Plain English with Derek Thompson - Why the Future of AI Should Terrify and Thrill You

Episode Date: October 11, 2022

This is the 100th episode of Plain English! I don’t know how that’s possible. Thanks to all of you who have listened. This has been a ton of work and a ton of fun. I’m still figuring out what this show is; how to balance news and tech gossip and big society questions and war coverage. There are days I think I know exactly what I’m doing and days I think I know even less than when I started out. And I just want to say to all the folks who have, on any medium, offered negative feedback or positive feedback: I’m reading it. Today, we're joined again by our first-ever guest, Kevin Roose from the 'New York Times,' to talk about Elon vs. Twitter and the deep implications of the year's astonishing breakthroughs in AI. If you have questions, observations, or ideas for future episodes, email us at [e-mail]. You can find us on TikTok at www.tiktok.com/@plainenglish_Host: Derek Thompson Guest: Kevin Roose Producer: Devon Manze Learn more about your ad choices. Visit podcastchoices.com/adchoices

I'm Yossi Sallick, and I'm the host of Bandsplain, a show where we explain cult bands and iconic artists by going deep into their histories and discographies. We're back with a brand new season at our brand new home, the Ringer podcast network, tackling a whole new batch of artists, from grunge gods to power pop pioneers to new metal legends and many, many more. Listen to new episodes every Thursday, only on Spotify. Today, it is the 100th episode of Plain English. I don't know how that's possible. Thank you to all of you who've listened and said you love the show, said you hated the show. I've never done this before. Hosting a talk podcast is something I have never done.

It has been a ton of work and a ton of fun. I'm still figuring things out. I'm still figuring out what this show is, how to balance news and tech gossip and future science and big society questions and war coverage. And to be totally honest, there are days I think I know exactly what I'm doing. days, I think I know even less than when I started out. And the truth is that I think this is just a good moment, episode 100, to say to all the folks who have on any medium offered negative feedback or positive feedback, email, Twitter, Reddit, I am reading you. I am trying to make this show better. And I am really sincerely grateful to all of you for everything that you've written,

even the rude stuff. Because you know what? Sometimes rude people are right. If you like this show, if you enjoy or have learned from any of the first 100 episodes, you can do us one very specific solid. You can give us five stars on Spotify, or if you listen on Apple Podcasts, leave a five-star review. I would really appreciate that. So this show launched in mid-November, 2021. Do you remember mid-November 2021? It was an entirely different world, the literal peak of the stock market, the literal peak of crypto valuations. Inflation? Nope. Interest rate, increase. Not really. Nope. NFT buzz was huge. I did this show with the New York Times, Kevin Roos, my very first podcast episode called The Future is Gonna Be Weird as Hell. And it was all about these future

technologies like crypto and the metaverse that were all the rage last November. And if you go back and listen, I think you'll find that our analysis holds up pretty well. We were, I think, duly skeptical that cryptocurrencies and metaverse technologies as they existed were going to make a big impact on the world, even though we retained a certain amount of curiosity about their long-term potential. Well, today, Kevin Ruse is back, and this episode could have also been called The Future is Going to be Weird as Hell, or the Future is still going to be weird as hell, but for a different reason, I think the most important story of 2022, more important than anything happening may be in Ukraine or the economy is the rise of AI tools. That sounds crazy,

because there are people dying in Ukraine. There is war. There is famine. There's rising costs for essentials like energy and food. But wars end. Inflation peters out. Interest rates go up and down. But technology can leave a mark on the world that doesn't go away.

The invention of cars in the 1880s is more with us today than the end. empire struggles of the late 19th century. And if you look around the world right now, you look at the organization OpenAI, they release two of the most important tools, GPT3, which synthesizes language responses to prompts, and Dali, which creates original art based on any text. If this episode is successful,

I think you'll come away with an almost odd sense of how bizarre and strange and incredible and potentially scary the future could be in a world where these kind of technologies grow and become more a part of our everyday life. In today's episode, we talk about how these AI tools work in the first place, why this summer was a coming out party for this technology, why it should thrill us, why it should terrify us,

and how AI could change our relationship to the internet, to society, to truth, and to ourselves. For the 100th time, I'm Derek Thompson. This is plain English. Kevin Ruse, welcome back to the podcast. Wow, what a journey it has been. I am so happy to be here and close the circle on 100 episodes. Congratulations.

A hundred episodes. You were my first guest. You are my 100th guest. It is an honor to have you bookend at the first 100 episodes. If you recall, our first episode was called The Future Will Be Weird as Hell. And we talked about implications for crypto and the Metaverse. Since then, at the time, Bitcoin's price was 69,000. Now it's 19,000.

Crypto is crashed. The Metaverse is somewhere, not here. The markets melted down. Inflation took over. Interest rates soared. Growth stocks plunged. Netflix, Peloton, Zoom. All the pandemic darlings fell down to Earth.

And also, Musk Twitter happened. And so that's kind of where I want to start. The Musk Twitter acquisition, to me, is kind of like a television show that I enjoyed for a few weeks. like, wait, the showrunners, like, haven't really thought this one through. You know, like, when you watch a season of TV and it's like a 10 to 15, like, episode arc, but you're like, this has to have pitched as a three-hour movie. There's not more than three hours of, like, actual plot here. That's kind of where I got to with Elon and Twitter.

But we had the stirring season finale last week. Elon announced that he will likely go ahead with the deal and buy Twitter for $44 billion. Kevin, as someone who tuned out of the show for the last six, eight weeks or so, what did I miss? Where do we stand now? Yeah, so we are in this very strange endgame now of sort of phase one of this story. So you can think of it as a multi-season prestige TV show. We are now, I believe, coming to the end of season one, which is, you know, the latest that's happened is that Elon Musk sent a letter to Twitter very unexpectedly.

saying that he intended to close this deal, that he was willing to pay the original price, $44 billion, $54 and 20 cents a share, for Twitter, and that he was going to attempt to close this deal. Twitter, and that happened days before, a trial was supposed to begin in the Delaware Chancery Court where Twitter was going to force him to live up to his original intent and buy the company for $44 billion.

dollars. So now in the last couple of days, what we've seen is that these sides, there's obviously a lot of distrust between Twitter and Elon Musk because of Elon spending months trying to get his way out of this deal. And so Twitter has said, essentially, we're not, we're not, you know, calling off the trial until we get our money in, until you buy the company. And just yesterday, the trial was stayed, the Delaware Chancery Corps. stayed the trial and gave the two parties, Elon Musk and Twitter, until October 28th to officially close this deal. So either that will happen. Elon Musk will pay $44 billion and get control of Twitter before October 28th,

or they'll go to trial. And that process will happen sometime in November. So three parts. Part one, Elon says, I'm going to buy you Twitter. Part two, he says, never mind. I don't want to buy you. Part three, he says, never mind, never mind. I'm actually going to buy you.

It's very obvious to me why we went from part one to part two. The market crashed, the expected value of Twitter. If you lined it up with similar companies that were social media companies, advertising-based companies like Snapchat, should have been much lower than the $44 billion that he had pledged to pay. He was trying to either wriggle out of the deal or negotiate some kind of lower price. It became very clear at some point that the Delaware Chancery Court was not going to allow that. You have a good sense of why we went from part two to part three in the story?

Why did he say, never mind to the never mind? Well, yeah, that's the big obvious question here, is what caused this sort of capitulation? He's not known for, you know, giving up and rolling over, especially when it comes to this deal. And my best guess at this point, and we'll need more reporting to confirm this, is basically that he just got told, look, you're going to lose. You're going to lose this case. It's going to drag you into the public, you know, more of your texts and emails and phone calls and et cetera are going to be dragged out in court. It's going to be embarrassing

to you and your friends. And at the end of the day, the court's going to make you pay for $44 billion anyway. So you might as well do it now, frame it as a victory, try to sort of make lemonade out of lemons and get this over with. Right. It's like you're going to pay $44 billion anyway. You can either pay $20 million in lawyer fees and embarrass all your friends and then pay $44 billion dollars or you cannot pay 20 million dollars in lawyer fees and just pay the 44 billion now. That seems to me to be, as you said, we're going to get more reporting on this, but that seems to me to be likely why he ultimately capitulated. I just want to talk a little bit briefly about like what happens to Twitter now. I have three doors for you to choose from when it comes to a Twitter

owned by Elon. Door number one is nothing much changes. But everyone is so aware of Elon Musk's ownership of Twitter that we pretend like everything bad about the app is Elon's fault. So we're like, I saw an anti-vax tweet. And it's like, yeah, maybe that's Elon's thumb on the button trying to make anti-vax tweets go a little bit more viral. Or maybe it's just that people who are anti-vax have been on the site for many, many months. So door one is nothing changes, but we blame Elon for the pre-existence of Twitter's badness anyway. Door number two is what a lot of, I think, liberal journalists are worried about. It's like reactionary Valhalla. Like all the forces of far-right mayhem are unleashed under the world. Twitter becomes proud boy city. Every day in the

app feels like a men's rights rally. That's what a lot of people are afraid of Twitter becoming. Door number three is what Elon says Twitter might become under his ownership. It becomes the everything app. Elon does it. He turns Twitter and is some joyous, incredibly useful combination of a town square and a news platform and an all-purpose messaging app and a Venmo thing on top. Like what we chat is for China, Twitter becomes for America. So here are three doors that you can choose from or invent some other door because all these doors are totally fake. Everyone Convechtes about the same old shit, number one, two, proud boy city, three, the everything app. What do you think is the future of Twitter?

My likeliest guess, and I don't have super high confidence in this guess because I've been wrong about this deal and Elon Musk at, you know, in many points, is some combination of doors one and two. I think that there will be turmoil, certainly. I think that he's going to clean house at the company. There will be lots and lots of turnover among executives and employees, He's just people who don't want to work for him, and that that will result in not just lots of coverage of the turmoil there, but in actual product failures. Like, there will be, you know, Twitter is not a super, like,

they don't have a super stable infrastructure. A lot of the site is, like, held together with bubble gum and scotch tape. And so if you have, like, a significant number of employees departing, like, stuff is going to start breaking. And that's going to happen at the same. time as all of these people who have been banned from Twitter in the past are let back on. That is one thing that, like Elon has said very clearly in public and in his private text messages that he is going to do is to reverse all permanent bans or most permanent bans of

Twitter users, including Donald Trump, including, you know, including people, you know, right-wing figures who have been banned for hate speech and harassment and misinformation and things like that. So I think there will be those two things happening simultaneously. I also think the first door is very likely, which is that no matter what, like, he is going to become the de facto customer service department of Twitter. And like, I don't think he understands how annoying that is going to be for him, because it's not just going to be people on the left who are saying, like, this sucks and, and, you know, it's Elon's fault. It's also going to be his friends who are saying, you know, oh, I didn't get verified, or I, you know, I'm being shadow banned because my tweets

aren't getting as much reach. Like, that is absolutely going to happen. And I think it's going to make him pretty miserable. The known unknown here is who's actually going to run Twitter? Elon Musk does not have time to actually, on an hour-to-hour basis, be the chief executive of Twitter. He's a CEO of SpaceX. He's running Tesla. He's running a bunch of other stuff. He's trying to create brain brain computer interfaces with Neurlink. He's trying to create robots at Tesla. There's a lot going on in his sort of weird emporium of future tech science stuff.

He's not going to have that much time to worry about ShadowBans. So do you have any insight into who from the Elon Metaverse will end up as being de facto in charge of Twitter? I don't. I mean, what we saw in the text messages with him and his friends that came out of this as part of this court proceeding

is that there are a lot of people who would gladly stick their hands up in his orbit to be the CEO of Twitter with Jason Calcanus, the podcaster and investor, volunteered to be the CEO, Twitter, very magnanimously. Matthias Dapner, the owner of Politico, also volunteered to run it for him.

So I'm sure there will be any number of people who will be glad to take that on. The question is, like, do any of them have actual expertise and a vision for this company, or are they just trying to suck up to Elon Musk and use his ownership as a way to increase their own power? I have no idea who's going to run it. I have no idea what's going to happen.

Are you kidding? I have absolutely no interest in ruining my life. The only more efficient way I could possibly ruin my life is by running for political office. Those are two of the more high prestige ways to absolutely, without a doubt, fuck over your life and everyone who you love. Like, become the customer service department

and lead executive of the most famously hated global app in the world or run for political office and make sure that like your entire life is just taking hate from everybody. It's actually the same. It is the same job. In a weird way, like the future of being CEO of Twitter is a kind of political appointment. It is in part because a lot of the incoming hate is going to be explicitly political. Oh, why is this group not getting more attention? Why is this group getting too much attention? I mean, what a friggin mess. Do you want to run it?

Does Casey want to run it? No, I'm not volunteering. I can't think of anything that would make me more miserable. But I think we should say, in the interest of seriousness, I actually think that there will be people. I mean, it can be hard to see from our little media bubble. But, like, Elon Musk has plenty of people who work for him at his other companies. Very talented engineers.

Like, they build rockets. They build cars. Clearly there are talented people who are going to work for Elon Musk because he's Elon Musk. So I actually, like, I've talked to some people this week and basically have said, like, you know, they've told me, like, look, this is recruiting is not an issue that we're worried about. So someone will step in to run this company for him. And, you know, engineers will sign up to replace the ones who leave because they don't want to work for Elon Musk. I actually, I think that's like that's something that is probably overstated in the media is, you know, clearly there are a lot of people who already. desperately want to work for Elon Musk, and I think there will be more people who want to do it at Twitter.

Yeah, and I think to a certain extent, Elon will de facto run Twitter the way that, like, the Pope de facto ran the Holy Roman Empire, like not overseeing day-to-day decisions, not overseeing actual politics and policy, but having the ear of whoever is actually running it so that if anything starts to break or goes the wrong way, Elon says, actually, maybe you should do this for me, and then that thing will be accomplished by the deputy. Anyway, that's definitely enough Elon Twitter. news will evolve there, or as it's happened in the last two months, not really evolved there. And we'll just stay in some bizarre Delaware Chancery purgatory. I want to talk about AI because I think that far more than Elon and Twitter, the story of the

year when we look back 20 years, is going to be that this year and the summer in particular was an extraordinary inflection point in the frontier of AI. Last year, we've seen the emergence of two incredible categories of artificial intelligence. You've got language models like GPT3 where you basically ask an app, a program, a question, or give it a statement, and like some kind of disembodied linguistic superbrain, it can answer complicated questions about the world

or take a sliver of text and write a full story in the voice of someone. It can synthesize books. It is this kind of, like I said, disembodied linguistic super brain that's really extraordinary and has awesome implications. And then maybe even more famously and more remarkably is this explosion of text to image AI, like Dolly and stable diffusion.

And these are programs that take prompts and turn them into extraordinary art. So, for example, if you say, I want a picture of a cat wearing a cowboy hat, holding a scepter, riding a unicorn on Venus in the post-impressionist Van Gogh style, it can produce not just one, not just two, but several. images in precisely that mode, depicting precisely that bizarre description of things. Explain to me first, like I'm five, at a high level, how these AI image generators work. What is the general concept behind the magic? Yeah, it's a great question, and it's something that, you know, in order to properly explain it,

we would probably need an hour and several computer science PhDs here with us to explain how it works. It's very, very complicated. But I'll boil it out into like a very simplified, like, thing that, you know, people who actually understand this stuff will probably write in and say, like, oh, you took out, you know, 17 of the most important steps. But like, here is my, like, maybe not explain like I'm five, but maybe, like, explain like I'm 15 version of it.

So basically, it's a two-step process. So first, there's an AI model that, the one that runs Dali is called Clip. and it has been trained on hundreds of millions of image text pairs. So essentially they went out, they licensed a bunch of data and scraped a bunch of images from the open web. And these would be things like, you know, social media images that have captions or stock photos that have captions or newswire photos that have captions. So they take, I think it was hundreds of millions of images and captions, shove them into this giant AI model and then basically tell it to figure out all of the relationships between not just these

images and these texts examples, but figure out the relationships between all text and all images. It's a little more complicated than that, but basically that's how it works. And then once that model is sufficiently good and trained, then anything, any text that you throw at it, like, you know, monkey riding a bicycle on Mars will give you something that's reasonably good. And by the way, it's not just like taking elements from images that exist. It's not taking a monkey from one photo and a motorcycle from one photo. It's actually like generating these things entirely anew. And so they are unique creations. So, okay, that's the first step. Step one, this thing called clip.

And then there's another model, a second model that's called a diffusion model. And this one is pretty cool, actually. So basically the way it works is it takes images and then it adds random noise to them. So like you can imagine like turning up, you know, the static on a TV screen or something. And it does that gradually. And eventually it's all noise. And then it reverses that process. So it takes the image from noise.

back to an image that is, that is photo quality. And it eventually learns how to take images that are static and make them into images. It would be like, you can kind of compare it to like if it learned to build a car by watching tons and tons of videos of people dismantling cars and then reversing those things. So it learns to make images out of noise. So basically, you put in your text prompt into the first model, monkey riding motorcycle on Mars,

and then it generates something. And then the second model, the diffusion model, kind of refines that image and turns it into the thing that you ultimately get on your computer screen. It's pretty cool. That's lovely. One thing that you reminded me of is I remember in my favorite college English class, I learned about the story of Ariadne's Thread. and the Minotaur. And the way that the labyrinth of the Minotaur was solved was that Ariadne carried the thread through the labyrinth to the center of the labyrinth where the Minotaur lives so that the thread could be followed back out in order to escape from the labyrinth. And so that's

that's like the metaphor that occurred to me as you were talking about how you can figure out how to do something generatively by becoming an expert in its opposite in a weird way. And like, It's such a lovely and interesting thought experiment. And I just want to double down on one thing that you said, because I do think it's misunderstood. These are original creations. This is not pastiche. Like whatever my stupid introductory example was,

the rodent cowboy hat, it's not as if Adali or Stable Defusion is taking a rodent that they found on Google image, then putting it with the hat that they found on like Bing image search. It's literally drawing, or drawing is actually too anthropomorphic. It's in, it's synthesizing these things. uniquely for my prompt, which is spooky and weird and kind of wonderful. One question that I have for you that I haven't seen explored as much as I wanted to be

is that I think it's underratedly amazing that all these AI tools came out around the same time. Like GPT3 comes out and then like Lambda comes out from Google and they do very similar things. Or Dali comes out in the text image world from OpenAI and then stable diffusion comes out. out. And this is an example of something in tech history that's called simultaneous invention, like Newton and Leibniz, basically invented calculus at the same time. Alexander Graham Bell and Elijah Gray invented or submitted their patent for the telephone in the exact same day. Do you have any explanation for why so many different groups seem to have reached the frontier at the exact same time? I mean, it's like this one summer was the Cambrian explosion of all

of these different tools. Like, why were they all secretly at the same place at the same time. Yeah, it's a great question. And I think there are a couple explanations for it. Sort of a few things that kind of converged within the past, really the past two years. So one thing is that hardware just got a lot better. So this is sort of a, you know, a classic Moore's Law phenomenon in the, you know, in the tech industry, like processors, GPUs, these things get better over time. But specifically for this kind of AI, it relies on things called GPs, which are basically the graphics cards inside computers. That's what runs these models. And 10 years ago, if you wanted to build an AI, like a neural network, you'd go out and you'd

literally buy just a bunch of graphics cards, like the same ones that you'd buy if you were like building a high-end gaming PC. And then you use those to run your AI models, and it worked pretty well. But now, like 10 years later, companies like Nvidia, these chip companies, they make hardware that is specifically designed to run AI models. And that has made these models much faster, much more capable. It's brought the cost down to be able to train one of these models. The other thing, so another thing that has happened, so hardware is one piece of it. There's also just what they call the hyper-scalers. So these companies like Google, like Open AI, like Facebook, they got much richer because of things like search ads for Google and targeted social media.

ads for Facebook, and they used those profits from their main businesses to essentially do a land grab in AI. So Google used the profits from search ads to hire like hundreds of, you know, maybe thousands of AI experts, like the top people in the field, gave them essentially unlimited budgets and access to their massive troves of data and compute power. Facebook did the same thing. And so they were all kind of racing, armed with billions and billions of dollars in capital and all trying to work on some of the same problems. So that's another structural piece of it. The third, and I think actually the most important thing that happened in the last few years, is that computers figured out language. So in 2019, OpenAI came out with this model called GPT2, which was the predecessor to,

GPT3, and it was basically, as you said, this like disembodied linguistic superbrain. It was massively trained on tons and tons of examples using one of the biggest computers in the world, and it just became extremely good at predicting and transforming text. And that started what's been called the large language model era of AI. And then once GPT2 and GPT3 then came out, people started realizing, wait a minute, there are all these other things that can be turned into language problems. So instead of asking the large language model to finish a sentence, what if you asked it to finish a line of code? What if you asked it to finish a song or draw picture or predict the 3D structure of a protein? Or the most recent example that just literally happened in the last week, Meta and Google both came out with text to video generators.

So the same principle, the same thing you can do with Dali for static photos and images you can now do with movies. So basically everything, once computers had kind of figured out language, then the researchers started realizing, wait a minute, everything can be a language problem. There's even this one really amazing paper that just came out of Google where the researchers took a robot, like a physical hardware robot. And traditionally the way these things are programmed is like, you know, in sort of if-then statements. If there is a drink that spills on the counter, then you go over this many inches to the sink and you grab the sponge and you come back

and you use your robot arm at such and such an angle. And instead of programming it that way, they turned the robot into a language model. They started asking it, how would you clean up a spill on the counter? Is it A, B, C, or D? The robot, chose, and then it executes those instructions. So the robot basically became the body of the

large language model. So this is an incredible breakthrough, and it's something that has underlied a lot of the last couple years of progress. That was a fantastic answer, and it did better than I could possibly do to explain why I think these AI breakthroughs might be the single most important story of 2022. Inflation is going to go away. Interest rates will go up, and then they'll come down, these tech breakthroughs could be the opening innings, I think, of just an absolutely wild revolution. And in the next few questions, I want to choose to be excited about this revolution and then scared. And just sort of like, you know, bifurcate my emotions. Like, let's be excited first.

And then just like, let's save all of our fear for the next question. All right, let's be excited. Like, holy fucking shit, are you telling me that I could just like tell a machine to like make an animated scene of a movie and it's going to do it? that I can just say, I want a song in the key of C-sharp that sounds like early cold play, but actually the voice is a female, and it can, like, begin to put together a truly novel. Again, these are synthetic generative products. They are new, not borrowed. It can create a new song and therefore help me as a songwriter.

Like, what are you, Kevin, most excited about in terms of actually using one of these, like, AI disembodied super brain assistance to help you be more creative in your life. Yeah, I mean, there's unlimited possibilities, right? So I am a writer and a podcaster, and I also enjoy art. So I've made some stuff in Dali and these other image models that I'm very excited about. I actually am having one of my creations, like printed in frame to put up on the wall of my office. I also have some ideas about things that I would like to do with AI when it comes to, for example, making videos.

I've talked to a number of, I've interviewed a number of people who are already using this stuff in their day jobs. One person is a sort of filmmaker and visual special effects designer who is using this stuff already in storyboarding, in doing some sort of primitive special effects. effects with it. So, I mean, everyone is now going to be able to make not just art, but movies. They're going to be able to write in ways that they've never been able to write. I used a GPT3-based model last year to write a book review for the New York Times book review. And it was a little bit of a stunt because the book was about AI and I disclosed it. But like, if I hadn't told my editors that this is what I was doing, like, I'm pretty sure they would have had no idea. So that's the excitement.

part, we're all going to get not just better at this stuff, but faster. I mean, one of the most interesting pieces of data about the effects of these programs so far was with something called co-pilot, which came out of GitHub. And it's basically this idea of using a large language model to complete lines of code. And so it doesn't replace the programmer, but it basically goes inside the programming environment. And it sort of auto-completes when it, you know, it sort of senses what you're trying to do. It says, like, would you like to just have us do the rest? And they've done some initial tests on that, and they've found that it makes programmers about 50% faster. So just imagine that

extrapolated to every creative industry. And you get a sense for, like, just the pace and the scale that this is going to enable as far as trying out new ideas, you know, making new creations. we're just going to have sort of a Cambrian explosion of new creativity. My daydream is that I have this parallel life on Earth 2, where rather than go into journalism, I tried to become a musician, and I'm not very good at playing piano, and I can't play guitar, and I've written a few songs,

but I haven't really shared them with more than one or two people. But sometimes, you know, I'm not trying to be a wife guy here, But sometimes I like playing, like, a song that my wife says she likes, like learn, like, the really simple, like, triad chord structure and then, like, sing it back to her. One thing that really excites me as, you know, someone who has this sort of, like, parallel life daydream as musician, well, my wife really likes Billy Eilish and Taylor Swift. So for her birthday, I could ask, like, the Dolly or GBT3 of original music writing,

write me a song that goes no higher than F-sharp, because I'm a, you know, a 10 or two, that goes no higher than F-sharp and is played or orchestrated in the style of Taylor Swift and Billy Eilish, and is happy but wistful. And like, it could actually help me write an original song. Like, there's little things like this

that are just like so exciting to me. And they're exciting not only for me, you know, look, I'm 36. I'm never going to become a professional musician ever, period. Not with that attitude. Not with that attitude you want. But like an 11-year-old who has the same interest in music, he or she can now spend 10 years of their life becoming a professional musician with these tools. Like, does it excite you to think like how these tools are going to be used in education?

Or again, you're going to be able to have this super brain alongside you to become a better writer. You know, write my essay in the style of, Who knows? John Updike, right? If you like John Updike, you want to write with really beautiful crystalline metaphors,

rewrite my essay in the style of John Updike, and in seeing your own words translated through this algorithm, you might be able to become a better writer, a better music, a better musician, a better artist.

I mean, it's these sort of educational benefits that actually have me really, really thrilled. Yeah, it's super exciting, and it's, you know, it's really going to expand not just the work that we're able to do, but our ability to focus on the parts of our jobs

that we actually like. So, like, one early example of AI kind of transforming our industry is that, you know, I used to spend, when I was a young journalist 10 years ago, I used to spend a lot of time transcribing interviews. It was like kind of a drag. And then this software came along that could just do it.

And I think the same thing goes for podcasts. I mean, there used to have to be a producer, if you were working on a certain kind of podcast, who would sit in the meeting and type out everything that everyone said. And then, you know, now that's all instantaneous, basically, and it happens through AI. So I think there's just a ton of potential, not just to allow for new forms of creativity, but to sort of take away the parts of the job that no one actually likes doing. Now let's be a little bit scared. I think there was a prediction of AI and Alisbury. algorithms and automation that said that the kind of jobs that are most likely to be replaced

by this next frontier technology are jobs that are routine-based. The jobs that were considered safe were the jobs that were creative. Jobs like, for example, writing, making music, drawing images, being a video game designer, being an artist, but it is precisely these higher-level creative tasks that this new crop of AI is potentially impinging on. Are you worried about the employment or disemployment effects of technology like this on the workforce? Absolutely. I mean, I think anyone who says that this is not going to displace workers in creative industries

is insane or is selling something or is willfully ignorant. I'm already hearing from game studios that say, you know, we can now do things that we would have had to hire 50 people to do. And you're right, it is a total failure of our predictive capabilities. I mean, I remember just a few years ago, you had this sort of narrative. And Andrew Yang was out there talking about how automation was going to get rid of truckers and retail cashiers and warehouse workers, industrial workers. And the whole narrative was that basically AI and automation were a working class.

problems. And instead, what happened is that it came for the creative class first. And so I, you know, I'm revising a lot of my own predictions about that. I think that we'll have fully automated game and movie studios before we'll have fully automated Amazon warehouses. I think that an AI generated song will reach number one on the billboard charts before a car makes it from New York to L.A. with no assistance. I think that we are radically changed. I think that we are radically changed. our idea of who is at risk of automation. And I think a lot of that is based on this misunderstanding of creative work. A lot of what we call creative work turns out to just be pattern recognition,

fast following. If you're a fashion designer, it's like, oh, what are other fashion houses doing this year? What hasn't come around in this many seasons? What's due for a resurgence? It's just pattern recognition and prediction. and turning ideas into media objects. So I think that's like sort of the reasoning flaw at the heart of this misunderstanding

is that a lot of what we call creative work actually is just sort of gussied up prediction and pattern recognition. I'm so fascinated by and motivated by the second thing that you said. I'm not exactly sure of the right way to think about it, but there's this idea called More of X-Paradox, which says that there's this irony in algorithms and AI and automation

that so-called hard tasks, like, say, building an algorithm that's smarter than the greatest chess player in the world, that seems hard. But it's actually like one of the first things that IBM could do. Meanwhile, like walking down the stairs, organizing a stack of paper,

these things seem really complicated, like a five-year-old can do them, but it's very difficult to build a robot that does them fluidly. And maybe the most poetic and possibly bullshit explanation for more of X-paradox, is that we are reverse running the history of evolution. Our brains are among the last parts, or our sophisticated, creative minds are among the last part of the human body to evolve potentially. And so our creativity is younger than

our ability to walk, evolutionarily speaking. And so as we sort of design through technology, the full capacity of the embodied human, we're moving against time and checking the box of creative brain before we fully check the box of manual skills and perfect walking ability. Have you ever heard any... I don't know if you have strong opinions

about more of X-Paradox or whether you've heard other explanations for why we've been able to at least partially solve these problems of AI art before we've clearly solved problems that seemed easier like self-driving cars or robots that can do things other than just affix the steering wheel.

Yeah, I mean, I think there are a lot of possible explanations of it. I agree that it's a real phenomenon. One possible explanation is just that it's the difference between bits and atoms, right? That, you know, things like hardware robots have to contend with, like, the physical world, right? They get dusty, you know, things aren't where they're supposed to be, the power goes out. Like, there are all these kinds of edge cases that can, you know, if you're talking about self-driving cars.

Famously, like, you know, people will say, like, we're 99% of the way there, but that last 1% is all that anyone cares about because those are the strange edge cases, the, you know, snowball gets stuck on the sensor, you know, that's where people crash and die. And so you really have to be like 99.999% there before anyone's going to get in a self-driving car. So there's more fault tolerance in some of these creative pursuant. suits, but it's also purely software-based. So you don't have to, like, exist.

Nothing has to come out of the computer screen, basically, to do that. And so this is really, like, I think related to a point that I've been thinking and talking about for years, which is that I think the most susceptible kinds of jobs to AI are not just the creative jobs, but the remote jobs, right? Because remote jobs are mostly software-based, a lot of them. And if you can make a job remote, if you can make it as abstract and sort of task-oriented, as a fully remote job has to be in order to be able to be done remotely,

you can also give that set of tasks to an AI a lot of the time. That's a really powerful idea, that remoteability is a proxy for replaceability in some version of the future that might be coming. Obviously, there's lots of jobs that I, think a remote that are hard to replace. We don't yet have AI that can do perfect software programming. We don't yet have AI that can write full articles for the New York Times. But one way that I think about technology is that it doesn't automate jobs so much as it automates tasks within that

job. So, for example, if you're a podcast maker, one of your tasks in say, you know, circa 2009 or 2010, was to transcribe interviews so that you could edit them and help your producer edit the tape. Now that job can just be outsourced to Rev.com. It can just be purely outsourced to a machine. But that doesn't change the necessity of a job of a podcast producer, a podcast host. It just automates the task within it. And so optimistically, one would hope that while this might have certain pressures in the white-collar economy, what it mostly does is akin to your very first example.

It makes creative jobs more fun because it amplifies creativity and accelerates the process by which you go from zero to final product, whether you're making a video game or making a piece of art or making a piece of music. It just speed runs the creativity process a little bit that you get to the end faster. That is certainly an optimistic way of looking at it. And I hope that you are right. I think that what we've seen throughout history, and I wrote a book about this a few years ago, about sort of AI and as part of that book, I went back and read a bunch of studies and books about previous waves of technological change. And it turns out that a lot of work is atomized. So it is just a collection of tasks.

And as soon as the people who are running the company, the CEO or the CTO or whatever, can, can automate those tasks, they generally do. So if you think about how many tasks are involved in your or my job, there's probably like seven or eight main ones. And some of those can be automated probably today. And some of them may not be automated for 10 years. But I do think that we sort of lull ourselves into a false sense of security when we say that it's just going to help us.

It's just going to be our sort of assistant, or Emanuensis or whatever, and that it's not going to replace us. Because I think what we've seen is just that that's sort of a salve that we apply, and then the people in charge are the ones who are really deciding, okay, it's not, maybe it can only do, maybe the AI can only do five of the seven things that Derek or Kevin currently do in their jobs,

but actually that's probably good enough. Knowing what I know now, I feel like it's one of these situations where the stage of takeoff, that we're in with AI is kind of similar to where, say, like, you know, fintech apps, apps might have been in 2011 or, like, Bitcoin in 2010. Like, a part of me is, like, are we idiots for not seeing that this is the moment? That, like, if you want to, like, do something that is hugely influential, that, like, is a lever on the world and that could make you super damn rich, like, get into AI now

because, like, that plane is really starting to take off. If you were given the opportunity to solve one problem with this suite of AI tools that are coming into view, it could be in the disciplines of art and culturally talked about, it can be in using AI to come up with interesting new drug molecule combinations that can solve diseases in ways that humans previously hadn't synthesized. What is the space where if you were going to jump head, shoulders, and feet into AI, where would you jump? That's a great question. I mean, I'm a journalist, so naturally, like, that's where my expertise is, and I think there are going to be some interesting attempts to use these language models. I mean, I'm sort of amazed that there's no fully automated GPT3-based news outlet already. It's not something that has been. And just because I think maybe 15% of people listening might understand exactly what that is.

But like just paint a brief picture. What would a fully GPT3 automated news site kind of feel like? Well, it could look a couple different ways. I mean, the sort of light version of that would be, you know, instead of coming up with, you know, story ideas and then pitching them to an editor and then, you know, painstakingly reporting and writing them and then having a photo editor who selects the art that goes on the story and then, you know, someone who hits publish and then, you know, a team of people who are promoting it. Like that could, you know, most or all of that could be automated.

So I could come in one day and an AI could have already, you know, said, these are the 10 most important stories in the world. We've generated versions of them, you know, three versions for each story. Which one of these do you want to publish? Slap some headlines on them automatically. You know, generate the art automatically. Hit publish automatically, promote on social media, automatically that this would all just be seamless and is easy. And you could basically run an entire

newsroom with one person. That's that's one of the visions for this. The other vision that I've spent a lot of time thinking about recently is what I would call sort of dynamically dynamic sort of generated content. So if you go to a website right now, if you go to the New York Times, you know, dot com or the Atlantic.com, you know, no matter who you are, where in the world, you are looking at that website, you're going to see the same stories, right? You might see them in a slightly different order, or you might see, you know, different mix of stories depending on, you know, what browser are you using or mobile or whatever. But you're going to see the same stories. You're going to open them up. They're going to look the same. They're going to have the same arts. They're going to have

the same words. With AI and large language models, what you could theoretically do is to generate content on the fly for an audience of one. So when you go to a news website, it could figure out who you are, like where in the world you're located. It could make some predictions based on user behavior and pull from your social media profiles or whatever. And it could say,

this is the version of a Ukraine story that Derek would most resonate with Derek. And it could generate that spontaneously on the fly, a brand new story that no one's ever seen before. Choose a piece of art, generate a piece of art that is calculated to your specific taste. And it could do that all in the time that it takes you to load a news site today.

So you could effectively have billions of people each reading versions, reading and experiencing and watching versions of reality that are specifically tailored to them. Dude, that is terrifyingly plausible. And you know what company is obviously best suited to make this a reality. Think about all the questions we already ask Google. I ask Google news questions, right? What just happened in the UK? Who's the new UK Prime Minister?

I was doing a podcast a week ago about sort of some of the domino effects of Federal Reserve policy on Southern Asian currencies. Okay, they know the Google questions that I asked for that and the tabs that I opened up for that. But more than that, they know my consumer questions. I just decided that I want an Apple Watch. So I went around looking for which Apple Watch I should buy as a first time Apple Watch customer.

I need new slippers. So I went looking for slippers. All right. What's really annoying is having a bunch of ads that really have nothing to do with what I'm interested in, or actually are just like boxes where it's like, here's a picture of slippers. What's more interesting is to have the wirecutter article or something like it presented to me so that I already have this organization of, oh, here are the best kind of Apple watches for someone like you who's buying their first Apple Watch and doesn't swim a lot and doesn't like go for long runs. while that same news site, which is serving me dynamic advertisements in the form of content, is also serving me content that is fully written out. It's an article in GPT3, but based on the prompts that I've asked Google.

It's like ask Jeeves, but turned into a dynamic website paid for by the advertisement that's also based on my search behavior. I mean, I'm not, I'm sure there's a thousand problems with this, maybe, in terms of privacy. I'm sure there'd be some kind of blowback because it's a little bit creepy to have a website that so very clearly holds up a mirror to your online Googles. But something like this seems horrifyingly plausible to me. Oh, I'm sure there are startups that are building this right now. I'm sure there are, you know, someone will come out with this. And I think that's going to be a really interesting inflection point, not just in the digital media business, but in like kind of the history of society as,

a self-conceiving thing. So I have a person that I talk to a lot about this, is Jack Clark, who's a former journalist, who then went into AI and now is running a big AI company, and he's very, very smart about this stuff. And he sort of talks about this coming reality collapse, what he calls the reality collapse, which is that basically, up until this point,

society, at least since Gutenberg, has had sort of a centralized idea about what is not only happening in the world, but like, what is society? What are our institutions doing? Who are we? And so we're moving from that, he believes, and I tend to agree with him, into a world in which we'll all have these kind of like hyper-personalized, dynamically generated, kind of like sort of choose-your-own-adventure realities.

that's all going to be sort of overseen and conducted by AI. So, you know, it's not just that right-wing people will read right-wing news sites and left-wing people will read left-wing news sites. It's like everyone's sense of reality will be individually tailored to an AI's prediction of what will generate the most clicks or watchtime or product buys or anything from them. And I think that's a really important moment. And it's really, you know, this is sort of the same. scary part is that we actually have never run that experiment before, and it will be very,

very interesting. And I think there will be a lot of trouble that comes with that. Yeah, here's three categories of content that I just thought of. A magazine is human content edited by humans. TikTok is human content organized by algorithm. What's box number three? Algorithmic content organized by algorithm. Where all the TikToks or all the articles or all the images, as if it's an algorithmic Instagram, are all produced by these AIs that we're talking about. And also, there's a meta-AI, an orchestra conductor that's organizing the process by which or organization of which we confront. I mean, that's a really fascinating and weird idea that we will have, at that point, moved from a content ecosystem that is for humans and by

humans to a content ecosystem that is for humans by entirely AIs written by humans such that each individual's experience of the world and news and maybe even culture is so unique to them because all of these articles that they're seeing and ads that they're seeing are algorithmically invented for an audience of one. It's bizarre, and I agree, like, rather destabilizing because it cuts against, I think, certain assumptions of what a society is, a kind of meshed, shared experience of similar inputs. Everyone listening to the same Taylor Swift song, everyone watching the Super Bowl, everyone watching the same presidential debate and reading about it in the news, what if it all becomes a little bit different and mediated by AI that is designed for one. It's a weird thought.

And just to, it's a very weird thought. And just to add another layer of weird thought on top of it, because I think this all sounds kind of futuristic and like sci-fi when we talk about it. But like, I think this is going to happen very soon, this sort of inflection point where most of the things that we see and encounter on the internet are going to be AI generated. I think we may be there within the year for images. I think that it's entirely plausible if current trends hold that by the end of this year, the majority of new original images that are produced on the internet every day will be generated by Dali, Mid-Jurney, stable diffusion, just as a matter of just pure, numbers. I think there's a possibility that's already happened.

It's funny because episode one was, The Future will be Weirder than you think. The Future will be weird as hell based on crypto and the Metaverse. And now we're like, the future is going to be weird as hell, but actually because of AI, maybe we're right about the big picture. But every year, I need you to come back because a different animating technology will prove the point that the future is going to be weird as hell. Like next year is going to be like, you know, microdrones dotting the skies is actually the big thing, the big breakout technology of 2023. And that's why the future is going to be weird as hell.

But if only there were a podcast devoted to the idea that the future is going to be weird as hell that would track these developments on a weekly basis, God, I wish someone would start a podcast like that. Kevin, are you starting a podcast like this with The New York Times? Derek, as it turns out, I am. I have a new podcast called Hard Fork. It came out, the first episode, just came out. And it's me and Casey Newton, another fabulous tech journalist and a friend of mine. And we are going to do weekly sort of chat shows about this very thing, about the fact that the future is weird as hell.

And we're going to look deeply into things like AI-generated art. We're also going to look at other technological phenomena, and it's going to be really fun. So thank you. I can't believe we ended up circling all the way back to around to a promotion of my new podcast. But the future is going to be weird as hell. That's the last note that I had on my notepad. So I'm very glad that you yourself built the on-ramp to it. Kevin Ruse, the podcast is called Hard Fork.

Hard Fork. Hard Fork with Kevin Ruse and Casey Newton. Kevin, thank you so much. Thanks for having you. Thank you for listening. Plain English is produced by Devin Manzi. If you like the show, please go to Apple Podcast or Spotify. Give us a five-star rating.

Leave a review. And don't forget to check out our TikTok at Plain English underscore. That's at plain English underscore on TikTok.