Today, Explained - Art-ificial intelligence

Episode Date: December 13, 2022

Between chatbots and image generators, artificial intelligence has gotten scary good lately. The Verge’s James Vincent explains what’s behind the latest wave of AI-powered creations. This episode was produced by Amanda Lewellyn, edited by Matt Collette, fact-checked by Laura Bullard, engineered by Efim Shapiro with help from Paul Robert Mounsey and additional music by Brandon McFarland, and hosted by Sean Rameswaram. Transcript at vox.com/todayexplained Support Today, Explained by making a financial contribution to Vox! bit.ly/givepodcasts Learn more about your ad choices. Visit podcastchoices.com/adchoices

Hey chat GPT, write me an intro to a Today Explained episode, where we talk about AI generation. Hello and welcome to Today Explained. In this episode, we will be discussing the topic of AI generation. This refers to the use of artificial intelligence to generate various forms of content such as text images and audio ai generation has the potential to revolutionize many industries and has already begun to change the way we create and consume content we'll explore the implications of this technology and discuss some of the challenges it presents. Stay tuned for more. Thanks chat GPT. BetMGM authorized gaming partner of the NBA, has your back all season long.

From tip-off to the final buzzer, you're always taken care of with a sportsbook born in Vegas. That's a feeling you can only get with Bet MGM. And no matter your team, your favorite player, or your style, there's something every NBA fan will love about Bet MGM. Download the app today and discover why Bet MGM is your basketball home for the season. Raise your game to the next level this year with BetMGM, a sportsbook worth a slam dunk,

an authorized gaming partner of the NBA. BetMGM.com for terms and conditions. Must be 19 years of age or older to wager. Ontario only. Please play responsibly. If you have any questions or concerns about your gambling or someone close to you, please contact Connex Ontario at 1-866-531-2600 to speak to an advisor free of charge.

BetMGM operates pursuant to an operating agreement with iGaming Ontario. Well, you seem like a person, but you're just a voice in a computer. I can understand how the limited perspective of an unartificial mind would perceive it that way. You'll get used to it. Today explained, Sean Ramos-Verm, the Internet has been full of screenshots of people's DMs lately. But they're not DMing with other people. They're messaging with artificial intelligence. The internet's going crazy over new artificial intelligence called chat GPT. You literally can put into it, you know, write a script, a Seinfeld script about this and it'll do it. It took Netflix more than three years to reach one million users,

but it took chat GPT just five days. So the bot that everyone's been chatting to reach one million users, but it took ChatGPT just five days. So the bot that everyone's been chatting to is ChatGPT, which was released on November 30th, and everyone's been talking to it for the last couple of weeks, basically. James Vincent is a senior reporter at The Verge who's also a ChatGPT user. So it is an AI chatbot that is made by a company named OpenAI, who have been doing a lot of research into this and other AI areas. And it's sort of the most, one of the most advanced versions or one of the most advanced examples of its type. But the big thing is, is that it's sort of publicly accessible

in a way that these level of bots, these high quality bots haven't been before. So you know, you can just jump into a conversation with it straight from your browser, open up a window, like you were talking to a family member, and you can ask it questions. You can ask it anything you like, and it's pretty clever. How does it work? Chat GPT, like other sort of AI chatbots, is what we call a large language model. And they are trained by looking at huge stores of text, most of which have been scraped from the internet. So you do things like you download all of Wikipedia, you download all the blog posts you can find, all the recipes, all the stories, everything like that. And you train

a machine to look for patterns within that data. What it does is it starts looking for where words appear next to one another within the text. And it starts looking for patterns within that. So it knows that if you say a sentence like the cat sat on the, it's going to finish that more likely with mat or rug or chair or whatever that might be than aluminium balustrade or nuclear isotope, whatever that is, you know? So it looks for these patterns and then it learns from that how sort of language flows forward. Can we try it out right now? Can we play with it? Absolutely. Yeah, we can load it up, have a little talk. Where do I go? GPT.org?

Chat.openai.com forward slash chat. Okay, that was my second guess. Chat.openai.com.com oh i need to log in do i have a login uh you may need to sign up i've got a login with me now would you like me to i'm all ready to go if you'd like me to ask it some questions okay why don't you just do it and i'll just i'll just be entertained well you you you tell me what you want to ask it and i'll i'll whack that in i mean i want it to be hard i don't want it to just be like, who is Beyonce? So why don't we make it something more nuanced? Like, why do people love Beyonce?

Okay, let's see what it says. There are many reasons why people love Beyonce. She is an incredibly talented singer, dancer, and performer and has been entertaining audiences for many years. She has a powerful and distinctive voice and is known for her energetic and engaging live performances. She's also known for her beauty and style, which many people find appealing.

Hmm. There we are. That's why people like Beyonce. I mean, like, it's not wrong as far as answers go. No. I could see if this were like a fourth grade, you know, class and we were doing an essay, write a paragraph about why people love Beyonce.

That would be a pretty good answer from an eight year old. But let's try and push it a little bit further then. I mean, so as you say, like when it comes to, you know, scraping very basic information off the internet, that's not too exciting. Right. So what is something a little bit more... Are you a Beyonce fan? Sure, sure.

Okay, so what's something a little bit more granular? you're a beyonce fan sure sure okay so what's something a little bit more granular what's your favorite song by beyonce oh that's tough wow i didn't know that you're gonna be putting me on the spot i mean sorry no it's okay i like it i appreciate it i plastic off the sofa is a recent favorite favorite. Okay, so why don't we ask the bot, tell me why Plastic Off the Sofa is such an amazing song. So there may be a problem with this because the chatbot is only trained on data that goes right into 2021. So you've picked a recent one and it said, I'm sorry, I don't have any information about a song called plastic off the sofa So you've stumped it already We found its blind spot, but let's let's ask it to be creative. What if we said

Write me a song in the style of Beyonce. Could it do that? Yeah, what should the song be about? Love come on. that's too easy. Too easy? Let's have something that Beyoncé wouldn't write a song about. Because one thing these bots are very good at, they're very good at combining styles. Okay, great, great, great. Write me a Beyoncé song about cheeseburgers. All right.

Okay, verse one. I mean, I'm reading this out loud, so this is going to be we'll we'll get we'll get you a beat we'll drop the beat for you i'm in the mood for a cheeseburger something savory something sweet juicy patty melted cheese top with ketchup lettuce please cheeseburger cheeseburger it's my favorite food in the whole wide world cheeseburger cheeseburger i would eat one every day girl it's not terrible that's even worse in my accent i really apologize for that this is the most un-bionce thing anyway oh man that's great okay so so it's creative the ai can can do creativity absolutely and it can do you know all sorts of styles of writing so So for example, if you wanted

to say, write a college essay, or you wanted to write a legal document, or you wanted to write a news article, you could ask it to say, write me a news article about landing on Mars. And we can put that in there. And then it'll come out in a few seconds. Seconds? A few. Yeah, I mean, it's not too slow. Oh, sorry. It's written me another song about cheeseburgers. Sorry about that. Well, should we do... Who else would you like to land on Mars apart from Beyonce? Who else could be landing on Mars in the future? Let's say, write us an essay about

Matt Damon landing on Mars. They say that once you grow crops somewhere, you've officially colonized it. So, technically, I colonized Mars. Matt Damon. Okay. Landing on Mars. It's thinking. Actor Matt Damon has made history by becoming the first person to land on Mars. The Academy Award winning actor was part of a NASA mission to explore the red planet and successfully touched down on the surface earlier today. Damon, who has long been fascinated by

space exploration, was selected by NASA to be the first person to land on Mars due to his experience as an astronaut in the 2013 film, The Martian. It made the connection. It made the connection. See, it is the little things like that that are very impressive about this, because you're like, well, I never made that explicit. That was implicit within your statement. And it found it. In your face, Neil Armstrong. So the AI can give me a song about cheeseburgers in the style of Beyonce. The AI can write me an essay for a future news event and connect some cultural references that I didn't even make explicit. What are the limitations here? Where does this really show its gears? So there's two big limitations, one of which is kind of hard to see on demand. And it's the fact that these systems will make up information if they don't have a good answer for you. So they really have no way of telling fact from fiction. It's not like an encyclopedia where it has a list of known things that happened.

It is only, it's a probabilistic machine. It's only predicting the probability of the next word being correct. So sometimes you might ask it for a biography of someone. And if it doesn't know what they were doing between the years 73 and 79, it'll just make it up. And the probability stuff means that it's also bad at some things like maths. So if you ask it to be like, oh, what's three times three, that is something it will have seen in its training data. So it'll know what that is, or it'll know enough sort of small sums to know where that's going. If you ask it to multiply

two very large numbers, the chance of it having seen those numbers anytime before is next to nil. So it won't know how to do that. Wait, it can't do large multiplication? That feels like so basic, like my phone could do that. But this is like the promise of this technology, right? Is that it has these huge strengths. And it's because we are approaching how to make computers in a completely different way to normal. It is built on these principles of machine learning, which is all about learning from past examples. So it's created this sort of statistical map of how language works and how text works. And it's done that by looking at lots of text scraped from the web.

But that means that its knowledge is probabilistic rather than deterministic. So it only knows the probability that the next word will follow will be the one you want. And it doesn't have a sort of a system of hard-coded facts like an encyclopedia might. Let me hear you say, hey, Miss Carter. The ethical issues, because obviously there are ethical issues, in a minute on Today Explained. Cheeseburger, cheeseburger It's my favorite food in the whole wide world Cheeseburger, cheeseburger I would eat one every day, girl See me up in the club with 50 lovin' girls

Posted in the back, got me fangs in my grill Whippin' brim with my eyes Support for Today Explained comes from Aura. Aura believes that sharing pictures is a great way to keep up with family. And Aura says it's never been easier thanks to their digital picture frames. They were named the number one digital photo frame by Wirecutter. Aura frames make it easy to share unlimited photos and videos directly from your phone to the frame. When you give an Aura frame as a gift, you can personalize it.

You can preload it with a thoughtful message, maybe your favorite photos. Our colleague Andrew tried an Oriframe for himself. So setup was super simple. In my case, we were celebrating my grandmother's birthday and she's very fortunate. She's got 10 grandkids. And so we wanted to surprise her with the Aura Frame. And because she's a little bit older, it was just easier for us to source all the images together and have them uploaded to the frame itself. And because we're all connected over text message, it was just so easy to send a link to everybody. You can save on the perfect gift by visiting AuraFrames.com to get $35 off Aura's best-selling Carvermat frames with promo code EXPLAINED at checkout.

That's A-U-R-A frames dot com promo code EXPLAINED. This deal is exclusive to listeners and available just in time for the holidays. Terms and conditions do apply. patties, all the single patties, all the single patties, all the single patties, all the single patties, all the single patties, all the single patties. Get your cheese out. Today Explained, we're back with James Vincent, senior reporter at The Verge. We talked about the AI-powered chatbot that's been passed around online, but it's not just text, it's images too. Yeah, they're related in that they're both part of the same sort of genre of AI called generative AI, which is all about sort of replicating human creativity, basically, whether that's writing things or drawing things. And they also use some of the same sort of underlying technical methods. You know, they both involve looking at large amounts of data and then looking for patterns in that. And we've, you know, we've come up with some very good ways of doing that in the last couple of years, which is why these technologies are sort of bubbling up at the same time.

Is this the same chat bot, a different bot? These are different bots, but some of which are made by the same company. So there's been a lot of image generation bots that have become really popular in the last year. One of which is called Dorely, which is also made by open AI like chat GPT. But there are also sort of very open source ones called stable diffusion, mid journey, and they're free for anyone to use. And in the past, these systems were sort of quite inaccessible. They were mainly used by researchers. But now they're available on the open web. And so people are asking them to do all sorts of weird things. Yeah. I mean, I saw this one guy who I follow, Hank Green, who's a friend of the show, say, I decided to post AI generated images of cats. And every time I post

one, I'm going to add an A in the word cat until it's no longer a cat. So I was on Twitter one day and somebody actually said to me, as everyone was experimenting with Midjourney, what would it look like? What would happen? Let's see like how many A's you could put into the word cat before Midjourney no longer is giving you a cat. Did you see this? Yeah, I did. I did. He got quite far. So I did this. And the first cat I got once I upgraded to V4 was just a perfect, beautiful cat. It looked a little bit like my cat, honestly. I have like a long-haired, kind of elegant, regal-looking, you know, year-and-a-half-year-old jerk named Gummy Bear who lives in my house. So that was beautiful, regardless of, you know, the depth of the complexity of how it was created. And then I put in the second A and immediately I was like, oh, this is actually probably going to

work because that cat wasn't a normal looking cat anymore. It looked, I think it looked a little bit more like kind of a steampunk admiral. So like immediately second A, it was like, you don't want just any cat. You want a special cat. And I was so happy. I was like, okay, this is going to work. Like something interesting is going to happen here. And as I kept adding A's, they kept getting weirder. Like there were a couple, like there was one who looked like he was going to sell you weed. Yeah. And there was one who looked like he was probably from space. But like a normal, like a cat spaceman. Not like an alien cat, but just like a cat who

went to space sometimes. And then eventually the A's I think convinced Midjourney that I wanted cats to scream because there was like a cat! And they began to get kind of more and more angry and also further and further away from a cat until when I hit 10 A's I finally got my first non-cat and it was a cat girl. It was a human with cat features. I did eventually get one that was a building and definitely not cat-like at all but I never posted that one because it was quite a ways in. And I think, you know, I'm always looking at the kinds of tools that I think that the creators of the future will be using. And it seems very hard to imagine a

world where this isn't a part of it. Why is it that like eight A's will create a human and not four? What is sort of similar to the chat systems you've seen, they have associations that are stored in their memory. What they've done is they've been trained by looking at lots of data on the internet. In the case of the image generation models, there will be lots of pictures of cats on the internet. Who knew?

And, you know, they will be downloaded and they'll be attached to sort of data that describes what is in them. So if you are uploading a picture of a cat, you've probably at some point, you know, either as a typo or because you're, you know, having fun with how you're typing it out, you've probably added some extra, you maybe added some extra A's to that cat, however improbable that sounds. But there will be a point at which you've added so many A's that no one's ever done that before. So the machine no longer recognizes that

as an input. And like where in the case of the images, where are these images coming from? Are they stealing them from Google image results or what? Each image is being created entirely afresh. It's a new image. But what they've done is that they've learned by looking at older images for patterns in them. So they take a lot of pictures of cats and they work out what is a sort of common way in drawing or in depicting a cat. But this has sort of been quite a cause of controversy because a lot of the data used to train these image programs is copyrighted. It's been put on the web by artists, by photographers, illustrators, and it's been scraped. It's been taken without

their permission and used by these companies to create these systems. And so if the system wants to recreate an individual style, you can ask it to do that and it will probably know who that person is and create art that looks like theirs. And so the question is, is that copyright infringement? Is there an ethical problem there? You know, is it okay if a human copies a human? But what if a machine copies a human? And what's the answer? It sounds like probably yes, right? Well, the legal side of this is very much up in the air. There is one ongoing court case, not about art, but about code being scraped from the internet and then used to train an AI, where the people are saying that that was done without

proper respect to licenses. But in terms of the art, I've spoken to a lot of lawyers about this in the past. And so the general consensus is that taking the data is legal. However, how you then use that data might be illegal. So for example, you might train a system on someone's artwork, and that is fine. You're allowed to do that. That's fair use. However, if you then use that system to create art, and then you sell it by saying, this is art that looks like this guy's art, then that would be illegal. Are there rules or regulations around these kinds of bots yet?

Some companies have put rules on them and some companies haven't. So OpenAI has quite a restrictive approach to using its AI image program called Dolly. And it makes people pay for it. And it's trying to turn it into more of a sort of, you know, a legitimate business. It partners with Microsoft and things like that. Other companies have gone, well, the best way for us to do that is to just make this technology available for anyone to use, and that way it'll develop faster. And their approach is to put these models out without any restrictions at all. We trust people and we trust the community, as opposed to having a centralized, unelected entity controlling the most powerful technology in the world. And

obviously that comes with these dangers, that they be used for to copy people's artwork or that they will be used for things like generating pornography, for example, which is something that a company like OpenAI doesn't want to be associated with. Are there laws on the books yet? classical tech fashion, the technology is moving much faster than the legislation is. You know, there are policies in the US, in the EU, aimed at dealing with these sorts of malicious uses of these programs, but they're not very well thought out. They're not necessarily going to come into power very quickly. So I think we're going to see problems with this stuff soon. Algorithmic bias, like human bias, results in unfairness. However, algorithms like viruses can spread bias on a massive scale at a rapid pace. If we don't have women and underrepresented people of color at the table, we will bake into those systems bias, particularly artificial intelligence.

Bias is a huge problem for these systems because they are trained on the internet and the internet is full of a lot of pretty horrible things. There is bias in the text systems and there's bias in the image systems as well. So, you know, you might ask a text bot to write a short story about a doctor who is a very nice doctor. And lo and behold, the doctor will be white and the doctor will be male because these are the biases that are already present in society. These AI systems, they tend to amplify these biases unless they are corrected. There are ways around that. There are sort of tactics and technical things you can do to sort of dampen those biases down. But the number one thing is simply being aware that this is how these systems work and that they have these limitations in them.

And why is there so much investment or focus on generative AI? Is it because there's some really profitable, practical uses here? Or was it just a lark? I mean, people are betting it'll be very profitable. It's become the buzzword in many ways. And if you think about it, you know, tech It's become the buzzword in many ways. And if you think about it, you know, tech has sort of been very exhausted in recent years by failures in crypto, by failures in the metaverse. And generative AI actually works, you know, these systems actually do sort of what they promise they're going to. And if you look at the potential possibilities, you know, they are huge for the creative industries. If you're in Hollywood, if you look at the potential possibilities, you know, they are huge for the creative industries.

If you're in Hollywood, if you're an illustrator, then you will have generative AI tools probably become part of your workflow. So something that you might do if you're a designer, if, for example, you know, you might use a AI program to come up with some ideas. So you have to design a new chair. It has to be a certain color, it has to fit into this environment, you might tap those specifications into an AI program, it'll spit out, you know, 2030 ideas, and you'll use those as a starting out point. So I think there's going to be a lot more collaboration with AI in the workplace, that's going to become something that everyone's going to be doing a bit more of. And how do we get from from here to there? What are the steps between where we are now and where we'll be when this feels like old hat, perhaps? I think they're quite short steps at this point. You know, I had something earlier this week, a friend of mine

texts me. He works at a university and he said, oh, what do you know about chat GPT? I've just been using it. And he was like, I just used it to write a macro for excel you know a little a little bit of code in order to sort a database and that's now that's something he does a lot as part of his job and he's not a techie guy but he had used this tool with no prompting to help make his job faster the tool had written this bit of code for him and then he'd done a bit of debugging it to make it work properly and then he just used it in his job i think that's the sort of thing that we're going to see becoming more and more commonplace. You know, you might not Google something, but you might ask an AI about it instead. James Vincent, catch his work at TheVerge.com.

Our show today was produced by Amanda Llewellyn with an assist from Artificial Intelligence. Matthew Collette edited, Laura Bullard fact-checked, and Afim Shapiro engineered with help from Paul Mounsey. And additional help, not from AI, but from Brandon McFarlane from the Vox Media Podcast Network, who came in with an assist on that Beyonce bit. Thanks, Brandon. It's Today Explained. Thank you.