1: if someone else could have run the code and use 436 00:26:53,320 --> 00:26:56,560 Speaker 1: the same general pool of images and train the code, 437 00:26:56,840 --> 00:27:00,880 Speaker 1: they might have seen similar results, which means someone else 438 00:27:00,880 --> 00:27:03,480 Speaker 1: could have done the exact same thing that obvious did, 439 00:27:03,800 --> 00:27:08,359 Speaker 1: and so that raises questions as well. Maybe there's nothing 440 00:27:08,359 --> 00:27:11,919 Speaker 1: special about owning the machine. In other words, in the 441 00:27:11,960 --> 00:27:15,920 Speaker 1: digital world, using open source code to make something new 442 00:27:16,000 --> 00:27:19,240 Speaker 1: and then profit from it sell it. That happens regularly, 443 00:27:19,320 --> 00:27:21,760 Speaker 1: but again it's all on how you do it. If 444 00:27:21,800 --> 00:27:25,200 Speaker 1: you follow the general rules of etiquette, you're typically pretty good. 445 00:27:25,400 --> 00:27:28,000 Speaker 1: But if not, people think of that as being kind 446 00:27:28,000 --> 00:27:33,480 Speaker 1: of a jerk face. So it's not it's it's frowned 447 00:27:33,560 --> 00:27:37,879 Speaker 1: upon in the open source community. Broad is quoted in 448 00:27:37,920 --> 00:27:40,600 Speaker 1: a piece on The Verge as saying, quote, I'm more 449 00:27:40,680 --> 00:27:44,360 Speaker 1: concerned about the fact that actual artists using AI are 450 00:27:44,359 --> 00:27:47,760 Speaker 1: being deprived of the spotlight. It's a very bad first 451 00:27:47,800 --> 00:27:51,520 Speaker 1: impression for the field to have end quote. So he's 452 00:27:51,520 --> 00:27:55,280 Speaker 1: not saying he's upset and missing out on money, but 453 00:27:55,880 --> 00:28:00,920 Speaker 1: rather that the the whole field is getting is represented 454 00:28:01,520 --> 00:28:03,920 Speaker 1: The Verge piece also does a great job pointing out 455 00:28:04,000 --> 00:28:07,520 Speaker 1: how many in the AI digital art field feel that 456 00:28:07,600 --> 00:28:11,160 Speaker 1: Obvious is painting a misleading picture to use a pun 457 00:28:11,680 --> 00:28:13,720 Speaker 1: that if you were to look at the press release 458 00:28:14,520 --> 00:28:16,520 Speaker 1: that the group has put out and the way that 459 00:28:16,600 --> 00:28:19,360 Speaker 1: they've presented the art, it would seem as if these 460 00:28:19,359 --> 00:28:23,840 Speaker 1: programs were largely undirected or even fully autonomous, and they aren't. 461 00:28:24,440 --> 00:28:27,679 Speaker 1: Just because it's called unsupervised machine learning doesn't mean that 462 00:28:27,720 --> 00:28:31,080 Speaker 1: there's no human component. So there's a debate going on 463 00:28:31,840 --> 00:28:35,480 Speaker 1: within the digital art world on where in the spectrum 464 00:28:36,080 --> 00:28:40,960 Speaker 1: these algorithms should fall. Are they closer to being tools 465 00:28:41,000 --> 00:28:44,640 Speaker 1: like what a paint brush would be to a traditional painter, 466 00:28:45,520 --> 00:28:49,960 Speaker 1: or are they more closely connected to a collaborator, maybe 467 00:28:50,080 --> 00:28:53,680 Speaker 1: someone who's assisting a painter. But they certainly are not 468 00:28:53,800 --> 00:28:57,200 Speaker 1: fully autonomous robots. Now. In a way, this question of 469 00:28:57,240 --> 00:29:00,920 Speaker 1: ownership actually makes me think of an earlier incident involving 470 00:29:00,920 --> 00:29:05,760 Speaker 1: a different art form. It involved a monkey, a digital camera, 471 00:29:06,120 --> 00:29:09,000 Speaker 1: and a lawsuit. So back in two thousand and eleven, 472 00:29:09,240 --> 00:29:13,080 Speaker 1: a photographer named David Slater was working on an assignment 473 00:29:13,120 --> 00:29:18,000 Speaker 1: in Indonesia and that's where he met Naruto Naruto was 474 00:29:18,040 --> 00:29:22,520 Speaker 1: a seven year old crested macaque, so Naruto was a 475 00:29:22,520 --> 00:29:27,520 Speaker 1: monkey now. On this assignment, Naruto at one point grabbed 476 00:29:27,680 --> 00:29:32,200 Speaker 1: Slater's camera, and while handling Slater's camera, Naruto took a 477 00:29:32,280 --> 00:29:36,320 Speaker 1: photo of himself. So it's a monkey selfie, and it's 478 00:29:36,360 --> 00:29:38,480 Speaker 1: a great photo. If you've not seen it, you've got 479 00:29:38,480 --> 00:29:42,760 Speaker 1: to look up monkey selfie because it is amazing. The 480 00:29:42,800 --> 00:29:45,360 Speaker 1: monkey obviously didn't understand what it was doing, but the 481 00:29:45,400 --> 00:29:50,080 Speaker 1: selfie is just about perfect. So then this image goes 482 00:29:50,160 --> 00:29:53,480 Speaker 1: up online and it goes viral. It gets posted all 483 00:29:53,520 --> 00:29:58,000 Speaker 1: over the place, including on Wikipedia, and David Slater would 484 00:29:58,000 --> 00:30:00,640 Speaker 1: reach out to Wikipedia and say, hey, you can't just 485 00:30:00,720 --> 00:30:03,160 Speaker 1: put my photograph up on your site without asking for 486 00:30:03,240 --> 00:30:07,560 Speaker 1: permission or paying a licensing fee. The Wikipedia said, dude, 487 00:30:08,160 --> 00:30:11,520 Speaker 1: you didn't take the photograph. It doesn't belong to you. 488 00:30:12,000 --> 00:30:14,600 Speaker 1: It was taken on your camera, but you didn't snap 489 00:30:14,640 --> 00:30:18,560 Speaker 1: the picture. A monkey took the photos, so you don't 490 00:30:18,560 --> 00:30:22,160 Speaker 1: have copyright to that image. In fact, no one has 491 00:30:22,200 --> 00:30:26,360 Speaker 1: copyright to that image because news flash, animals can't hold 492 00:30:26,360 --> 00:30:31,040 Speaker 1: copyrights to any work. But then Peter ak, a People 493 00:30:31,080 --> 00:30:34,479 Speaker 1: for the Ethical Treatment of Animals, would sue David Slater 494 00:30:34,680 --> 00:30:38,920 Speaker 1: and a publishing company called Blurb for copyright infringement, saying, Hey, 495 00:30:39,040 --> 00:30:42,360 Speaker 1: Naruto took that photo, so Naruto should hold the copyright. 496 00:30:42,720 --> 00:30:46,160 Speaker 1: The judge in that case would ultimately say that animals 497 00:30:46,200 --> 00:30:51,040 Speaker 1: can't hold copyright, backing up what Wikipedia had said, and 498 00:30:51,160 --> 00:30:55,480 Speaker 1: that this whole argument was invalid. Peter appealed the decision 499 00:30:55,760 --> 00:30:57,840 Speaker 1: it went to or it was scheduled to go to 500 00:30:58,080 --> 00:31:01,600 Speaker 1: a higher court, but ultimately the various parties came to 501 00:31:01,640 --> 00:31:04,560 Speaker 1: a settlement out of court. And this is where I 502 00:31:04,640 --> 00:31:08,680 Speaker 1: kind of roll my eyes at Peter. But this situation, 503 00:31:08,720 --> 00:31:12,800 Speaker 1: while silly on the surface, raises questions that also applied 504 00:31:12,800 --> 00:31:15,880 Speaker 1: to artificial intelligence. In a case like this, who has 505 00:31:15,920 --> 00:31:19,680 Speaker 1: the right to use or exploit a work? Now, I 506 00:31:19,680 --> 00:31:23,480 Speaker 1: would argue than the case with artificial intelligence, it gets 507 00:31:23,600 --> 00:31:27,280 Speaker 1: even thornier than that. Right now, we're talking about paintings. 508 00:31:27,560 --> 00:31:30,240 Speaker 1: But as I said earlier, gain algorithms could produce all 509 00:31:30,360 --> 00:31:33,760 Speaker 1: sorts of different stuff, including text. So we could have 510 00:31:33,880 --> 00:31:37,440 Speaker 1: a computer generated novel or a screenplay in the future, 511 00:31:37,840 --> 00:31:41,840 Speaker 1: and sure, the first versions of those will probably be terrible, 512 00:31:42,200 --> 00:31:45,280 Speaker 1: And to be fair, we already have a surplus of 513 00:31:45,480 --> 00:31:49,560 Speaker 1: terrible books and terrible movies and terrible TV shows that 514 00:31:49,600 --> 00:31:51,680 Speaker 1: are made by real human beings. We don't we don't 515 00:31:51,720 --> 00:31:54,960 Speaker 1: need robots to make more of those, but we could 516 00:31:55,000 --> 00:31:57,680 Speaker 1: also end up with some that are interesting or that 517 00:31:57,840 --> 00:32:01,840 Speaker 1: say something surprising that people will value. In those cases, 518 00:32:01,960 --> 00:32:04,800 Speaker 1: who has a claim to that intellectual property? Who should 519 00:32:04,840 --> 00:32:07,240 Speaker 1: profit from it? Maybe it should be the person who 520 00:32:07,240 --> 00:32:09,600 Speaker 1: wrote the code in the first place. But if that's 521 00:32:09,640 --> 00:32:12,880 Speaker 1: the case, let's take this thought experiment in another direction. 522 00:32:13,240 --> 00:32:15,880 Speaker 1: Let's say someone creates code for an AI that does 523 00:32:15,920 --> 00:32:20,800 Speaker 1: something entirely different. There it's not generating any content. Let's 524 00:32:20,800 --> 00:32:24,200 Speaker 1: say it's the artificial intelligence you would need to power 525 00:32:24,280 --> 00:32:28,080 Speaker 1: an autonomous car. Now, let's say one of those cars 526 00:32:28,240 --> 00:32:31,560 Speaker 1: is found to have caused a really bad accident. So 527 00:32:31,560 --> 00:32:34,360 Speaker 1: should the person who wrote the code be held responsible? 528 00:32:35,160 --> 00:32:38,160 Speaker 1: What if the scenario that led up to the accident 529 00:32:38,320 --> 00:32:41,640 Speaker 1: was so unusual that no one would have ever predicted it. 530 00:32:42,360 --> 00:32:45,320 Speaker 1: Because it's one thing to overlook a common event, Like 531 00:32:45,400 --> 00:32:49,520 Speaker 1: if someone were to program an autonomous car and say, oh, crap, 532 00:32:49,600 --> 00:32:54,000 Speaker 1: I totally forgot about stop signs, that would be demonstrably bad, 533 00:32:54,280 --> 00:32:57,280 Speaker 1: And you could say, well, that is that is endangerment, 534 00:32:57,440 --> 00:33:01,080 Speaker 1: That is definitely not cool. But it's a totally different 535 00:33:01,080 --> 00:33:04,840 Speaker 1: thing if you just don't predict an accident that involves 536 00:33:04,880 --> 00:33:08,720 Speaker 1: a lot of unique factors, because those happen too. There's 537 00:33:08,720 --> 00:33:12,120 Speaker 1: stuff that happens on the road every single day that 538 00:33:12,200 --> 00:33:16,000 Speaker 1: happens in a way that nobody anticipated. And because we 539 00:33:16,080 --> 00:33:19,640 Speaker 1: have so many people driving so many cars on so 540 00:33:19,680 --> 00:33:23,120 Speaker 1: many roads under so many conditions on a daily basis, 541 00:33:23,720 --> 00:33:26,560 Speaker 1: it's inevitable that we're going to have moments where those 542 00:33:26,680 --> 00:33:29,640 Speaker 1: unique situations pop up and it would be impossible to 543 00:33:29,920 --> 00:33:35,080 Speaker 1: identify or predict them. So in those cases, would you 544 00:33:35,200 --> 00:33:38,480 Speaker 1: still hold hold someone who made the code responsible that 545 00:33:38,560 --> 00:33:41,680 Speaker 1: they weren't able to predict something that nobody could predict? 546 00:33:41,960 --> 00:33:47,080 Speaker 1: Or does that put them at an unreasonable standard? Is 547 00:33:47,120 --> 00:33:50,280 Speaker 1: it the fault of the car manufacturer? Is it the 548 00:33:50,320 --> 00:33:53,280 Speaker 1: fault of the person who designed the road. I mean, 549 00:33:53,280 --> 00:33:56,520 Speaker 1: there's so many different questions and we don't have all 550 00:33:56,520 --> 00:34:00,320 Speaker 1: the answers, But I think in this case, with the painting, 551 00:34:00,800 --> 00:34:04,760 Speaker 1: we have this high profile example of AI producing something. 552 00:34:05,440 --> 00:34:08,520 Speaker 1: It leads us to get into a deeper conversation about 553 00:34:08,520 --> 00:34:11,759 Speaker 1: those ideas, and my guess is we will ultimately come 554 00:34:11,840 --> 00:34:17,000 Speaker 1: up with answers that are not entirely satisfactory for all situations, 555 00:34:17,040 --> 00:34:20,120 Speaker 1: but maybe some people will even go so far as 556 00:34:20,160 --> 00:34:26,160 Speaker 1: to to vehemently disagree with him. But more importantly, we 557 00:34:26,200 --> 00:34:30,399 Speaker 1: will actually have maybe answers right So, yeah, it might 558 00:34:30,400 --> 00:34:33,239 Speaker 1: be answers that not everyone is happy with, but at 559 00:34:33,320 --> 00:34:35,400 Speaker 1: least they would be answers right now we have nothing. 560 00:34:36,080 --> 00:34:39,680 Speaker 1: So this is a good case study for us to say, 561 00:34:39,920 --> 00:34:43,759 Speaker 1: we've got to start thinking about this stuff because the 562 00:34:43,840 --> 00:34:47,319 Speaker 1: era of AI playing a more pivotal role in our 563 00:34:47,360 --> 00:34:49,560 Speaker 1: lives is right around the corner, and it would be 564 00:34:49,560 --> 00:34:52,680 Speaker 1: better for us to figure this out now rather than 565 00:34:52,760 --> 00:34:55,520 Speaker 1: have to react to it when it's too late later. 566 00:34:55,920 --> 00:34:58,120 Speaker 1: I'm curious to hear what you guys have to say 567 00:34:58,160 --> 00:35:01,120 Speaker 1: about this subject. Why don't you pop on over to 568 00:35:01,280 --> 00:35:05,000 Speaker 1: text Stuff podcast dot com. That's our website. Get in 569 00:35:05,080 --> 00:35:07,239 Speaker 1: touch with me and let me know what you think. 570 00:35:07,560 --> 00:35:09,960 Speaker 1: If you have suggestions for future episodes of tech Stuff, 571 00:35:10,480 --> 00:35:12,560 Speaker 1: I'd love to hear those two. Make sure you go 572 00:35:12,600 --> 00:35:15,080 Speaker 1: over to t public dot com slash tech stuff. Check 573 00:35:15,120 --> 00:35:18,000 Speaker 1: out our our store. There lots of cool things over there. 574 00:35:18,239 --> 00:35:22,120 Speaker 1: Get yourself something fun for the holidays, because every purchase 575 00:35:22,160 --> 00:35:23,840 Speaker 1: you make goes to help the show, and I greatly 576 00:35:23,880 --> 00:35:27,719 Speaker 1: appreciate it, and I'll talk to you again really soon 577 00:35:33,600 --> 00:35:36,000 Speaker 1: for more on this and bathands of other topics, because 578 00:35:36,040 --> 00:35:47,200 Speaker 1: it how stuff works. Dot com