[AUDIO PLAYBACK - Clip from War Games] - Wow. We got something. We're in. - Greetings. - Hello. - Shall we play a game? - Oh, love to. Let's play global thermonuclear war. [END PLAYBACK] [MUSIC PLAYING] AMEET DOSHI: You are listening to WREK Atlanta, and this is global thermonuclear. No, it's Lost in the Stacks, the research library rock and roll radio show. I'm Ameet Doshi in the studio with Abby and Fred and Edmond. Where each week on Lost in the Stacks, we pick a theme and then use it to create a mix of music and library talk. Whichever you're here for, we hope you dig it. FRED RASCOE: Our show today, Ameet, is called Skynet Fail, AI in Libraries. And I think we'll only be skimming the surface of this unwieldy and diffuse topic. AMEET DOSHI: You know, Fred, if only we had some way of compiling all of that disparate scholarship together, identifying the best, most reliable sources, and then providing useful guidance about how to tackle the subject. FRED RASCOE: You mean like in an algorithm? AMEET DOSHI: Algorithm. Yes, of course. All glory to the algorithm. FRED RASCOE: All hail the algorithm. If you, our listeners, want to join the conversation, the hashtag for this show is LITS 405 for Lost in the Stacks episode 405. Feel free to tweet your thoughts, questions, or solutions to the 8 Queens puzzle with that hashtag. AMEET DOSHI: Our songs today are about being cautious, the qualities that make us human, and machines being in control. Whenever I'm tempted to scoff at the notion of machines controlling my own human brain, I have to stop to remember that I haven't been more than an arm's reach away from the computing power of a laptop or a smartphone in, let's just say, several years. Computers aren't just things we own, they're the things we need. So let's start our show today with "My Computer" by Prince, right here on Lost in the Stacks. Prince, yes. [MUSIC PLAYING] Goodbye. "My Computer" by Prince. That was such a great way to kick this show off. Thank you, Charlie. This is Lost in the Stacks. Joining us on the show is Georgia Tech Professor Edmond Chow from the College of Computing, right here at Georgia Tech. Professor Chow, Edmund, welcome back to the show. EDMOND CHOW: Thank you. It's a pleasure. AMEET DOSHI: It was seven years ago to the date. EDMOND CHOW: Is it really? AMEET DOSHI: No. [LAUGHS] But I like throwing that in. And today we're talking about AI and libraries. And based on your Wikipedia entry, I know that you work in the area of large scale computing. And I guess the question that we might want to begin with is, do these kinds of systems, these artificial intelligence systems, inherently require large scale computing? My sense is probably yes, but maybe you could give us some more insight into this. EDMOND CHOW: Well, these systems generally involve large scale data. It was discovered some years ago that with large amounts of data, you can discover very interesting patterns that can be used to train an intelligent system. So with the large scale data, you have to compute on it. And that's where the large scale computing comes from. So further, the computations are very complicated. They involve solving some very difficult optimization problems, meaning that we're trying to find exactly what the model is for this artificial system. So with the amount of data and these complicated optimization problems, we do have large scale computations involved. And to give you an idea, I wouldn't be surprised if computations take on the order of weeks to train one of these systems on a large computer. FRED RASCOE: When an artificial intelligence system such as-- I'm going to say that-- I'm going to use Siri as an example. Is that a valid example, first of all? EDMOND CHOW: I think it's one of the first ones that we became aware of. FRED RASCOE: OK, right. And, like Ameet said, you have your own Wikipedia page. Not all of our guests have their own Wikipedia page. But when I ask Siri a question, usually the first response that she pulls. EDMOND CHOW: It, Fred, it. FRED RASCOE: Oh yeah, I'm anthropomorphizing, aren't I? When it pulls an answer, it's usually pulling straight from Wikipedia. So you just mentioned that it takes, weeks to do the calculations to get that to work. But once it's working, what kind of calculations are happening? Because it's very instantaneous once it's actually happening for the user. EDMOND CHOW: Right, right. I think the hardest part, at least historically, has been for systems like Siri to understand your speech. So there is a lot happening behind the scenes in this training phase. This is where I'm referring to weeks of computation, where the system is fed lots of samples of speech and is trained on what words are, where they appear in sentences. And then once the training is done and the model is created, in other words, solving that optimization problem, then the process is relatively quick. You just send a sentence over to some servers, in this case Apple, and your speech would be fed into the system, and then out comes the result. And then on top of that, there's natural language processing and all of that to understand what you actually said. FRED RASCOE: When you say optimization problem, can you define what that is when artificial intelligence is solving an optimization problem? EDMOND CHOW: Sure. So one of the most popular methods in machine learning these days is neural networks. And here we're constructing a model that emulates what is going on in the human brain. And we need to assign numerical values to each component of this model. And to determine what these values are, we have to solve an optimization problem. We're trying to minimize the difference between what the model would predict versus what we would observe in real life. FRED RASCOE: So it's like comparing what a human would do when compared with-- when confronted with a sheet of text. What a human would do to comprehend that versus a computer. EDMOND CHOW: That's right. We're trying to minimize that difference. AMEET DOSHI: There's another somewhat jargony term, but I think it might be in the popular domain now or the popular consciousness. And that's the idea of a Turing machine. And a lot of people are probably familiar with the film about Turing called The Imitation Game, where he built this machine. But his goal was to optimize the machine. Is that accurate? Is that an accurate reading of some of the seminal kinds of thinking about AI? EDMOND CHOW: Yeah, so in AI, there are hosts of techniques being used, and optimization is the foundation of the ones that are most successful in that we use recently. So what Alan Turing was doing and what early practitioners were doing, those are a whole host of things. Many were trying to also emulate what the human brain does. And other approaches are much more abstract. FRED RASCOE: OK, we're speaking with Edmond Chow of the Georgia Tech College of Computing about AI, and I think we're going to dive into how AI factors into the library world pretty soon. We'll be back with more after a music set. ABBY: File this set under Q335.A78582. [MUSIC PLAYING] You just heard "Transmitters" by Sonic Youth, and before that was "Death to Los Campesinos" by Los Campesinos. Songs about machines being in control. FRED RASCOE: Today on Lost in the Stacks, we're discussing AI, Artificial Intelligence, in libraries with Georgia Tech Professor Edmond Chow. And off air, we were just chatting in the studio here, and Edmond kind of pointed out that the first segment was a little serious, it was a kind of serious tone, serious mood. EDMOND CHOW: Like a lecture. FRED RASCOE: Yeah. We don't want that at Lost In The Stacks. That's not our vibe, I don't think. AMEET DOSHI: No, no, our vibe is definitely not lecture. FRED RASCOE: OK. So you can program, it turns out, an artificial intelligence to write jokes. So just to balance things out, the serious with the light, let's hear a joke that an artificial intelligence wrote. You ready? AMEET DOSHI: I'm ready. FRED RASCOE: What do you call a cat does it take to screw in a light bulb? [LAUGHTER] That setup again. What do you call a cat does it take to screw in a light bulb? AMEET DOSHI: Does not compute. FRED RASCOE: They could worry the banana. [LAUGHTER] AMEET DOSHI: That's genius, though, I get it. FRED RASCOE: What did the new ants say after a dog? AMEET DOSHI: I don't know. FRED RASCOE: You really don't. It was a pirate. Get it? AMEET DOSHI: No. EDMOND CHOW: That's not a joke. ABBY: This is probably why we don't quite have robot comedy stand up shows yet. FRED RASCOE: Right. EDMOND CHOW: But just wait. AMEET DOSHI: Yeah, just wait. And I would pay to see robot stand up. It could happen. FRED RASCOE: I mean, those jokes, I'm guessing, I don't know a lot about how those were programmed and then constructed, the system that programmed those and constructed those jokes. But it's taking in some human had to put in the word ants and the word dog and the word banana, and some sort of sentence structure rules, and it spit it out into this just serendipitous combination of words that kind of resembles what we identify as a joke. And serendipity is a big part of library culture, because in libraries, at least when the days of-- remember when we had actual books? AMEET DOSHI: When you could get lost in the stacks? FRED RASCOE: You could go up and you could grab a book, pull it off the stacks, and then there's something next to it and something on the aisle over. And you could just make all these random combinations. But we're going to have to-- since books are going more digital, Edmond, we're going to have to make those serendipitous connections a different way. How can we use AI to make serendipitous connections in our library collections? EDMOND CHOW: Well, AI systems are very good at making recommendations, and certainly making recommendations was one of the very first things that these systems were doing well when the internet or the world wide web was available. FRED RASCOE: That's like, if you bought this, you might like this. Is that what-- EDMOND CHOW: You watched this movie. You might like this movie. And that's being taken to an extreme. If you bought this artwork, you might like this artwork. So AI systems are very good at that. And I think that this also applies to making recommendations of what books to read, what journal articles to take a look at that's related to your research or research of people that you're following. But you know what? The serendipity is lost, because you don't think of that as being serendipitous. So the pleasure of discovering something is lost. FRED RASCOE: Because it's served to you, rather than you going in and reaching out and taking something random yourself? AMEET DOSHI: So it's more about the materiality, though. I mean, in a sense, we can't recreate that feeling of being in the stacks with the physical connection with paper and the smell of paper breaking down, I guess. EDMOND CHOW: All that mold. AMEET DOSHI: Yeah, all that mold. But I would challenge the assertion that the only way to achieve serendipity in the digital is to be served something, because you can certainly be shown a digital shelf, bookshelf. There are many of these around, of course. But it is such a vexing issue, because true serendipity is random. You're not even in the stack, because by making that choice to go into the QC, you're already bounded. But true serendipity is like the book taps you on the shoulder and you weren't expecting it. EDMOND CHOW: Right. That's a challenge for how we should create AI systems in the future to facilitate that type of serendipity. And maybe it leads to creativity. So, for instance, at the simplest level, you can make sure that your AI systems have some kind of diversity. In other words, it's going to suggest things that are immediately obvious, but then throw in things that might be a little bit off the beaten track or just loosely related, and see how you react to that. So that's at the very basic level. And then at a higher level, you can have environments in which the serendipitous types of encounters could be facilitated. And this could be, for instance, showing you what would be on the stacks, which I know the library website already has something like that. It's not quite the same as being in this. AMEET DOSHI: Smelling the mold. Yeah. EDMOND CHOW: But certainly a lot of the pleasure of it, a lot of the pleasure of being in the library is lost that way. But trying to preserve this serendipity or trying to recreate that serendipity, I think, is something that is within reach. FRED RASCOE: Well, maybe there's something more than serendipity then that AI offers. I'm thinking analysis of the corpus of electronic texts or their other applications that maybe we're not even imagining yet. Are there other applications that you can maybe conceptualize that aren't happening now but maybe could? EDMOND CHOW: Yeah, so technology can really help. In a way, the old serendipity, which is just walking around in the library and looking at books on the shelves and seeing what taps us on the shoulder, that's a pleasurable but relatively inefficient way of facilitating serendipity. Whereas machines reading books can really find a lot of connections that we might not find or we might not be able to put in linear order, like in call number order. So combining topics discovered automatically by a computer, finding the relationships between these topics, and also using the information of past users, of what did past users go to next after they looked at a specific book or journal article. That type of thing offers a lot more richness than we've had before. AMEET DOSHI: I'm imagining our colleague, Charlie, driving a nail through Moby Dick. And if you don't what I'm talking about, you can find out on the internet. You're listening to Lost in the Stacks, and we'll be back with more on the left side of the hour. RACHEL ROSENFELD: This is Rachel Rosenfeld coming to you from Hollywood, California, where I am the Archival Acquisitions Processor for the Academy of Motion Picture Arts and Sciences, Margaret Herrick Library. And you are listening to Lost in the Stacks on WREK Atlanta. [MUSIC PLAYING] AMEET DOSHI: Our show today is called AI in libraries. Here's a quote from Larry Page, Google co-founder from the year 2000. Quote, "Artificial intelligence would be the ultimate version of Google, the ultimate search engine that would understand everything on the web. It would understand exactly what you wanted and it would give you the right thing. We're nowhere near doing that now. However, we can get incrementally closer to that. And that is basically what we're working on." End quote. And by we, he means you. While we check our Gmail, file this set under TK7820.B63. [MUSIC PLAYING] ABBY: You just heard "The Life Machine" by Tubeway Army and before that, "Like Humans Do" by David Byrne. Those were songs about the lines between human and inhuman. FRED RASCOE: Welcome back to Lost in the Stacks. Our show today is called Skynet Fail, AI In Libraries. And our guest today is Edmond Chow, a Georgia Tech professor. Have we lightened it up enough, I guess, from that first? EDMOND CHOW: It's going to go lighter. FRED RASCOE: It's going to get even lighter. I think it might get lighter, because I want to talk about creativity and AI. In some of the AI systems that we have now, Siri, the recommendation machines that you mentioned that we interact with online, it's an imitation of what humans do naturally. It doesn't seem like it's a naturally created thing. And actually, we were talking during the music set, in between interview segments. And Abby, our board op, a computer science student, had a pretty good analogy about how computations can be described. ABBY: Yeah, I was just thinking when you were talking about or asking about how it takes so long to do the data processing side of AI. And to me, that's just a lot like, for example, if you were learning a new language. It's going to take you hours and hours and hours of practice to learn the associations, the meanings, the context. But ideally, when you're fluent, you know how to respond when someone speaks to you in a different language, and you know how to respond differently if it's a party versus a professional meeting. And so I feel like that's kind of where the long term data learning turns into immediate responses from Siri. FRED RASCOE: Right. So Siri, the programmers behind Siri, have figured that out for just like basic questions and pulling from Wikipedia and other online sources. But how close are we to making the leap to real creativity, writing real poems, not just an imitation or using jumbled up-- like those jokes that I used in the second segment. How close are we? EDMOND CHOW: Depends on what you mean by real. FRED RASCOE: Oh, man, are we real? Are we just holograms Ameet? What's going on? AMEET DOSHI: I mean, before I have my first cup of coffee. FRED RASCOE: Maybe we're imitations. Maybe we are imitations. AMEET DOSHI: Yeah, this is the idea of emergence. Is this the emergent computing area that we're starting to touch on? EDMOND CHOW: So right now, there's a lot of research that's going on that's producing very useful systems. But they're not trying to actually be intelligent or creative. It's an imitation of what humans do by studying lots of examples. So as you mentioned, two months ago, there was a piece of art created that was auctioned at Christie's and went for almost half a million. AMEET DOSHI: It went for a lot. EDMOND CHOW: Almost half a million dollars. And the question is whether or not that work is creative. FRED RASCOE: I wonder if the creative spark is like, we can train a computer to write something that we could identify as poetry or paint a picture that's something we-- OK, this thing has painted an abstract piece of work. But is the real essence of artificial intelligence when that system can just-- it can occur to itself, oh, you know what I need to do? Paint a picture or write a poem. Is that the line when we have achieved real artificial intelligence? EDMOND CHOW: I think there will be different levels of that. And certainly right now, systems can produce things that are beautiful in different ways, beautiful paintings, poems. I'm sure they're out there. Music, certainly. So systems can already produce these things. And it remains a challenge to AI researchers to develop systems that can actually go beyond that. Maybe something you want to call transcendence. AMEET DOSHI: Transcendence, yeah. Oh, it's the Johnny Depp movie. Have you seen this? FRED RASCOE: I have not. AMEET DOSHI: Yeah, you are in the majority. But I think that was the idea is that his brain gets downloaded into a computer and becomes the-- takes over the planet for a brief period of time before heroics emerge. I think for me, it's the creativity that's transcends time seems to be-- that stands the test of time. That I think might be the true test. Creating a piece of poetry that makes sense in 2018 is very different than something that's created in 2018 and is still beautiful in 2118. That is the true test to me, but maybe not. I don't know. EDMOND CHOW: I think a lot of artists feel offended when you say that computers can replace them, produce art, poetry. But we don't need-- the goal doesn't have to be creating systems with-- at least not necessarily. In the short term, there are great tools. So for instance, could use these systems to help inspire your creativity. In other words, just like a writer might use a thesaurus, you might use a system that suggests metaphors to you or suggests melodies. So they're very useful assistants right now. FRED RASCOE: When you said some artists might get offended if they said AI could replace them, I can tell you librarians have thought about this, and librarians get their backs up a little bit when you suggest that maybe AI will replace a librarian. EDMOND CHOW: Yeah, we don't need to help us find related articles. FRED RASCOE: I will tell you, having gone through the Library Information School master's program in the early 2000s, Google was the enemy at that time. And the predominating rhetoric was that Google could never duplicate what a librarian does. And now fast forward about 12, 14 years, and librarians turn to Google first. And so I wonder if you see is artificial intelligence going to have to get back to that creative spark, but in the library and reference mode before it can take over what a librarian does? Or do you see stages of AI taking over the different functions of-- EDMOND CHOW: It's definitely going to be in stages. So for instance, even doctors are using these systems to help in initial diagnoses, things like that. FRED RASCOE: Is there professional concern in the medical industry? EDMOND CHOW: Oh, definitely. I think that there's concern across the board whenever you're using these types of systems. There could be biases in these systems that you're not aware of. Therefore, for instance, if you're looking for materials, you're just steered towards certain types of materials and not others. And that can limit what we can do collectively if we rely on these systems too much. AMEET DOSHI: Well, as usual, on Lost in the Stacks, we are just getting into it, and now unfortunately, we have to end the segment. We've been speaking with Professor Edmond Chow from the Georgia Tech College of Computing about artificial intelligence and poetry as well. EDMOND CHOW: And the libraries. AMEET DOSHI: And the libraries. Dr. Chow, thank you for being on the show. EDMOND CHOW: My pleasure. Thank you. ABBY: File this set under QA76.9.C66G46. SPEAKER: Greetings, music lovers. [MUSIC PLAYING] (SINGING) Chains are formed from broken things Looking for a link Cold and clear ABBY: You just heard "Lone Bell" by Mount Eerie. Before that, "The Big Setup" by The Mad Scene. Those were songs about using caution. [MUSIC PLAYING] AMEET DOSHI: Today's show is called AI in Libraries. Fred, any final thoughts? FRED RASCOE: OK, well, in keeping with our goal of keeping it a little light, I went to poemgenerator.org.uk. And the artificial intelligence at that website wrote a poem just for us. Would you like to hear it? AMEET DOSHI: Sure. FRED RASCOE: OK. Edmond, would you like to hear it? EDMOND CHOW: Of course. FRED RASCOE: This is called Library by poemgenerator.org.uk. I should note that I put in some terms library and similar terms to that, and this is the poem that it generated. "Because I could not house for library It did kindly house for me Does the library make you shiver Does it All that is incomplete is not bibliographic Bibliographic, by all account is complete A bibliographic is skilled A bibliographic is downright A bibliographic is standalone, however I cannot help but stop and look at the all knowing Bodleian Never forget the wise and omniscient Bodleian Pay attention to the repository The repository is the most big sepulcher of all Are you upset by how broad it is Does it tear you apart to see the repository so large scale" AMEET DOSHI: poemgenerator.org.uk, a bibliographic is downright. Roll the credits. [MUSIC PLAYING] ABBY: Lost in the Stacks is a collaboration between WREK Atlanta and the Georgia Tech Library, produced by Charlie Bennett, Ameet Doshi, Wendy Hagenmaier, and Fred Rascoe. FRED RASCOE: Abby was our engineer today, and the show was brought to you in part by The Collective, a library conference designed to create collaborations between next generation academic librarians, archivists, and library staff. AMEET DOSHI: Find out more at thelibrarycollective.org, where you can see the program for next year's conference, register, and plan your travel. FRED RASCOE: Legal counsel and an upgrade to our Hal 9,000 computer were provided by the Burrus Intellectual Property Law Group in Atlanta, Georgia. AMEET DOSHI: Special thanks to Edmond for being on the show, to AI poem generators everywhere, and thanks, as always, to each and every one of you for listening. FRED RASCOE: Find us online at lostinthestacks.org. And you can subscribe to our podcast on Apple Podcasts, Google Play, and plenty of other places we don't even about. AMEET DOSHI: Next week on Lost in the Stacks, we reconnect with the Electronic Frontier Foundation. It's EFF part two. FRED RASCOE: Time for our last song today. And we close out this episode with a song that asks a question that is perhaps difficult to answer. As more and more of our actions are augmented by or outsourced to computing technologies, where is the line between what a human decides and what a machine decides for us? So let's end with a track from 1932, the good old pre-digital days, by the way. And it's called "Was That The Human Thing To Do" by the Boswell Sisters, backed up by the Dorsey Brothers, by the way. Right here on Lost in the Stacks. Have a great weekend, everyone. Downright. [MUSIC PLAYING] (SINGING) Never thought that anyone in their right mind Could ever treat another human so unkind Didn't you sneak away and leave a note behind Was that the human thing to do Always thought that yours was such a heart of gold But after I was sold on all the tales you told Didn't you let your kisses turn from hot to cold Was that the human