WEBVTT - Creative Writing Machines, with Mike Sharples 0:00:03.040 --> 0:00:05.360 Welcome to Stuff to Blow Your Mind production of My 0:00:05.480 --> 0:00:14.920 Heart Radio. Hey you welcome to Stuff to Blow Your Mind. 0:00:15.160 --> 0:00:18.320 My name is Robert Lamb and I'm Joe McCormick. Joe. 0:00:18.360 --> 0:00:21.400 While you were out, I conducted an interview with educational 0:00:21.480 --> 0:00:25.759 technologist Mike Sharple's on the book he wrote with Rafael 0:00:25.840 --> 0:00:31.040 Perezi Perez, Story Machines, How Computers Have Become Creative Writers. 0:00:31.480 --> 0:00:34.919 It's a fascinating read about the nature of storytelling, our 0:00:35.080 --> 0:00:38.320 history of attempting to instill the spirit of storytelling into 0:00:38.360 --> 0:00:41.360 the machine, where we seem to be going with this technology, 0:00:41.520 --> 0:00:45.360 and quite remarkably where we are already. Well, I can't 0:00:45.400 --> 0:00:48.200 wait to hear this interview. All right, Well, without further ado, 0:00:48.320 --> 0:00:52.200 let's jump right in. Hi, Mike, welcome to the show. 0:00:52.479 --> 0:00:55.360 I'm here. It's good to be at. The book, co 0:00:55.520 --> 0:01:00.560 authored with Rafael Perez Perez is Story Machines, How Computers 0:01:00.560 --> 0:01:05.640 Have Become Creative Writers, Publishing July five. It's a terrific read, 0:01:05.840 --> 0:01:07.679 but it has to be stressed from the outside. That's 0:01:07.720 --> 0:01:10.800 the storytelling isn't just a pastime that humans engage in. 0:01:11.319 --> 0:01:17.160 Storytelling is something greater? Right? Yeah, So storytelling we suggest 0:01:17.200 --> 0:01:20.360 in the book is something that's fundamental to human existence 0:01:21.000 --> 0:01:25.640 and has been for millennia. It suggested that instead of 0:01:26.120 --> 0:01:30.440 language coming first and then storytelling involving after that, perhaps 0:01:30.440 --> 0:01:35.880 it's the other way around. That perhaps storytelling started as 0:01:35.920 --> 0:01:41.000 a way of human communication through mine, through expressive gestures, 0:01:41.160 --> 0:01:44.520 and then language followed from that. So we want to 0:01:44.520 --> 0:01:49.160 make the point that storytelling is fundamental at a neural level, 0:01:49.560 --> 0:01:52.160 it's how we make sense of the world. It's at 0:01:52.160 --> 0:01:56.200 a cognitive level about how we create narratives to explain 0:01:56.200 --> 0:01:59.840 our existence, and in the social world, it's the stories 0:02:00.000 --> 0:02:02.480 we tell to each other that makes us who we are. 0:02:02.760 --> 0:02:06.240 So when we're talking about machines engaging in storytelling or 0:02:06.320 --> 0:02:09.600 story creation, we're really getting deep into human creativity and 0:02:09.680 --> 0:02:15.440 human identity. Then, yeah, we are. And it's both a 0:02:15.480 --> 0:02:21.560 fascinating insight into human creativity and it's also something the threat. 0:02:21.919 --> 0:02:27.120 So for centuries, writers have been fascinated about this idea 0:02:27.200 --> 0:02:31.040 that a that a machine might take over their craft, 0:02:31.400 --> 0:02:35.160 that a machine might become a storyteller. It goes right 0:02:35.160 --> 0:02:40.040 back to Jonathan Swift, for example, in Gulliver's Travels Coming 0:02:40.080 --> 0:02:46.320 Across a Weird Academy, where apprentices were manipulating a story 0:02:46.360 --> 0:02:51.440 machine which churned out academic texts right through to modern 0:02:51.440 --> 0:02:54.720 writers such as Roll Dow. He wrote a short story 0:02:55.080 --> 0:02:58.880 about how an author was selling his soul to a 0:02:58.960 --> 0:03:03.480 machine that generated short stories for him. So it's become 0:03:03.560 --> 0:03:07.600 something that particularly professional authors have been fascinated by this 0:03:07.720 --> 0:03:10.760 idea that a machine might be as creative as a 0:03:10.880 --> 0:03:15.240 human and also giving us insight into human creativity. Now, 0:03:15.760 --> 0:03:18.480 of course it's one thing to imagine these machines. You 0:03:18.480 --> 0:03:23.120 point out that they're they're necessary precursors in linguistics, literary analysis, 0:03:23.120 --> 0:03:25.480 and other fields to even get to the point of 0:03:25.520 --> 0:03:29.560 considering asking a machine or or making machine they can 0:03:29.600 --> 0:03:32.360 write a story or novel for you. Yeah, and I 0:03:32.400 --> 0:03:34.960 think there are there are a number of ways we 0:03:35.000 --> 0:03:38.240 can approach this. So one of them is to look 0:03:38.280 --> 0:03:42.160 at language. So we are language machines. We we have 0:03:42.280 --> 0:03:46.320 been trained in how to manipulate language as humans, and 0:03:47.000 --> 0:03:50.400 we can now design machines that could copy that in 0:03:50.400 --> 0:03:53.040 a very expressive way. So I'm sure we'll come onto that. 0:03:53.160 --> 0:03:57.960 But AI systems such as GPT three that are expert wordsmiths. 0:03:58.000 --> 0:04:02.760 So that's one route and other route is simulating characters. 0:04:03.120 --> 0:04:08.200 So some of the newer computer games now have stories 0:04:08.280 --> 0:04:13.480 nonplayer characters that can't tell stories to the human players. 0:04:14.120 --> 0:04:16.960 And then the third way is to build models of 0:04:17.200 --> 0:04:21.000 the creative mind. And that's what my colleague Raphael pre 0:04:21.279 --> 0:04:24.960 Perez has been doing for many years in his Mexican program, 0:04:25.080 --> 0:04:28.039 is trying to build a model of the human creative mind. 0:04:28.240 --> 0:04:32.080 So the different routes for coming at modeling storytelling and 0:04:32.160 --> 0:04:35.680 understanding storytelling with machines. And the one that you mentioned 0:04:35.720 --> 0:04:40.000 that I think is very fascinating is Vladimir props morphology 0:04:40.040 --> 0:04:44.080 of the folk tale. This idea of taking apart of 0:04:44.080 --> 0:04:46.840 folk tale tradition and in figuring out like what are 0:04:46.839 --> 0:04:50.120 the basic strokes, what are the basic elements, and thus 0:04:50.200 --> 0:04:54.080 creating I guess, sort of the palette for recreating stories. Yeah, 0:04:54.120 --> 0:04:59.160 this was back in the early twentieth century that Vladimir 0:04:59.200 --> 0:05:04.200 prop was part of a group of Russian linguists who 0:05:04.279 --> 0:05:10.760 and folk tale academics who became interested in the morphology 0:05:10.920 --> 0:05:15.359 the structure of folk tales, and he realized that Russian 0:05:15.360 --> 0:05:19.440 folk tales had this very similar structure, just as fairy 0:05:19.480 --> 0:05:23.719 tales in other Western traditions have, and so he set 0:05:23.720 --> 0:05:26.960 about trying to ride a set of rules that would 0:05:27.000 --> 0:05:31.360 both analyze these folk tales and show their underlying structure, 0:05:31.440 --> 0:05:34.000 but also could be turned around to generate them. That 0:05:34.160 --> 0:05:37.839 if you use these rules essentially as what we now 0:05:37.880 --> 0:05:42.760 call programs or algorithms, they could be employed to create 0:05:42.839 --> 0:05:47.039 new folk tales. And so he set the foundations for 0:05:47.120 --> 0:05:52.520 the structural analysis of stories way back what a hundred 0:05:52.560 --> 0:05:56.560 and twenty years ago now, And interestingly, some of the 0:05:56.680 --> 0:06:02.640 earliest computer programs to generate story were based on props formalism. 0:06:02.680 --> 0:06:04.559 It was really if you look at it, it's really 0:06:04.600 --> 0:06:08.880 like a very pair down computer algorithm. Now answer some 0:06:08.960 --> 0:06:13.840 of the first actual text generating machines. You mentioned a 0:06:13.880 --> 0:06:17.920 few different examples of this. Which are the most important 0:06:17.960 --> 0:06:20.160 to mention or which one is the most important to mention? 0:06:20.520 --> 0:06:23.320 So one of the earliest ones was by Christopher Stracci, 0:06:23.480 --> 0:06:27.760 who was a colleague of volenteering working in Manchester, and 0:06:28.000 --> 0:06:31.920 he developed a very very early computer program on one 0:06:31.960 --> 0:06:38.640 of the first prototype computers that generated love letters Victorian 0:06:38.760 --> 0:06:42.000 love letters. Um, and we can speculate on why he 0:06:42.080 --> 0:06:45.400 would want to write a program to generate love letters, 0:06:45.400 --> 0:06:47.400 but he did and pin them up on the wall 0:06:47.440 --> 0:06:51.120 of his lab. And then since then there's been a 0:06:51.279 --> 0:06:55.000 number of people who have particularly taken a linguistic approach, 0:06:55.600 --> 0:06:59.160 and so one of the pioneers was a linguist called 0:06:59.160 --> 0:07:05.560 Sheldon Clade, and Sheldon Klein had this big, grandiose project 0:07:05.920 --> 0:07:13.160 to try and model human language production as a as 0:07:13.160 --> 0:07:16.200 a as an algorithm, as a computer program, and then 0:07:16.560 --> 0:07:21.000 through this trying to understand the origins of language. And so, 0:07:21.400 --> 0:07:24.920 along with a group of colleagues, many PhD students, he 0:07:25.080 --> 0:07:28.800 produced one of the earliest programs that generated stories. It 0:07:28.960 --> 0:07:32.960 generated murder mysteries, a sort of country house Agatha Christie 0:07:32.960 --> 0:07:37.040 type mysteries, and it was part of this big project 0:07:37.240 --> 0:07:41.120 to formalize language in a generative way. The problem was 0:07:41.160 --> 0:07:47.640 the program the stories that generated were pretty trivial. There 0:07:47.800 --> 0:07:50.640 was people gathered at the country house, there was a murder, 0:07:51.480 --> 0:07:56.120 somebody tried to investigate it the end, and so although 0:07:56.160 --> 0:07:58.640 it was seen as a novelty at the time, it 0:07:58.760 --> 0:08:02.560 wasn't really risk acted for the Great linguistic project that 0:08:02.680 --> 0:08:06.800 lasted for twenty or thirty years in the nine nineties, 0:08:07.600 --> 0:08:10.160 and there were a number of attempts to write entire 0:08:10.280 --> 0:08:14.600 novels by computer and probably The most interesting one was 0:08:14.640 --> 0:08:20.600 by Geckll Scott French, who programmed a Mac computer to 0:08:21.280 --> 0:08:25.360 generate a novel in the style of Jacqueline Susan. I've 0:08:25.400 --> 0:08:28.760 got it here. It's called Justice Once, and it is 0:08:28.960 --> 0:08:33.200 a complete published novel in the style of the pot 0:08:33.240 --> 0:08:39.160 boilant author called Jacqueline Susan, and it seems to be genuine. 0:08:39.200 --> 0:08:41.560 There's a picture of him with his Mac computer on 0:08:41.600 --> 0:08:45.160 the cover of the book. And it took him eight 0:08:45.280 --> 0:08:50.000 years to design this very early AI system that would 0:08:50.040 --> 0:08:54.640 mimic the style of this author, and he engaged that 0:08:54.760 --> 0:09:00.160 eyed dialogue with the program to generate an entire three 0:09:00.200 --> 0:09:04.640 hundred page novel. So that was probably the first, I'm 0:09:05.040 --> 0:09:08.439 right up to this day, greatest example of story writing 0:09:08.440 --> 0:09:12.600 with machine. What's fascinating now is that what took him 0:09:12.720 --> 0:09:15.720 eight years to do could now be done in a 0:09:15.720 --> 0:09:22.960 few seconds with the most recent AI UM generator programs. Yeah, 0:09:23.080 --> 0:09:25.080 in the book you you write them just to read 0:09:25.120 --> 0:09:27.400 a quick quote quote. It took just twenty years ago 0:09:27.440 --> 0:09:29.520 from a program that wrote love letters to one that 0:09:29.559 --> 0:09:33.319 created complete short stories. Uh, and a further twenty years 0:09:33.480 --> 0:09:36.120 to a published three page novel in partnership with the 0:09:36.120 --> 0:09:39.959 computer Uh, it's fascinating to think about. You know this, 0:09:40.200 --> 0:09:45.080 this the technological advancement during that time in broad strokes, 0:09:45.120 --> 0:09:47.920 like what I mean, what were the key advancements going 0:09:47.960 --> 0:09:50.760 on here that made this possible. Well, the first advancement 0:09:51.000 --> 0:09:55.600 was to be able to have interactive computer systems that 0:09:55.760 --> 0:10:00.440 you could program in a high level programming which and 0:10:00.480 --> 0:10:04.120 that's what Christopher Straitchy and Alan Turing were working on, 0:10:05.640 --> 0:10:10.120 and that they could then demonstrate this in generating very 0:10:10.200 --> 0:10:16.080 simple love letters. The next stage was to be able 0:10:16.160 --> 0:10:21.079 to produce grammars, generative grammars, and this goes back to 0:10:21.120 --> 0:10:25.000 the work of prop who realized that you could have 0:10:25.040 --> 0:10:30.360 grammars that didn't just generate individual sentences, but could generate 0:10:30.520 --> 0:10:33.320 entire stories. That you could describe the structure of a 0:10:33.360 --> 0:10:37.280 story in terms of something like a formal grammar, and 0:10:37.640 --> 0:10:41.439 that in THEES there were a number of projects to 0:10:42.160 --> 0:10:46.200 put that grammar into a program that would generate a 0:10:46.200 --> 0:10:48.880 short story. That's what Sheldon Klein and his team did. 0:10:49.840 --> 0:10:53.720 And then the next step beyond that was to write 0:10:53.920 --> 0:10:59.000 symbolic AI programs that modeled the style of a particular writer, 0:10:59.480 --> 0:11:03.040 and that was the great achievement of um Scott French. 0:11:03.320 --> 0:11:08.800 It was a symbolic AI expert system of style. And 0:11:08.840 --> 0:11:12.040 then we come right up to date. And there are 0:11:12.080 --> 0:11:19.200 programs like GPT three, which are hugely competent and well 0:11:19.280 --> 0:11:22.760 trained language systems, So they aren't rule based systems. They 0:11:22.800 --> 0:11:29.480 don't have something like the Sheldon client grammar inside them. 0:11:29.600 --> 0:11:34.480 They have been trained on billions of pieces of text 0:11:34.960 --> 0:11:40.280 and they have many millions of interconnections that create an 0:11:40.320 --> 0:11:45.000 internal language model which they then use to generate texts 0:11:45.040 --> 0:11:48.240 in particular styles. You can start it in a style 0:11:48.440 --> 0:11:50.559 and it will continue in that style. You can give 0:11:50.600 --> 0:11:52.840 it an instruction of what sort of story to write 0:11:53.000 --> 0:11:57.240 and it will continue in that story. But the really 0:11:57.280 --> 0:11:59.600 important thing to say is that there are two different 0:11:59.640 --> 0:12:02.560 sorts of AI. So the sort of AI that's got 0:12:02.559 --> 0:12:06.480 French used was writing explicit rules. That's why it took 0:12:06.520 --> 0:12:10.760 him eight years to code these rules to imitate one 0:12:10.840 --> 0:12:17.880 person style. The GPT three type AI transformer programs induce 0:12:18.480 --> 0:12:22.440 infer those rules from being trained on billions and billions 0:12:22.440 --> 0:12:24.800 of pieces of texts. So there are two different sorts 0:12:24.800 --> 0:12:29.240 of AI. Both had their strengths and weaknesses. UH And 0:12:29.640 --> 0:12:31.800 one of the fascinating things of the future is whether 0:12:31.840 --> 0:12:34.200 we can put them together, whether we can merge those 0:12:34.240 --> 0:12:37.400 two different sorts of AI into the sort of universal 0:12:37.520 --> 0:12:43.080 story machine. Wow. And so GPT three is the current model, 0:12:43.200 --> 0:12:46.720 Is that correct? Or are we had four yet? So 0:12:46.800 --> 0:12:49.800 there isn't a GPT four yet, although I'm sure there's 0:12:49.840 --> 0:12:54.080 one in the pipeline that continually revising GPT three. And 0:12:54.200 --> 0:12:57.480 just for those who don't know GPT, the GPT models 0:12:57.640 --> 0:13:01.720 were developed by and still being developed by company called 0:13:01.760 --> 0:13:05.959 open Ai that was founded by a group of entrepreneurs 0:13:06.000 --> 0:13:10.160 including Elon Musk and others, and that company was set 0:13:10.280 --> 0:13:16.280 up to explore the opportunities of AI for social good. 0:13:16.800 --> 0:13:20.520 It has developed a number of different programs. There's one 0:13:20.559 --> 0:13:24.240 that it's developed for art called Dally, which can do 0:13:24.320 --> 0:13:28.520 the same for art and images as GPT three can 0:13:28.640 --> 0:13:34.040 for for words and stories. But the GPT three now 0:13:34.080 --> 0:13:40.760 it's its third generation. In essence, what it's been, what's 0:13:40.760 --> 0:13:43.560 how it works is it's been trained on billions of 0:13:43.559 --> 0:13:48.200 pieces of text. It uses those texts to form an 0:13:48.240 --> 0:13:54.000 internal model of both the surface structure of language but 0:13:54.080 --> 0:13:58.319 also the internal structure language essentially how the world works. 0:13:58.840 --> 0:14:04.760 And then initially the early GPT models were sentenced computers 0:14:04.880 --> 0:14:08.520 like very highly trained, suped up sentenced computers of the 0:14:08.559 --> 0:14:11.240 sort that you've got on your mobile phone. But they 0:14:11.240 --> 0:14:14.160 can look back at the last five words or so. 0:14:14.640 --> 0:14:19.280 They haven't a big attention window, so they know what 0:14:19.440 --> 0:14:22.320 they've written before, and they use this to continue in 0:14:22.360 --> 0:14:25.040 the same style, the same structure. So they give a 0:14:25.280 --> 0:14:35.600 very plausible simulation, a very plausible indication of coherent language, 0:14:35.960 --> 0:14:39.680 as if it were being written by a human. Now, 0:14:40.160 --> 0:14:44.440 the problem is it's highly believable, but it doesn't have 0:14:44.600 --> 0:14:49.160 any sense of itself. So these systems can't reflect on 0:14:49.240 --> 0:14:53.320 what they've written. They can't look back and say, does 0:14:53.360 --> 0:14:57.520 this makes sense? Does this fit with a good model 0:14:57.520 --> 0:15:01.840 of the world? Is it legal on his decent? It doesn't. 0:15:02.440 --> 0:15:05.680 In essence, it's they're a moral They don't have any 0:15:05.760 --> 0:15:10.560 internal sense of what's right and morality. And so as 0:15:10.600 --> 0:15:15.880 story machines, they're great because they can tell fascinating, plausible, engaging, 0:15:16.280 --> 0:15:21.320 sometimes even poignant stories. But for other purposes like writing 0:15:21.320 --> 0:15:25.520 newspaper articles or writing student essays, then they can be 0:15:25.600 --> 0:15:28.400 dangerous because they don't know what they've written, and they 0:15:28.440 --> 0:15:32.280 don't know whether what they've written is decent, honest, and 0:15:32.360 --> 0:15:41.560 truthful than now. Our listeners can can actually get a 0:15:41.560 --> 0:15:45.720 taste of this by going to story slash machines dot net. 0:15:46.800 --> 0:15:49.640 You have an interface here where you can put in 0:15:49.800 --> 0:15:53.040 just a title or also a title and in some 0:15:53.120 --> 0:15:55.440 text like an opening line of a of a story. 0:15:56.040 --> 0:15:59.280 And this is powered by GPT three. Correct. Yeah, that's right. 0:15:59.440 --> 0:16:04.080 So what I did was wrote a website with an 0:16:04.120 --> 0:16:09.360 interface to the most recent GPT three language model. So 0:16:09.400 --> 0:16:12.920 it basically provides a very simple way for you to 0:16:13.160 --> 0:16:16.120 write the title of a story, a few opening words, 0:16:17.000 --> 0:16:21.480 press the create button, and it comes back with about 0:16:21.480 --> 0:16:25.760 a two d and fifty word introduction to a short story. 0:16:25.920 --> 0:16:28.640 So it gives you a taste of what GPT three 0:16:28.680 --> 0:16:31.120 can do. And yeah, I really recommend that you go 0:16:31.160 --> 0:16:34.880 and try it. Um. I launched it just a couple 0:16:34.920 --> 0:16:38.720 of days ago, so people have already been generating fascinating 0:16:38.760 --> 0:16:42.400 short stories using that, So so go and give it 0:16:42.440 --> 0:16:47.600 a try. It's story hyphen machines dot net. Yeah I was. 0:16:47.680 --> 0:16:50.040 I was firing this up when I was reading the 0:16:50.080 --> 0:16:53.720 book yesterday, and yeah, eventually I was having to bring 0:16:53.760 --> 0:16:55.800 my son over and showed to him. I was really 0:16:56.160 --> 0:17:00.680 impressed toying around with it like trying having a title 0:17:00.800 --> 0:17:04.119 that sort of suggested a certain genre, maybe throwing in 0:17:04.200 --> 0:17:08.720 a first sentence that that was the record that the 0:17:08.760 --> 0:17:11.880 reference to certain genre. And then also I was really 0:17:11.920 --> 0:17:14.560 impressed when I would say, put in a title and 0:17:14.600 --> 0:17:17.200 then put in the first line of an existing short 0:17:17.240 --> 0:17:19.960 story by an author. I think I used Clark Ashton 0:17:20.000 --> 0:17:24.400 Smith as an example, and I used um Corny McCarthy. 0:17:24.400 --> 0:17:26.240 I put in a line of Corny McCarthy's work, and 0:17:26.240 --> 0:17:28.800 I was impressed at how well it seemed to model 0:17:29.400 --> 0:17:34.480 a portion of writing as if it were by these authors. Yeah, exactly. Um, 0:17:34.520 --> 0:17:36.800 I'll give you an example, because it's quite quite good 0:17:36.840 --> 0:17:38.760 to show an example. You can cut it out if 0:17:38.760 --> 0:17:41.560 you want to. So so here's an example one that 0:17:41.640 --> 0:17:45.440 I did on Ernest Hemingway in London. The fog came 0:17:45.480 --> 0:17:47.920 in from the river and it was cold for April. 0:17:48.480 --> 0:17:51.320 I walked up past the British Museum and turned down 0:17:51.359 --> 0:17:55.000 towards Holburn. I had a lunch date with Harriet the Criterion, 0:17:55.200 --> 0:17:57.919 and I was early. It was one of those heavy 0:17:58.080 --> 0:18:00.760 grade days that seemed to press down on the city, 0:18:01.160 --> 0:18:04.919 weighing it down. The traffic moved slowly through the streets, 0:18:04.960 --> 0:18:08.119 and the people hurried along with their heads down trying 0:18:08.160 --> 0:18:10.720 to get out of the cold. I glanced in the 0:18:10.720 --> 0:18:14.040 windows of the shops as I walked, looking for something 0:18:14.080 --> 0:18:17.560 to buy Harry for his birthday. I knew he'd appreciate 0:18:17.640 --> 0:18:20.239 a good bottle of scotch, but I wanted to get 0:18:20.320 --> 0:18:23.240 him something else as well. That's just the start of 0:18:23.320 --> 0:18:26.880 as a story in Ernest Hemingway style, and you can 0:18:26.880 --> 0:18:29.960 give it whatever style you want. It's I found it 0:18:30.000 --> 0:18:35.120 works best when you do a weird juxtaposition. So there 0:18:35.240 --> 0:18:38.399 was someone just a few hours ago did one on 0:18:38.960 --> 0:18:43.520 the Sad Sandwich, and it was a really poignant, sad 0:18:43.560 --> 0:18:48.199 story about a poor, neglected sandwich. So try doing some 0:18:48.200 --> 0:18:51.440 some juxtapositions at words like that. Yes, absolutely, and I'd 0:18:51.480 --> 0:18:54.840 love to hear back from from listeners after they've toyed 0:18:54.880 --> 0:18:58.960 around with us and explored it. But of course, recognizing 0:18:59.000 --> 0:19:03.000 the power of of this technology, Uh yeah, we we 0:19:03.080 --> 0:19:07.000 certainly get into this, this area of of anxiety perhaps, 0:19:07.000 --> 0:19:12.040 but also hope and opportunity. Um I guess on the 0:19:12.080 --> 0:19:14.960 anxiety side of things. The first place my mind went 0:19:15.000 --> 0:19:18.760 as I remember seeing Max Tegmark reference a kind of 0:19:19.600 --> 0:19:22.400 an illustration that was like a topography of human abilities 0:19:22.440 --> 0:19:25.960 and jobs with the idea that at the higher elevations 0:19:25.960 --> 0:19:28.760 were going to be more protected from the rising sea 0:19:28.880 --> 0:19:32.600 levels of AI. So chess and jeopardy were already in 0:19:32.640 --> 0:19:36.680 the water, uh, speech recognition, investment, and social interaction there 0:19:36.800 --> 0:19:40.320 in the lowlands they're going next. And then in this 0:19:40.400 --> 0:19:43.919 particular image, we had science as the highest peak, just 0:19:44.080 --> 0:19:47.439 above the peaks of book writing and AI design. So 0:19:47.520 --> 0:19:50.680 I was just wondering, do you feel like this representation 0:19:51.720 --> 0:19:54.679 was accurate or and or have the water has just 0:19:54.840 --> 0:19:58.600 risen so high already? I think every risen high already. 0:19:58.640 --> 0:20:01.240 I think the ones that are going to be protected 0:20:01.359 --> 0:20:05.120 from the rising tide of AI are probably the caring professions, um, 0:20:05.480 --> 0:20:13.120 the nurses, the child careers. But any profession that works 0:20:13.200 --> 0:20:16.760 with words, he's going to find both an opportunity and 0:20:16.760 --> 0:20:20.440 a threat, I think in generative AI. So the opportunity 0:20:21.000 --> 0:20:24.800 is that it's a new kind of tool. The tools 0:20:24.800 --> 0:20:27.359 that we've had up to now have tended to be 0:20:27.560 --> 0:20:34.000 ones that slow down writing, um, like the saurus or 0:20:35.080 --> 0:20:40.679 spell corrector that they make you pause and check. You 0:20:40.760 --> 0:20:42.760 can do them at the end, but there's always a 0:20:42.800 --> 0:20:46.560 temptation to look up a word to slow down. What's 0:20:46.600 --> 0:20:49.320 different about these is that they can be used to 0:20:49.359 --> 0:20:52.160 speed up your writing. You can, and I've tried doing this, 0:20:52.359 --> 0:20:57.080 writing a paragraph when it starts to dry up, them 0:20:57.080 --> 0:20:59.680 handing it over to the machine to write the next paragraph. 0:21:00.119 --> 0:21:04.800 I'm not a fiction writer. I found it quite empowering 0:21:05.359 --> 0:21:11.080 to use GPT three as a writing buddy. So I 0:21:11.080 --> 0:21:13.680 would write a paragraph, it would write the next paragraph 0:21:13.760 --> 0:21:16.639 and probably take it off in some unexpected direction that 0:21:16.680 --> 0:21:19.520 I would then have to follow, and it might introduce 0:21:19.560 --> 0:21:24.120 a new character, a plot twist, and so both as 0:21:24.160 --> 0:21:27.760 a tool for budding writers and also perhaps a prop 0:21:28.080 --> 0:21:33.440 for professional writers, particularly ones with deadlines to meet, then 0:21:33.480 --> 0:21:38.120 that's the opportunity. I think the threat is the inverse 0:21:38.160 --> 0:21:40.479 side of that, that if you either see it as 0:21:40.480 --> 0:21:45.320 a crutch that rather than trying to do your own writing, 0:21:45.359 --> 0:21:47.840 you just hand it over to the machine, then it's 0:21:47.920 --> 0:21:51.560 very easy to become lazy. And also, as I've said, 0:21:51.720 --> 0:21:54.439 they are a moral machine. So if you're trying to 0:21:54.480 --> 0:22:00.200 do scientific writing, or you're trying to do accurate journalists them, 0:22:00.920 --> 0:22:05.879 then beware because they may well throw in some entirely 0:22:06.000 --> 0:22:11.520 fake research study, some entirely inaccurate, fake reference, um, perhaps 0:22:11.640 --> 0:22:19.960 reference to some completely non nonsensical or inaccurate event that's 0:22:19.960 --> 0:22:22.679 happened in the world. So you've always got to be 0:22:22.720 --> 0:22:25.560 aware of the facts that it throws up. And there's 0:22:25.560 --> 0:22:28.560 a good reason for that that it isn't a fact checker. 0:22:29.320 --> 0:22:32.840 It isn't a Wikipedia or even a Google search. It 0:22:33.000 --> 0:22:38.359 is a language machine. It's loves, you know, anthropomorphic sense. 0:22:38.359 --> 0:22:42.720 It loves playing with words, but those words don't necessarily 0:22:42.760 --> 0:22:47.120 make sense. So if you're going to use tools like 0:22:47.280 --> 0:22:50.600 GPT three as AIGs for writing, then you have to 0:22:50.640 --> 0:22:56.119 be very careful that you cross check the facts and 0:22:56.600 --> 0:23:01.560 the output of that machine to make sure that it 0:23:01.720 --> 0:23:06.800 is accurate and honest and truthful. My mind also went 0:23:06.880 --> 0:23:11.920 to um the possibility of of of almost accidental plagiarism 0:23:12.720 --> 0:23:14.960 because I put in the first line of a Clark 0:23:15.000 --> 0:23:18.520 Ashton Smith story and it it threw in a fascinating 0:23:18.560 --> 0:23:21.960 plot twist that was not in the original story, and 0:23:22.240 --> 0:23:24.600 I don't think I've ever seen in a story, and 0:23:24.600 --> 0:23:26.679 so part of me was wondering like, well, well, you know, 0:23:26.760 --> 0:23:28.640 I should I should latch onto this, maybe I could 0:23:28.760 --> 0:23:31.040 use this. But then the other part, you know, the 0:23:31.119 --> 0:23:33.320 lights coming on in my mind were saying, but hold 0:23:33.359 --> 0:23:36.439 on this. Just because I haven't read it doesn't mean 0:23:36.480 --> 0:23:39.600 it doesn't already exist out there. Might Uh we run 0:23:39.640 --> 0:23:44.720 into situations where, uh, the AI is is reproducing something 0:23:45.200 --> 0:23:47.600 you know, perhaps you honestly, if we want to use 0:23:47.680 --> 0:23:51.399 that term when in in in actuality, it may exist 0:23:51.400 --> 0:23:53.600 out there in some story or another. Yeah, I mean, 0:23:53.640 --> 0:23:55.960 I think the first thing to note is that it's 0:23:56.119 --> 0:24:00.600 not working at this sort of word and sad tense level. 0:24:00.800 --> 0:24:03.520 So it's not copying bits of text from the web 0:24:03.680 --> 0:24:08.359 or from published books. It's working below that, basically at 0:24:08.400 --> 0:24:12.000 the phone name level. It's putting together pieces of words, 0:24:12.880 --> 0:24:16.920 but it's putting together these pieces of words in a 0:24:17.000 --> 0:24:22.000 hugely proficient way. So I've tried taking the output of 0:24:22.040 --> 0:24:26.920 GPT three and doing Google search on phrases and sentences, 0:24:27.600 --> 0:24:32.640 and you don't find them. So it seems like they 0:24:32.760 --> 0:24:41.680 are genuinely producing novel pieces of text. So, for example, 0:24:41.920 --> 0:24:44.400 if students are going to use these for writing essays, 0:24:44.840 --> 0:24:48.960 which is already happening, they're already companies that are advertising 0:24:49.000 --> 0:24:52.159 the services of AI generators for students to write their essays. 0:24:52.960 --> 0:24:57.520 Plagiarism checkers won't detect them. I've tried putting them through 0:24:57.560 --> 0:25:04.920 plagiarism checkers and they of sort of nineties five originality. 0:25:05.280 --> 0:25:08.520 So they're not copying bits of text from the web. 0:25:08.960 --> 0:25:13.679 They are genuinely generating new language. Now, of course, there 0:25:13.720 --> 0:25:18.000 are phrases that may pre exist, and if you give 0:25:18.440 --> 0:25:25.080 highly constrained styles like Shakespeare's sonnets, then it may come 0:25:25.200 --> 0:25:29.800 up with previous lines from other Shakespeare sonnets. But providing 0:25:29.840 --> 0:25:33.040 you give it a broad enough brief, providing you give 0:25:33.080 --> 0:25:35.239 it a general enough style, or even if it's from 0:25:35.240 --> 0:25:41.199 a particular author, then it will generate original text. And 0:25:41.240 --> 0:25:44.720 it's still a bit scariest too. You know, I've talked 0:25:44.720 --> 0:25:46.720 about this with other people and said, but surely it's 0:25:46.760 --> 0:25:50.000 copying from the web. No, it isn't. It's generating. It's 0:25:50.080 --> 0:25:53.720 generating new text in the style of that author, or 0:25:53.760 --> 0:25:58.399 in the style of that piece of fiction or piece 0:25:58.400 --> 0:26:02.280 of journalism. So you already touched on like this the 0:26:02.840 --> 0:26:08.400 collaborative possibilities here. But having touched on in school papers 0:26:08.480 --> 0:26:11.320 and such, what do you think are the the educational 0:26:11.440 --> 0:26:16.600 opportunities with this technology. I think the main educational opportunities 0:26:16.640 --> 0:26:23.600 are for beginning writers. It's a way to explore expressivity 0:26:23.640 --> 0:26:27.199 and creativity. One of the problems when you're beginning writing 0:26:27.280 --> 0:26:32.240 is you tend to see everything as being a linear process. 0:26:32.640 --> 0:26:35.720 You write some words, you write some more words. There's 0:26:35.720 --> 0:26:38.520 a flow of writing. It's very difficult to get out 0:26:38.520 --> 0:26:40.919 of that flow and to think about alternative ways of 0:26:40.960 --> 0:26:44.800 expressing something, how it might be different. And what machines 0:26:45.760 --> 0:26:48.560 light GPT three can do is help you to see 0:26:48.600 --> 0:26:52.760 another way of continuing, another way of expressing your ideas. 0:26:53.080 --> 0:26:55.600 It will look back over the last five words or 0:26:55.640 --> 0:26:59.000 so that you've written and perhaps take it in new directions. 0:26:59.600 --> 0:27:04.760 So it's a way for buddying writers to explore possibilities. 0:27:05.119 --> 0:27:08.679 And you can take what you've written so far and 0:27:08.800 --> 0:27:11.520 press the create button a number of times, and each 0:27:11.560 --> 0:27:13.880 time it will take your writing in a different direction. 0:27:14.200 --> 0:27:18.359 So that's one way. Another way is for in a 0:27:18.440 --> 0:27:21.719