WEBVTT - Chocolate Chicken Chicken Cake 0:00:04.600 --> 0:00:08.360 Sleepwalkers is a production of our Heart Radio and unusual productions. 0:00:12.360 --> 0:00:16.079 So I'm here for a surprise poetry reading. It's about 0:00:16.120 --> 0:00:21.120 to start. The silence is hardly final. Somewhere in the street, 0:00:21.400 --> 0:00:24.600 I can see the trees begin to rise and fall 0:00:24.720 --> 0:00:27.319 for the light of the dark thing above me. The 0:00:27.400 --> 0:00:29.600 dream is like a shiny black hair, and the sun 0:00:29.760 --> 0:00:32.800 is like a dream. I stand up and watch the 0:00:32.800 --> 0:00:35.800 sun shine on a single day, and the sun has 0:00:35.840 --> 0:00:39.320 a chance to accomplish from the springs of my own delight. 0:00:40.760 --> 0:00:45.960 Kind of haunting, abstract, yes, but beautiful too. And crucially, 0:00:47.400 --> 0:00:50.360 when I read this, I felt, as you just did. 0:00:50.400 --> 0:00:53.120 I hope that it is beautiful. I found it evocative 0:00:53.240 --> 0:00:55.280 of experience that I've had it in the past. I 0:00:55.280 --> 0:00:58.279 found it nostalgic um and I'm like, oh my god, 0:00:58.320 --> 0:01:01.720 I'm having real human being emotions. That was filmmaker Oscar Sharp, 0:01:02.000 --> 0:01:05.639 and that poem wasn't written by him or by anyone else. 0:01:06.200 --> 0:01:10.119 It was written by a computer, a machine poet. We're 0:01:10.120 --> 0:01:13.160 more and more worried about robots coming to take our jobs. 0:01:13.400 --> 0:01:16.160 And though perhaps few would regret less hips to poets 0:01:16.160 --> 0:01:20.280 shambly around Brooklyn, somehow machines in the creative world are 0:01:20.480 --> 0:01:24.800 especially uncanny, even frightening, because poetry and music and humor 0:01:24.840 --> 0:01:27.560 are supposed to be the things that define our humanity, 0:01:27.600 --> 0:01:30.720 aren't they. In this episode, we look at how AI 0:01:30.920 --> 0:01:33.520 is being used in the creative arts, and in doing so, 0:01:33.959 --> 0:01:36.920 we understand a lot more about how this often intimidating 0:01:36.959 --> 0:01:55.520 technology actually works. I'm Moza Lush and welcome to Sleepwalkers. Hi, 0:01:56.040 --> 0:02:01.120 Hi Karen So did Oscar's poems spring your delight? I 0:02:01.160 --> 0:02:04.040 would have preferred it in your British act? Then I think, well, 0:02:04.040 --> 0:02:06.040 I can't blame you for that. It did remind me 0:02:06.080 --> 0:02:08.360 of Um. My friend is in a band called the 0:02:08.520 --> 0:02:12.040 x Ambassadors and they partnered with this producer called Alex 0:02:12.160 --> 0:02:15.040 the Kid who was asked by IBM to make a 0:02:15.080 --> 0:02:17.799 song using Watson. And the song is not bad. The 0:02:17.840 --> 0:02:21.200 song is not bad. Um, the sounds good. But basically, 0:02:21.840 --> 0:02:24.519 the way in which they used Watson was that they 0:02:24.680 --> 0:02:29.640 crunched the twenty six thousand songs from the top one charts. 0:02:30.400 --> 0:02:32.359 It's from over what time period, like over a few 0:02:32.400 --> 0:02:35.960 years presume, I don't know, But the point is that 0:02:36.000 --> 0:02:39.640 they used them to discover patterns in the songs. What 0:02:39.800 --> 0:02:42.120 makes the top hundred song. Basically that's right, and then 0:02:42.160 --> 0:02:44.560 that's reproduce it right, which is interesting because I think 0:02:44.560 --> 0:02:46.640 anybody could kind of tell you what makes the top 0:02:46.680 --> 0:02:49.040 on hundred song right. They're called earworms. But when you 0:02:49.040 --> 0:02:52.200 think about a data set of twenty six thousand, like, 0:02:52.320 --> 0:02:54.200 no human being can listen to that many songs and 0:02:54.400 --> 0:02:57.320 do any productive after they hear it. So in this episode, 0:02:57.360 --> 0:03:00.359 we're going to look at AI and art and different 0:03:00.400 --> 0:03:03.800 kinds of fields to really understand how computers crunched data 0:03:03.919 --> 0:03:06.840 to crack open this creative code. But first I want 0:03:06.880 --> 0:03:09.400 to go back to the algorithm who wrote the poem 0:03:09.440 --> 0:03:11.920 we heard at the beginning of the episode, because that 0:03:12.000 --> 0:03:15.239 algorithm also wrote the film that Steven Spielberg of the 0:03:15.280 --> 0:03:18.799 machine learning world, and the algorithm is called Benjamin. So 0:03:18.880 --> 0:03:22.080 we're going to meet Benjamin and a few people not 0:03:22.240 --> 0:03:26.480 named Benjamin. I was a speech trator for John Kerry, 0:03:27.040 --> 0:03:30.000 Tim Gaitner, and Barack Obama, um, not in that order, 0:03:30.080 --> 0:03:32.640 And I'm essentially a ghost writer and a photographer who 0:03:32.760 --> 0:03:37.040 learned to code one giant Frankenstein monster. That's Ross Goodwin 0:03:37.480 --> 0:03:41.240 and that Frankenstein monster it goes by Benjamin and it's 0:03:41.280 --> 0:03:44.400 the work of Ross Goodwin and Oscar Sharp, who you'll 0:03:44.400 --> 0:03:47.680 remember from that poetry reading, but neither of them actually 0:03:47.800 --> 0:03:51.800 named Benjamin. Well, they named itself, or rather, there was 0:03:51.840 --> 0:03:53.520 a piece of paper that came out of it that 0:03:53.600 --> 0:03:55.840 said my name is Benjamin on it. I read that 0:03:56.040 --> 0:03:58.360 in response to a question that was put to it, 0:03:58.760 --> 0:04:02.480 and a room full of people went, oh, a program 0:04:02.520 --> 0:04:06.080 that names itself is rather uncanny. And Oscar had been 0:04:06.120 --> 0:04:08.400 chasing that uncanny nous ever since he was at n 0:04:08.560 --> 0:04:11.280 y U for graduate school. Whenever I met anyone who 0:04:11.280 --> 0:04:14.160 good program, I would grab them by the pels and 0:04:14.280 --> 0:04:16.440 yell into the face. Can you can you build something 0:04:16.480 --> 0:04:19.520 that can write like people talk in some way? And 0:04:19.600 --> 0:04:22.720 one day in class Oscar notices there's this sneakerhead and 0:04:22.720 --> 0:04:25.000 he's sitting on his laptop and his laptop is writing 0:04:25.160 --> 0:04:30.080 without him touching it. And I'm like, oh, so we go, 0:04:30.320 --> 0:04:33.360 So we go for coffee, and that gut coffee was. 0:04:33.400 --> 0:04:35.560 It was a lengthy coffee. We're still having that coffee. 0:04:35.560 --> 0:04:39.719 We're still having such a cold coffee. By now, you 0:04:39.800 --> 0:04:41.960 might not believe it if it happened in a film, 0:04:42.080 --> 0:04:44.800 but Oscar had stumbled on exactly the person he was 0:04:44.839 --> 0:04:47.680 looking for. Oscar came to me and he said, I 0:04:47.720 --> 0:04:50.239 want to make a movie from a computer generated screenplay. 0:04:50.360 --> 0:04:52.120 And I said, you know, of course that sounds amazing. 0:04:52.200 --> 0:04:54.280 Let's do it. But let's figure out how we're going 0:04:54.279 --> 0:04:57.480 to generate the screenplay, because that's a nuanced process with 0:04:57.520 --> 0:04:59.679 lots of stabs, and we need to consider like every 0:04:59.680 --> 0:05:03.760 part it. So Oscar volunteered himself to teach me all 0:05:03.880 --> 0:05:07.800 the things about storytelling and narrative and filmmaking. He turned 0:05:07.800 --> 0:05:11.360 me onto like Vladimir prop Joseph campbell Um, all these 0:05:11.400 --> 0:05:16.000 theories of storytelling, and so they begin to experiment. I 0:05:16.040 --> 0:05:19.400 tried a bunch of prototypes that used like various structures 0:05:19.440 --> 0:05:23.159 that had been postulated by these theorists over time, and 0:05:23.360 --> 0:05:27.680 the output was not interesting. Despite following the rules laid 0:05:27.680 --> 0:05:31.680 out by narrative theorists, Ross couldn't get anything good just 0:05:31.760 --> 0:05:35.599 telling his programs what a story should contain. So a 0:05:35.680 --> 0:05:39.000 year passes, Oscar moves to l A and when I 0:05:39.040 --> 0:05:42.320 get this email from Ross and it's the results in 0:05:42.360 --> 0:05:43.960 one of those experiments that he wants me to read, 0:05:44.200 --> 0:05:47.240 and read it he did. Rossity mailed the poem from 0:05:47.240 --> 0:05:50.480 the beginning of the episode, the room is blown away 0:05:50.480 --> 0:05:53.240 from the door and the stones are beginning to shine. 0:05:53.760 --> 0:05:55.839 I immediately was like, oh my god, I don't know 0:05:55.839 --> 0:05:57.400 how he's doing this. But he said, I don't know 0:05:57.400 --> 0:06:00.120 what technology you're using right now, but can we it 0:06:00.200 --> 0:06:03.680 for screenplay? And so they did, and not just a screenplay, 0:06:03.720 --> 0:06:06.680 they actually produced a short film called Sunspring and they 0:06:06.720 --> 0:06:09.839 even got Thomas middle Ditch, the lead on Hbos Silaken Valley, 0:06:10.000 --> 0:06:15.599 to star in it. Principle is completely constructed. Of the 0:06:15.600 --> 0:06:18.320 same time, it's all about you. To be true, you 0:06:18.360 --> 0:06:20.279 didn't even watch the movie with the rest of the base. 0:06:20.360 --> 0:06:22.880 I don't know, I don't care. I know it's a 0:06:22.920 --> 0:06:26.279 consequence whatever you need to know about the presence of 0:06:26.320 --> 0:06:29.120 the story. I'm a little bit of a boy on 0:06:29.160 --> 0:06:32.680 the floor. So what do you think, Carol? It kind 0:06:32.680 --> 0:06:34.440 of reminds me of when my parents used to take 0:06:34.480 --> 0:06:37.520 me to like a bad production of Macbeth or as 0:06:37.560 --> 0:06:41.880 you like it. Traumatic. You're there and you're seven or eight, 0:06:42.200 --> 0:06:44.880 and you want to understand what's going on, and so 0:06:45.160 --> 0:06:47.640 you kind of pay as close attention as you possibly 0:06:47.680 --> 0:06:50.520 can to what the actors are doing because you have 0:06:50.720 --> 0:06:53.840 no idea what the dialogue means. Yeah, I mean that 0:06:53.920 --> 0:06:57.279 to me is quite impressive because a machine can create 0:06:57.360 --> 0:07:00.120 something which has enough of the elements in common the 0:07:00.240 --> 0:07:02.400 film that we can talk about a real film. You 0:07:02.440 --> 0:07:05.080 can't say it's not a film absolutely. Of course. What's 0:07:05.080 --> 0:07:07.680 different is it didn't take Benjamin very long at all 0:07:07.800 --> 0:07:10.800 to make it. Once you press the button fraction of 0:07:10.800 --> 0:07:12.760 a second there was a couple of seconds perpase, maybe 0:07:12.800 --> 0:07:14.480 maybe a couple of seconds total, actually a fraction of 0:07:14.480 --> 0:07:17.880 a second per page. That's right. After months of agonizing 0:07:17.920 --> 0:07:21.720 over centuries of storytelling theory, the final output only took 0:07:21.800 --> 0:07:25.640 a couple of seconds. So what was Ross's breakthrough? To 0:07:25.760 --> 0:07:28.080 understand we turned to one of the most famous AI 0:07:28.120 --> 0:07:32.760 scientists in the world, Sebastian Throne. Recently, something magical had 0:07:32.760 --> 0:07:37.800 happened recently. The feat has discovered was called machine learning. 0:07:38.080 --> 0:07:41.680 With AI, computers can now find their own rules. They 0:07:41.720 --> 0:07:45.440 are called neural networks. They're comprised of hundreds of millions 0:07:45.440 --> 0:07:48.280 of little vase sample processing units, and those units are 0:07:48.360 --> 0:07:51.800 modeled after what a neurons do in our physical brains. 0:07:52.040 --> 0:07:54.640 You just give them examples, very much like the way 0:07:54.680 --> 0:07:58.600 we we waste children. We don't give our children rules 0:07:58.640 --> 0:08:01.560 for every contingency in life. In the first eight years 0:08:01.560 --> 0:08:04.720 of education. We let them learn, They experience the world, 0:08:05.120 --> 0:08:07.880 and they loan behold. They make their own rules. And 0:08:07.880 --> 0:08:09.560 we are now in the world where computers can do 0:08:09.680 --> 0:08:12.360 the same thing. And this means machine learning can be 0:08:12.440 --> 0:08:16.160 used in all kinds of different fields. Sebastian himself applied 0:08:16.160 --> 0:08:19.240 the technology at Google, where he led the initial development 0:08:19.280 --> 0:08:21.840 of their self driving car. When you want to read 0:08:21.880 --> 0:08:24.720 a book, a book on like what the car should 0:08:24.760 --> 0:08:27.640 do in every situation, that rule book is really complicated 0:08:27.680 --> 0:08:29.680 and it can promise you no matter how many years 0:08:29.760 --> 0:08:33.040 you spent writing it, it's not gonna work. But when 0:08:33.080 --> 0:08:36.360 you give the machine the ability to learn its own rules, 0:08:36.640 --> 0:08:40.040 it is actually able to surpass how people can drive. 0:08:41.200 --> 0:08:45.079 We'll hear more from Sebastian later, but machine learning mL 0:08:45.520 --> 0:08:48.040 is the engine that drives almost all of the excitement 0:08:48.080 --> 0:08:52.280 about AI today, from identifying targets on the battlefield to 0:08:52.600 --> 0:08:56.640 understanding genetic diseases. And it's also what allowed Ross and 0:08:56.640 --> 0:08:59.959 Oscar to create a usable movie script. Rather than laying 0:09:00.000 --> 0:09:04.199 down storytelling rules, they simply showed Benjamin hundreds of examples 0:09:04.520 --> 0:09:12.160 and the algorithm found patterns and learned for itself more sleepwalkers. 0:09:12.360 --> 0:09:23.160 After the break, you're like, oh, did we essentially we 0:09:23.240 --> 0:09:27.000 teach this algorithm anything else about screenplay other than just 0:09:27.040 --> 0:09:30.280 putting in a bunch of screenplays, right, And that's the 0:09:30.320 --> 0:09:33.680 way that machine learning works. What is happening in a 0:09:33.760 --> 0:09:36.439 deep learning algorithm of this kind is it's building an 0:09:36.480 --> 0:09:41.199 extraordinarily complicated mathematical formula by reading all of this stuff 0:09:41.240 --> 0:09:43.439 over and over again, like the auto complete on your phone. 0:09:43.440 --> 0:09:46.160 The neural that is actually sampling from a probability distribution 0:09:46.200 --> 0:09:49.640 of which letters, bass or punctuation become next. So the 0:09:49.679 --> 0:09:53.240 script for sun Spring was essentially the most mathematically probable 0:09:53.280 --> 0:09:56.800 Sci Fi script except Ross and Oscar did have one 0:09:56.880 --> 0:10:01.040 important lever of creative control. The other of parameters that 0:10:01.080 --> 0:10:04.480 you're probably wondering about, there's one called the temperature is 0:10:04.480 --> 0:10:08.240 the riskiness of those next letter predictions. What Ross is 0:10:08.280 --> 0:10:12.200 describing is almost like a dial for creativity. Turn it 0:10:12.240 --> 0:10:14.680 up to a really high temperature, and the neural net 0:10:14.720 --> 0:10:17.880 is going to be extra creative and start making up words, 0:10:17.960 --> 0:10:22.600 babbling at a very high temperature. It's essentially drunk. Low temperature, 0:10:22.600 --> 0:10:25.200 it's going to be very repetitive and possible even begin 0:10:25.240 --> 0:10:28.760 to plagiarize its source material. So it'll be very repetitive. 0:10:28.760 --> 0:10:30.480 It'll be like the streets and the streets, and the 0:10:30.520 --> 0:10:34.400 streets and the streets. It's essentially went working for network television. Yeah, exactly. 0:10:34.480 --> 0:10:36.280 So we wanted it to be sort of in the middle. 0:10:37.000 --> 0:10:39.280 In the middle is where we found the best output 0:10:39.440 --> 0:10:44.280 and the most I think usable output, and Sunspring was born. 0:10:45.960 --> 0:10:49.320 So Benjamin Ross and Oscar right together now they write 0:10:49.320 --> 0:10:53.080 poetry and movies and sometimes what Benjamin spits out is good. 0:10:53.600 --> 0:10:55.360 Often they have to sift through it to find the 0:10:55.400 --> 0:10:59.440 best stuff. But he's prolific and he never ever suffers 0:10:59.440 --> 0:11:04.520 from writers. Look, so Kara was telling us earlier about 0:11:04.559 --> 0:11:07.320 Alex the kidd and using AI to make music, and 0:11:07.320 --> 0:11:09.199 that's something I want to understand a bit more about. 0:11:09.559 --> 0:11:13.640 So Julian went on a little bit of an expedition. Yes, 0:11:13.720 --> 0:11:15.920 I did. I've been seeing a lot of articles lately 0:11:15.960 --> 0:11:18.040 about AI and the arts, and I've been pretty curious 0:11:18.040 --> 0:11:20.760 about music specifically. We might take it for granted, but 0:11:20.920 --> 0:11:24.040 music is this primal emotional thing that's been with us forever. 0:11:24.080 --> 0:11:27.240 It might even predate language. But now Warner Music Group 0:11:27.320 --> 0:11:30.360 made history in April two nineteen, is the first major 0:11:30.440 --> 0:11:33.800 label to sign an AI to a record deal. Yeah, 0:11:33.840 --> 0:11:37.720 they signed this bot called Endel, which makes ambient noises 0:11:37.760 --> 0:11:39.840 based on where you are and what the weather is 0:11:39.880 --> 0:11:41.760 and what time of day it is. When I think 0:11:41.800 --> 0:11:44.240 of this kind of music, I think of those Spotify 0:11:44.320 --> 0:11:48.880 playlists like Peaceful Piano and Blissed Out Dinner Party, which 0:11:48.920 --> 0:11:52.440 would become extremely popular. It's not the same thing as Beyonce, No, 0:11:52.640 --> 0:11:56.280 definitely not. But Warner Music Group signed Endl to generate 0:11:56.320 --> 0:11:59.600 twenty albums of ambient music. And now that we live 0:11:59.600 --> 0:12:03.120 in a world where aies can get record deals, what 0:12:03.240 --> 0:12:05.559 does this mean for artists? What does this mean for 0:12:05.600 --> 0:12:08.920 even just music as we know it? Well, in my 0:12:09.000 --> 0:12:12.000 quest to find out, I visited this company called Amper. 0:12:12.640 --> 0:12:15.840 My name is Drew Silverstein. I am the co founder 0:12:15.920 --> 0:12:18.720 and CEO of Amper Music. Amper is an AI music 0:12:18.760 --> 0:12:22.120 company that Drew says will enable anyone to create music. 0:12:22.360 --> 0:12:24.320 In fact, the only things you need to know are 0:12:24.600 --> 0:12:27.160 the genre of music you want to create, the mood 0:12:27.440 --> 0:12:29.400 you'd like to convey, and the length of your piece 0:12:29.400 --> 0:12:31.760 of music. That's all you know. You can create a 0:12:31.760 --> 0:12:34.160 brand new, unique piece of music in a matter of seconds. 0:12:34.280 --> 0:12:39.679 So the big question is should musicians worry about computers 0:12:39.720 --> 0:12:43.400 taking their job? Well, let's try it and see. So 0:12:43.760 --> 0:12:48.200 what do you want to do? Cinematic, documentary, folk cinematic cinematic, 0:12:48.640 --> 0:12:53.960 minimal percussion or quirky percussion. It's rendering a song right now. 0:12:54.000 --> 0:13:00.520 And here we go. We've got something. I'm were deep 0:13:00.559 --> 0:13:04.439 in the forest of Nicaragua. There's a breed of jaguar. 0:13:05.360 --> 0:13:07.960 You might have heard of it. It's called the take 0:13:08.000 --> 0:13:12.480 a Killer panther. Look, here's the thing. I don't know 0:13:12.760 --> 0:13:17.439 what the difference between that and music is. I really don't. Yeah, 0:13:17.440 --> 0:13:20.760 so you're woud Yeah, all right, So there's that. And 0:13:21.000 --> 0:13:24.520 this isn't the only AI music app out there. Another 0:13:24.600 --> 0:13:28.440 major player is called Magenta, and big surprise, they're at Google. 0:13:28.720 --> 0:13:31.160 Magenta are using AI to create a ton of new tools. 0:13:31.240 --> 0:13:34.360 From a piano genie that makes it impossible to play 0:13:34.400 --> 0:13:37.439 bad notes to something that can generate drum loops, or 0:13:37.520 --> 0:13:40.560 something that can even play piano duets with you. You 0:13:40.559 --> 0:13:44.560 can even translate raw audio to a piano score. Raw audio, 0:13:44.640 --> 0:13:47.480 like if I play just something raw on the piano, 0:13:47.880 --> 0:13:53.760 raw audio like oh, literal raw audio, literal raw audio. 0:13:54.520 --> 0:13:57.040 And Magenta has also trained a neural network just like 0:13:57.360 --> 0:14:00.240 Ross and Oscar, only instead of sci fi scripts, they 0:14:00.320 --> 0:14:04.559 trained on over four hundred performances by skilled pianists. They 0:14:04.600 --> 0:14:07.080 fed it into the neural network and let me play 0:14:07.120 --> 0:14:09.480 one of the piano experts. First. This is a real 0:14:09.640 --> 0:14:17.440 piano player, all right, so nice? Right? Yeah, okay, ready 0:14:17.440 --> 0:14:28.680 for the AI. What that's a computer that was all 0:14:28.680 --> 0:14:30.920 a computer. I didn't ever play a human one. That's 0:14:30.920 --> 0:14:33.800 a computer that was trained by a human playing piano. 0:14:34.480 --> 0:14:37.080 And then how do you make a computer come up 0:14:37.120 --> 0:14:39.560 with that? Right? So even though it's not a screenplay, 0:14:39.600 --> 0:14:41.520 it's still data that you can feed a neural network 0:14:41.560 --> 0:14:44.440 with to find patterns. And in this case, Magenta used 0:14:44.440 --> 0:14:47.680 a data set from the Yamahai Piano competition. So human 0:14:47.720 --> 0:14:50.480 pianists played on these digital keyboards which recorded the nuances 0:14:50.480 --> 0:14:52.640 of their performance, like how long they hit notes, and 0:14:52.680 --> 0:14:54.960 it recorded all that information into a digital score that 0:14:55.000 --> 0:14:57.600 a computer could interpret. And we've actually had that technology 0:14:57.640 --> 0:14:59.840 for a while now, it's called MIDI. But training and 0:15:00.000 --> 0:15:02.080 on network on the data is new. See. The thing 0:15:02.120 --> 0:15:05.520 that I come back to is that a computer doesn't 0:15:05.600 --> 0:15:09.040 know it's playing music, so much of watching a musical 0:15:09.080 --> 0:15:13.480 performance is knowing that this is coming from someone who 0:15:13.560 --> 0:15:18.280 is emoting. Right, Yeah, there's actually there's an emotional communication happening, right, 0:15:18.360 --> 0:15:21.960 that's right. I do think though the future is not 0:15:22.200 --> 0:15:26.240 rejecting this. It's better to imagine what would Stravinsky have 0:15:26.440 --> 0:15:29.040 done with this kind of technology, because Stravinsky is still 0:15:29.080 --> 0:15:42.600 a musical genius. Right, Yeah, Definitely it's cool to listen 0:15:42.640 --> 0:15:45.840 to those musical examples of machine learning because you can 0:15:45.880 --> 0:15:50.720 really hear how the algorithm is reinterpreting existing material. Of course, 0:15:51.200 --> 0:15:54.040 listening to the output is one thing. Tasting it is 0:15:54.120 --> 0:15:58.440 quite another. The problem was that somebody had told me 0:15:58.480 --> 0:16:00.520 that they had made the recipe for Stan, that it 0:16:00.640 --> 0:16:03.360 was good, and what it was as a recipe called 0:16:03.440 --> 0:16:08.160 chocolate baked and serves. That's Janelle Shane. She's a research 0:16:08.200 --> 0:16:11.240 scientist and the author of a blog called AI Weirdness. 0:16:11.800 --> 0:16:14.080 She's talking about a recipe written by Ai that she 0:16:14.120 --> 0:16:19.680 actually cooked and eight. It starts out as a perfectly ordinary, 0:16:19.760 --> 0:16:24.000 flowerless chocolate brownie all the way until the very last ingredient, 0:16:24.400 --> 0:16:26.760 which is a cup of horse badish. I knew I 0:16:26.800 --> 0:16:28.480 was in trouble when I opened the oven door and 0:16:28.520 --> 0:16:33.760 my eyes just started watering. It was yeah, it was terrible. 0:16:34.080 --> 0:16:36.960 On her blog, Jenelle experiments with putting AI to a 0:16:37.080 --> 0:16:40.800 range of tasks, from writing new pickup lines to naming 0:16:40.840 --> 0:16:44.720 Halloween costumes, and often her experiments with machine learning are 0:16:44.720 --> 0:16:49.640 pretty revealing about us. It plays into this thought experiment, 0:16:49.760 --> 0:16:52.560 what would an alien think of our world? It takes 0:16:52.560 --> 0:16:57.400 something that's very ordinary and mixes it up into this 0:16:57.520 --> 0:17:01.560 thing that sounds like the original, but the meaning has 0:17:01.600 --> 0:17:05.240 been completely changed. Chopped whipping cream may be an ingredient 0:17:05.320 --> 0:17:08.840 and fold water, enrolled it into cubes, or spread the 0:17:08.840 --> 0:17:12.080 butter in the refrigerator. That's another direction that came up with. 0:17:12.640 --> 0:17:15.560 Remember Ross and Oscar playing with the creativity setting for 0:17:15.600 --> 0:17:19.439 their scripts. Janelle plays with herbot's temperature too, so I 0:17:19.480 --> 0:17:22.160 can turn it up and the neural net may choose 0:17:22.160 --> 0:17:24.880 its second best or third best guess as to what 0:17:25.080 --> 0:17:28.200 letter comes next. And if I turn the creativity all 0:17:28.200 --> 0:17:31.440 the way down, then everything maybe something like the the 0:17:32.560 --> 0:17:36.200 the or recipes may be just you know, one teaspoon 0:17:36.240 --> 0:17:38.960 of vanilla over and over and over again, because that's 0:17:39.040 --> 0:17:44.640 just a very likely ingredient. It's really interesting with the 0:17:44.680 --> 0:17:48.879 recipes to turn down the creativity and see what it 0:17:48.920 --> 0:17:52.320 comes up with as the most quintessential recipes. At the 0:17:52.359 --> 0:17:55.320 lowest setting, you may not get hole Strandish brownies, but 0:17:55.400 --> 0:17:57.320 you do get a clear picture of what we eat 0:17:57.640 --> 0:18:00.399 and who we are. I look at what kinds of 0:18:00.760 --> 0:18:03.200 recipe titles that comes up with. There are things like 0:18:03.760 --> 0:18:07.919 chocolate chicken chicken cake, and another one that's chocolate chocolate 0:18:08.000 --> 0:18:10.520 chocolate chocolate cake. And there was a lot of cheese 0:18:10.560 --> 0:18:14.080 in these recipes too, so it's kind of revealing about 0:18:14.119 --> 0:18:18.880 what sorts of things we cook with. Then we like chocolate, 0:18:18.920 --> 0:18:21.960 cheese and chicken apparently. But then I did the same 0:18:22.000 --> 0:18:26.920 experiment with recipes from Bone Appetite, and then the most 0:18:26.920 --> 0:18:31.280 common ingredients that kept using were cilantro and pomegranate juice. 0:18:32.160 --> 0:18:35.000 So these algorithms essentially hold up a mirror to the 0:18:35.080 --> 0:18:37.840 data sets that we give them. They do, yeah, they 0:18:37.880 --> 0:18:41.800 reflect the data sets back to us in really weird ways, 0:18:42.160 --> 0:18:45.280 and they can absolutely pick up whatever bias there is 0:18:45.320 --> 0:18:48.240 in a input data set. And I think what we're 0:18:48.280 --> 0:18:52.359 discovering is just how prevalent that bias is and how 0:18:52.400 --> 0:18:55.840 easy it is for neural networks to latch onto that 0:18:55.880 --> 0:18:59.720 bias and copy it as a handy tool toward copying 0:18:59.720 --> 0:19:02.800 whatever where the humans are doing. They say that the 0:19:02.800 --> 0:19:05.679 way to a person's heart is through their stomach. But 0:19:05.760 --> 0:19:09.160 Janelle didn't stop at chocolate, bakes and surfs. She's also 0:19:09.200 --> 0:19:12.680 turned AI onto some more direct roots. I really liked 0:19:12.720 --> 0:19:15.000 the pickup lines. And there are all these puns and 0:19:15.040 --> 0:19:17.879 all this wordplay that it didn't have any way to 0:19:17.880 --> 0:19:20.159 grab hold of and figure out how to use. But 0:19:20.600 --> 0:19:25.880 I think what it produced this sort of charming surrealism 0:19:26.000 --> 0:19:30.480 and kind of garble nonsensical. I think it's an improvement 0:19:30.640 --> 0:19:34.080 on every single one of the originals. My very favorite 0:19:34.080 --> 0:19:37.240 one is you look like a thing and I love you. 0:19:37.240 --> 0:19:39.400 You are so beautiful that you make me feel better 0:19:39.480 --> 0:19:42.679 to see you. Or you must be a tringle because 0:19:42.680 --> 0:19:46.800 you're the only thing here. Are you a camera? Because 0:19:46.840 --> 0:19:50.480 I want to see the most beautiful than you. Yeah, 0:19:49.640 --> 0:19:54.400 I'll definitely lie with you. No one's have a used 0:19:54.480 --> 0:19:56.960 real pickup line on me? Use one on you know? 0:19:57.640 --> 0:20:01.280 Do I look like someone who would receive a pickup Well, 0:20:01.320 --> 0:20:04.880 here's one of them. I don't know you. That's good 0:20:05.119 --> 0:20:07.760 a lot of girls are into that. Are you a candle? 0:20:07.880 --> 0:20:12.639 Because you're so hard of the looks with you. So 0:20:12.640 --> 0:20:15.680 in effect, what the algorithm is doing is highlighting patterns 0:20:15.680 --> 0:20:18.360 in the data. I mean, there's sound structural like pickup 0:20:18.400 --> 0:20:20.960 lines about the words themselves don't make any sense. The 0:20:21.000 --> 0:20:24.440 machines are reflecting their creators and spitting back something which 0:20:24.880 --> 0:20:27.560 resembles pick up lines and makes us think a little 0:20:27.560 --> 0:20:31.320 bit more carefully about what a pickup line is. And 0:20:31.359 --> 0:20:34.280 while training Benjamin Ross and Oscar found the same thing 0:20:34.840 --> 0:20:38.680 as the algorithm learned patterns revealed bias present in our cinema. 0:20:40.600 --> 0:20:43.199 When you train an algorithm like Benjamin on millions and 0:20:43.240 --> 0:20:45.879 millions in this case of synopsis from the Internet of 0:20:45.920 --> 0:20:49.119 the movies, the synopsis that come out have certain patterns 0:20:49.160 --> 0:20:51.520 in them. For example, they mentioned men full times more 0:20:51.520 --> 0:20:54.680 often than they mentioned women. But you you learn other 0:20:54.760 --> 0:20:57.120 things about it than that. You learned that the most 0:20:57.119 --> 0:20:59.080 common phrase in the in the output is a young 0:20:59.119 --> 0:21:02.040 man in a small town. So what does a filmmaker 0:21:02.080 --> 0:21:04.600 like Oscar learn from this? I used to call this 0:21:04.640 --> 0:21:06.960 project the average movie projects. And the reason I called 0:21:06.960 --> 0:21:09.359 it that is the theory was for me, if you 0:21:09.359 --> 0:21:11.520 could make the right kind of algorithm that the movie 0:21:11.560 --> 0:21:14.200 that you would make that would be the theoretically perfect 0:21:14.280 --> 0:21:17.240 movie would also be the most boring movie ever made, 0:21:17.480 --> 0:21:19.120 and that it would. It would it would be by 0:21:19.119 --> 0:21:21.399 definition all of the things that were the most clich 0:21:21.640 --> 0:21:23.399 because that's what cliche means, is the thing that that 0:21:23.480 --> 0:21:26.400 you can rely on to work. And why do that? 0:21:26.560 --> 0:21:28.679 Because the thing I'm most interested in is doing the 0:21:28.720 --> 0:21:30.360 thing that we haven't done yet. I want to move 0:21:30.400 --> 0:21:33.440 the form forward. Seeing all of these biases and assumptions 0:21:33.480 --> 0:21:36.639 that are baked into our movies and our snacks doesn't 0:21:36.680 --> 0:21:39.560 mean we're doomed to repeat them. In fact, the awareness 0:21:39.560 --> 0:21:42.800 can be liberating. That's what's helped me, I think, is 0:21:42.840 --> 0:21:46.600 seeing Benjamin's capacity to show me more directly what it 0:21:46.680 --> 0:21:49.160 is that are our patents, are our habits, and then 0:21:49.359 --> 0:21:51.600 I can ask more easily how to move forward from that. 0:21:53.720 --> 0:22:05.000 We'll get there after the break. So we've heard about 0:22:05.080 --> 0:22:08.240 Janelle Shane using AI to reveal bias, and Ross and 0:22:08.280 --> 0:22:10.879 Oscar using it to help them think more creatively about 0:22:10.920 --> 0:22:13.919 filmmaking as well as how it can be applied to music. 0:22:14.520 --> 0:22:17.200 And that's the great promise of AI. We may worry 0:22:17.200 --> 0:22:20.159 about replacing jobs, but it can augment our lives in 0:22:20.200 --> 0:22:23.520 so many ways. At least that's how Sebastian Throne sees it. 0:22:24.480 --> 0:22:27.359 I would say the term AI is a bit deceptive 0:22:27.359 --> 0:22:29.760 because it sets up computers to be on equal power 0:22:29.760 --> 0:22:33.000 with people. I see it to be stronger where we 0:22:33.040 --> 0:22:36.480 are weak, and weaker where you're strong. It's not a 0:22:36.680 --> 0:22:40.680 technology that will replace us, as it's not really empowers 0:22:40.680 --> 0:22:44.080 but what might that empowerment look like beyond bias detection 0:22:44.240 --> 0:22:48.800 and piano playing well. In seventeen, Sebastian published a paper 0:22:48.840 --> 0:22:52.840 in Nature on using AI to diagnose skin cancer using 0:22:52.840 --> 0:22:56.600 just an iPhone. So in medicine, you can think of 0:22:56.640 --> 0:23:00.960 your iPhone that can find skin cancer as turning regular 0:23:01.400 --> 0:23:04.879 physicians or anybody in the world into an expert on 0:23:04.960 --> 0:23:07.320 day one, because now they have the superpower to be 0:23:07.359 --> 0:23:10.359 able to distinguish something that previously would have taken tens 0:23:10.359 --> 0:23:13.199 of years to learn. The same is true for the 0:23:13.240 --> 0:23:16.239