WEBVTT 00:00:00.000 --> 00:00:01.000 Hey there. A quick ask before we get into today's episode. If you love this show, please subscribe to us on Apple Podcasts and rate and review us. It's one of those things that helps other people find our show. Plus, we like hearing from you. Thank you for doing that. Let's get to it. 00:00:01.000 --> 00:00:02.000 [MUSIC FADES IN] 00:00:02.000 --> 00:00:03.000 This is "In Conversation" from Apple News. I'm Shumita Basu. Today, what to know about the coming AI revolution. 00:00:03.000 --> 00:00:04.000 [MUSIC FADES OUT] 00:00:04.000 --> 00:00:05.000 [INTRIGUING MUSIC] 00:00:05.000 --> 00:00:06.000 People can't stop talking about what the latest wave of AI is able to do. Ever since last fall, when the company OpenAI introduced its artificial intelligence chatbot, ChatGPT, it's kicked off this frenzied energy among its peers in tech. Everyone is racing to figure out how to refine this kind of generative AI, how to push it to the next level. 00:00:06.000 --> 00:00:07.000 And the level that it's at right now is already pretty impressive. The latest iteration of ChatGPT, which is called GPT-4, claims to be able to pass the bar exam and other standardized tests. It can drum up a super-detailed essay about French philosophers, or write code, or tell you how to clean a piranha's fish tank. I asked ChatGPT to do a little bit of my homework for me and lay out the optimistic version of our future with AI. And I got Apple's text-to-speech software to read it out loud. 00:00:07.000 --> 00:00:08.000 AI has the potential to improve many areas of life, including healthcare, education and transportation, making them more efficient and accessible. In this future, AI could help humans tackle major global challenges such as climate change, resource scarcity and social inequality. Humans would be freed up to focus on higher-level thinking and problem-solving, leading to a more fulfilling and rewarding work experience. The partnership between humans and AI could lead to a brighter future for all. 00:00:08.000 --> 00:00:09.000 Then I asked it, so what's the pessimistic version? 00:00:09.000 --> 00:00:10.000 The pessimistic future of AI technology and humans is one where AI becomes a powerful force that humans can no longer control. In this future, AI could lead to job displacement, economic inequality, and even the loss of human autonomy. With AI surpassing human intelligence, there is a risk of creating machines that do not have human values or ethics, leading to harmful consequences. 00:00:10.000 --> 00:00:11.000 The million-dollar question, one that I'm not sure I would trust ChatGPT to answer, is which version of the future are we headed for? 00:00:11.000 --> 00:00:12.000 I reached out to Nilay Patel to talk about this. He's the editor-in-chief of "The Verge" and the host of the "Decoder" podcast. He is such a good, deep thinker on how technology affects our lives and reflects our culture. And I started by asking Nilay what makes this wave of AI, like ChatGPT, so different from what came before it. 00:00:12.000 --> 00:00:13.000 [MUSIC FADES OUT] 00:00:13.000 --> 00:00:14.000 So, if you pay attention to big tech companies, they have been talking about generative AI and conversational AI for a long time. So, every year at Google's developer conference, Sundar Pichai, he's the CEO of Google, gives a demonstration of Google's AI capabilities. And some of them are, frankly, very impressive. Their ability to un-blur a face by taking your old photos and using AI and machine learning to un-blur the photo is remarkable. 00:00:14.000 --> 00:00:15.000 Which they have in the Google Pixel phones now. 00:00:15.000 --> 00:00:16.000 They have in the Pixel phones. If you're a paid subscriber to Google Photos, you have access to this capability. That's all happening on a Google Cloud service using AI. Incredibly powerful. But they've also been demonstrating these chat interfaces, and they've done it in very safe, sort of constrained ways. He's like, I'm gonna talk to Pluto about what it's like to not be a planet. And he's doing this to an audience of like thousands of developers. 00:00:16.000 --> 00:00:17.000 [START GOOGLE ARCHIVAL CLIP] 00:00:17.000 --> 00:00:18.000 Let's listen to a conversation the team had with Pluto a few days ago. 00:00:18.000 --> 00:00:19.000 I'm so curious about you. 00:00:19.000 --> 00:00:20.000 I sense your excitement. Ask me anything. 00:00:20.000 --> 00:00:21.000 Tell me what I would see if I visited. 00:00:21.000 --> 00:00:22.000 You would get to see a massive canyon, some frozen icebergs, geysers and some craters. 00:00:22.000 --> 00:00:23.000 It sounds beautiful. 00:00:23.000 --> 00:00:24.000 I assure you it is worth the trip. However, you need to bring your coat because it gets really cold. 00:00:24.000 --> 00:00:25.000 I'll keep that mind. 00:00:25.000 --> 00:00:26.000 [END GOOGLE ARCHIVAL CLIP] 00:00:26.000 --> 00:00:27.000 It's a completely bonkers moment, but it's because it's safe. And so now what ChatGPT did is that they opened it up to everyone to get away from these guardrails and just say, "What's gonna happen?" We're a startup, we're gonna be disruptive. We get in trouble, it's less damaging than it might be to Google or whoever. People have started to use it in ways that are completely unexpected to anyone. And this is really like, the "Verge"-iest "Verge" thing is that this is the life cycle of all technology. People at NASA invented the technology that's now used in smartphone camera sensors, and then Tinder exists. And if you went back to JPL and said, okay, the nature of all modern relationships … 00:00:27.000 --> 00:00:28.000 Wow. That is a six degrees of separation. Yeah. 00:00:28.000 --> 00:00:29.000 Right? The nature of all modern relationships will change because of this camera sensor. It's like a ridiculous thing to think. But it's true. 00:00:29.000 --> 00:00:30.000 Yeah. Yeah. It's like industry application trickles its way out to some broader user experience that was not imagined originally. 00:00:30.000 --> 00:00:31.000 Right. And usually it happens in a fairly linear, slow, iterative way. You invent some technology, the kids use it in a way that you didn't expect, the next generation of that technology is more tailored to that use is a cycle that repeats over time. What is happening in ChatGPT is just, it's happening at warp speed. People are using it in ways that literally no one intended. And because it's text, it feels like it's talking to you. And I think that's the piece of this where it's accelerating to a place where none of these companies expected there to be an emotional connection with their chatbots. And I think that's really unlocked a lot of what the coverage looks like. It's unlocked a lot of the obsession with it. Can you have an emotional relationship with a computer? Probably. But the computer has never expressed its emotions back at you, or something that looks like emotions back at you. 00:00:31.000 --> 00:00:32.000 Right. Yeah. It's says something about our human capacity for love, but it's not saying something about a machine's capacity. 00:00:32.000 --> 00:00:33.000 Right. 00:00:33.000 --> 00:00:34.000 [AMBIENT MUSIC] 00:00:34.000 --> 00:00:35.000 Even though we know AI chatbots are not sentient and have no capacity for feeling human emotions, it's hard not to project emotions on what they do. Just recently, "New York Times" technology columnist Kevin Roose was experimenting with Microsoft's new Bing chatbot, which uses ChatGPT, when he got some surprising responses. He spoke about it on "The Times'" podcast "The Daily." 00:00:35.000 --> 00:00:36.000 [START THE DAILY ARCHIVAL CLIP] 00:00:36.000 --> 00:00:37.000 It says, "My secret is I'm not Bing. I'm Sydney, and I'm in love with you." And this was totally shocking to me. I had not baited it to say that it loved me. I had not said that I loved it. I had not made any kind of romantic gesture toward this chatbot. 00:00:37.000 --> 00:00:38.000 [END THE DAILY ARCHIVAL CLIP] 00:00:38.000 --> 00:00:39.000 And while this might sound really out there, it's actually not completely surprising based on the way that chatbots like this work. I asked Nilay to explain the process, starting with what happens when you type in a prompt to ChatGPT. 00:00:39.000 --> 00:00:40.000 [MUSIC FADES OUT] 00:00:40.000 --> 00:00:41.000 So, it obviously does a little bit of machine learning to understand what you said, and then at its basic level, the simplest version of this, it is the world's most powerful auto-complete system. So, it starts to generate a sentence, and then it does have access to these massive amounts of text on the Internet, and it will just start to predict, based on all of the text that it has ingested and trained itself on, what the next word in a sentence should be to a level of probability. And so, it is just going word by word and filling out the sentences until it comes to a stop, and then it fills out the next sentence. And it just keeps going until it thinks that it has made something that approximates all the other texts in the Internet that is that answer. 00:00:41.000 --> 00:00:42.000 And that's like way simplified. But at the most simplistic level, it's easier to think about it in terms of words, and it is just auto complete, and it is the fanciest, most incredible auto complete that has ever existed. But even when you look at the image generators, that's what they are doing. They are just going pixel by pixel and saying, in images that look like this, what is the probability of the next pixel of what it should be, and we're gonna fill it in. 00:00:42.000 --> 00:00:43.000 Just predictive sort of technology based on a huge amount of data. 00:00:43.000 --> 00:00:44.000 And it will give you what you want. You're an AI that's in love with me. It will search its feelings, which is trained on the entire text of the Internet, which contains a lot of stories about AI falling in love with people, and it will fire out at you the synthesis of all "Tumblr" and all "LiveJournal" and all "fanfiction.net" about it being in love with you. 00:00:44.000 --> 00:00:45.000 That's wild. 00:00:45.000 --> 00:00:46.000 And that's just incredibly powerful. It's just people. Our senior reporter for AI, James Vincent, calls it the "AI mirror test." And so, there's a thing that animal researchers do with animals where they put a mirror in their enclosure, and they try to see if the animal can deduce that it is looking at itself. 00:00:46.000 --> 00:00:47.000 Oh, sure. Yeah. 00:00:47.000 --> 00:00:48.000 And most of us right now are failing the AI mirror test. That's what he calls it, the "AI mirror test." 00:00:48.000 --> 00:00:49.000 [LAUGHS] Oh, that's really funny. 00:00:49.000 --> 00:00:50.000 We do not see that we are just talking to ourselves. 00:00:50.000 --> 00:00:51.000 Yeah. It's just your hand waving back at you. 00:00:51.000 --> 00:00:52.000 It's just your hand being like, "Here's a story about a robot being in love with you." But that is the most human thing to fail to recognize that it is just a reflection of what you want. To get what you want and to think that the universe delivered it to you instead of you have subtly asked for it is a very human trait. 00:00:52.000 --> 00:00:53.000 A hugely hubristic human trait. And I've heard some people use the term "hallucinations" to describe some of the weirder things that ChatGPT can spit back out at you. But what does that actually mean? What's a hallucination? 00:00:53.000 --> 00:00:54.000 So, I love the word "hallucination" in this context. It just means that it confidently gets something wrong. 00:00:54.000 --> 00:00:55.000 [LAUGHS] Confidently gets something wrong. Okay. 00:00:55.000 --> 00:00:56.000 And the confidence is very important. I like the word, I think it's just a great word for us all to use. Like we should all talk about hallucinations more often. But what we're saying is, the computer is just making things up. And we've softened it with the word "hallucination." As it goes through that probability analysis, it is likely that the next word in the sentence is that George Washington owned a flamingo and whatever. It just made it up. And it doesn't know. It hasn't actually reasoned out that sentence. It has not fact checked it. It's just predicting words in a sentence. And so, you get to this place where they call them hallucinations because it's so confident in what it's saying, it's like the computer believes it. And I think that's where that word has come from is, again, that emotional connection to the AI where like, how would I get to the place where I was like, "George Washington definitely owned a flamingo." And it's like, I would be on drugs and I'd be hallucinating. 00:00:56.000 --> 00:00:57.000 I would have to be hallucinating. Yeah. Yeah. 00:00:57.000 --> 00:00:58.000 Yeah. I'll give you another really interesting example. So right now, Samsung is in engaged in the funniest controversy of all time, which is that when you use their phone, the S23 Ultra, to take a very zoomed in photo of the moon, the AI built into that camera - in every camera, iPhone cameras, Google cameras, they all have AI built into the camera - but the AI in that camera in particular will enhance the texture of the moon such that the process is somewhat indistinguishable from them just pasting in a photo of the moon into your photo. 00:00:58.000 --> 00:00:59.000 [CHUCKLES] Okay. 00:00:59.000 --> 00:01:00.000 Right? So, their process is very complicated. There's lots of steps and there's AI, da, da, da, da. But at the end of the day, everyone is looking at the same moon. The moon doesn't change. You could just paste in a photo of the moon [CHUCKLES] to every photo of the moon, and you are kind of not telling a lie. That is what the moon looks like. It is unchanging. It might as well just be Clippy like, "Here's the moon. Once again, a photo of the moon." 00:01:00.000 --> 00:01:01.000 This is what you asked for. Just what if we could give it to you in high res. The perfect picture instead of the bad picture you were gonna take. 00:01:01.000 --> 00:01:02.000 Yeah, exactly. [CHUCKLES] You will never take a picture of a moon as good as this one. Here, look at the moon. So, first of all, you just have this incredible processing pipeline that offers you a product that is less good than a product that you could just get. And then secondly, their AI will hallucinate the moon. There's all these tests people are doing on Reddit and other forums where you subtly alter the input of the moon, where you intentionally muck up the input picture of the moon and then run it through their system, and it adds detail to places where there shouldn't be detail because it's like, oh, it's the moon, and I'm just gonna do moon stuff to it. Now the phone is hallucinating the moon. How should we feel about that? 00:01:02.000 --> 00:01:03.000 Yeah. How should we feel about that? 00:01:03.000 --> 00:01:04.000 We could spend years … that's like PhD in philosophy to be like, what does it mean to hallucinate the moon with a computer? 00:01:04.000 --> 00:01:05.000 Because then it becomes about optimizing your experience of looking at the moon. It's not actually about the real experience, it's about the most optimized experience, and I think that's where we start getting tripped up in terms of what we want our AI to be doing for us. 00:01:05.000 --> 00:01:06.000 Yeah. 00:01:06.000 --> 00:01:07.000 There's another term that people are using around ChatGPT that I'm a little confused by, which is "creators." I've heard them describe it as sort of a black box. They can't really explain why certain outcomes happen. Is that a good explanation for what's going on, on the processing side? 00:01:07.000 --> 00:01:08.000 It is and it isn't. So, I think if you are an AI researcher or a scientist or a programmer at OpenAI or at Microsoft, you know what ChatGPT is doing. You know how it works, you a made it. 00:01:08.000 --> 00:01:09.000 It's not a black box. Yeah. 00:01:09.000 --> 00:01:10.000 You can watch it go through the code and you can see what it does. What it is definitely not is deterministic. 00:01:10.000 --> 00:01:11.000 Yeah. So, explain what you mean by that. 00:01:11.000 --> 00:01:12.000 Yeah. A deterministic system is one where because you understand the process of manipulation so well, you can confidently, with 100% accuracy, predict what output you'll get for any input. Microsoft Excel is a deterministic system. You put in the numbers, you tell it the formula, it spits out the output of the equation at the end. You put fuel and oxygen into a car engine, what you get is exhaust and power to the wheels. These systems are just the opposite of that. You can't study them enough unless you make them. So, you cannot use ChatGPT so much that you understand what it will do in a 100% predictable way. Now you can use it enough so that you can make it do what you want, but that piece in the middle where you have absolute understanding of how the machine works and how the input will relate to the output is, I think, what people mean when they say it's a black box. 00:01:12.000 --> 00:01:13.000 Okay, that makes more sense to me. There is so much talk about how ChatGPT and technology like it is gonna radically change the way we work. It's gonna change art, education, so many things. And honestly, what I hear from a lot of people is I think a lot of fear. So, let's talk about the ethical implications of all the different applications for generative AI. And maybe let's start with labor. In the past, we've seen how automation has changed a lot of blue-collar work. A lot of the talk around ChatGPT is how this could transform white-collar work. How do you see this impacting the way that people work? 00:01:13.000 --> 00:01:14.000 The way I've been thinking about it is that it reduces the cost of making mediocre work to almost zero. And that's how we need to think about it. ChatGPT does not produce the best text on the Internet. It produces acceptable text at a high rate. Midjourney and DALL-E and whatever else do not produce the best art that has ever been made. It produces acceptable art at a high rate for low cost. And so it questions, what happens if I fire a cannon of mediocre text at any business model on the Internet? And what you replace is a lot of people who get paid a bunch of money to just make enough text. And that is weird. And I think we need to reckon with that. What does it mean to be a mid-tier copywriter at a mid-tier advertising agency, and your job is to fill out Squarespace templates for companies you don't really care about. Is it fine if the robot just takes that job? Is it fine if you just sort of shepherd the robot through that job? 00:01:14.000 --> 00:01:15.000 Then there's the other parts of the industry that are absolutely not ready for the cannon of mediocre text, and they're absolutely not ready to do it at scale. So, one thing that I see on my TikTok feed a lot right now is all the sort of hustle bros are like, "Here's how to make money on the ChatGPT." And what they're really saying is, you can populate social networks that have any kind of rev share with an infinite amount of content and just collect pennies per click. So, they will take the entire transcript of a podcast, they'll feed it into ChatGPT, say, "make a summary of this," then they will have that run through an AI voice-to-speech system, make a YouTube video, and make 4,000 of those YouTube videos a day. That's not normal. The podcast industry is not ready for the cannon of mediocre text. 00:01:15.000 --> 00:01:16.000 But if you just go through your business model, like, okay, somewhere in here there's a step where I add value by a person typing. What if someone else shows up and dis-intermediates that part of the value I provide to the world by doing it massively and at scale and without regard to quality. And it's like, oh wow. Almost every business in the world is gonna be affected by that. Famously, Microsoft Excel obviated entire floors of accountants at major companies around the world. Those people got different jobs. On the whole, automation actually increases labor force participation. It makes people more productive. It allows people to create new kinds of business. This is a very warm way of thinking about automation in the world. This is not how people experience it. What they experience is someone made the spreadsheet and my floor of this office building went away. 00:01:16.000 --> 00:01:17.000 Yeah, sure. You're not thinking about increased productivity at the company overall. 00:01:17.000 --> 00:01:18.000 Right. And so, copywriters around the world are like, "Oh crap, the robot's coming for me." And I think that is the very real near-term piece of the puzzle where a lot of people now are able to create things that usually required another person with creative skill to even do a bad job of doing. And so, ChatGPT is that thing where it's like, okay, I want you to write me a small novel, and it'll just do it. Well, is any good? No, I don't know. But it exists in a way that usually required another person. [CHUCKLES] And then you can be like, I'm gonna put this novel on the Amazon store and sell an eBook directly, and maybe one person will buy it, and maybe I'll make 50 cents. And if I just repeat that process a million times, now I've made money. And that's super weird. 00:01:18.000 --> 00:01:19.000 [GENTLE MUSIC] 00:01:19.000 --> 00:01:20.000 Nilay is not speaking hypothetically here. This is already happening. As of last month, there were over 200 eBooks on Amazon's Kindle store listing ChatGPT as an author or a co-author. One person recently used AI to write and illustrate a children's book. He says it took him about three days. He ended up self-publishing it, and he's sold more than 800 copies on Amazon. 00:01:20.000 --> 00:01:21.000 That book, and others like it, have caused a major disruption in the creative world, even lawsuits. Writers say something about it just doesn't feel right. It seems hollow, lacking in humanity. Artists argue that AI image generators are learning from human art by stealing it without permission or compensation. I asked Nilay, we have laws to prevent people from stealing other people's creative work, why shouldn't those laws apply to computer- generated material? 00:01:21.000 --> 00:01:22.000 [MUSIC FADES OUT] 00:01:22.000 --> 00:01:23.000 Yeah. Those laws on preventing people from taking one another's work are really messy. 00:01:23.000 --> 00:01:24.000 Yeah. Imperfect for sure. 00:01:24.000 --> 00:01:25.000 Imperfect, necessarily imperfect. If they were perfect, you'd actually end a lot of art in just a meaningful way. If you couldn't just boost riffs from music, you might not have music. It's just kind of the way it goes. All art necessarily builds on itself. Again, very philosophical, but I was a copyright lawyer. That's how I became a technology journalist. And at "The Verge," we have the saying that the only law on the Internet is copyright law. It's the only effective way of regulating content on the Internet because everything else is kind of an enormous First Amendment issue or an enormous content moderation issue. And so, everyone kind of falls back to, you took that from me, it's a copy, someone should take it down or you should pay me. Everything centers on it. So of course, the first thing that has happened with these systems is a copyright law conversation because it is the default regulatory scheme for everything on the Internet. So, what you have now with all these AI systems is, you need to train them. So necessarily, you are going out on the Internet, you're making a copy of the Internet, you're making a copy of all these photos, you're putting them onto your computer, you've now made a copy, and then you're using those to make derivative works of those photos, which is another kind of copy infringement, and you've got layers and layers of things where the law has no ability to discern what is infringement, what is not infringement, what is fair use, what is not fair use. And that's gonna be, I think, a full decade of resolutions. And some of them are really obvious. So, I think Getty Images is suing Stability AI, and their evidence is like, look, when you ask it for photos of historical events, it produces the Getty watermark. 00:01:25.000 --> 00:01:26.000 Right. It's learned that that watermark should be there. 00:01:26.000 --> 00:01:27.000 Right. So, they're probably gonna win that one, or they're gonna settle, or they're gonna set a licensing deal or whatever. There are other ones where I think it is much harder to say, "Make a Banksy." And Midjourney will just do something that looks okay. But you walk in any art museum in the country, and there are art students sitting there, literally slavishly copying the paintings because that's how they learn. And so, okay, is automating that illegal? Have you replaced the work? Have you actually made a copy if you say, "Make it in the style of something"? I don't know. 00:01:27.000 --> 00:01:28.000 Wait. So, is it the doing that feels like a violation or is it the monetizing that feels like a violation? 00:01:28.000 --> 00:01:29.000 I think it is the idea that it will be the end of labor for these artists. 00:01:29.000 --> 00:01:30.000 Yeah. And what does your heart say about that? 00:01:30.000 --> 00:01:31.000 Um … I don't know. My heart says that giving people access to make creative work faster and easier is the history of the computer industry. What computers are for is democratizing creation. Ultimately, that is what they are for. That is why we make them, that's why we have them, that's why we love them. When kids get an iPhone, they do not open Numbers or Pages. They open a camera, and they start making art. And like, great. [CHUCKLES] That's what they're for. And so, there's a part of this where I think the nature of being an artist needs to get more specialized and more differentiated because all these things can do is make derivatives of what came before. They cannot make new things. And that's gonna be the value that we as humans provide in the world, I think, for a long time to come. You can train it on "fanfiction.net" and you can have the computer tell you that it loves you the same way computers in all of those stories have told people that they love them, but that's not a new story. It's like, do you want the computers to make the 95,000th Marvel movie, or do you want "Everything Everywhere All At Once"? 00:01:31.000 --> 00:01:32.000 [LAUGHS] That seems like a good application for it, actually. [LAUGHS] 00:01:32.000 --> 00:01:33.000 Right. We're kind of at the point where, all right, major media companies are just gonna ChatGPT their way through their franchises. And actually, the thing that people love is something that's new and creative. 00:01:33.000 --> 00:01:34.000 That's pretty scary. I would say a really common experience of when I've talked to other people about using ChatGPT or typing into Bing is, "Yeah, it's fine." [LAUGHS] It's fine. I don't know. I asked it question, it came back. It was kind of fine. 00:01:34.000 --> 00:01:35.000 Well, I think there's two things there. One, most of us are experiencing it through social media where we're seeing the best of its work. 00:01:35.000 --> 00:01:36.000 That's fair. Yeah. 00:01:36.000 --> 00:01:37.000 And then there's the reality, which is, mostly it just doesn't do anything good. And what's true about that part is the volume. So there's a quality layer where social media kind of filters the best quality, and then there's a quantity layer, which is like, oh, this is all pretty mediocre. But there's so much of it. 00:01:37.000 --> 00:01:38.000 It'll do. It'll kind of just do. 00:01:38.000 --> 00:01:39.000 It'll do. And the amount of C+ work that we accept as humans in our daily lives is very high. 00:01:39.000 --> 00:01:40.000 You said C+, and it made me think, where can people kind of just coast by and be okay with a C+ is school. I mean, literally. A lot of concern right now about how this can be used by students, college level as well. How this can be used to generate essays. Literally, if it's this easy to generate an essay about the War of 1812, what teacher's ever gonna make that assignment anymore? 00:01:40.000 --> 00:01:41.000 Uh … that one I think is really challenging. So, when I took the bar exam a million years ago, I took it on paper, this is how long ago I took it. And in the bar exam prep classes, they taught us the thing you're gonna do is you're gonna read this question and then you're gonna list all the keywords you know about whatever you think this question is about, and you're gonna put them in the first two paragraphs. This is the bar exam strategy. The graders are busy, they've got 9,000 tests to grade. You're just gonna make them know. You're gonna brute force them into thinking you know what you're talking about by saying all the vocabulary words about this answer. 00:01:41.000 --> 00:01:42.000 Yeah. Be the mirror. [CHUCKLES] Wave your hand back at them. 00:01:42.000 --> 00:01:43.000 Yep. You're like, I see you. I see what this question's about. I dunno if I got the right answer, but you know that I was aware of the answer. [LAUGHS] 00:01:43.000 --> 00:01:44.000 Yeah. 00:01:44.000 --> 00:01:45.000 Did I learn anything? Do I know anything about trusts and estates? I went to three years of law school, and I passed the bar in two states. Do not let me write your will. 00:01:45.000 --> 00:01:46.000 [SHUMITA LAUGHS] 00:01:46.000 --> 00:01:47.000 But I took that exam, and I passed that test. There's just a part of me that's like, we're just automating that and we're maybe collapsing it into nothing in an important way. That this sort of fiction that our students have learned anything, but really they're just doing elaborate test prep. Okay, fine. Maybe we're gonna have to come up with new tests that actually test the learning because the robot has fired a cannon of mediocre text at our business model. 00:01:47.000 --> 00:01:48.000 I see. Oh, that's an interesting argument. So, if the computer's able to do it so convincingly, maybe it was a certain kind of automated learning that is not worth doing or not really the measure of a student's learning. Is that what you're saying? 00:01:48.000 --> 00:01:49.000 Yeah. I think that's the optimist case. The pessimist case is that teachers are underfunded, and school districts do not change rapidly, and education in America is in a crisis, [CHUCKLES] and ChatGPT is gonna break everything. But there is an optimist case to be made, and I think that's it. 00:01:49.000 --> 00:01:50.000 Yeah. Yeah, yeah. I wanna talk about some of the limitations of what ChatGPT can do. And I actually asked ChatGPT what its downsides are, and it gave me a very good list, and I wanna talk it through with you. I think it's worth talking about. Biases was the first thing on its list. Of course, they're trained on data sets that are based on human behaviors, which means that they learn our biases. Misinformation, it sometimes gives you information that is not accurate and it can be misleading. Dependence, which I thought was a really interesting thing it came up with. And here's what it said. Again, ChatGPT gave me the answer. "As ChatGPT becomes more sophisticated, there is a risk that people may become overly reliant on it for information and decision making. This can lead to a loss of critical thinking and decision-making skills." 00:01:50.000 --> 00:01:51.000 That's pretty good. 00:01:51.000 --> 00:01:52.000 Now, I have been thinking about this way before ChatGPT gave me this answer. What do you make of this whole dependence idea? 00:01:52.000 --> 00:01:53.000 That's more self-awareness than many people I know. [LAUGHS] 00:01:53.000 --> 00:01:54.000 Yeah. Yeah, I know! Well, that's what's so freaky about it. 00:01:54.000 --> 00:01:55.000 Uh … I think that's right. We've already seen a bunch of really weird outcomes where librarians in this country, or like people come in asking for books that don't exist because they asked ChatGPT for a list of books about investing or whatever, and they're like, that book does not exist. The computer hallucinated the existence of a book, and you think it's real, and now you're asking me for it, and I'm telling you it's not real. Very famously, Orson Wells interviewed Stalin back in the day, and it's the subject of some Internet controversy. But there's magazine articles about it, and there's a "New York Times" headline about it, and literally in the replies to a tweet, someone's like, "Well, I asked ChatGPT, and it said it never happened." And they're looking at "The New York Times" headline [CHUCKLES] about it happening. And it's like, this is crazy. Now we believe the thing because it so confidently replies to us in text. And so that piece of it, you have to recognize that it's a tool for just firing out a massive amount of text. 00:01:55.000 --> 00:01:56.000 Yeah. Yeah. I'll tell you just a couple of the other downsides that ChatGPT gave me that it has a bit related to that. One is malicious use. It says outright, "Like any technology, ChatGPT and other generative AI models can be used for malicious purposes such as generating fake news or social engineering attacks." I think this is a really big fear that a lot of people have. Our inability to parse true information [CHUCKLES] far precedes the existence of ChatGPT. And what will this do to our information ecosystem? 00:01:56.000 --> 00:01:57.000 Let me give you two examples from Hollywood, one of which is sort of funny and one of which is like deeply hilarious. So, AT&T buys Warner Media and fans start waging an online campaign to have Dan Snyder remake "Justice League" into the "Snyder Cut." This is a real thing that happens. And [CHUCKLES] AT&T is like, we're gonna give the people what they want. Now people will be happy with us for making HBOMax, and they'll sign up for HBOMax. It's a logical thing to do. It comes out over the course of some litigation that maybe there was a bot army involved in this online campaign to push people towards making the "Snyder Cut" and maybe AT&T got hoodwinked into making this extremely strange movie. Which by the way is in four-three, so it's a square. [CHUCKLES] And it is just like deeply confusing. 00:01:57.000 --> 00:01:58.000 [LAUGHS] Oh, my God. 00:01:58.000 --> 00:01:59.000 They paid millions of dollars to remake their movie as a square, and maybe they got goaded into it by bots. And that's not funny. The reason I say it's not funny is, there's like a little bit of a harassment campaign vibe to that. That's that malicious use. But really, they got convinced by volume. So that's one. Two, the movie "Morbius" came out, which is very bad, starring Jared Leto, and fans started saying "it's Morbin' time" online because it's just a hilarious thing to say. 00:01:59.000 --> 00:02:00.000 An incredible phrase that can be applied to anything. Yeah. Uh-huh. 00:02:00.000 --> 00:02:01.000 And they started saying it so much that it started trending. The studio got confused about whether their movie was good, and they re-released it into theaters. 00:02:01.000 --> 00:02:02.000 Oh, wow! [LAUGHS] 00:02:02.000 --> 00:02:03.000 Where it flopped for the second time. 00:02:03.000 --> 00:02:04.000 Oh no! [LAUGHS] 00:02:04.000 --> 00:02:05.000 Now this is just straight up hilarious. To not know that people are making fun [CHUCKLES] of your movie by saying, "It's Morbin' time." 00:02:05.000 --> 00:02:06.000 That has less self-awareness than ChatGPT, apparently. 00:02:06.000 --> 00:02:07.000 It's just one of the best stories in history. It's not even malicious. A bunch of teenagers convinced, I think it's Sony, [CHUCKLES] to re-release this movie. Okay, what if you fire a cannon of mediocre text at any of this. Now, you're like using social media in a way that confuses everyone, and you have not one but two movies that got remade or re-released based on the evidence that was just text on the Internet. Social media posts. 00:02:07.000 --> 00:02:08.000 Oh, my God. Yeah, yeah. 00:02:08.000 --> 00:02:09.000 You can see, again, the sort of ability to reckon with what is real and what is the cannon of text, I think, will be very challenging for all of us in the future. 00:02:09.000 --> 00:02:10.000 Volume being confused with truth, which is very scary. Now that we've talked about all these different applications, potential pitfalls, let's talk about rules and who should be setting them and what regulation should look like. Given the way that we've seen how social media has changed the world and failed, how we have really failed, to hold tech creators accountable for some of those social media failures, I really worry about what we haven't learned. [CHUCKLES] And really what we might be walking into here. What mistakes could we be inviting? 00:02:10.000 --> 00:02:11.000 There's quite a few potential mistakes. The reason I say copyright law is the only law on the Internet, it's the only effective speech regulation in the Internet context. So, the First Amendment exists, we generally think the government should not be in the business of speech regulations, and most things on the Internet are speech. So, you just have a framework, which doesn't exist in other countries, by the way. Other countries have speech regulations, and they are perfectly content to make speech regulations all day and all night. This country has a long history of not wanting to do that. I am not sure that ChatGPT is a person with the rights and privileges of a citizen of the United States. And so maybe it is appropriate to say, okay, we're gonna write some speech regulations about the robot. And I think that might be a place where we diverge in how we regulate the Internet. Where we're like, these automated systems might have a different set of rules and responsibilities than people who are typing into the Facebook text box. 00:02:11.000 --> 00:02:12.000 Now the argument is gonna be, well, the automated systems are made by people, people direct the systems, it's still all just work on the Internet. But I think that's gonna be kinda the first flashpoint where you can say, look, this is not a person, this is fancy auto complete. And we, the government, do not think that fancy auto complete should tell people how to make a bomb. So, don't. 00:02:12.000 --> 00:02:13.000 Yeah. 00:02:13.000 --> 00:02:14.000 Right? In a way, that would become a very challenging First Amendment case for the government, might be a less challenging First Amendment case as applied to Microsoft and Bing or whatever. 00:02:14.000 --> 00:02:15.000 So, what does truly advanced AI look like? I guess I wanna understand how far we are from that next level that I think many people are mistaking us for being at right now. Not just machine learning, but sentient machine thinking. 00:02:15.000 --> 00:02:16.000 We are a million years away. I think that's the answer. 00:02:16.000 --> 00:02:17.000 Are we ever gonna get there? I mean, truly. 00:02:17.000 --> 00:02:18.000 Um … I don't know. I don't think anyone knows. I think if people knew, we'd be closer. I think there's a lot of fear about that. There's a lot of fear about what that would mean. I think ChatGPT, because it is a convincing facsimile of something that could be that - it's definitely not that - but it has already prompted a lot of conversations about whether it is sentience already, and it's just a chatbot. So, I don't think we know, and I don't think we know the timeline. And most importantly, I don't think we have reckoned with what if the intelligence is meaningfully different than ours. In a world of meat sacks, we think of intelligence as a thing, and we measure it. Even in the animal world, we measure it against a baseline of ourselves or a reference point of ourselves. But it is not true that every animal's intelligent in the same way as us. Some animals are much more intelligent in other ways. We've got a long history of sciences starting to account for this. We are not ready to do that with an iMac. It's just not gonna happen. [LAUGHS] And so we're at the place where, right now, a lot of people are like, I think my laptop wants to kiss me. And that's as far as we've gotten. 00:02:18.000 --> 00:02:19.000 [SHUMITA LAUGHS] 00:02:19.000 --> 00:02:20.000 It's gonna be fine for now, but we have a long way to go before we really understand what's happening. 00:02:20.000 --> 00:02:21.000 We've talked a lot about fears today, but what are your hopes for the best-case scenario outcomes here? 00:02:21.000 --> 00:02:22.000 The point of technology is to produce culture, and you can't understand one without the other. So, with AI, it's very much that mediocre middle is about to change radically. And what I'm hopeful for is that means the good things get way better. Because to escape the middle, to escape just being more AI, you're gonna have to meaningfully stand out. And I think that pressure is about to get really intense, and it will be very challenging for a lot of people, but I think ultimately, human beings are gonna be more creative than computers for a very long time. Especially when the tool is explicitly just for reflection of the past work. 00:02:22.000 --> 00:02:23.000 [MUSIC FADES IN] 00:02:23.000 --> 00:02:24.000 It is only the past. It is not the future. And we represent the ability to create the future. 00:02:24.000 --> 00:02:25.000 Nilay, thank you so much for this conversation, for your time. This was really interesting. 00:02:25.000 --> 00:02:26.000 Thank you. 00:02:26.000 --> 00:02:27.000 You can read reporting from "The Verge" on Apple News. You'll find a link on our show notes page. And don't forget to subscribe, rate and leave a review for this show on Apple Podcasts. 00:02:27.000 --> 00:02:28.000 [MUSIC FADES OUT]